reliability-based design in geotechnical engineering
TRANSCRIPT
Reliability-Based Design inGeotechnical Engineering
Also available from Taylor amp Francis
Geological HazardsFred Bell Hb ISBN 0419-16970-9
Pb ISBN 0415-31851-3
Rock Slope EngineeringDuncan Wyllie and Chris Mah Hb ISBN 0415-28000-1
Pb ISBN 0415-28001-X
Geotechnical ModellingDavid Muir Wood Hb ISBN 9780415343046
Pb ISBN 9780419237303
Soil LiquefactionMike Jefferies and Ken Been Hb ISBN 9780419161707
Advanced Unsaturated SoilMechanics and EngineeringCharles WW Ng and Bruce Menzies Hb ISBN 9780415436793
Advanced Soil Mechanics 3rd editionBraja Das Hb ISBN 9780415420266
Pile Design and Construction Practice5th editonMichael Tomlinson and John Woodward Hb ISBN 9780415385824
Reliability-Based Design inGeotechnical Engineering
Computations and Applications
Kok-Kwang Phoon
First published 2008by Taylor amp Francis2 Park Square Milton Park Abingdon Oxon OX14 4RN
Simultaneously published in the USA and Canadaby Taylor amp Francis270 Madison Avenue New York NY 10016 USA
Taylor amp Francis is an imprint of the Taylor amp Francis Groupan informa business
copy Taylor amp Francis
All rights reserved No part of this book may be reprinted or reproducedor utilised in any form or by any electronic mechanical or other meansnow known or hereafter invented including photocopying and recordingor in any information storage or retrieval system without permission inwriting from the publishers
The publisher makes no representation express or implied with regardto the accuracy of the information contained in this book and cannotaccept any legal responsibility or liability for any efforts or omissions thatmay be made
British Library Cataloguing in Publication DataA catalogue record for this book is availablefrom the British Library
Library of Congress Cataloging in Publication DataPhoon Kok-Kwang
Reliability-based design in geotechnical engineering computations andapplicationsKok-Kwang Phoon
p cmIncludes bibliographical references and indexISBN 978-0-415-39630-1 (hbk alk paper) ndash ISBN 978-0-203-93424-1
(e-book) 1 Rock mechanics 2 Soil mechanics 3 Reliability I Title
TA706P48 20086241prime51ndashdc22 2007034643
ISBN10 0-415-39630-1 (hbk)IBSN10 0-213-93424-5 (ebk)
ISBN13 978-0-415-39630-1 (hbk)ISBN13 978-0-203-93424-1 (ebk)
ISBN 0-203-93424-5 Master e-book ISBN
ldquoTo purchase your own copy of this or any of Taylor amp Francis or Routledgersquoscollection of thousands of eBooks please go to wwweBookstoretandfcoukrdquo
This edition published in the Taylor amp Francis e-Library 2008
Contents
List of contributors vii
1 Numerical recipes for reliability analysis ndash a primer 1KOK-KWANG PHOON
2 Spatial variability and geotechnical reliability 76GREGORY B BAECHER AND JOHN T CHRISTIAN
3 Practical reliability approach using spreadsheet 134BAK KONG LOW
4 Monte Carlo simulation in reliability analysis 169YUSUKE HONJO
5 Practical application of reliability-based design indecision-making 192ROBERT B GILBERT SHADI S NAJJAR YOUNG-JAE CHOI AND
SAMUEL J GAMBINO
6 Randomly heterogeneous soils under static anddynamic loads 224RADU POPESCU GEORGE DEODATIS AND JEAN-HERVEacute PREacuteVOST
7 Stochastic finite element methods in geotechnicalengineering 260BRUNO SUDRET AND MARC BERVEILLER
8 Eurocode 7 and reliability-based design 298TREVOR L L ORR AND DENYS BREYSSE
vi Contents
9 Serviceability limit state reliability-based design 344KOK-KWANG PHOON AND FRED H KULHAWY
10 Reliability verification using pile load tests 385LIMIN ZHANG
11 Reliability analysis of slopes 413TIEN H WU
12 Reliability of levee systems 448THOMAS F WOLFF
13 Reliability analysis of liquefaction potential of soilsusing standard penetration test 497CHARNG HSEIN JUANG SUNNY YE FANG AND DAVID KUN LI
Index 527
List of contributors
Gregory B Baecher is Glenn L Martin Institute Professor of Engineeringat the University of Maryland He holds a BSCE from UC Berkeley andPhD from MIT His principal area of work is engineering risk manage-ment He is co-author with JT Christian of Reliability and Statistics inGeotechnical Engineering (Wiley 2003) and with DND Hartford ofRisk and Uncertainty in Dam Safety (Thos Telford 2004) He is recipientof the ASCE Middlebrooks and State-of-the-Art Awards and a memberof the US National Academy of Engineering
Marc Berveiller received his masterrsquos degree in mechanical engineering fromthe French Institute in Advanced Mechanics (Clermont-Ferrand France2002) and his PhD from the Blaise Pascal University (Clermont-FerrandFrance) in 2005 where he worked on non-intrusive stochastic finiteelement methods in collaboration with the French major electrical com-pany EDF He is currently working as a research engineer at EDF onprobabilistic methods in mechanics
Denys Breysse is a Professor of Civil Engineering at Bordeaux 1 UniversityFrance He is working on randomness in soils and building materialsteaches geotechnics materials science and risk and safety He is the chair-man of the French Association of Civil Engineering University Members(AUGC) and of the RILEM Technical Committee TC-INR 207 on Non-Destructive Assessment of Reinforced Concrete Structures He has alsocreated (and chaired 2003ndash2007) a national research scientific networkon Risk Management in Civil Engineering (MR-GenCi)
Young-Jae Choi works for Geoscience Earth amp Marine Services Inc inHouston Texas He received his Bachelor of Engineering degree in 1995from Pusan National University his Master of Science degree from theUniversity of Colorado at Boulder in 2002 and his PhD from TheUniversity of Texas at Austin in 2007 He also has six years of practicalexperience working for Daelim Industry Co Ltd in South Korea
viii List of contributors
John T Christian is a consulting geotechnical engineer in WabanMassachusetts He holds BSc MSc and PhD degrees from MIT Hisprincipal areas of work are applied numerical methods in soil dynamicsand earthquake engineering and reliability methods A secondary inter-est has been the evolving procedures and standards for undergraduateeducation especially as reflected in the accreditation process He was the39th Karl Terzaghi Lecturer of the ASCE He is recipient of the ASCEMiddlebrooks and the BSCE Desmond Fitzgerald Medal and a memberof the US National Academy of Engineering
George Deodatis received his Diploma in Civil Engineering from theNational Technical University of Athens in Greece He holds MS andPhD degrees in Civil Engineering from Columbia University He startedhis academic career at Princeton University where he served as a Post-doctoral Fellow Assistant Professor and eventually Associate ProfessorHe subsequently moved to Columbia University where he is currently theSantiago and Robertina Calatrava Family Professor at the Department ofCivil Engineering and Engineering Mechanics
Sunny Ye Fang is currently a consulting geotechnical engineer withArdaman amp Associates Inc She received her BS degree from NingboUniversity Ningbo China MSCE degree from Zhejiang UniversityHangzhou Zhejiang China and PhD degree in Civil Engineering fromClemson University South Carolina Dr Fang has extensive consultingprojects experience in the field of geoenvironmental engineering
Samuel J Gambino has ten years of experience with URS Corporation inOakland California His specialties include foundation design tunnelinstallations and dams reservoirs and levees Samuel received his bach-elors degree in civil and environmental engineering from the Universityof Michigan at Ann Arbor in 1995 his masters degree in geotechnicalengineering from the University of Texas at Austin in 1998 and has helda professional engineers license in the State of California since 2002
Robert B Gilbert is the Hudson Matlock Professor in Civil Architecturaland Environmental Engineering at The University of Texas at AustinHe joined the faculty in 1993 Prior to that he earned BS (1987) MS(1988) and PhD (1993) degrees in civil engineering from the Universityof Illinois at Urbana-Champaign He also practiced with Golder Asso-ciates Inc as a geotechnical engineer from 1988 to 1993 His expertise isthe assessment evaluation and management of risk in civil engineeringApplications include building foundations slopes pipelines dams andlevees landfills and groundwater and soil remediation systems
Yusuke Honjo is currently a professor and the head of the Civil Engineer-ing Department at Gifu University in Japan He holds ME from Kyoto
List of contributors ix
University and PhD from MIT He was an associate professor and divi-sion chairman at Asian Institute of Technology (AIT) in Bangkok between1989 and 1993 and joined Gifu University in 1995 He is currentlythe chairperson of ISSMGE-TC 23 lsquoLimit state design in geotechnicalengineering practicersquo He has published more than 100 journal papersand international conference papers in the area of statistical analyses ofgeotechnical data inverse analysis and reliability analyses of geotechnicalstructures
Charng Hsein Juang is a Professor of Civil Engineering at Clemson Universityand an Honorary Chair Professor at National Central University TaiwanHe received his BS and MS degrees from National Cheng Kung Universityand PhD degree from Purdue University Dr Juang is a registered Profes-sional Engineering in the State of South Carolina and a Fellow of ASCEHe is recipient of the 1976 Outstanding Research Paper Award ChineseInstitute of Civil and Hydraulic Engineering the 2001 TK Hsieh Awardthe Institution of Civil Engineers United Kingdom and the 2006 BestPaper Award the Taiwan Geotechnical Society Dr Juang has authoredmore than 100 refereed papers in geotechnical related fields and is proudto be selected by his students at Clemson University as the Chi EpsilonOutstanding Teacher in 1984
Fred H Kulhawy is Professor of Civil Engineering at Cornell Univer-sity Ithaca New York He received his BSCE and MSCE from NewJersey Institute of Technology and his PhD from University of Californiaat Berkeley His teaching and research focuses on foundations soil-structure interaction soil and rock behavior and geotechnical computerand reliability applications and he has authored over 330 publicationsHe has lectured worldwide and has received numerous awards fromASCE ADSC IEEE and others including election to Distinguished Mem-ber of ASCE and the ASCE Karl Terzaghi Award and Norman Medal Heis a licensed engineer and has extensive experience in geotechnical practicefor major projects on six continents
David Kun Li is currently a consulting geotechnical engineer with GolderAssociates Inc He received his BSCE and MSCE degrees from ZhejiangUniversity Hangzhou Zhejiang China and PhD degree in Civil Engi-neering from Clemson University South Carolina Dr Li has extensiveconsulting projects experience on retaining structures slope stabilityanalysis liquefaction analysis and reliability assessments
Bak Kong Low obtained his BS and MS degrees from MIT and PhD degreefrom UC Berkeley He is a Fellow of the ASCE and a registered pro-fessional engineer of Malaysia He currently teaches at the NanyangTechnological University in Singapore He had done research while onsabbaticals at HKUST (1996) University of Texas at Austin (1997)
x List of contributors
and Norwegian Geotechnical Institute (2006) His research interest andpublications can be found at httpalummiteduwwwbklow
Shadi S Najjar joined the American University of Beirut (AUB) as an Assis-tant Professor in Civil Engineering in September 2007 Dr Najjar earnedhis BE and ME in Civil Engineering from AUB in 1999 and 2001 respec-tively In 2005 he graduated from the University of Texas at Austin witha PhD in civil engineering Dr Najjarrsquos research involves analytical stud-ies related to reliability-based design in geotechnical engineering Between2005 and 2007 Dr Najjar taught courses on a part-time basis in severalleading universities in Lebanon In addition he worked with Polytech-nical Inc and Dar Al-Handasah Consultants (Shair and Partners) as ageotechnical engineering consultant
Trevor Orr is a Senior Lecturer at Trinity College Dublin He obtainedhis PhD degree from Cambridge University He has been involved inEurocode 7 since the first drafting committee was established in 1981In 2003 he was appointed Chair of the ISSMGE European Technical Com-mittee 10 for the Evaluation of Eurocode 7 and in 2007 was appointeda member of the CEN Maintenance Group for Eurocode 7 He is theco-author of two books on Eurocode 7
Kok-Kwang Phoon is Director of the Centre for Soft Ground Engineering atthe National University of Singapore His main research interest is relatedto development of risk and reliability methods in geotechnical engineer-ing He has authored more than 120 scientific publications includingmore than 20 keynoteinvited papers and edited 15 proceedings He is thefounding editor-in-chief of Georisk and recipient of the prestigious ASCENormal Medal and ASTM Hogentogler Award Webpage httpwwwengnusedusgcivilpeoplecvepkkpkkhtml
Radu Popescu is a Consulting Engineer with URS Corporation and aResearch Professor at Memorial University of Newfoundland CanadaHe earned PhD degrees from the Technical University of Bucharest andfrom Princeton University He was a Visiting Research Fellow at PrincetonUniversity Columbia University and Saitama University (Japan) andLecturer at the Technical University of Bucharest (Romania) andPrinceton University Radu has over 25 years experience in computationaland experimental soil mechanics (dynamic soil-structure interaction soilliquefaction centrifuge modeling site characterization) and publishedover 100 articles in these areas In his research he uses the tools ofprobabilistic mechanics to address various uncertainties manifested in thegeologic environment
Jean H Prevost is presently Professor of Civil and Environmental Engineer-ing at Princeton University He is also an affiliated faculty at the Princeton
List of contributors xi
Materials Institute in the department of Mechanical and AerospaceEngineering and in the Program in Applied and Computational Math-ematics He received his MSc in 1972 and his PhD in 1974 from StanfordUniversity He was a post-doctoral Research Fellow at the NorwegianGeotechnical Institute in Oslo Norway (1974ndash1976) and a ResearchFellow and Lecturer in Civil Engineering at the California Institute ofTechnology (1976ndash1978) He held visiting appointments at the EcolePolytechnique in Paris France (1984ndash1985 2004ndash2005) at the EcolePolytechnique in Lausanne Switzerland (1984) at Stanford University(1994) and at the Institute for Mechanics and Materials at UCSD (1995)He was Chairman of the Department of Civil Engineering and Opera-tions Research at Princeton University (1989ndash1994) His principal areasof interest include dynamics nonlinear continuum mechanics mixturetheories finite element methods XFEM and constitutive theories He isthe author of over 185 technical papers in his areas of interest
Bruno Sudret has a masterrsquos degree from Ecole Polytechnique (France 1993)a masterrsquos degree in civil engineering from Ecole Nationale des Ponts etChausseacutees (1995) and a PhD in civil engineering from the same insti-tution (1999) After a post-doctoral stay at the University of Californiaat Berkeley in 2000 he joined in 2001 the RampD Division of the Frenchmajor electrical company EDF He currently manages a research group onprobabilistic engineering mechanics He is member of the Joint Commit-tee on Structural Safety since 2004 and member of the board of directorsof the International Civil Engineering Risk and Reliability Associationsince 2007 He received the Jean Mandel Prize in 2005 for his work onstructural reliability and stochastic finite element methods
Thomas F Wolff is an Associate Dean of Engineering at Michigan StateUniversity From 1970 to 1985 he was a geotechnical engineer with theUS Army Corps of Engineers and in 1986 he joined MSU His researchand consulting has focused on design and reliability analysis of damslevees and hydraulic structures He has authored a number of Corps ofEngineersrsquo guidance documents In 2005 he served on the ASCE LeveeAssessment Team in New Orleans and in 2006 he served on the InternalTechnical Review (ITR) team for the IPET report which analyzed theperformance of the New Orleans Hurricane Protection System
Tien H Wu is a Professor Emeritus at Ohio State University He receivedhis BS from St Johns University Shanghai China and MS and PhDfrom University of Illinois He has lectured and conducted researchat Norwegian Geotechnical Institute Royal Institute of Technology inStockholm Tonji University in Shanghai National Polytechnical Insti-tute in Quito Cairo University Institute of Forest Research in ChristChurch and others His teaching research and consulting activities
xii List of contributors
involve geotechnical reliability slope stability soil properties glaciologyand soil-bioengineering He is the author of two books Soil Mechanicsand Soil Dynamics He is an Honorary Member of ASCE and has receivedASCErsquos State-of-the-Art Award Peck Award and Earnest Award and theUS Antarctic Service Medal He was the 2008 Peck Lecturer of ASCE
Limin Zhang is an Associate Professor of Civil Engineering and AssociateDirector of Geotechnical Centrifuge Facility at the Hong Kong Universityof Science and Technology His research areas include pile foundationsdams and slopes centrifuge modelling and geotechnical risk and relia-bility He is currently secretary of Technical Committee TC23 on lsquoLimitState Design in Geotechnical Engineeringrsquo and member of TC18 on lsquoDeepFoundationsrsquo of the ISSMGE and Vice Chair of the International Press-InAssociation
Chapter 1
Numerical recipes forreliability analysis ndash a primer
Kok-Kwang Phoon
11 Introduction
Currently the geotechnical community is mainly preoccupied with the tran-sition from working or allowable stress design (WSDASD) to Load andResistance Factor Design (LRFD) The term LRFD is used in a loose wayto encompass methods that require all limit states to be checked using aspecific multiple-factor format involving load and resistance factors Thisterm is used most widely in the United States and is equivalent to LimitState Design (LSD) in Canada Both LRFD and LSD are philosophicallyakin to the partial factors approach commonly used in Europe although adifferent multiple-factor format involving factored soil parameters is usedOver the past few years Eurocode 7 has been revised to accommodate threedesign approaches (DAs) that allow partial factors to be introduced at thebeginning of the calculations (strength partial factors) or at the end of thecalculations (resistance partial factors) or some intermediate combinationsthereof The emphasis is primarily on the re-distribution of the original globalfactor safety in WSD into separate load and resistance factors (or partialfactors)
It is well accepted that uncertainties in geotechnical engineering design areunavoidable and numerous practical advantages are realizable if uncertain-ties and associated risks can be quantified This is recognized in a recentNational Research Council (2006) report on Geological and Geotechni-cal Engineering in the New Millennium Opportunities for Research andTechnological Innovation The report remarked that ldquoparadigms for dealingwith hellip uncertainty are poorly understood and even more poorly practicedrdquoand advocated a need for ldquoimproved methods for assessing the potentialimpacts of these uncertainties on engineering decisions helliprdquo Within the arenaof design code development increasing regulatory pressure is compellinggeotechnical LRFD to advance beyond empirical re-distribution of the orig-inal global factor of safety to a simplified reliability-based design (RBD)framework that is compatible with structural design RBD calls for a willing-ness to accept the fundamental philosophy that (a) absolute reliability is an
2 Kok-Kwang Phoon
unattainable goal in the presence of uncertainty and (b) probability theorycan provide a formal framework for developing design criteria that wouldensure that the probability of ldquofailurerdquo (used herein to refer to exceedingany prescribed limit state) is acceptably small Ideally geotechnical LRFDshould be derived as the logical end-product of a philosophical shift in mind-set to probabilistic design in the first instance and a simplification of rigorousRBD into a familiar ldquolook and feelrdquo design format in the second The needto draw a clear distinction between accepting reliability analysis as a nec-essary theoretical basis for geotechnical design and downstream calibrationof simplified multiple-factor design formats with emphasis on the formerwas highlighted by Phoon et al (2003b) The former provides a consis-tent method for propagation of uncertainties and a unifying frameworkfor risk assessment across disciplines (structural and geotechnical design)and national boundaries Other competing frameworks have been suggested(eg λ-method by Simpson et al 1981 worst attainable value method byBolton 1989 Taylor series method by Duncan 2000) but none has thetheoretical breadth and power to handle complex real-world problems thatmay require nonlinear 3D finite element or other numerical approaches forsolution
Simpson and Yazdchi (2003) proposed that ldquolimit state design requiresanalysis of un-reality not of reality Its purpose is to demonstrate thatlimit states are in effect unreal or alternatively that they are lsquosufficientlyunlikelyrsquo being separated by adequate margins from expected statesrdquo It isclear that limit states are ldquounlikelyrdquo states and the purpose of design is toensure that expected states are sufficiently ldquofarrdquo from these limit states Thepivotal point of contention is how to achieve this separation in numericalterms (Phoon et al 1993) It is accurate to say that there is no consensus onthe preferred method and this issue is still the subject of much heated debatein the geotechnical engineering community Simpson and Yazdchi (2003)opined that strength partial factors are physically meaningful because ldquoit isthe gradual mobilisation of strength that causes deformationrdquo This is con-sistent with our prevailing practice of applying a global factor of safety tothe capacity to limit deformations Another physical justification is that vari-ations in soil parameters can create disproportionate nonlinear variations inthe response (or resistance) (Simpson 2000) In this situation the author feltthat ldquoit is difficult to choose values of partial factors which are applicableover the whole range of the variablesrdquo The implicit assumption here is thatthe resistance factor is not desirable because it is not practical to prescribea large number of resistance factors in a design code However if main-taining uniform reliability is a desired goal a single partial factor for sayfriction angle would not produce the same reliability in different problemsbecause the relevant design equations are not equally sensitive to changes inthe friction angle For complex soilndashstructure interaction problems apply-ing a fixed partial factor can result in unrealistic failure mechanisms In fact
Numerical recipes for reliability analysis 3
given the complexity of the limit state surface it is quite unlikely for thesame partial factor to locate the most probable failure point in all designproblems These problems would not occur if resistance factors are cali-brated from reliability analysis In addition it is more convenient to accountfor the bias in different calculation models using resistance factors How-ever for complex problems it is possible that a large number of resistancefactors are needed and the design code becomes too unwieldy or confusing touse Another undesirable feature is that the engineer does not develop a feelfor the failure mechanism if heshe is only required to analyze the expectedbehavior followed by application of some resistance factors at the end ofthe calculation
Overall the conclusion is that there are no simple methods (factoredparameters or resistances) of replacing reliability analysis for sufficientlycomplex problems It may be worthwhile to discuss if one should insiston simplicity despite all the known associated problems The more recentperformance-based design philosophy may provide a solution for thisdilemma because engineers can apply their own calculations methods forensuring performance compliance without being restricted to following rigiddesign codes containing a few partial factors In the opinion of the authorthe need to derive simplified RBD equations perhaps is of practical impor-tance to maintain continuity with past practice but it is not necessary andit is increasingly fraught with difficulties when sufficiently complex prob-lems are posed The limitations faced by simplified RBD have no bearing onthe generality of reliability theory This is analogous to arguing that limita-tions in closed-form elastic solutions are related to elasto-plastic theory Theapplication of finite element softwares on relatively inexpensive and power-ful PCs (with gigahertz processors a gigabyte of memory and hundreds ofgigabytes ndash verging on terabyte ndash of disk) permit real-world problems to besimulated on an unprecedented realistic setting almost routinely It sufficesto note here that RBD can be applied to complex real-world problems usingpowerful but practical stochastic collocation techniques (Phoon and Huang2007)
One common criticism of RBD is that good geotechnical sense and judg-ment would be displaced by the emphasis on probabilistic analysis (Boden1981 Semple 1981 Bolton 1983 Fleming 1989) This is similar to the on-going criticism of numerical analysis although this criticism seems to havegrown more muted in recent years with the emergence of powerful and user-friendly finite element softwares The fact of the matter is that experiencesound judgment and soil mechanics still are needed for all aspects of geotech-nical RBD (Kulhawy and Phoon 1996) Human intuition is not suited forreasoning with uncertainties and only this aspect has been removed from thepurview of the engineer One example in which intuition can be misleadingis the common misconception that a larger partial factor should be assignedto a more uncertain soil parameter This is not necessarily correct because
4 Kok-Kwang Phoon
the parameter may have little influence on the overall response (eg capacitydeformation) Therefore the magnitude of the partial factor should dependon the uncertainty of the parameter and the sensitivity of the response tothat parameter Clearly judgment is not undermined instead it is focusedon those aspects for which it is most suited
Another criticism is that soil statistics are not readily available becauseof the site-specific nature of soil variability This concern only is true fortotal variability analyses but does not apply to the general approach whereinherent soil variability measurement error and correlation uncertainty arequantified separately Extensive statistics for each component uncertain-ties have been published (Phoon and Kulhawy 1999a Uzielli et al 2007)For each combination of soil type measurement technique and correlationmodel the uncertainty in the design soil property can be evaluated systemat-ically by combining the appropriate component uncertainties using a simplesecond-moment probabilistic approach (Phoon and Kulhawy 1999b)
In summary we are now at the point where RBD really can be used asa rational and practical design mode The main impediment is not theoret-ical (lack of power to deal with complex problems) or practical (speed ofcomputations availability of soil statistics etc) but the absence of simplecomputational approaches that can be easily implemented by practitionersMuch of the controversies reported in the literature are based on qualitativearguments If practitioners were able to implement RBD easily on their PCsand calculate actual numbers using actual examples they would gain a con-crete appreciation of the merits and limitations of RBD Misconceptions willbe dismissed definitively rather than propagated in the literature generatingfurther confusion The author believes that the introduction of powerful butsimple-to-implement approaches will bring about a greater acceptance ofRBD amongst practitioners in the broader geotechnical engineering commu-nity The Probabilistic Model Code was developed by the Joint Committeeon Structural Safety (JCSS) to achieve a similar objective (Vrouwenvelderand Faber 2007)
The main impetus for this book is to explain RBD to students andpractitioners with emphasis on ldquohow to calculaterdquo and ldquohow to applyrdquoPractical computational methods are presented in Chapters 1 3 4 and 7Geotechnical examples illustrating reliability analyses and design are pro-vided in Chapters 5 6 8ndash13 The spatial variability of geomaterials isone of the distinctive aspects of geotechnical RBD This important aspectis covered in Chapter 2 The rest of this chapter provides a primer onreliability calculations with references to appropriate chapters for follow-up reading Simple MATLAB codes are provided in Appendix A andat httpwwwengnusedusgcivilpeoplecvepkkprob_libhtml By focus-ing on demonstration of RBD through calculations and examples thisbook is expected to serve as a valuable teaching and learning resource forpractitioners educators and students
Numerical recipes for reliability analysis 5
12 General reliability problem
The general stochastic problem involves the propagation of input uncer-tainties and model uncertainties through a computation model to arriveat a random output vector (Figure 11) Ideally the full finite-dimensionaldistribution function of the random output vector is desired although par-tial solutions such as second-moment characterizations and probabilities offailure may be sufficient in some applications In principle Monte Carlosimulation can be used to solve this problem regardless of the complexi-ties underlying the computation model input uncertainties andor modeluncertainties One may assume with minimal loss of generality that a com-plex geotechnical problem (possibly 3D nonlinear time and constructiondependent etc) would only admit numerical solutions and the spatialdomain can be modeled by a scalarvector random field Monte Carlosimulation requires a procedure to generate realizations of the input andmodel uncertainties and a numerical scheme for calculating a vector ofoutputs from each realization The first step is not necessarily trivial andnot completely solved even in the theoretical sense Some ldquouser-friendlyrdquomethods for simulating random variablesvectorsprocesses are outlined inSection 13 with emphasis on key calculation steps and limitations Detailsare outside the scope of this chapter and given elsewhere (Phoon 2004a2006a) In this chapter ldquouser-friendlyrdquo methods refer to those that canbe implemented on a desktop PC by a non-specialist with limited pro-gramming skills in other words methods within reach of the generalpractitioner
Model errors
Computationmodel
Uncertain modelparameters
Uncertainoutput vector
Uncertaininput vector
Figure 11 General stochastic problem
6 Kok-Kwang Phoon
The second step is identical to the repeated application of a determin-istic solution process The only potential complication is that a particularset of input parameters may be too extreme say producing a near-collapsecondition and the numerical scheme may become unstable The statisticsof the random output vector are contained in the resulting ensemble ofnumerical output values produced by repeated deterministic runs Fentonand Griffiths have been applying Monte Carlo simulation to solve soil-interaction problems within the context of a random field since the early1990s (eg Griffiths and Fenton 1993 1997 2001 Fenton and Griffiths1997 2002 2003) Popescu and co-workers developed simulation-basedsolutions for a variety of soil-structure interaction problems particularlyproblems involving soil dynamics in parallel Their work is presented inChapter 6 of this book
For a sufficiently large and complex soil-structure interaction problemit is computationally intensive to complete even a single run The rule-of-thumb for Monte Carlo simulation is that 10pf runs are needed to estimatea probability of failure pf within a coefficient of variation of 30 Thetypical pf for a geotechnical design is smaller than one in a thousand andit is expensive to run a numerical code more than ten thousand times evenfor a modest size problem This significant practical disadvantage is wellknown At present it is accurate to say that a computationally efficient andldquouser-friendlyrdquo solution to the general stochastic problem remains elusiveNevertheless reasonably practical solutions do exist if the general stochasticproblem is restricted in some ways for example accept a first-order esti-mate of the probability of failure or accept an approximate but less costlyoutput Some of these probabilistic solution procedures are presented inSection 14
13 Modeling and simulation of stochastic data
131 Random variables
Geotechnical uncertainties
Two main sources of geotechnical uncertainties can be distinguished Thefirst arises from the evaluation of design soil properties such as undrainedshear strength and effective stress friction angle This source of geotechni-cal uncertainty is complex and depends on inherent soil variability degreeof equipment and procedural control maintained during site investigationand precision of the correlation model used to relate field measurementwith design soil property Realistic statistical estimates of the variability ofdesign soil properties have been established by Phoon and Kulhawy (1999a1999b) Based on extensive calibration studies (Phoon et al 1995) threeranges of soil property variability (low medium high) were found to be
Numerical recipes for reliability analysis 7
sufficient to achieve reasonably uniform reliability levels for simplified RBDchecks
Geotechnical parameter Property variability COV ()
Undrained shear strength Low 10ndash30Medium 30ndash50High 50ndash70
Effective stress friction angle Low 5ndash10Medium 10ndash15High 15ndash20
Horizontal stress coefficient Low 30ndash50Medium 50ndash70High 70ndash90
In contrast Reacutethaacuteti (1988) citing the 1965 specification of the AmericanConcrete Institute observed that the quality of concrete can be evaluated inthe following way
Quality COV ()
Excellent lt 10Good 10ndash15Satisfactory 15ndash20Bad gt 20
It is clear that the coefficients of variations or COVs of natural geomaterialscan be much larger and do not fall within a narrow range The ranges ofquality for concrete only apply to the effective stress friction angle
The second source arises from geotechnical calculation models Althoughmany geotechnical calculation models are ldquosimplerdquo reasonable predictionsof fairly complex soilndashstructure interaction behavior still can be achievedthrough empirical calibrations Model factors defined as the ratio of themeasured response to the calculated response usually are used to correctfor simplifications in the calculation models Figure 12 illustrates modelfactors for capacity of drilled shafts subjected to lateral-moment loadingNote that ldquoSDrdquo is the standard deviation ldquonrdquo is the sample size and ldquopADrdquois the p-value from the AndersonndashDarling goodness-of-fit test (gt 005 impliesacceptable lognormal fit) The COVs of model factors appear to fall between30 and 50
It is evident that a geotechnical parameter (soil property or model factor)exhibiting a range of values possibly occurring at unequal frequencies isbest modeled as a random variable The existing practice of selecting onecharacteristic value (eg mean ldquocautiousrdquo estimate 5 exclusion limit etc)is attractive to practitioners because design calculations can be carried outeasily using only one set of input values once they are selected However thissimplicity is deceptive The choice of the characteristic values clearly affects
04 12 20 28 36
HhHu(Reese)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 142 SD = 014 COV = 029
n = 74 pAD = 0186
Hyperbolic capacity (Hh)
04 12 20 28 36HhHu(Broms)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 228 SD = 085 COV = 037
n = 74 pAD = 0149
04 12 20 28 36HhHu(Hansen)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 098 SD = 032 COV = 033
n = 77 pAD = 0229
04 12 20 28 36HhHu(Randolph amp Houlsby)HhHu(Randolph amp Houlsby)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 132 SD = 038 COV = 029
n = 74 pAD = 0270
04 12 20 28 36
HhHu(Simplified Broms)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 130 SD = 049 COV = 038
n = 77 pAD = 0141
00 08 16 24 32
HhHu(Reese)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 092 SD = 027 COV = 029
n = 72 pAD = 0633
Lateral on moment limit (HL)Capacity model
Reese (1958)(undrained)
Broms (1964a)(undrained)
Randolph amp Houlsby(1984) (undrained)
Hansen (1961)(drained)
Broms (1964b)(drained)
00 08 16 24 32HhHu(Broms)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 149 SD = 057 COV = 038
n = 72 pAD = 0122
00 08 16 24 32HhHu(Hansen)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 067 SD = 026 COV = 038
n = 75 pAD = 0468
00 08 16 24 32
HhHu(Simplified Broms)
0
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 088 SD = 036 COV = 041
n = 75 pAD = 0736
00 08 16 24 320
01
02
03
04
Rel
ativ
e F
requ
ency
Mean = 085 SD = 024 COV = 028
n = 72 pAD = 0555
Figure 12 Capacity model factors for drilled shafts subjected to lateral-moment loading(modified from Phoon 2005)
Numerical recipes for reliability analysis 9
the overall safety of the design but there are no simple means of ensur-ing that the selected values will achieve a consistent level of safety In factthese values are given by the design point of a first-order reliability anal-ysis and they are problem-dependent Simpson and Driscoll (1998) notedin their commentary to Eurocode 7 that the definition of the characteristicvalue ldquohas been the most controversial topic in the whole process of draftingEurocode 7rdquo If random variables can be included in the design process withminimal inconvenience the definition of the characteristic value is a mootissue
Simulation
The most intuitive and possibly the most straightforward method for per-forming reliability analysis is the Monte Carlo simulation method It onlyrequires repeated execution of an existing deterministic solution process Thekey calculation step is to simulate realizations of random variables This stepcan be carried in a general way using
Y = Fminus1(U) (11)
in which Y is a random variable following a prescribed cumulative distribu-tion F(middot) and U is a random variable uniformly distributed between 0 and 1(also called a standard uniform variable) Realizations of U can be obtainedfrom EXCELtrade under ldquoTools gt Data Analysis gt Random Number Gener-ation gt Uniform Between 0 and 1rdquo MATLABtrade implements U using theldquorandrdquo function For example U = rand(101) is a vector containing 10 real-izations of U Some inverse cumulative distribution functions are availablein EXCEL (eg norminv for normal loginv for lognormal betainv for betagammainv for gamma) and MATLAB (eg norminv for normal logninv forlognormal betainv for beta gaminv for gamma) More efficient methods areavailable for some probability distributions but they are lacking in generality(Hastings and Peacock 1975 Johnson et al 1994) A cursory examinationof standard probability texts will reveal that the variety of classical proba-bility distributions is large enough to cater to almost all practical needs Themain difficulty lies with the selection of an appropriate probability distribu-tion function to fit the limited data on hand A complete treatment of thisimportant statistical problem is outside the scope of this chapter Howeverit is worthwhile explaining the method of moments because of its simplicityand ease of implementation
The first four moments of a random variable (Y) can be calculated quitereliably from the typical sample sizes encountered in practice Theoreticallythey are given by
micro =int
yf (y)dy = E(Y) (12a)
10 Kok-Kwang Phoon
σ 2 = E[(Y minusmicro)2] (12b)
γ1 = E[(Y minusmicro)3]σ 3 (12c)
γ2 = E[(Y minusmicro)4]σ 4 minus 3 (12d)
in which micro = mean σ 2 = variance (or σ = standard deviation) γ1 =skewness γ2 = kurtosis f (middot) = probability density function = dF(y)dy andE[middot] = mathematical expectation The practical feature here is that momentscan be estimated directly from empirical data without knowledge of theunderlying probability distribution [ie f (y) is unknown]
y =
nsumi=1
yi
n(13a)
s2 = 1n minus 1
nsumi=1
(yi minus y)2 (13b)
g1 = n(n minus 1)(n minus 2)s3
nsumi=1
(yi minus y)3 (13c)
g2 = n(n + 1)(n minus 1)(n minus 2)(n minus 3)s4
nsumi=1
(yi minus y)4 minus 3(n minus 1)2
(n minus 2)(n minus 3)(13d)
in which n = sample size (y1y2 yn) = data points y = sample means2 = sample variance g1 = sample skewness and g2 = sample kurtosis Notethat the MATLAB ldquokurtosisrdquo function is equal to g2 + 3 If the sample sizeis large enough the above sample moments will converge to their respectivetheoretical values as defined by Equation (12) under some fairly general con-ditions The majority of classical probability distributions can be determineduniquely by four or less moments In fact the Pearson system (Figure 13)reduces the selection of an appropriate probability distribution function tothe determination of β1 = g2
1 and β2 = g2 + 3 Calculation steps with illus-trations are given by Elderton and Johnson (1969) Johnson et al (1994)provided useful formulas for calculating the Pearson parameters based onthe first four moments Reacutethaacuteti (1988) provided some β1 and β2 values ofSzeged soils for distribution fitting using the Pearson system (Table 11)
Johnson system
A broader discussion of distribution systems (in which the Pearson sys-tem is one example) and distribution fitting is given by Elderton and
0 02 04 06 08 1
β1
β 2
12 14 16 18
VI
VIV
III
Limit of possible distributions
I
2
VII
N
II
1
2
3
4
5
6
7
8
Figure 13 Pearson system of probability distribution functions based on β1 = g21 and
β2 = g2 + 3 N = normal distribution
Table 11 β1 and β2 values for some soil properties from Szeged Hungary (modifiedfrom Reacutethaacuteti 1988)
Soillayer
Watercontent
Liquidlimit
Plasticlimit
Plasticityindex
Consistencyindex
Voidratio
Bulkdensity
Unconfinedcompressivestrength
β1S1 276 412 681 193 013 650 128 002S2 001 074 034 049 002 001 009 094S3 003 096 014 085 000 130 289 506S4 005 034 013 064 003 013 106 480S5 110 002 292 000 098 036 010 272
β2S1 739 810 1230 492 334 1119 398 186S2 345 343 327 269 317 267 387 393S3 762 513 432 446 331 552 1159 1072S4 717 319 347 357 461 403 814 1095S5 670 231 915 217 513 414 494 674
NotePlasticity index (Ip) = wL minus wP in which wL = liquid limit and wP = plastic limit consistency index(Ic)= (wL minusw)Ip = 1minus IL in which w = water content and IL = liquidity index unconfined compressivestrength (qu) = 2su in which su = undrained shear strength
12 Kok-Kwang Phoon
Johnson (1969) It is rarely emphasized that almost all distribution functionscannot be generalized to handle correlated random variables (Phoon 2006a)In other words univariate distributions such as those discussed above cannotbe generalized to multivariate distributions Elderton and Johnson (1969)discussed some interesting exceptions most of which are restricted to bivari-ate cases It suffices to note here that the only convenient solution availableto date is to rewrite Equation (11) as
Y = Fminus1[Φ(Z)] (14)
in which Z = standard normal random variable with mean = 0 andvariance = 1 and Φ(middot) = standard normal cumulative distribution function(normcdf in MATLAB or normsdist in EXCEL) Equation (14) is calledthe translation model It requires all random variables to be related to thestandard normal random variable The significance of Equation (14) iselaborated in Section 132
An important practical detail here is that standard normal random vari-ables can be simulated directly and efficiently using the BoxndashMuller method(Box and Muller 1958)
Z1 =radic
minus2ln(U1)cos(2πU2) (15)
Z2 =radic
minus2ln(U1)sin(2πU2)
in which Z1Z2 = independent standard normal random variables andU1U2 = independent standard uniform random variables Equation (15)is computationally more efficient than Equation (11) because the inversecumulative distribution function is not required Note that the cumulativedistribution function and its inverse are not available in closed-form for mostrandom variables While Equation (11) ldquolooksrdquo simple there is a hiddencost associated with the numerical evaluation of Fminus1(middot) Marsaglia and Bray(1964) proposed one further improvement
1 Pick V1 and V2 randomly within the unit square extending between minus1and 1 in both directions ie
V1 = 2U1 minus 1
V2 = 2U2 minus 1(16)
2 Calculate R2 = V21 + V2
2 If R2 ge 10 or R2 = 00 repeat step (1)
Numerical recipes for reliability analysis 13
3 Simulate two independent standard normal random variables using
Z1 = V1
radicminus2ln(R2)
R2
Z2 = V2
radicminus2ln(R2)
R2
(17)
Equation (17) is computationally more efficient than Equation (15) becausetrigonometric functions are not required
A translation model can be constructed systematically from (β1 β2) usingthe Johnson system
1 Assume that the random variable is lognormal ie
ln(Y minus A) = λ+ ξZ Y gt A (18)
in which ln(middot) is the natural logarithm ξ2 = ln[1 + σ 2(microminus A)2] λ =ln(microminus A) minus 05ξ2 micro = mean of Y and σ 2 = variance of Y The ldquolog-normalrdquo distribution in the geotechnical engineering literature typicallyrefers to the case of A = 0 When A = 0 it is called the ldquoshiftedlognormalrdquo or ldquo3-parameter lognormalrdquo distribution
2 Calculate ω = exp(ξ2)3 Calculate β1 = (ω minus 1)(ω + 2)2 and β2 = ω4 + 2ω3 + 3ω2 minus 3 For the
lognormal distribution β1 and β2 are related as shown by the solid linein Figure 14
4 Calculate β1 = g21 and β2 = g2 + 3 from data If the values fall close to
the lognormal (LN) line the lognormal distribution is acceptable5 If the values fall below the LN line Y follows the SB distribution
ln(
Y minus AB minus Y
)= λ+ ξZ B gt Y gt A (19)
in which λξAB = distribution fitting parameters6 If the values fall above the LN line Y follows the SU distribution
ln
⎡⎣(Y minus A
B minus A
)+radic
1 +(
Y minus AB minus A
)2⎤⎦= sinhminus1
(Y minus AB minus A
)= λ+ ξZ
(110)
Examples of LN SB and SU distributions are given in Figure 15 Simulationof Johnson random variables using MATLAB is given in Appendix A1Carsel and Parrish (1988) developed joint probability distributions for
0 02 04 06 08 1
β1
β 2
12 14 16 18
SU
Limit of possible distributions
Pearson III
Pearson V
SB
2
1
2
3
4
5
6
7
8
LN
N
Figure 14 Johnson system of probability distribution functions based on β1 = g21 and
β2 = g2 + 3 N = normal distribution LN = lognormal distribution
00
01
02
03
04
05
06
07LNSBSU
08
1 2 3 4 5
Figure 15 Examples of LN (λ = 100ξ = 020) SB (λ = 100ξ = 036 A = minus300B = 500) and SU (λ= 100ξ = 009 A = minus188 B = 208) distributions withapproximately the same mean and coefficient of variation asymp 30
Numerical recipes for reliability analysis 15
parameters of soilndashwater characteristic curves using the Johnson system Themain practical obstacle is that the SB Johnson parameters (λξAB) arerelated to the first four moments (ys2g1g2) in a very complicated way(Johnson 1949) The SU Johnson parameters are comparatively easier toestimate from the first four moments (see Johnson et al 1994 for closed-form formulas) In addition both Pearson and Johnson systems require theproper identification of the relevant region (eg SB or SU in Figure 14)which in turn determines the distribution function [Equation (19) or (110)]before one can attempt to calculate the distribution parameters
Hermite polynomials
One-dimensional Hermite polynomials are given by
H0(Z) = 1
H1(Z) = Z
H2(Z) = Z2 minus 1
H3(Z) = Z3 minus 3Z
Hk+1(Z) = ZHk(Z) minus kHkminus1(Z)
(111)
in which Z is a standard normal random variable (mean = 0 andvariance = 1) Hermite polynomials can be evaluated efficiently using therecurrence relation given in the last row of Equation (111) It can be provenrigorously (Phoon 2003) that any random variable Y (with finite variance)can be expanded as a series
Y =infinsum
k=0
akHk(Z) (112)
The numerical values of the coefficients ak depend on the distribution of YThe key practical advantage of Equation (112) is that the randomness ofY is completely accounted for by the randomness of Z which is a knownrandom variable It is useful to observe in passing that Equation (112) maynot be a monotonic function of Z when it is truncated to a finite number ofterms The extrema are located at points with zero first derivatives but non-zero second derivatives Fortunately derivatives of Hermite polynomials canbe evaluated efficiently as well
dHk(Z)dZ
= kHkminus1(Z) (113)
16 Kok-Kwang Phoon
The Hermite polynomial expansion can be modified to fit the first fourmoments as well
Y = y + s k[Z + h3(Z2 minus 1) + h4(Z3 minus 3Z)] (114)
in which y and s = sample mean and sample standard deviation of Y respec-tively and k = normalizing constant = 1(1 + 2h2
3 + 6h24)05 It is important
to note that the coefficients h3 and h4 can be calculated from the sampleskewness (g1) and sample kurtosis (g2) in a relatively straightforward way(Winterstein et al 1994)
h3 = g1
6
(1 minus 0015
∣∣g1
∣∣+ 03g21
1 + 02g2
)(115a)
h4 = (1 + 125g2)13 minus 110
(1 minus 143g2
1
g2
)1minus01(g2+3)08
(115b)
The theoretical skewness (γ1) and kurtosis (γ2) produced by Equation (114)are (Phoon 2004a)
γ1 = k3(6h3 + 36h3h4 + 8h33 + 108h3h2
4) (116a)
γ2 = k4(3 + 48h43 + 3348h4
4 + 24h4 + 1296h34 + 60h2
3 + 252h24
+ 2232h23h2
4 + 576h23h4) minus 3 (116b)
Equation (115) is determined empirically by minimizing the error [(γ1 minusg1)2 + (γ2 minus g2)2] subjected to the constraint that Equation (114) is amonotonic function of Z It is intended for cases with 0 lt g2 lt 12 and0 le g2
1 lt 2g23 (Winterstein et al 1994) It is possible to minimize the errorin skewness and kurtosis numerically using the SOLVER function in EXCELrather than applying Equation (115)
In general the entire cumulative distribution function of Y F(y) canbe fully described by the Hermite expansion using the following simplestochastic collocation method
1 Let (y1y2 yn) be n realizations of Y The standard normal data iscalculated from
zi = Φminus1F(yi) (117)
Numerical recipes for reliability analysis 17
2 Substitute yi and zi into Equation (112) For a third-order expansionwe obtain
yi = a0 + a1zi + a2(z2i minus 1) + a3(z3
i minus 3zi) = a0 + a1hi1
+ a2hi2 + a3hi3 (118)
in which (1 hi1hi2hi3) are Hermite polynomials evaluated at zi3 The four unknown coefficients (a0a1a2a3) can be determined using
four realizations of Y (y1y2y3y4) In matrix notation we write
Ha = y (119)
in which H is a 4 times 4 matrix with ith row given by (1 hi1hi2hi3) y isa 4 times 1 vector with ith component given by yi and a is a 4 times 1 vectorcontaining the unknown numbers (a0a1a2a3)prime This is known as thestochastic collocation method Equation (119) is a linear system andefficient solutions are widely available
4 It is preferable to solve for the unknown coefficients using regression byusing more than four realizations of Y
(HprimeH)a = Hprimey (120)
in which H is an ntimes4 matrix and n is the number of realizations Notethat Equation (120) is a linear system amenable to fast solution as well
Calculation of Hermite coefficients using MATLAB is given in Appendix A2Hermite coefficients can be calculated with ease using EXCEL as well(Chapter 9) Figure 16 demonstrates that Equation (112) is quite efficient ndashit is possible to match very small probabilities (say 10minus4) using a third-orderHermite expansion (four terms)
Phoon (2004a) pointed out that Equations (14) and (112) are theoreti-cally identical Equation (112) appears to present a rather circuitous routeof achieving the same result as Equation (14) The key computational differ-ence is that Equation (14) requires the costly evaluation of Fminus1(middot) thousandsof times in a low-probability simulation exercise (typical of civil engineeringproblems) while Equation (117) requires less than 100 costly evaluationsof minus1F(middot) followed by less-costly evaluation of Equation (112) thousandsof times Two pivotal factors govern the computational efficiency of Equa-tion (112) (a) cheap generation of standard normal random variables usingthe BoxndashMuller method [Equation (15) or (17)] and (b) relatively smallnumber of Hermite terms
The one-dimensional Hermite expansion can be easily extended to ran-dom vectors and processes The former is briefly discussed in the next
18 Kok-Kwang Phoon
100
Theoretical4-term Hermite
Theoretical
4-term Hermite
Cum
ulat
ive
dist
ribut
ion
func
tion
10minus2
10minus4
10minus6minus1 0 1 2
Value3 4 5
08
07
06
05
Pro
babi
lity
dens
ity fu
nctio
n
04
03
02
01
0minus1 0 1 2
Value3 4 5
Theoretical
4-term Hermite
100
Theoretical4-term Hermite
Cum
ulat
ive
dist
ribut
ion
func
tion
10minus2
10minus4
10minus60 1 2 3
Value4 5 6
08
07
06
05
Pro
babi
lity
dens
ity fu
nctio
n
04
03
02
01
00 1 2 3
Value4 5 6
Figure 16 Four-term Hermite expansions for Johnson SB distribution with λ = 100ξ = 036 A = minus300 B = 500 (top row) and Johnson SU distribution withλ = 100ξ = 009 A = minus188 B = 208 (bottom row)
section while the latter is given elsewhere (Puig et al 2002 Sakamoto andGhanem 2002 Puig and Akian 2004) Chapters 5 7 9 and elsewhere(Sudret and Der Kiureghian 2000 Sudret 2007) present applications ofHermite polynomials in more comprehensive detail
132 Random vectors
The multivariate normal probability density function is available analyticallyand can be defined uniquely by a mean vector and a covariance matrix
f (X) = |C|minus 12 (2π )minus
n2 exp
[minus05(X minus micro)primeCminus1(X minus micro)
](121)
in which X= (X1X2 Xn)prime is a normal random vector with n componentsmicro is the mean vector and C is the covariance matrix For the bivariate
Numerical recipes for reliability analysis 19
(simplest) case the mean vector and covariance matrix are given by
micro =micro1micro2
C =[
σ 21 ρσ1σ2
ρσ1σ2 σ 22
] (122)
in which microi and σi = mean and standard deviation of Xi respectively andρ = product-moment (Pearson) correlation between X1 and X2
The practical usefulness of Equation (121) is not well appreciated Firstthe full multivariate dependency structure of a normal random vector onlydepends on a covariance matrix (C) containing bivariate information (cor-relations) between all possible pairs of components The practical advantageof capturing multivariate dependencies in any dimension (ie any number ofrandom variables) using only bivariate dependency information is obviousNote that coupling two random variables is the most basic form of depen-dency and also the simplest to evaluate from empirical data In fact there areusually insufficient data to calculate reliable dependency information beyondcorrelations in actual engineering practice Second fast simulation of cor-related normal random variables is possible because of the elliptical form(X minus micro)primeCminus1(X minus micro) appearing in the exponent of Equation (121) Whenthe random dimension is small the following Cholesky method is the mostefficient and robust
X = LZ + micro (123)
in which Z = (Z1Z2 Zn)prime contains uncorrelated normal random com-ponents with zero means and unit variances These components can besimulated efficiently using the BoxndashMuller method [Equations (15) or (17)]The lower triangular matrix L is the Cholesky factor of C ie
C = LLprime (124)
Cholesky factorization can be roughly appreciated as taking the ldquosquarerootrdquo of a matrix The Cholesky factor can be calculated in EXCEL usingthe array formula MAT_CHOLESKY which is provided by a free add-inat httpdigilanderliberoitfoxesindexhtm MATLAB produces Lprime usingchol (C) [note Lprime (transpose of L) is an upper triangular matrix] Choleskyfactorization fails if C is not ldquopositive definiterdquo The important practical ram-ifications here are (a) the correlation coefficients in the covariance matrixC cannot be selected independently and (b) an erroneous C is automaticallyflagged when Cholesky factorization fails When the random dimension ishigh it is preferable to use the fast fourier transform (FFT) as described inSection 133
20 Kok-Kwang Phoon
As mentioned in Section 131 the multivariate normal distribution playsa central role in the modeling and simulation of correlated non-normalrandom variables The translation model [Equation (14)] for one randomvariable (Y) can be generalized in a straightforward to a random vector(Y1Y2 Yn)
Yi = Fminus1i [Φ(Xi)] (125)
in which (X1X2 Xn) follows a multivariate normal probability densityfunction [Equation (121)] with
micro =
⎧⎪⎪⎪⎨⎪⎪⎪⎩
00
0
⎫⎪⎪⎪⎬⎪⎪⎪⎭
C =
⎡⎢⎢⎢⎣
1 ρ12 middot middot middot ρ1nρ21 1 middot middot middot ρ2n
ρn1 ρn2 middot middot middot 1
⎤⎥⎥⎥⎦
Note that ρij = ρji ie C is a symmetric matrix containing only n(n minus 1)2distinct entries Each non-normal component Yi can follow any arbitrarycumulative distribution function Fi(middot) The cumulative distribution func-tion prescribed to each component can be different ie Fi(middot) = Fj(middot) Thesimulation of correlated non-normal random variables using MATLAB isgiven in Appendix A3 It is evident that ldquotranslationmrdquo can be modifiedto simulate any number of random components as long as a compatiblesize covariance matrix is specified For example if there are three randomcomponents (n = 3) C should be a 3 times 3 matrix such as [1 08 04 081 02 04 02 1] This computational simplicity explains the popularityof the translation model As mentioned previously not all 3 times 3 matri-ces are valid covariance matrices An example of an invalid covariancematrix is C = [108 minus 08081 minus 02minus08 minus 021] An attempt to executeldquotranslationmrdquo produces the following error messages
Error using ==gt cholMatrix must be positive definite
Hence simulation from an invalid covariance matrix will not take placeThe left panel of Figure 17 shows the scatter plot of two correlated nor-mal random variables with ρ = 08 An increasing linear trend is apparentA pair of uncorrelated normal random variables will not exhibit any trendin the scatter plot The right panel of Figure 17 shows the scatter plot of twocorrelated non-normal random variables There is an increasing nonlineartrend The nonlinearity is fully controlled by the non-normal distributionfunctions (SB and SU in this example) In practice there is no reason forthe scatter plot (produced by two columns of numbers) to be related to the
Numerical recipes for reliability analysis 21
4
2
0
minus2
minus4minus4 minus2 0
X1 = normal distribution
X2 =
nor
mal
dis
trib
utio
n
2 4 01
2
3
4
Y2 =
SU
dis
trib
utio
n
Y1 = SB distribution
5
6
1 2 3 4 5
Figure 17 Scatter plots for correlated normal random variables with zero means and unitvariances (left) and correlated non-normal random variables with 1st compo-nent = Johnson SB distribution (λ = 100ξ = 036A = minus300B = 500) and2nd component = Johnson SU distribution (λ = 100ξ = 009 A = minus188B = 208) (right)
distribution functions (produced by treating each column of numbers sepa-rately) Hence it is possible for the scatter plot produced by the translationmodel to be unrealistic
The simulation procedure described in ldquotranslationmrdquo cannot be applieddirectly to practice because it requires the covariance matrix of X as aninput The empirical data can only produce an estimate of the covariancematrix of Y It can be proven theoretically that these covariance matricesare not equal although they can be approximately equal in some cases Thesimplest solution available so far is to express Equation (125) as Hermitepolynomials
Y1 = a10H0(X1) + a11H1(X1) + a12H2(X1) + a13H3(X1) +middotmiddot middotY2 = a20H0(X2) + a21H1(X2) + a22H2(X2) + a23H3(X2) +middotmiddot middot (126)
The relationship between the observed correlation coefficient (ρY1Y2) and the
underlying normal correlation coefficient (ρ) is
ρY1Y2=
infinsumk=1
ka1ka2kρk
radicradicradicradic( infinsumk=1
ka21k
)(infinsum
k=1ka2
2k
) (127)
This complication is discussed in Chapter 9 and elsewhere (Phoon 2004a2006a)
22 Kok-Kwang Phoon
From a practical viewpoint [only bivariate information needed cor-relations are easy to estimate voluminous n-dimensional data compiledinto mere n(n minus 1)2 coefficients etc] and a computational viewpoint(fast robust simple to implement) it is accurate to say that the multivariatenormal model and multivariate non-normal models generated using thetranslation model are already very good and sufficiently general for manypractical scenarios The translation model is not perfect ndash there are impor-tant and fundamental limitations (Phoon 2004a 2006a) For stochasticdata that cannot be modeled using the translation model the hunt forprobability models with comparable practicality theoretical power andsimulation speed is still on-going Copula theory (Schweizer 1991) pro-duces a more general class of multivariate non-normal models but it isdebatable at this point if these models can be estimated empirically andsimulated numerically with equal ease for high random dimensions Theonly exception is a closely related but non-translation approach based onthe multivariate normal copula (Phoon 2004b) This approach is outlinedin Chapter 9
133 Random processes
A natural probabilistic model for correlated spatial data is the randomfield A one-dimensional random field is typically called a random processA random process can be loosely defined as a random vector with an infi-nite number of components that are indexed by a real number (eg depthcoordinate z) We restrict our discussion to a normal random process X(z)Non-normal random processes can be simulated from a normal randomprocess using the same translation approach described in Section 132
The only computational aspect that requires some elaboration is thatsimulation of a process is usually more efficient in the frequency domainRealizations belonging to a zero-mean stationary normal process X(z) canbe generated using the following spectral approach
X(z) =infinsum
k=1
σk(Z1k sin2π fkz+Z2k cos2π fkz) (128)
in which σk =radic2S(fk)f f is the interval over which the spectral densityfunction S(f ) is discretized fk = (2k minus 1)f 2 and Z1k and Z2k are uncor-related normal random variables with zero means and unit variances Thesingle exponential autocorrelation function is commonly used in geostatis-tics R(τ ) = exp(minus2|τ |δ) in which τ is the distance between data points andδ the scale of fluctuation R(τ ) can be estimated from a series of numbers saycone tip resistances sampled at a vertical spacing of z using the method of
Numerical recipes for reliability analysis 23
moments
R(τ = jz) asymp 1(n minus j minus 1)s2
nminusjsumi=1
(xi minus x)(xi+j minus x) (129)
in which xi = x(zi)zi = i(z) x = sample mean [Equation (13a] s2 =sample variance [Equation (13b)] and n = number of data points The scaleof fluctuation is estimated by fitting an exponential function to Equation(129) The spectral density function corresponding to the single exponentialcorrelation function is
S(f ) = 4δ(2π f δ)2 + 4
(130)
Other common autocorrelation functions and their corresponding spectraldensity functions are given in Table 12 It is of practical interest to note thatS(f ) can be calculated numerically from a given target autocorrelation func-tion or estimated directly from x(zi) using the FFT Analytical solutions suchas those shown in Table 12 are convenient but unnecessary The simulationof a standard normal random process with zero mean and unit variance usingMATLAB is given in Appendix A4 Note that the main input to ldquoranpro-cessmrdquo is R(τ ) which can be estimated empirically from Equation (129)No knowledge of S(f ) is needed Some realizations based on the five com-mon autocorrelation functions shown in Table 12 with δ = 1 are given inFigure 18
Table 12 Common autocorrelation and two-sided power spectral density functions
Model Autocorrelation R(τ ) Two-sided power spectraldensity S(f )
Scale offluctuation δ
Singleexponential
exp(minusa |τ |) 2a
ω2 + a22a
Binary noise1 minus a |τ | |τ | le 1a0 otherwise
sin2 (ω2a)
a (ω2a)21a
Cosineexponential
exp(minusa |τ |)cos(aτ ) a(
1
a2 + (ω+ a)2+ 1
a2 + (ωminus a)2
)1a
Second-orderMarkov
(1 + a |τ |)exp(minusa |τ |) 4a3
(ω2 + a2)24a
Squaredexponential
exp[minus(aτ )2
] radicπ
aexp
(minus ω2
4a2
) radicπ
a
ω = 2π f
4
2
0
0 5 10 15 20
Depth z
25 30 35 40
X(z
)
minus2
minus4
(a)
4
2
0
0 5 10 15 20
Depth z
25 30 35 40
X(z
)
minus2
minus4
(b)
4
2
0
0 5 10 15 20
Depth z
25 30 35 40
X(z
)
minus2
minus4
(c)
Figure 18 Simulated realizations of normal process with mean = 0 variance = 1 scaleof fluctuation = 1 based on autocorrelation function = (a) single exponential(b) binary noise (c) cosine exponential (d) second-order Markov and (e) squareexponential
Numerical recipes for reliability analysis 25
4
2
0
0 5 10 15 20
Depth z
25 30 35 40
X(z
)
minus2
minus4
(d)
4
2
0
0 5 10 15 20
Depth z
25 30 35 40
X(z
)
minus2
minus4
(e)
Figure 18 Contrsquod
In geotechnical engineering spatial variability is frequently modeled usingrandom processesfields This practical application was first studied in geol-ogy and mining under the broad area of geostatistics The physical premisethat makes estimation of spatial patterns possible is that points close in spacetend to assume similar values The autocorrelation function (or variogram)is a fundamental tool describing similarities in a statistical sense as a func-tion of distance [Equation (129)] The works of G Matheron DG Krigeand FP Agterberg are notable Parallel developments also took place inmeteorology (LS Gandin) and forestry (B Mateacutern) Geostatistics is math-ematically founded on the theory of random processesfields developed byAY Khinchin AN Kolmogorov P Leacutevy N Wiener and AM Yaglomamong others The interested reader can refer to books by Cressie (1993)and Chilegraves and Delfiner (1999) for details VanMarcke (1983) remarked thatall measurements involve some degree of local averaging and that randomfield models do not need to consider variations below a finite scale becausethey are smeared by averaging VanMarckersquos work is incomplete in one
26 Kok-Kwang Phoon
crucial aspect This crucial aspect is that actual failure surfaces in 2D or 3Dproblems will automatically seek to connect ldquoweakestrdquo points in the randomdomain It is the spatial average of this failure surface that counts in practicenot simple spatial averages along pre-defined surfaces (or pre-specified mea-surement directions) This is an exceptionally difficult problem to solve ndash atpresent simulation is the only viable solution Chapter 6 presents extensiveresults on emergent behaviors resulting from spatially heterogeneous soilsthat illustrate this aspect quite thoroughly
Although the random processfield provides a concise mathematical modelfor spatial variability it poses considerable practical difficulties for statisti-cal inference in view of its complicated data structure All classical statisticaltests are invariably based on the important assumption that the data areindependent (Cressie 1993) When they are applied to correlated data largebias will appear in the evaluation of the test statistics (Phoon et al 2003a)The application of standard statistical tests to correlated soil data is thereforepotentially misleading Independence is a very convenient assumption thatmakes a large part of mathematical statistics tractable Statisticians can goto great lengths to remove this dependency (Fisher 1935) or be contentwith less-powerful tests that are robust to departures from the indepen-dence assumption In recent years an alternate approach involving the directmodeling of dependency relationships into very complicated test statisticsthrough Monte Carlo simulation has become feasible because desktop com-puting machines have become very powerful (Phoon et al 2003a Phoonand Fenton 2004 Phoon 2006b Uzielli and Phoon 2006)
Chapter 2 and elsewhere (Baecher and Christian 2003 Uzielli et al2007) provide extensive reviews of geostatistical applications in geotechnicalengineering
14 Probabilistic solution procedures
The practical end point of characterizing uncertainties in the design inputparameters (geotechnical geo-hydrological geometrical and possibly ther-mal) is to evaluate their impact on the performance of a design Reliabilityanalysis focuses on the most important aspect of performance namely theprobability of failure (ldquofailurerdquo is a generic term for non-performance)This probability of failure clearly depends on both parametric and modeluncertainties The probability of failure is a more consistent and completemeasure of safety because it is invariant to all mechanically equivalent def-initions of safety and it incorporates additional uncertainty informationThere is a prevalent misconception that reliability-based design is ldquonewrdquo Allexperienced engineers would conduct parametric studies when confidencein the choice of deterministic input values is lacking Reliability analysismerely allows the engineer to carry out a much broader range of para-metric studies without actually performing thousands of design checks with
Numerical recipes for reliability analysis 27
different inputs one at a time This sounds suspiciously like a ldquofree lunchrdquobut exceedingly clever probabilistic techniques do exist to calculate theprobability of failure efficiently The chief drawback is that these tech-niques are difficult to understand for the non-specialist but they are notnecessarily difficult to implement computationally There is no consensuswithin the geotechnical community whether a more consistent and com-plete measure of safety is worth the additional efforts or if the significantlysimpler but inconsistent global factor of safety should be dropped Reg-ulatory pressure appears to be pushing towards RBD for non-technicalreasons The literature on reliability analysis and RBD is voluminous Someof the main developments in geotechnical engineering are presented inthis book
141 Closed-form solutions
There is a general agreement in principle that limit states (undesirable statesin which the system fails to perform satisfactorily) should be evaluatedexplicitly and separately but there is no consensus on how to verify thatexceedance of a limit state is ldquosufficiently unlikelyrdquo in numerical terms Dif-ferent opinions and design recommendations have been made but there is alack of discussion on basic issues relating to this central idea of ldquoexceeding alimit staterdquo A simple framework for discussing such basic issues is to imag-ine the limit state as a boundary surface dividing sets of design parameters(soil load andor geometrical parameters) into those that result in satisfac-tory and unsatisfactory designs It is immediately clear that this surface canbe very complex for complex soil-structure interaction problems It is alsoclear without any knowledge of probability theory that likely failure sets ofdesign parameters (producing likely failure mechanisms) cannot be discussedwithout characterizing the uncertainties in the design parameters explicitlyor implicitly Assumptions that ldquovalues are physically boundedrdquo ldquoall valuesare likely in the absence of informationrdquo etc are probabilistic assump-tions regardless of whether or not this probabilistic nature is acknowledgedexplicitly If the engineer is 100 sure of the design parameters then onlyone design check using these fully deterministic parameters is necessary toensure that the relevant limit state is not exceeded Otherwise the situationis very complex and the only rigorous method available to date is reliabilityanalysis The only consistent method to control exceedance of a limit stateis to control the reliability index It would be very useful to discuss at thisstage if this framework is logical and if there are alternatives to it In fact ithas not been acknowledged explicitly if problems associated with strengthpartial factors or resistance factors are merely problems related to simpli-fication of reliability analysis If there is no common underlying purpose(eg to achieve uniform reliability) for applying partial factors and resis-tance factors then the current lack of consensus on which method is better
28 Kok-Kwang Phoon
cannot be resolved in any meaningful way Chapter 8 provides some usefulinsights on the reliability levels implied by the empirical soil partial factorsin Eurocode 7
Notwithstanding the above on-going debate it is valid to question if reli-ability analysis is too difficult for practitioners The simplest example is toconsider a foundation design problem involving a random capacity (Q) anda random load (F) The ultimate limit state is defined as that in which thecapacity is equal to the applied load Clearly the foundation will fail if thecapacity is less than this applied load Conversely the foundation should per-form satisfactorily if the applied load is less than the capacity These threesituations can be described concisely by a single performance function P asfollows
P = Q minus F (131)
Mathematically the above three situations simply correspond to the condi-tions of P = 0P lt 0 and P gt 0 respectively
The basic objective of RBD is to ensure that the probability of failure doesnot exceed an acceptable threshold level This objective can be stated usingthe performance function as follows
pf = Prob(P lt 0) le pT (132)
in which Prob(middot) = probability of an event pf =probability of failure andpT = acceptable target probability of failure A more convenient alternativeto the probability of failure is the reliability index (β) which is defined as
β = minusΦminus1(pf) (133)
in which Φminus1(middot) = inverse standard normal cumulative function The func-tion Φminus1(middot) can be obtained easily from EXCEL using normsinv (pf) orMATLAB using norminv (pf) For sufficiently large β simple approximateclosed-form solutions for Φ(middot) and Φminus1(middot) are available (Appendix B)
The basic reliability problem is to evaluate pf from some pertinent statisticsof F and Q which typically include the mean (microF or microQ) and the standarddeviation (σF or σQ) and possibly the probability density function A simpleclosed-form solution for pf is available if Q and F follow a bivariate normaldistribution For this condition the solution to Equation (132) is
pf = Φ
⎛⎜⎝minus microQ minusmicroFradic
σ 2Q +σ 2
F minus 2ρQFσQσF
⎞⎟⎠= Φ (minusβ) (134)
in which ρQF = product-moment correlation coefficient between Q and FNumerical values for Φ(middot) can be obtained easily using the EXCEL func-tion normsdist(minusβ) or the MATLAB function normcdf(minusβ) The reliability
Numerical recipes for reliability analysis 29
indices for most geotechnical components and systems lie between 1 and5 corresponding to probabilities of failure ranging from about 016 to3 times 10minus7 as shown in Figure 19
Equation (134) can be generalized to a linear performance function con-taining any number of normal components (X1X2 Xn) as long as theyfollow a multivariate normal distribution function [Equation (121)]
P = a0 +nsum
i=1
aiXi (135)
pf = Φ
⎛⎜⎜⎜⎜⎝minus
a0 +nsum
i=1aimicroiradic
nsumi=1
nsumj=1
aiajρijσiσj
⎞⎟⎟⎟⎟⎠ (136)
in which ai = deterministic constant microi = mean of Xiσi = standard devia-tion of Xiρij = correlation between Xi and Xj (note ρii = 1) Chapters 811 and 12 present some applications of these closed-form solutions
Equation (134) can be modified for the case of translation lognormals ieln(Q) and ln(F) follow a bivariate normal distribution with mean of ln(Q) =λQ mean of ln(F) =λF standard deviation of ln(Q) = ξQ standard deviation
Unsatisfactory
Hazardous1E+00
1Eminus01
1Eminus02
1Eminus03
Pro
bab
ility
of
failu
re
1Eminus04
1Eminus05
0 1 2 3
Reliability index4 5
1Eminus06
1Eminus07
PoorBelow average
Above average
Good
High
asymp10minus7
asymp10minus5
asymp10minus3
asymp2asymp7
asymp16
asymp5times10minus3
Figure 19 Relationship between reliability index and probability of failure (classificationsproposed by US Army Corps of Engineers 1997)
30 Kok-Kwang Phoon
of ln(F) = ξF and correlation between ln(Q) and ln(F) = ρprimeQF
pf = Φ
⎛⎜⎝minus λQ minusλFradic
ξ2Q + ξ2
F minus 2ρprimeQFξQξF
⎞⎟⎠ (137)
The relationships between the mean (micro) and standard deviation (σ ) of alognormal and the mean (λ) and standard deviation (ξ ) of the equivalentnormal are given in Equation (18) The correlation between Q and F (ρQF)is related to the correlation between ln(Q) and ln(F) (ρprime
QF) as follows
ρQF =exp(ξQξFρ
primeQF) minus 1radic
[exp(ξ2Q) minus 1][exp(ξ2
F ) minus 1](138)
If ρprimeQF = ρQF = 0 (ie Q and F are independent lognormals) Equation (137)
reduces to the following well-known expression
β =ln(
microQmicroF
radic1+COV2
F1+COV2
Q
)radic
ln[(
1 + COV2Q
)(1 + COV2
F
)] (139)
in which COVQ = σQmicroQ and COVF = σFmicroF If there are physical groundsto disallow negative values the translation lognormal model is more sensibleEquation (139) has been used as the basis for RBD (eg Rosenblueth andEsteva 1972 Ravindra and Galambos 1978 Becker 1996 Paikowsky2002) The calculation steps outlined below are typically carried out
1 Consider a typical Load and Resistance Factor Design (LRFD) equation
φQn = γDDn + γLLn (140)
in which φ = resistance factor γD and γL = dead and live load fac-tors and Qn Dn and Ln = nominal values of capacity dead loadand live load The AASHTO LRFD bridge design specifications rec-ommended γD = 125 and γL = 175 (Paikowsky 2002) The resistancefactor typically lies between 02 and 08
2 The nominal values are related to the mean values as
microQ = bQQn
microD = bDDn
microL = bLLn
(141)
Numerical recipes for reliability analysis 31
in which bQ bD and bL are bias factors and microQ microD and microL are meanvalues The bias factors for dead and live load are typically 105 and 115respectively Figure 12 shows that bQ (mean value of the histogram) canbe smaller or larger than one
3 Assume typical values for the mean capacity (microQ) and DnLn A reason-able range for DnLn is 1 to 4 Calculate the mean live load and meandead load as follows
microL = (φbQmicroQ)bL(γD
DnLn
+ γL
) (142)
microD =(
Dn
Ln
)microLbD
bL(143)
4 Assume typical coefficients of variation for the capacity dead load andlive load say COVQ = 03 COVD = 01 and COVL = 02
5 Calculate the reliability index using Equation (139) with microF = microD +microL and
COVF =radic
(microDCOVD)2 + (microLCOVL)2
microD +microL(144)
Details of the above procedure are given in Appendix A4 For φ = 05 γD =125 γL = 175 bQ = 1bD = 105bL = 115 COVQ = 03 COVD = 01COVL = 02 and DnLn = 2 the reliability index from Equation (139) is299 Monte Carlo simulation in ldquoLRFDmrdquo validates the approximationgiven in Equation (144) It is easy to show that an alternate approximationCOV2
F = COV2D+ COV2
L is erroneous Equation (139) is popular becausethe resistance factor in LRFD can be back-calculated from a target reliabilityindex (βT) easily
φ =bQ(γDDn + γLLn)
radic1+COV2
F1+COV2
Q
(bDDn + bLLn)expβT
radicln[(
1 + COV2Q
)(1 + COV2
F
)] (145)
In practice the statistics of Q are determined by comparing load test resultswith calculated values Phoon and Kulhawy (2005) compared the statisticsof Q calculated from laboratory load tests with those calculated from full-scale load tests They concluded that these statistics are primarily influencedby model errors rather than uncertainties in the soil parameters Hence it islikely that the above lumped capacity approach can only accommodate the
32 Kok-Kwang Phoon
ldquolowrdquo COV range of parametric uncertainties mentioned in Section 131To accommodate larger uncertainties in the soil parameters (ldquomediumrdquo andldquohighrdquo COV ranges) it is necessary to expand Q as a function of the gov-erning soil parameters By doing so the above closed-form solutions are nolonger applicable
142 First-order reliability method (FORM)
Structural reliability theory has a significant impact on the development ofmodern design codes Much of its success could be attributed to the advent ofthe first-order reliability method (FORM) which provides a practical schemeof computing small probabilities of failure at high dimensional space spannedby the random variables in the problem The basic theoretical result wasgiven by Hasofer and Lind (1974) With reference to time-invariant reliabil-ity calculation Rackwitz (2001) observed that ldquoFor 90 of all applicationsthis simple first-order theory fulfills all practical needs Its numerical accuracyis usually more than sufficientrdquo Ang and Tang (1984) presented numer-ous practical applications of FORM in their well known book ProbabilityConcepts in Engineering Planning and Design
The general reliability problem consists of a performance func-tion P(y1y2 yn) and a multivariate probability density functionf (y1y2 yn) The former is defined to be zero at the limit state less thanzero when the limit state is exceeded (ldquofailrdquo) and larger than zero otherwise(ldquosaferdquo) The performance function is nonlinear for most practical prob-lems The latter specifies the likelihood of realizing any one particular setof input parameters (y1y2 yn) which could include material load andgeometrical parameters The objective of reliability analysis is to calculatethe probability of failure which can be expressed formally as follows
pf =int
Plt0f(y1y2 yn
)dy1dy2 dyn (146)
The domain of integration is illustrated by a shaded region in the left panelof Figure 110a Exact solutions are not available even if the multivariateprobability density function is normal unless the performance function islinear or quadratic Solutions for the former case are given in Section 141Other exact solutions are provided in Appendix C Exact solutions are veryuseful for validation of new reliability codes or calculation methods Theonly general solution to Equation (146) is Monte Carlo simulation A simpleexample is provided in Appendix A5 The calculation steps outlined beloware typically carried out
1 Determine (y1y2 yn) using Monte Carlo simulation Section 132has presented fairly general and ldquouser friendlyrdquo methods of simulatingcorrelated non-normal random vectors
Numerical recipes for reliability analysis 33
βminuspoint
β
zlowast = (z1lowastz2
lowast)
PLlt 0
P lt 0FAIL
P gt 0SAFE PLgt 0
Actual P = 0
Approx PL=0
fz1z2(z1z2)
fy1y2(y1y2)
z2
z1
y2
y1
G(y1y2) = 0
Figure 110 (a) General reliability problem and (b) solution using FORM
2 Substitute (y1y2 yn) into the performance function and count thenumber of cases where P lt 0 (ldquofailurerdquo)
3 Estimate the probability of failure using
pf = nf
n(147)
in which nf = number of failure cases and n =number of simulations4 Estimate the coefficient of variation of pf using
COVpf=radic
1 minus pf
pfn(148)
For civil engineering problems pf asymp 10minus3 and hence (1 ndash pf) asymp 1 The samplesize (n) required to ensure COVpf
is reasonably small say 03 is
n = 1 minus pf
pfCOV2pf
asymp 1pf(03)2
asymp 10pf
(149)
It is clear from Equation (149) that Monte Carlo simulation is not prac-tical for small probabilities of failure It is more often used to validateapproximate but more efficient solution methods such as FORM
The approximate solution obtained from FORM is easier to visualize ina standard space spanned by uncorrelated Gaussian random variables withzero mean and unit standard deviation (Figure 110b) If one replaces theactual limit state function (P = 0) by an approximate linear limit state func-tion (PL = 0) that passes through the most likely failure point (also called
34 Kok-Kwang Phoon
design point or β-point) it follows immediately from rotational symmetryof the circular contours that
pf asymp Φ(minusβ) (150)
The practical result of interest here is that Equation (146) simply reduces toa constrained nonlinear optimization problem
β = minradic
zprimez forz G(z) le 0 (151)
in which z = (z1z2 zn)prime The solution of a constrained optimization prob-lem is significantly cheaper than the solution of a multi-dimensional integral[Equation (146)] It is often cited that the β-point is the ldquobestrdquo linearizationpoint because the probability density is highest at that point In actuality thechoice of the β-point requires asymptotic analysis (Breitung 1984) In shortFORM works well only for sufficiently large β ndash the usual rule-of-thumb isβ gt 1 (Rackwitz 2001)
Low and co-workers (eg Low and Tang 2004) demonstrated that theSOLVER function in EXCEL can be easily implemented to calculate thefirst-order reliability index for a range of practical problems Their studiesare summarized in Chapter 3 The key advantages to applying SOLVER forthe solution of Equation (151) are (a) EXCEL is available on almost allPCs (b) most practitioners are familiar with the EXCEL user interface and(c) no programming skills are needed if the performance function can becalculated using EXCEL built-in mathematical functions
FORM can be implemented easily within MATLAB as well Appendix A6demonstrates the solution process for an infinite slope problem (Figure 111)The performance function for this problem is
P =[γ(H minus h
)+ h(γsat minus γw
)]cosθ tanφ[
γ(H minus h
)+ hγsat]sinθ
minus 1 (152)
in which H = depth of soil above bedrock h = height of groundwater tableabove bedrock γ and γsat = moist unit weight and saturated unit weight ofthe surficial soil respectively γw = unit weight of water (981 kNm3) φ =effective stress friction angle and θ = slope inclination Note that the heightof the groundwater table (h) cannot exceed the depth of surficial soil (H)and cannot be negative Hence it is modeled by h = H times U in which U =standard uniform variable The moist and saturated soil unit weights are notindependent because they are related to the specific gravity of the soil solids(Gs) and the void ratio (e) The uncertainties in γ and γsat are characterizedby modeling Gs and e as two independent uniform random variables Thereare six independent random variables in this problem (HUφβe and Gs)
Numerical recipes for reliability analysis 35
ROCKh
SOIL
H
θγ
γsat
Figure 111 Infinite slope problem
Table 13 Summary of probability distributions for input random variables
Variable Description Distribution Statistics
H Depth of soil above bedrock Uniform [28] mh = H times U Height of water table U is uniform [0 1]φ Effective stress friction angle Lognormal mean = 35 cov = 8θ Slope inclination Lognormal mean = 20 cov = 5γ Moist unit weight of soil γsat Saturated unit weight of soil γw Unit weight of water Deterministic 981 kNm3
γ = γw (Gs + 02e)(1 + e) (assume degree of saturation = 20 for ldquomoistrdquo) γsat = γw (Gs + e)(1 + e) (degree of saturation = 100)Assume specific gravity of solids = Gs = uniformly distributed [25 27] and void ratio = e = uniformlydistributed [03 06]
and their probability distributions are summarized in Table 13 The first-order reliability index is 143 The reliability index calculated from MonteCarlo simulation is 157
Guidelines for modifying the code to solve other problems can besummarized as follows
1 Specify the number of random variables in the problem using theparameter ldquomrdquo in ldquoFORMmrdquo
2 The objective function ldquoobjfunmrdquo is independent of the problem3 The performance function ldquoPfunmrdquo can be modified in a straight-
forward way The only slight complication is that the physical vari-ables (eg HUφβe and Gs) must be expressed in terms of thestandard normal variables (eg z1z2 z6) Practical methods for
36 Kok-Kwang Phoon
converting uncorrelated standard normal random variables to correlatednon-normal random variables have been presented in Section 132
Low and Tang (2004) opined that practitioners will find it easier to appre-ciate and implement FORM without the above conversion procedureNevertheless it is well-known that optimization in the standard space is morestable than optimization in the original physical space The SOLVER optionldquoUse Automatic Scalingrdquo only scales the elliptical contours in Figure 110abut cannot make them circular as in Figure 110b Under some circum-stances SOLVER will produce different results from different initial trialvalues Unfortunately there are no automatic and dependable means of flag-ging this instability Hence it is extremely vital for the user to try differentinitial trial values and partially verify that the result remains stable Theassumption that it is sufficient to use the mean values as the initial trialvalues is untrue
In any case it is crucial to understand that a multivariate probabil-ity model is necessary for any probabilistic analysis involving more thanone random variable FORM or otherwise It is useful to recall thatmost multivariate non-normal probability models are related to the mul-tivariate normal probability model in a fundamental way as discussed inSection 132 Optimization in the original physical space does not elim-inate the need for the underlying multivariate normal probability modelif the non-normal physical random vector is produced by the translationmethod It is clear from Section 132 that non-translation methods exist andnon-normal probability models cannot be constructed uniquely from corre-lation information alone Chapters 9 and 13 report applications of FORMfor RBD
143 System reliability based on FORM
The first-order reliability method (FORM) is capable of handling any non-linear performance function and any combination of correlated non-normalrandom variables Its accuracy depends on two main factors (a) the cur-vature of the performance function at the design point and (b) the numberof design points If the curvature is significant the second-order reliabilitymethod (SORM) (Breitung 1984) or importance sampling (Rackwitz 2001)can be applied to improve the FORM solution Both methods are relativelyeasy to implement although they are more costly than FORM The calcu-lation steps for SORM are given in Appendix D Importance sampling isdiscussed in Chapter 4 If there are numerous design points FORM canunderestimate the probability of failure significantly At present no solutionmethod exists that is of comparable simplicity to FORM Note that problemscontaining multiple failure modes are likely to produce more than one designpoint
Numerical recipes for reliability analysis 37
Figure 112 illustrates a simple system reliability problem involving twolinear performance functions P1 and P2 A common geotechnical engineer-ing example is a shallow foundation subjected to inclined loading It isgoverned by bearing capacity (P1) and sliding (P2) modes of failure Thesystem reliability is formally defined by
pf = Prob[(P1 lt 0) cup (P2 lt 0)] (153)
There are no closed-form solutions even if P1 and P2 are linear and if theunderlying random variables are normal and uncorrelated A simple estimatebased on FORM and second-order probability bounds is available and is ofpractical interest The key calculation steps can be summarized as follows
1 Calculate the correlation between failure modes using
ρP2P1= α1 middotα2 = α11α21 +α12α22 = cosθ (154)
in which αi = (αi1αi2) = unit normal at design point for ith performancefunction Referring to Figure 110b it can be seen that this unit normalcan be readily obtained from FORM as αi1 = zlowast
i1βi and αi2 = zlowasti2βi
with (zlowasti1z
lowasti2) = design point for ith performance function
P2(P1 lt 0) cap (P2 lt 0)
fz1z2(z1z2)
z2
z1
θ
P1
β1
β2
α1
α2
Figure 112 Simple system reliability problem
38 Kok-Kwang Phoon
2 Estimate p21 = Prob[(P2 lt 0) cap (P1 lt 0)] using first-order probabilitybounds
pminus21 le p21 le p+
21
max[P(B1)P(B2)] le p21 le P(B1) + P(B2)(155)
in which P(B1) = Φ(minusβ1)Φ[(β1 cosθ minus β2)sinθ ] and P(B2) =Φ(minusβ2)Φ[(β2 cosθ minus β1)sinθ ] Equation (155) only applies forρP2P1
gt 0 Failure modes are typically positively correlated becausethey depend on a common set of loadings
3 Estimate pf using second-order probability bounds
p1 + max[(p2 minus p+21)0] le pf le min[p1 + (p2 minus pminus
21)1] (156)
in which pi = Φ(minusβi)
The advantages of the above approach are that it does not require infor-mation beyond what is already available in FORM and generalization to nfailure modes is quite straightforward
pf ge p1 + max[(p2 minus p+21)0]+ max[(p3 minus p+
31 minus p+320]+ middot middot middot
+ max[(pn minus p+n1 minus p+
n2 minusmiddotmiddot middotminus p+nnminus1)0] (157a)
pf le p1 + minp1 +[p2 minus pminus21]+ [p3 minus max(pminus
31pminus32)]+ middot middot middot
+ [pn minus max(pminusn1p
minusn2 middot middot middotpminus
nnminus1)]1 (157b)
The clear disadvantages are that no point probability estimate is availableand calculation becomes somewhat tedious when the number of failuremodes is large The former disadvantage can be mitigated using the followingpoint estimate of p21 (Mendell and Elston 1974)
a1 = 1
Φ(minusβ1)radic
2πexp
(minusβ2
1
2
)(158)
p21 asymp Φ
⎡⎢⎣ ρP1P2
a1 minusβ2radic1 minusρP1P2
2a1(a1 minusβ1)
⎤⎥⎦Φ(minusβ1) (159)
The accuracy of Equation (159) is illustrated in Table 14 Bold numbersshaded in gray are grossly inaccurate ndash they occur when the reliability indicesare significantly different and the correlation is high
Numerical recipes for reliability analysis 39
Table 14 Estimation of p21 using probability bounds point estimate and simulation
Performancefunctions
Correlation Probability bounds Point estimate Simulation
β1 = 1 01 (29 58) times10minus2 31 times10minus2 31 times10minus2
β2 = 1 05 (45 89) times10minus2 63 times10minus2 63 times10minus2
09 (06 13) times10minus1 12 times10minus1 12 times10minus1
β1 = 1 01 (48 93) times10minus3 50 times10minus3 50 times10minus3
β2 = 2 05 (11 18) times10minus2 13 times10minus2 13 times10minus2
09 (22 23) times10minus2 23 times10minus2 23 times10minus2
β1 = 1 01 (33 61) times10minus4 34 times10minus4 35 times10minus4
β2 = 3 05 (10 13) times10minus3 10 times10minus3 10 times10minus3
09 (13 13) times10minus3 05 times10minus3 13 times10minus3
β1 = 1 01 (86 157) times10minus6 89 times10minus6 85 times10minus6
β2 = 4 05 (28 32) times10minus5 23 times10minus5 28 times10minus5
09 (32 32) times10minus5 007 times10minus5 32 times10sminus5
β1 = 2 01 (80 160) times10minus4 87 times10minus4 88 times10minus4
β2 = 2 05 (28 56) times10minus3 41 times10minus3 41 times10minus3
09 (07 15) times10minus2 14 times10minus2 13 times10minus2
β1 = 2 01 (59 115) times10minus5 63 times10minus5 60 times10minus5
β2 = 3 05 (38 62) times10minus4 45 times10minus4 45 times10minus4
09 (13 13) times10minus3 12 times10minus3 13 times10minus3
β1 = 2 01 (17 32) times10minus6 18 times10minus6 25 times10minus6
β2 = 4 05 (16 22) times10minus5 16 times10minus5 15 times10minus5
09 (32 32) times10minus5 05 times10minus5 32 times10minus5
β1 = 3 01 (45 90) times10minus6 49 times10minus6 45 times10minus6
β2 = 3 05 (56 112) times10minus5 82 times10minus5 80 times10minus5
09 (33 66) times 10minus4 63 times 10minus4 60 times10minus4
Sample size = 5000000
144 Collocation-based stochastic response surfacemethod
The system reliability solution outlined in Section 143 is reasonably prac-tical for problems containing a few failure modes that can be individuallyanalyzed by FORM A more general approach that is gaining wider atten-tion is the spectral stochastic finite element method originally proposed byGhanem and Spanos (1991) The key element of this approach is the expan-sion of the unknown random output vector using multi-dimensional Hermitepolynomials as basis functions (also called a polynomial chaos expansion)The unknown deterministic coefficients in the expansion can be solved usingthe Galerkin or collocation method The former method requires significant
40 Kok-Kwang Phoon
modification of the existing deterministic numerical code and is impossibleto apply for most engineers with no access to the source code of their com-mercial softwares The collocation method can be implemented using a smallnumber of fairly simple computational steps and does not require the mod-ification of existing deterministic numerical code Chapter 7 and elsewhere(Sudret and Der Kiureghian 2000) discussed this important class of methodsin detail Verhoosel and Gutieacuterrez (2007) highlighted challenging difficultiesin applying these methods to nonlinear finite element problems involvingdiscontinuous fields It is interesting to observe in passing that the outputvector can be expanded using the more well-known Taylor series expansionas well The coefficients of the expansion (partial derivatives) can be calcu-lated using the perturbation method This method can be applied relativelyeasily to finite element outputs (Phoon et al 1990 Quek et al 1991 1992)but is not covered in this chapter
This section briefly explains the key computational steps for the morepractical collocation approach We recall in Section 132 that a vector of cor-related non-normal random variables Y = (Y1Y2 Yn)prime can be related to avector of correlated standard normal random variables X = (X1X2 Xn)primeusing one-dimensional Hermite polynomial expansions [Equation (126)]The correlation coefficients for the normal random vector are evaluatedfrom the correlation coefficients of the non-normal random vector usingEquation (127) This method can be used to construct any non-normal ran-dom vector as long as the correlation coefficients are available This is indeedthe case if Y represents the input random vector However if Y representsthe output random vector from a numerical code the correlation coefficientsare unknown and Equation (126) is not applicable Fortunately this practi-cal problem can be solved by using multi-dimensional Hermite polynomialswhich are supported by a known normal random vector with zero meanunit variance and uncorrelated or independent components
The multi-dimensional Hermite polynomials are significantly more com-plex than the one-dimensional version For example the second-order andthird-order forms can be expressed respectively as follows (Isukapalli1999)
Y asympao+nsum
i=1
aiZi +nsum
i=1
aii
(Z2
i minus1)+
nminus1sumi=1
nsumjgti
aijZiZj (160a)
Y asympao+nsum
i=1
aiZi +nsum
i=1
aii
(Z2
i minus1)+
nsumi=1
aiii
(Z3
i minus3Zi
)+
nminus1sumi=1
nsumjgti
aijZiZj
+nsum
i=1
nsumj=1j =i
aijj
(ZiZ
2j minusZi
)+
nminus2sumi=1
nminus1sumjgti
nsumkgtj
aijkZiZjZk (160b)
Numerical recipes for reliability analysis 41
For n = 3 Equation (160) produces the following expansions (indices forcoefficients are re-labeled consecutively for clarity)
Y asymp ao + a1Z1 + a2Z2 + a3Z3 + a4(Z21 minus 1) + a5(Z2
2 minus 1) + a6(Z23 minus 1)
+ a7Z1Z2 + a8Z1Z3 + a9Z2Z3 (161a)
Y asymp ao + a1Z1 + a2Z2 + a3Z3 + a4(Z21 minus 1) + a5(Z2
2 minus 1) + a6(Z23 minus 1)
+ a7Z1Z2 + a8Z1Z3 + a9Z2Z3 + a10(Z31 minus 3Z1) + a11(Z3
2 minus 3Z2)
+ a12(Z33 minus 3Z3) + a13(Z1Z2
2 minus Z1) + a14(Z1Z23 minus Z1)
+ a15(Z2Z21 minus Z2) + a16(Z2Z2
3 minus Z2) + a17(Z3Z21 minus Z3)
+ a18(Z3Z22 minus Z3) + a19Z1Z2Z3 (161b)
In general N2 and N3 terms are respectively required for the second-orderand third-order expansions (Isukapalli 1999)
N2 = 1 + 2n + n(n minus 1)2
(162a)
N3 = 1 + 3n + 3n(n minus 1)2
+ n(n minus 1)(n minus 2)6
(162b)
For a fairly modest random dimension of n = 5 N2 and N3 terms arerespectively equal to 21 and 56 Hence fairly tedious algebraic expressionsare incurred even at a third-order truncation One-dimensional Hermitepolynomials can be generated easily and efficiently using a three-term recur-rence relation [Equation (111)] No such simple relation is available formulti-dimensional Hermite polynomials They are usually generated usingsymbolic algebra which is possibly out of reach of the general practi-tioner This major practical obstacle is currently being addressed by Lianget al (2007) They have developed a user-friendly EXCEL add-in to gener-ate tedious multi-dimensional Hermite expansions automatically Once themulti-dimensional Hermite expansions are established their coefficients canbe calculated following the steps described in Equations (118)ndash(120) Twopractical aspects are noteworthy
1 The random dimension of the problem should be minimized to reducethe number of terms in the polynomial chaos expansion The spectraldecomposition method can be used
X = PD12Z + micro (163)
in which D = diagonal matrix containing eigenvalues in the leading diag-onal and P = matrix whose columns are the corresponding eigenvectors
42 Kok-Kwang Phoon
If C is the covariance matrix of X the matrices D and P can be calcu-lated easily using [PD] = eig(C) in MATLAB The key advantage ofreplacing Equation (123) by Equation (163) is that an n-dimensionalcorrelated normal vector X can be simulated using an uncorrelated stan-dard normal vector Z with a dimension less than n This is achieved bydiscarding eigenvectors in P that correspond to small eigenvalues A sim-ple example is provided in Appendix A7 The objective is to simulate a3D correlated normal vector following a prescribed covariance matrix
C =⎡⎣ 1 09 02
09 1 0502 05 1
⎤⎦
Spectral decomposition of the covariance matrix produces
D =⎡⎣0045 0 0
0 0832 00 0 2123
⎤⎦ P =
⎡⎣ 0636 0467 0614
minus0730 0108 06750249 minus0878 0410
⎤⎦
Realizations of X can be obtained using Equation (163) Results areshown as open circles in Figure 113 The random dimension can bereduced from three to two by ignoring the first eigenvector correspond-ing to a small eigenvalue of 0045 Realizations of X are now simulatedusing
⎧⎨⎩
X1X2X3
⎫⎬⎭=⎡⎣ 0467 0614
0108 0675minus0878 0410
⎤⎦[radic0832 0
0radic
2123
]Z1Z2
Results are shown as crosses in Figure 113 Note that three correlatedrandom variables can be represented reasonably well using only twouncorrelated random variables by neglecting the smallest eigenvalue andthe corresponding eigenvector The exact error can be calculated theo-retically by observing that the covariance of X produced by Equation(163) is
CX = (PD12)(D12Pprime) = PDPprime (164)
If the matrices P and D are exact it can be proven that PDPprime = C iethe target covariance is reproduced For the truncated P and D matrices
Numerical recipes for reliability analysis 43
Random dimension
ThreeTwo
3
2
1
0X3
X2
minus1
minus2
minus3minus3 minus2 minus1 0 2 2 3
Random dimension
ThreeTwo
3
2
1
0X3
X2
minus1
minus2
minus3minus3 minus2 minus1 0 2 2 3
Figure 113 Scatter plot between X1 and X2 (top) and X2 and X3 (bottom)
discussed above the covariance of X is
CX =⎡⎣ 0467 0614
0108 0675minus0878 0410
⎤⎦[0832 0
0 2123
]
times[0467 0108 minus08780614 0675 0410
]=⎡⎣0982 0921 0193
0921 0976 05080193 0508 0997
⎤⎦
44 Kok-Kwang Phoon
The exact error in each element of CX can be clearly seen by comparingwith the corresponding element in C In particular the variances ofX (elements in the leading diagonal) are slightly reduced For randomprocessesfields reduction in the random dimension can be achievedusing the KarhunenndashLoeve expansion (Phoon et al 2002 2004) It canbe shown that the spectral representation given by Equation (128) isa special case of the KarhunenndashLoeve expansion (Huang et al 2001)The random dimension can be further reduced by performing sensitiv-ity analysis and discarding input parameters that are ldquounimportantrdquoFor example the square of the components of the unit normal vec-tor α shown in Figure 112 are known as FORM importance factors(Ditlevsen and Madsen 1996) If the input parameters are independentthese factors indicate the degree of influence exerted by the corre-sponding input parameters on the failure event Sudret (2007) discussedthe application of Sobolrsquo indices for sensitivity analysis of the spectralstochastic finite element method
2 The number of output values (y1y2 ) in Equation (120) should beminimized because they are usually produced by costly finite elementcalculations Consider a problem containing two random dimensionsand approximating the output as a third-order expansion
yi = ao + a1zi1 + a2zi2 + a3(z2i1 minus 1) + a4(z2
i2 minus 1) + a5zi1zi2
+ a6(z3i1 minus 3zi1) + a7(z3
i2 minus 3zi2) + a8(zi1z2i2 minus zi1)
+ a9(zi2z2i1 minus zi2) (165)
Phoon and Huang (2007) demonstrated that the collocation points(zi1zi2) are best sited at the roots of the Hermite polynomial that is oneorder higher than that of the Hermite expansion In this example theroots of the fourth-order Hermite polynomial are plusmnradic
(3plusmnradic6) Although
zero is not one of the roots it should be included because the standardnormal probability density function is highest at the origin Twenty-five collocation points (zi1zi2) can be generated by combining the rootsand zero in two dimensions as illustrated in Figure 114 The roots ofHermite polynomials (up to order 15) can be calculated numerically asshown in Appendix A8
145 Subset simulation method
The collocation-based stochastic response surface method is very efficientfor problems containing a small number of random dimensions and perfor-mance functions that can be approximated quite accurately using low-orderHermite expansions However the method suffers from a rapid proliferation
Numerical recipes for reliability analysis 45
Z1
Z2
Figure 114 Collocation points for a third-order expansion with two random dimensions
of Hermite expansion terms when the random dimension andor order ofexpansion increase A critical review of reliability estimation procedures forhigh dimensions is given by Schueumlller et al (2004) One potentially practicalmethod that is worthy of further study is the subset simulation method (Auand Beck 2001) It appears to be more robust than the importance samplingmethod Chapter 4 and elsewhere (Au 2001) present a more extensive studyof this method
This section briefly explains the key computational steps involved inthe implementation Consider the failure domain Fm defined by the con-dition P(y1y2 yn) lt 0 in which P is the performance function and(y1y2 yn) are realizations of the uncertain input parameters The ldquofail-urerdquo domain Fi defined by the condition P(y1y2 yn) lt ci in which ci isa positive number is larger by definition of the performance function Weassume that it is possible to construct a nested sequence of failure domainsof increasing size by using an increasing sequence of positive numbers iethere exists c1 gt c2 gt gt cm = 0 such that F1 sup F2 sup Fm As shown inFigure 115 for the one-dimensional case it is clear that this is always possibleas long as one value of y produces only one value of P(y) The performancefunction will satisfy this requirement If one value of y produces two valuesof P(y) say a positive value and a negative value then the physical systemis simultaneously ldquosaferdquo and ldquounsaferdquo which is absurd
The probability of failure (pf ) can be calculated based on the above nestedsequence of failure domains (or subsets) as follows
pf = Prob(Fm) = Prob(Fm|Fmminus1)Prob(Fmminus1|Fmminus2)
times Prob(F2|F1)Prob(F1) (166)
46 Kok-Kwang Phoon
P(y)
y
C1
C2
F2 F2
F1
Figure 115 Nested failure domains
At first glance Equation (166) appears to be an indirect and more tediousmethod of calculating pf In actuality it can be more efficient because theprobability of each subset conditional on the previous (larger) subset canbe selected to be sufficiently large say 01 such that a significantly smallernumber of realizations is needed to arrive at an acceptably accurate resultWe recall from Equation (149) that the rule of thumb is 10pf ie only100 realizations are needed to estimate a probability of 01 If the actualprobability of failure is 0001 and the probability of each subset is 01 it isapparent from Equation (166) that only three subsets are needed implying atotal sample size = 3times100 = 300 In contrast direct simulation will require100001 = 10000 realizations
The typical calculation steps are illustrated below using a problemcontaining two random dimensions
1 Select a subset sample size (n) and prescribe p = Prob(Fi|Fiminus1) Weassume p = 01 and n = 500 from hereon
2 Simulate n = 500 realizations of the uncorrelated standard normal vector(Z1Z2)prime The physical random vector (Y1Y2)prime can be determined fromthese realizations using the methods described in Section 132
3 Calculate the value of the performance function gi = P(yi1yi2) associ-ated with each realization of the physical random vector (yi1yi2)prime i =12 500
4 Rank the values of (g1g2 g500) in ascending order The value locatedat the (np + 1) = 51st position is c1
Numerical recipes for reliability analysis 47
5 Define the criterion for the first subset (F1) as P lt c1 By construction atstep (4) P(F1) = npn = p The realizations contained in F1 are denotedby zj = (zj1zj2)prime j = 12 50
6 Simulate 1p = 10 new realizations from zj using the followingMetropolis-Hastings algorithm
a Simulate 1 realization using a uniform proposal distribution withmean located at micro = zj and range = 1 ie bounded by micro plusmn 05 Letthis realization be denoted by u = (u1u2)prime
b Calculate the acceptance probability
α = min(
10I(u)φ(u)φ(micro)
)(167)
in which I(u) = 1 if P(u1u2) lt c1 I(u) = 0 if P(u1u2) ge c1φ(u) = exp(minus05uprimeu) and φ(micro) = exp(minus05microprimemicro)
c Simulate 1 realization from a standard uniform distributionbounded between 0 and 1 denoted by v
d The first new realization is given by
w = u if v lt α
w = micro if v ge α(168)
Update the mean of the uniform proposal distribution in step (a) asmicro = w (ie centred about the new realization) and repeat the algo-rithm to obtain a ldquochainrdquo containing 10 new realizations Fiftychains are obtained in the same way with initial seeds at zj j = 1 50 It can be proven that these new realizations would followProb(middot|F1) (Au 2001)
7 Convert these realizations to their physical values and repeat Step (3)until ci becomes negative (note the smallest subset should correspondto cm = 0)
It is quite clear that the above procedure can be extended to any randomdimension in a trivial way In general the accuracy of this subset simula-tion method depends on ldquotuningrdquo factors such as the choice of np and theproposal distribution in step 6(a) (assumed to be uniform with range = 1)A study of some of these factors is given in Chapter 4 The optimal choice ofthese factors to achieve minimum runtime appears to be problem-dependentAppendix A9 provides the MATLAB code for subset simulation The perfor-mance function is specified in ldquoPfunmrdquo and the number of random variablesis specified in parameter ldquomrdquo Figure 116 illustrates the behavior of thesubset simulation method for the following performance function
P = 2 + 3radic
2 minus Y1 minus Y2 = 2 + 3radic
2 + ln[Φ(minusZ1)]+ ln[Φ(minusZ2)] (169)
minus1 0 1 2 3 4minus1
0
1
2
3
4
Z1
Z2
P lt 0
P lt c1
minus1 0 1 2 3 4minus1
0
1
2
3
4
Z1
Z2
P lt 0
P lt c1
Figure 116 Subset simulation method for P = 2 + 3radic
2 minus Y1 minus Y2 in which Y1 and Y2 areexponential random variables with mean = 1 first subset (top) and secondsubset (bottom)
Numerical recipes for reliability analysis 49
in which Y1Y2 = exponential random variables with mean = 1 and Z1Z2 =uncorrelated standard normal random variables The exact solution for theprobability of failure is 00141 (Appendix C) The solution based on subsetsimulation method solution is achieved as follows
1 Figure 116a Simulate 500 realizations (small open circles) and identifythe first 50 realizations in order of increasing P value (big open circles)The 51st smallest P value is c1 = 2192
2 These realizations satisfy the criterion P lt c1 (solid line) by construc-tion The domain above the solid line is F1
3 Figure 116b Simulate 10 new realizations from each big open circle inFig 116a The total number is 10 times 50 = 500 realizations (small opencircles)
4 There are 66 realizations (big open squares) satisfying P lt 0 (dashedline) The domain above the dashed line is F2 F2 is the actual failuredomain Hence Prob(F2|F1) = 66500 = 0132
5 The estimated probability of failure is pf = Prob(F2|F1)P(F1) = 0132times01 = 00132
Note that only 7 realizations in Figure 116a lie in F2 In contrast 66 real-izations in Figure 116b lie in F2 providing a more accurate pf estimate
15 Conclusions
This chapter presents general and user-friendly computational methods forreliability analysis ldquoUser-friendlyrdquo methods refer to those that can be imple-mented on a desktop PC by a non-specialist with limited programmingskills in other words methods within reach of the general practitionerThis chapter is organized under two main headings describing (a) how tosimulate uncertain inputs numerically and (b) how to propagate uncertaininputs through a physical model to arrive at the uncertain outputs Althoughsome of the contents are elaborated in greater detail in subsequent chaptersthe emphasis on numerical implementations in this chapter should shortenthe learning curve for the novice The overview will also provide a usefulroadmap to the rest of this book
For the simulation of uncertain inputs following arbitrary non-normalprobability distribution functions and correlation structure the translationmodel involving memoryless transform of the multivariate normal probabil-ity distribution function can cater for most practical scenarios Implemen-tation of the translation model using one-dimensional Hermite polynomialsis relatively simple and efficient For stochastic data that cannot be modeledusing the translation model the hunt for probability models with comparablepracticality theoretical power and simulation speed is still on-going Copulatheory produces a more general class of multivariate non-normal models but
50 Kok-Kwang Phoon
it is debatable at this point if these models can be estimated empirically andsimulated numerically with equal ease for high random dimensions Theonly exception is a closely related but non-translation approach based onthe multivariate normal copula
The literature is replete with methodologies on the determination of uncer-tain outputs from a given physical model and uncertain inputs The mostgeneral method is Monte Carlo simulation but it is notoriously tediousOne may assume with minimal loss of generality that a complex geotechni-cal problem (possibly 3D nonlinear time-and construction-dependent etc)would only admit numerical solutions and the spatial domain can be modeledby a scalarvector random field For a sufficiently large and complex problemit is computationally intensive to complete even a single run At present it isaccurate to say that a computationally efficient and ldquouser-friendlyrdquo solutionto the general stochastic problem remains elusive Nevertheless reasonablypractical solutions do exist if the general stochastic problem is restrictedin some ways for example accept a first-order estimate of the probabilityof failure or accept an approximate but less costly output The FORM isby far the most efficient general method for estimating the probability offailure of problems involving one design point (one dominant mode of fail-ure) There are no clear winners for problems beyond the reach of FORMThe collocation-based stochastic response surface method is very efficientfor problems containing a small number of random dimensions and perfor-mance functions that can be approximated quite accurately using low-orderHermite expansions However the method suffers from a rapid prolifera-tion of Hermite expansion terms when the random dimension andor order ofexpansion increase The subset simulation method shares many of the advan-tages of the Monte Carlo simulation method such as generality and tractabil-ity at high dimensions without requiring too many runs At low randomdimensions it does require more runs than the collocation-based stochas-tic response surface method but the number of runs is probably acceptableup to medium-scale problems The subset simulation method has numerousldquotuningrdquo factors that are not fully studied for geotechnical problems
Simple MATLAB codes are provided in the Appendix A to encouragepractitioners to experiment with reliability analysis numerically so as to gaina concrete appreciation of the merits and limitations Theory is discussedwhere necessary to furnish sufficient explanations so that users can modifycodes correctly to suit their purpose and to highlight important practicallimitations
Appendix A ndash MATLAB codes
MATLAB codes are stored in text files with extension ldquomrdquo ndash they are calledM-files M-files can be executed easily by typing the filename (eg hermite)within the command window in MATLAB The location of the M-file should
Numerical recipes for reliability analysis 51
be specified using ldquoFile gt Set Path helliprdquo M-files can be opened and editedusing ldquoFile gt Open helliprdquo Programming details are given under ldquoHelp gt
Contents gt MATLAB gt Programmingrdquo The M-files provided below areavailable at httpwwwengnusedusgcivilpeoplecvepkkprob_libhtml
A1 Simulation of Johnson distributions
Simulation of Johnson distributions Filename Johnsonm Simulation sample size nn = 100000 Lognormal with lambda xilambda = 1xi = 02Z = normrnd(0 1 n 1)X = lambda + ZxiLNY = exp(X) SB with lambda xi A Blambda = 1xi = 036A = -3B = 5X = lambda + ZxiSBY = (exp(X)B+A)(exp(X)+1) SU with lambda xi A Blambda = 1xi = 009A = -188B = 208X = lambda + ZxiSUY = sinh(X)(B-A)+A Plot probability density functions[f x] = ksdensity(LNY)plot(xf)hold[f x] = ksdensity(SBY)plot(xfrsquoredrsquo)[f x] = ksdensity(SUY)plot(xfrsquogreenrsquo)
52 Kok-Kwang Phoon
A2 Calculation of Hermite coefficients using stochasticcollocation method
Calculation of Hermite coefficients using stochasticcollocation method Filename herm_coeffsm Number of realizations nn = 20 Example Lognormal with lambda xilambda = 0xi = 1Z = normrnd(0 1 n 1)X = lambda + ZxiLNY = exp(X) Order of Hermite expansion mm = 6 Construction of Hermite matrix HH = zeros(nm+1)H(1) = ones(n1)H(2) = Zfor k = 3m+1H(k) = ZH(k-1) - (k-2)H(k-2)end Hermite coefficients stored in vector aK = HrsquoHf = HrsquoLNYa = inv(K)f
A3 Simulation of correlated non-normal randomvariables using translation method
Simulation of correlated non-normal random variablesusing translation method Filename translationm Number of random dimension nn = 2 Normal covariance matrix
Numerical recipes for reliability analysis 53
C = [1 08 08 1] Number of realizations mm = 100000 Simulation of 2 uncorrelated normal random variableswith mean = 0 and variance = 1Z = normrnd(0 1 m n) Cholesky factorizationCF = chol(C) Simulation of 2 correlated normal random variableswith with mean = 0 and covariance = CX = ZCF Example simulation of correlated non-normal with Component 1 = Johnson SBlambda = 1xi = 036A = -3B = 5W(1) = lambda + X(1)xiY(1) = (exp(W(1))B+A)(exp(W(1))+1) Component 2 = Johnson SUlambda = 1xi = 009A = -188B = 208W(2) = lambda + X(2)xiY(2) = sinh(W(2))(B-A)+A
A4 Simulation of normal random process using thetwo-sided power spectral density function S(f)
Simulation of normal random process using thetwo-sided power spectral density function S(f) Filename ranprocessm Number of data points based on power of 2 NN = 512
54 Kok-Kwang Phoon
Depth sampling interval delzdelz = 02 Frequency interval delfdelf = 1Ndelz Discretization of autocorrelation function R(tau) Example single exponential functiontau = zeros(1N)tau = -(N2-1)delzdelz(N2)delzd = 2 scale of fluctuationa = 2dR = zeros(1N)R = exp(-aabs(tau)) Numerical calculation of S(f) using FFTH = zeros(1N)H = fft(R) Notes 1 f=0 delf 2delf (N2-1)delf correspondsto H(1) H(2) H(N2) 2 Maximum frequency is Nyquist frequency= N2delf= 12delz 3 Multiply H (discrete transform) by delz to getcontinuous transform 4 Shift R back by tau0 ie multiply H byexp(2piiftau0)f = zeros(1N)S = zeros(1N) Shuffle f to correspond with frequencies orderingimplied in Hf(1) = 0for k = 2N2+1f(k)= f(k-1)+delfendf(N2+2) = -f(N2+1)+delffor k = N2+3Nf(k) = f(k-1)+delfendtau0 = (N2-1)delz
Numerical recipes for reliability analysis 55
for k = 1NS(k) = delzH(k)exp(2piif(k)tau0) i isimaginary numberendS = real(S) remove possible imaginary parts due toroundoff errors Shuffle S to correspond with f in increasing orderf = -(N2-1)delfdelf(N2)delftemp = zeros(1N)for k = 1N2-1temp(k) = S(k+N2+1)endfor k = N2Ntemp(k)=S(k-N2+1)endS = tempclear temp Simulation of normal process using spectralrepresentation mean of process = 0 and variance of process = 1 Maximum possible non-periodic process lengthLmax = 12fmin = 1delf Minimum frequency fmin = delf2Lmax = 1delf Simulation length L = bLmax with b lt 1L = 05Lmax Number of simulated data points nz Depth sampling interval dz Depth coordinates z(1) z(2) z(nz)nz = round(Ldelz)dz = Lnzz = dzdzL Number of realisations mm = 10000 Number of positive frequencies in the spectralexpansion nf = N2
56 Kok-Kwang Phoon
nf = N2 Simulate uncorrelated standard normal randomvariablesrandn(rsquostatersquo 1)Z = randn(m2nf)sigma = zeros(1nf)wa = zeros(1nf) Calculate energy at each frequency using trapezoidalrule 2S(f) for k = 1nfsigma(k) = 205(S(k+nf-1)+S(k+nf))delfwa(k) = 052pi(f(k+nf-1)+f(k+nf))endsigma = sigmaˆ05 Calculate realizations with mean = 0 andvariance = 1X = zeros(mnz)X = Z(1nf)diag(sigma)cos(warsquoz)+Z(nf+12nf)diag(sigma)sin(warsquoz)
A5 Reliability analysis of Load and Resistance FactorDesign (LRFD)
Reliability analysis of Load and Resistance FactorDesign (LRFD) Filename LRFDm Resistance factor phiphi = 05 Dead load factor gD Live load factor gLgD = 125gL = 175 Bias factors for Q D LbR = 1bD = 105bL = 115 Ratio of nominal dead to live load loadratioloadratio = 2 Assume mR = bRRn mD = bDDn mL = bLLn Design equation phi(Rn) = gD(Dn) + gL(Ln)
Numerical recipes for reliability analysis 57
mR = 1000mL = phibRmRbL(gDloadratio+gL)mD = loadratiomLbDbL Coefficients of variation for R D LcR = 03cD = 01cL = 02 Lognormal X with mean = mX and coefficent ofvariation = cXxR = sqrt(log(1+cRˆ2))lamR = log(mR)-05xRˆ2xD = sqrt(log(1+cDˆ2))lamD = log(mD)-05xDˆ2xL = sqrt(log(1+cLˆ2))lamL = log(mL)-05xLˆ2 Simulation sample size nn = 500000Z = normrnd(0 1 n 3)LR = lamR + Z(1)xRR = exp(LR)LD = lamD + Z(2)xDD = exp(LD)LL = lamL + Z(3)xLL = exp(LL) Total load = Dead load + Live LoadF = D + L Mean of FmF = mD + mL Coefficient of variation of F based on second-momentapproximationcF = sqrt((cDmD)ˆ2+(cLmL)ˆ2)mF Failure occurs when R lt Ffailure = 0for i = 1nif (R(i) lt F(i)) failure = failure+1 endend
58 Kok-Kwang Phoon
Probability of failure = no of failuresnpfs = failuren Reliability indexbetas = -norminv(pfs) Closed-form lognormal solutiona1 = sqrt(log((1+cRˆ2)(1+cFˆ2)))a2 = sqrt((1+cFˆ2)(1+cRˆ2))beta = log(mRmFa2)a1pf = normcdf(-beta)
A6 First-order reliability method
First-order reliability method Filename FORMm Number of random variables mm = 6 Starting guess is z0 = 0z0 = zeros(1 m) Minimize objective function exitflag = 1 for normal terminationoptions = optimset(rsquoLargeScalersquorsquooffrsquo)[zfvalexitflagoutput] =fmincon(objfunz0[][][][][][]Pfunoptions) First-order reliability indexbeta1 = fvalpf1 = normcdf(-beta1)
Objective function for FORM Filename objfunmfunction f = objfun(z) Objective function = distance from the origin z = vector of uncorrelated standard normal randomvariables f = norm(z)
Numerical recipes for reliability analysis 59
Definition of performance function P Filename Pfunmfunction [c ceq] = Pfun(z) Convert standard normal random variables to physicalvariables Depth of rock Hy(1) = 2+6normcdf(z(1)) Height of water tabley(2) = y(1)normcdf(z(2)) Effective stress friction angle (radians)xphi = sqrt(log(1+008ˆ2))lamphi = log(35)-05xphiˆ2y(3) = exp(lamphi+xphiz(3))pi180 Slope inclination (radians)xbeta = sqrt(log(1+005ˆ2))lambeta = log(20)-05xbetaˆ2y(4) = exp(lambeta+xbetaz(4))pi180 Specific gravity of solidsGs = 25+02normcdf(z(5)) Void ratioe = 03+03normcdf(z(6)) Moist unit weighty(5) = 981(Gs+02e)(1+e) Saturated unit weighty(6) = 981(Gs+e)(1+e) Nonlinear inequality constraints c lt 0c = (y(5)(y(1)-y(2))+y(2)(y(6)-981))cos(y(4))tan(y(3))((y(5)(y(1)-y(2))+y(2)y(6))sin(y(4)))-1 Nonlinear equality constraintsceq = []
60 Kok-Kwang Phoon
A7 Random dimension reduction using spectraldecomposition
Random dimension reduction using spectraldecomposition Filename eigendecompm Example Target covarianceC = [1 09 02 09 1 05 02 05 1] Spectral decomposition[PD] = eig(C) Simulation sample size nn = 10000 Simulate standard normal random variablesZ = normrnd(01n3) Simulate correlated normal variables following CX = Psqrt(D)ZrsquoX = Xrsquo Simulate correlated normal variables without 1steigen-componentXT = P(1323)sqrt(D(2323))Z(23)rsquoXT=XTrsquo Covariance produced by ignoring 1st eigen-componentCX = P(1323)D(2323)P(1323)rsquo
A8 Calculation of Hermite roots using polynomial fit
Calculation of Hermite roots using polynomial fit Filename herm_rootsm Order of Hermite polynomial m lt 15m = 5 Specification of z values for fittingz = (-412m4)rsquo Number of fitted points nn = length(z)
Numerical recipes for reliability analysis 61
Construction of Hermite matrix HH = zeros(nm+1)H(1) = ones(n1)H(2) = zfor k = 3m+1H(k) =zH(k-1)-(k-2)H(k-2)end Calculation of Hermite expansiony = H(m+1) Polynomial fitp = polyfit(zym) Roots of hermite polynomialr = roots(p) Validation of rootsnn = length(r)H = zeros(nnm+1)H(1) = ones(nn1)H(2) = rfor k = 3m+1H(k) =rH(k-1)-(k-2)H(k-2)endnorm(H(m+1))
A9 Subset Markov Chain Monte Carlo method
Subset Markov Chain Monte Carlo method Filename subsetm copy (2007) Kok-Kwang Phoon Number of random variables mm = 2 Sample size nn = 500 Probability of each subsetpsub = 01 Number of new samples from each seed samplens = 1psub
62 Kok-Kwang Phoon
Simulate standard normal variable zz = normrnd(01nm)stopflag = 0pfss = 1while (stopflag == 0) Values of performance functionfor i=1n g(i) = Pfun(z(i)) end Sort vector g to locate npsub smallest values[gsortindex]=sort(g) Subset threshold gt = npsub+1 smallest value of g if gt lt 0 exit programgt = gsort(npsub+1)if (gt lt 0)
i=1while gsort(i)lt0
i=i+1endpfss=pfss(i-1)nstopflag = 1break
elsepfss = pfsspsub
end npsub seeds satisfying Pfun lt gt stored inz(index(1npsub)) Simulate 1psub samples from each seed to get nsamples for next subsetw = zeros(nm)for i=1npsub
seed = z(index(i))for j=1ns
Proposal density standard uniform with mean =seedu = rand(1m)u = (seed-05)+ u Calculate acceptance probabilitypdf2 = exp(-05sum(seedˆ2))pdf1 = exp(-05sum(uˆ2))
Numerical recipes for reliability analysis 63
I = 0if Pfun(u)ltgt I=1 endalpha = min(1Ipdf1pdf2) Accept u with probability = alphaV = randif V lt alpha
w(ns(i-1)+j)= uelse
w(ns(i-1)+j)= seedend
seed = w(ns(i-1)+j)end
endz=wend Reliability indexbetass = -norminv(pfss)
Appendix B ndash Approximate closed-form solutionsfor Φ(middot) and Φminus1(middot)The standard normal cumulative distribution functionΦ(middot) can be calculatedusing normcdf in MATLAB For sufficiently large β it is also possible to esti-mate Φ(minusβ) quite accurately using the following asymptotic approximation(Breitung 1984)
Φ(minusβ) sim (2π )minus12exp(minusβ22)βminus1 (B1)
An extensive review of estimation methods for the standard normal cumu-lative distribution function is given by Johnson et al (1994) Comparisonbetween Equation (B1) and normcdf(middot) is given in Figure B1 It appearsthat β gt 2 is sufficiently ldquolargerdquo Press et al (1992) reported an extremelyaccurate empirical fit with an error less than 12 times 10minus7 everywhere
t = 1
1 + 025radic
2β05t times exp(minus05β2minus126551223 + t
times (100002368 + t times (037409196 +Φ(minusβ) asymp t times (009678418 + t
times (minus018628806 + t times (027886807 + t times (minus113520398 + t
times (148851587 + t times (minus082215223 + t times 017087277))))))))) (B2)
64 Kok-Kwang Phoon
0 1 2 3 4 510minus7
10minus6
10minus5
10minus4
10minus3
10minus2
10minus1
100
Reliability index β
Pro
babi
lity
of fa
ilure
pf
normcdfEq (B1)
Figure B1 Comparison between normcdf(middot) and asymptotic approximation of the standardnormal cumulative distribution function
However it seems difficult to obtain β = minusΦminus1(pf) which is of signifi-cant practical interest as well Equation (B1) is potentially more useful inthis regard To obtain the inverse standard normal cumulative distributionfunction we rewrite Equation (B1) as follows
β2 + 2ln(radic
2πpf) + 2ln(β) = 0 (B3)
To solve Equation (B3) it is natural to linearize ln(β) using Taylor seriesexpansion ln(β) = minusln[1 + (1β minus 1)] asymp 1 minus 1β with β gt 05 Using thislinearization Equation (B3) simplifies to a cubic equation
β3 + 2[In(radic
2πpf) + 1]β minus 2 = 0 (B4)
The solution is given by
Q = 2[ln(radic
2πpf) + 1]3
R = 1
β =(
R +radic
Q3 + R2)13
+(
R minusradic
Q3 + R2)13
pf lt exp(minus1)radic
2π
(B5)
Numerical recipes for reliability analysis 65
0 1 2 3 4 510minus7
10minus6
10minus5
10minus4
10minus3
10minus2
10minus1
100
Reliability index β
Pro
babi
lity
of fa
ilure
pf
normcdfEq (B5)Eq (B7)
Figure B2 Approximate closed-form solutions for inverse standard normal cumulativedistribution function
Equation (B5) is reasonably accurate as shown in Figure B2 An extremelyaccurate and simpler linearization can be obtained by fitting ln(β) = b0 +b1β
over the range of interest using linear regression Using this linearizationEquation (B3) simplifies to a quadratic equation
β2 + 2b1β + 2[ln(radic
2πpf) + b0] = 0 (B6)
The solution is given by
β = minusb1 +radic
b21 minus 2[ln(
radic2πpf) + b0] pf lt exp(b2
12 minus b0)radic
2π (B7)
The parameters b0 = 0274 and b1 = 0266 determined from linear regressionover the range 2 le β le 6 were found to produce extremely good results asshown in Figure B2 To the authorrsquos knowledge Equations (B5) and (B7)appear to be original
Appendix C ndash Exact reliability solutions
It is well known that exact reliability solutions are available for problemsinvolving the sum of normal random variables or the product of lognormalrandom variables This appendix provides other exact solutions that areuseful for validation of new reliability codescalculation methods
66 Kok-Kwang Phoon
C1 Sum of n exponential random variables
Let Y be exponentially distributed with the following cumulative distributionfunction
F(y) = 1 minus exp(minusyb) (C1)
The mean of Y is b The sum of n independent identically distributed expo-nential random variables follows an Erlang distribution with mean = nband variance = nb2 (Hastings and Peacock 1975) Consider the followingperformance function
P = nb +αbradic
n minusnsum
i=1
Yi = c minusnsum
i=1
Yi (C2)
in which c and α are positive numbers The equivalent performance functionin standard normal space is
P = c +nsum
i=1
bln[Φ(minusZi)
](C3)
The probability of failure can be calculated exactly based on the Erlangdistribution (Hastings and Peacock 1975)
pf = Prob(P lt 0)
= Prob
(nsum
i=1
Y gt c
)(C4)
= exp(minus c
b
)nminus1sumi=0
(cb)i
i
Equation (C2) with b = 1 is widely used in the structural reliabilitycommunity to validate FORMSORM (Breitung 1984 Rackwitz 2001)
C2 Probability content of an ellipse in n-dimensionalstandard normal space
Consider the ldquosaferdquo elliptical domain in 2D standard normal space shownin Figure C1 The performance function is
P = α2 minus(
Z1
b1
)2
minus(
Z2
b2
)2
(C5)
Numerical recipes for reliability analysis 67
Z2
Z1
R + r = αb2
(Z1b1)2 + (Z2b2)2 = α2
R - r = αb1
R + r
R - r
Figure C1 Elliptical safe domain in 2D standard normal space
in which b2 gtb1 gt0 Let R =α(b1+b2)2 and r =α(b2minusb1)2 Johnson andKotz (1970) citing Ruben (1960) provided the following the exact solutionfor the safe domain
Prob(P gt 0) = Prob(χ prime22 (r2) le R2) minus Prob(χ prime2
2 (R2) le r2) (C6)
in which χ prime22 (r2) = non-central chi-square random variable with two degrees
of freedom and non-centrality parameter r2 The cumulative distributionfunction Prob(χ prime2
2 (r2) le R2) can be calculated using ncx2cdf(R22r2) inMATLAB
There is no simple generalization of Equation (C6) to higher dimen-sions Consider the following performance function involving n uncorrelatedstandard normal random variables (Z1Z2 Zn)
P = α2 minus(
Z1
b1
)2
minus(
Z2
b2
)2
minusmiddotmiddot middot(
Zn
bn
)2
(C7)
in which bn gt bnminus1 gt middot middot middot gt b1 gt 0 Johnson and Kotz (1970) citing Ruben(1963) provided a series solution based on the sum of central chi-squarecumulative distribution functions
Prob(P gt 0) =infinsum
r=0
er Prob[χ2
n+2r le (bnα)2] (C8)
in which χ2ν (middot) = central chi-square random variable with ν degrees of free-
dom The cumulative distribution function Prob(χ2ν le y) can be calculated
68 Kok-Kwang Phoon
using chi2cdf(yν) in MATLAB The coefficients er are calculated as
e0 =nprod
j=1
(bj
bn
)
er = (2r)minus1rminus1sumj=0
ejhrminusj r ge 1
(C9)
hr =nsum
j=1
(1 minus
b2j
b2n
)r
(C10)
It is possible to calculate the probabilities associated with n-dimensional non-central ellipsoidal domains as well The performance function is given by
P = α2 minus(
Z1 minus a1
b1
)2
minus(
Z2 minus a2
b2
)2
minusmiddotmiddot middot(
Zn minus an
bn
)2
(C11)
in which bn gt bnminus1 gt gt b1 gt 0 Johnson and Kotz (1970) citing Ruben(1962) provided a series solution based on the sum of non-central chi-squarecumulative distribution functions
Prob(P gt 0) =infinsum
r=0
erProb
⎡⎣χ prime2
n+2r
⎛⎝ nsum
j=1
a2j
⎞⎠le (bnα
)2⎤⎦ (C12)
The coefficients er are calculated as
e0 =nprod
j=1
(bj
bn
)
er = (2r)minus1rminus1sumj=0
ejhrminusj r ge 1
(C13)
h1 =nsum
j=1(1 minus a2
j )
(1 minus
b2j
b2n
)
hr =nsum
j=1
⎡⎣(1 minus
b2j
b2n
)r
+ rb2
na2
j b2j
(1 minus
b2j
b2n
)rminus1⎤⎦ r ge 2
(C14)
Remark It is possible to create a very complex performance function byscattering a large number of ellipsoids in nD space The parameters in
Numerical recipes for reliability analysis 69
Equation (C11) can be visualized geometrically even in nD space Hencethey can be chosen in a relatively simple way to ensure that the ellipsoidsare disjoint One simple approach is based on nesting (a) encase the firstellipsoid in a hyper-box (b) select a second ellipsoid that is outside the box(c) encase the first and second ellipsoids in a larger box and (d) select athird ellipsoid that is outside the larger box Repeated application will yielddisjoint ellipsoids The exact solution for such a performance function ismerely the sum of Equation (C12) but numerical solution can be very chal-lenging An example constructed by this ldquocookie-cutterrdquo approach may becontrived but it has the advantage of being as complex as needed to testthe limits of numerical methods while retaining a relatively simple exactsolution
Appendix D ndash Second-order reliability method(SORM)
The second-order reliability method was developed by Breitung andothers in a series of papers dealing with asymptotic analysis (Breitung1984 Breitung and Hohenbichler 1989 Breitung 1994) The calcula-tion steps are illustrated below using a problem containing three randomdimensions
1 Let the performance function be defined in the standard normal spaceie P(z1z2z3) and let the first-order reliability index calculated fromEquation (150) be β = (zlowastprimezlowast)12 in which zlowast = (zlowast
1zlowast2z
lowast3)prime is the design
point2 The gradient vector at zlowast is calculated as
nablaP(zlowast) =⎧⎨⎩
partP(zlowast)partz1partP(zlowast)partz2partP(zlowast)partz3
⎫⎬⎭ (D1)
The magnitude of the gradient vector is
||nablaP(zlowast)|| = [nablaP(zlowast)primenablaP(zlowast)]12 (D2)
3 The probability of failure is estimated as
pf = Φ(minusβ)|J|minus12 (D3)
in which J is the following 2 times 2 matrix
J =[
1 00 1
]+ β
nablaP(zlowast)[
part2P(zlowast)partz21 part2P(zlowast)partz1partz2
part2P(zlowast)partz2partz1 part2P(zlowast)partz22
](D4)
70 Kok-Kwang Phoon
Z2
Z1
4
201
5 6
3
∆
∆
∆ ∆
Figure D1 Central finite difference scheme
Equations (D1) and (D4) can be estimated using the central finite differencescheme shown in Figure D1 Let Pi be the value of the performance functionevaluated at node i Then the first derivative is estimated as
partPpartz1
asymp P2 minus P1
2(D5)
The second derivative is estimated as
part2P
partz21
asymp P2 minus 2P0 + P1
∆2 (D6)
The mixed derivative is estimated as
part2Ppartz1partz2
asymp P4 minus P3 minus P6 + P5
4∆2 (D7)
References
Ang A H-S and Tang W H (1984) Probability Concepts in Engineering Plan-ning and Design Vol 2 (Decision Risk and Reliability) John Wiley and SonsNew York
Au S K (2001) On the solution of first excursion problems by simulation with appli-cations to probabilistic seismic performance assessment PhD thesis CaliforniaInstitute of Technology Pasadena
Au S K and Beck J (2001) Estimation of small failure probabilities in highdimensions by subset simulation Probabilistic Engineering Mechanics 16(4)263ndash77
Numerical recipes for reliability analysis 71
Baecher G B and Christian J T (2003) Reliability and Statistics in GeotechnicalEngineering John Wiley amp Sons New York
Becker D E (1996) Limit states design for foundations Part II ndash Developmentfor National Building Code of Canada Canadian Geotechnical Journal 33(6)984ndash1007
Boden B (1981) Limit state principles in geotechnics Ground Engineering14(6) 2ndash7
Bolton M D (1983) Eurocodes and the geotechnical engineer Ground Engineering16(3) 17ndash31
Bolton M D (1989) Development of codes of practice for design In Proceedings ofthe 12th International Conference on Soil Mechanics and Foundation EngineeringRio de Janeiro Vol 3 Balkema A A Rotterdam pp 2073ndash6
Box G E P and Muller M E (1958) A note on the generation of random normaldeviates Annals of Mathematical Statistics 29 610ndash1
Breitung K (1984) Asymptotic approximations for multinormal integrals Journalof Engineering Mechanics ASCE 110(3) 357ndash66
Breitung K (1994) Asymptotic Approximations for Probability Integrals LectureNotes in Mathematics 1592 Springer Berlin
Breitung K and Hohenbichler M (1989) Asymptotic approximations to mul-tivariate integrals with an application to multinormal probabilities Journal ofMultivariate Analysis 30(1) 80ndash97
Broms B B (1964a) Lateral resistance of piles in cohesive soils Journal of SoilMechanics and Foundations Division ASCE 90(SM2) 27ndash63
Broms B B (1964b) Lateral resistance of piles in cohesionless soils Journal of SoilMechanics and Foundations Division ASCE 90(SM3) 123ndash56
Carsel R F and Parrish R S (1988) Developing joint probability distributions ofsoil water retention characteristics Water Resources Research 24(5) 755ndash69
Chilegraves J-P and Delfiner P (1999) Geostatistics ndash Modeling Spatial UncertaintyJohn Wiley amp Sons New York
Cressie N A C (1993) Statistics for Spatial Data John Wiley amp Sons New YorkDitlevsen O and Madsen H (1996) Structural Reliability Methods John Wiley amp
Sons ChichesterDuncan J M (2000) Factors of safety and reliability in geotechnical engineer-
ing Journal of Geotechnical and Geoenvironmental Engineering ASCE 126(4)307ndash16
Elderton W P and Johnson N L (1969) Systems of Frequency Curves CambridgeUniversity Press London
Fenton G A and Griffiths D V (1997) Extreme hydraulic gradient statistics in astochastic earth dam Journal of Geotechnical and Geoenvironmental EngineeringASCE 123(11) 995ndash1000
Fenton G A and Griffiths D V (2002) Probabilistic foundation settlement on spa-tially random soil Journal of Geotechnical and Geoenvironmental EngineeringASCE 128(5) 381ndash90
Fenton G A and Griffiths D V (2003) Bearing capacity prediction of spatiallyrandom c- soils Canadian Geotechnical Journal 40(1) 54ndash65
Fisher R A (1935) The Design of Experiments Oliver and Boyd EdinburghFleming W G K (1989) Limit state in soil mechanics and use of partial factors
Ground Engineering 22(7) 34ndash5
72 Kok-Kwang Phoon
Ghanem R and Spanos P D (1991) Stochastic Finite Element A SpectralApproach Springer-Verlag New York
Griffiths D V and Fenton G A (1993) Seepage beneath water retaining structuresfounded on spatially random soil Geotechnique 43(6) 577ndash87
Griffiths D V and Fenton G A (1997) Three dimensional seepage through spa-tially random soil Journal of Geotechnical and Geoenvironmental EngineeringASCE 123(2) 153ndash60
Griffiths D V and Fenton G A (2001) Bearing capacity of spatially random soilthe undrained clay Prandtl problem revisited Geotechnique 54(4) 351ndash9
Hansen J B (1961) Ultimate resistance of rigid piles against transversal forcesBulletin 12 Danish Geotechnical Institute Copenhagen 1961 5ndash9
Hasofer A M and Lind N C (1974) An exact and invariant first order reliabilityformat Journal of Engineering Mechanics Division ASCE 100(EM1) 111ndash21
Hastings N A J and Peacock J B (1975) Statistical Distributions A Handbookfor Students and Practitioners Butterworths London
Huang S P Quek S T and Phoon K K (2001) Convergence study of the truncatedKarhunenndashLoeve expansion for simulation of stochastic processes InternationalJournal of Numerical Methods in Engineering 52(9) 1029ndash43
Isukapalli S S (1999) Uncertainty analysis of transport ndash transformation modelsPhD thesis The State University of New Jersey New Brunswick
Johnson N L (1949) Systems of frequency curves generated by methods oftranslation Biometrika 73 387ndash96
Johnson N L Kotz S and Balakrishnan N (1970) Continuous UnivariateDistributions Vol 2 Houghton Mifflin Company Boston
Johnson N L Kotz S and Balakrishnan N (1994) Continuous UnivariateDistributions Vols 1 and 2 John Wiley and Sons New York
Kulhawy F H and Phoon K K (1996) Engineering judgment in the evolution fromdeterministic to reliability-based foundation design In Uncertainty in the Geo-logic Environment ndash From Theory to Practice (GSP 58) Eds C D ShackelfordP P Nelson and M J S Roth ASCE New York pp 29ndash48
Liang B Huang S P and Phoon K K (2007) An EXCEL add-in implementationfor collocation-based stochastic response surface method In Proceedings of the1st International Symposium on Geotechnical Safety and Risk Tongji UniversityShanghai China pp 387ndash98
Low B K and Tang W H (2004) Reliability analysis using object-orientedconstrained optimization Structural Safety 26(1) 69ndash89
Marsaglia G and Bray T A (1964) A convenient method for generating normalvariables SIAM Review 6 260ndash4
Mendell N R and Elston R C (1974) Multifactorial qualitative traits geneticanalysis and prediction of recurrence risks Biometrics 30 41ndash57
National Research Council (2006) Geological and Geotechnical Engineering inthe New Millennium Opportunities for Research and Technological InnovationNational Academies Press Washington D C
Paikowsky S G (2002) Load and resistance factor design (LRFD) for deep founda-tions In Proceedings International Workshop on Foundation Design Codes andSoil Investigation in view of International Harmonization and Performance BasedDesign Hayama Balkema A A Lisse pp 59ndash94
Numerical recipes for reliability analysis 73
Phoon K K (2003) Representation of random variables using orthogonal poly-nomials In Proceedings of the 9th International Conference on Applicationsof Statistics and Probability in Civil Engineering Vol 1 Millpress RotterdamNetherlands pp 97ndash104
Phoon K K (2004a) General non-Gaussian probability models for first-order reli-ability method (FORM) a state-of-the-art report ICG Report 2004-2-4 (NGIReport 20031091-4) International Centre for Geohazards Oslo
Phoon K K (2004b) Application of fractile correlations and copulas to non-Gaussian random vectors In Proceedings of the 2nd International ASRANet(Network for Integrating Structural Safety Risk and Reliability) ColloquiumBarcelona (CDROM)
Phoon K K (2005) Reliability-based design incorporating model uncertainties InProceedings of the 3rd International Conference on Geotechnical EngineeringDiponegoro University Semarang Indonesia pp 191ndash203
Phoon K K (2006a) Modeling and simulation of stochastic data In Geo-Congress 2006 Geotechnical Engineering in the Information Technology AgeEds D J DeGroot J T DeJong J D Frost and L G Braise ASCE Reston(CDROM)
Phoon K K (2006b) Bootstrap estimation of sample autocorrelation functionsIn GeoCongress 2006 Geotechnical Engineering in the Information TechnologyAge Eds D J DeGroot J T DeJong J D Frost and L G Braise ASCE Reston(CDROM)
Phoon K K and Fenton G A (2004) Estimating sample autocorrelation func-tions using bootstrap In Proceedings of the 9th ASCE Specialty Conference onProbabilistic Mechanics and Structural Reliability Albuquerque (CDROM)
Phoon K K and Huang S P (2007) Uncertainty quantification using multi-dimensional Hermite polynomials In Probabilistic Applications in GeotechnicalEngineering (GSP 170) Eds K K Phoon G A Fenton E F Glynn C H JuangD V Griffiths T F Wolff and L M Zhang ASCE Reston (CDROM)
Phoon K K and Kulhawy F H (1999a) Characterization of geotechnicalvariability Canadian Geotechnical Journal 36(4) 612ndash24
Phoon K K and Kulhawy F H (1999b) Evaluation of geotechnical propertyvariability Canadian Geotechnical Journal 36(4) 625ndash39
Phoon K K and Kulhawy F H (2005) Characterization of model uncertainties forlaterally loaded rigid drilled shafts Geotechnique 55(1) 45ndash54
Phoon K K Becker D E Kulhawy F H Honjo Y Ovesen N K and Lo S R(2003b) Why consider reliability analysis in geotechnical limit state design InProceedings of the International Workshop on Limit State design in GeotechnicalEngineering Practice (LSD2003) Cambridge (CDROM)
Phoon K K Kulhawy F H and Grigoriu M D (1993) Observations onreliability-based design of foundations for electrical transmission line struc-tures In Proceedings International Symposium on Limit State Design inGeotechnical Engineering Vol 2 Danish Geotechnical Institute Copenhagenpp 351ndash62
Phoon K K Kulhawy F H and Grigoriu M D (1995) Reliability-based designof foundations for transmission line structures Report TR-105000 Electric PowerResearch Institute Palo Alto
74 Kok-Kwang Phoon
Phoon K K Huang S P and Quek S T (2002) Implementation of KarhunenndashLoeve expansion for simulation using a wavelet-Galerkin scheme ProbabilisticEngineering Mechanics 17(3) 293ndash303
Phoon K K Huang H W and Quek S T (2004) Comparison between KarhunenndashLoeve and wavelet expansions for simulation of Gaussian processes Computersand Structures 82(13ndash14) 985ndash91
Phoon K K Quek S T and An P (2003a) Identification of statistically homo-geneous soil layers using modified Bartlett statistics Journal of Geotechnical andGeoenvironmental Engineering ASCE 129(7) 649ndash59
Phoon K K Quek S T Chow Y K and Lee S L (1990) Reliability anal-ysis of pile settlement Journal of Geotechnical Engineering ASCE 116(11)1717ndash35
Press W H Teukolsky S A Vetterling W T and Flannery B P (1992) Numer-ical Recipes in C The Art of Scientific Computing Cambridge University PressNew York
Puig B and Akian J-L (2004) Non-Gaussian simulation using Hermite polynomialsexpansion and maximum entropy principle Probabilistic Engineering Mechanics19(4) 293ndash305
Puig B Poirion F and Soize C (2002) Non-Gaussian simulation using Hermitepolynomial expansion convergences and algorithms Probabilistic EngineeringMechanics 17(3) 253ndash64
Quek S T Chow Y K and Phoon K K (1992) Further contributions toreliability-based pile settlement analysis Journal of Geotechnical EngineeringASCE 118(5) 726ndash42
Quek S T Phoon K K and Chow Y K (1991) Pile group settlement a proba-bilistic approach International Journal of Numerical and Analytical Methods inGeomechanics 15(11) 817ndash32
Rackwitz R (2001) Reliability analysis ndash a review and some perspectives StructuralSafety 23(4) 365ndash95
Randolph M F and Houlsby G T (1984) Limiting pressure on a circular pileloaded laterally in cohesive soil Geotechnique 34(4) 613ndash23
Ravindra M K and Galambos T V (1978) Load and resistance factor design forsteel Journal of Structural Division ASCE 104(ST9) 1337ndash53
Reese L C (1958) Discussion of ldquoSoil modulus for laterally loaded pilesrdquoTransactions ASCE 123 1071ndash74
Reacutethaacuteti L (1988) Probabilistic Solutions in Geotechnics Elsevier New YorkRosenblueth E and Esteva L (1972) Reliability Basis for some Mexican Codes
Publication SP-31 American Concrete Institute DetroitRuben H (1960) Probability constant of regions under spherical normal distribu-
tion I Annals of Mathematical Statistics 31 598ndash619Ruben H (1962) Probability constant of regions under spherical normal distribu-
tion IV Annals of Mathematical Statistics 33 542ndash70Ruben H (1963) A new result on the distribution of quadratic forms Annals of
Mathematical Statistics 34 1582ndash4Sakamoto S and Ghanem R (2002) Polynomial chaos decomposition for the simu-
lation of non-Gaussian non-stationary stochastic processes Journal of EngineeringMechanics ASCE 128(2) 190ndash200
Numerical recipes for reliability analysis 75
Schueumlller G I Pradlwarter H J and Koutsourelakis P S (2004) A criticalappraisal of reliability estimation procedures for high dimensions ProbabilisticEngineering Mechanics 19(4) 463ndash74
Schweizer B (1991) Thirty years of copulas In Advances in Probability Distribu-tions with Given Marginals Kluwer Dordrecht pp 13ndash50
Semple R M (1981) Partial coefficients design in geotechnics Ground Engineering14(6) 47ndash8
Simpson B (2000) Partial factors where to apply them In Proceedings of theInternational Workshop on Limit State Design in Geotechnical Engineering(LSD2000) Melbourne (CDROM)
Simpson B and Driscoll R (1998) Eurocode 7 A Commentary ConstructionResearch Communications Ltd Watford Herts
Simpson B and Yazdchi M (2003) Use of finite element methods in geotech-nical limit state design In Proceedings of the International Workshop onLimit State design in Geotechnical Engineering Practice (LSD2003) Cambridge(CDROM)
Simpson B Pappin J W and Croft D D (1981) An approach to limit statecalculations in geotechnics Ground Engineering 14(6) 21ndash8
Sudret B (2007) Uncertainty propagation and sensitivity analysis in mechanicalmodels Contributions to structural reliability and stochastic spectral methods1rsquoHabilitation a Diriger des Recherches Universite Blaise Pascal
Sudret B and Der Kiureghian A 2000 Stochastic finite elements and reliabilitya state-of-the-art report Report UCBSEMM-200008 University of CaliforniaBerkeley
US Army Corps of Engineers (1997) Engineering and design introduction to proba-bility and reliability methods for use in geotechnical engineering Technical LetterNo 1110-2-547 Department of the Army Washington D C
Uzielli M and Phoon K K (2006) Some observations on assessment of Gaussianityfor correlated profiles In GeoCongress 2006 Geotechnical Engineering in theInformation Technology Age Eds D J DeGroot J T DeJong J D Frost andL G Braise ASCE Reston (CDROM)
Uzielli M Lacasse S Nadim F and Phoon K K (2007) Soil variability analysisfor geotechnical practice In Proceedings of the 2nd International Workshop onCharacterisation and Engineering Properties of Natural Soils Vol 3 Taylor andFrancis Singapore pp 1653ndash752
VanMarcke E H (1983) Random Field Analysis and Synthesis MIT PressCambridge MA
Verhoosel C V and Gutieacuterrez M A (2007) Application of the spectral stochasticfinite element method to continuum damage modelling with softening behaviourIn Proceedings of the 10th International Conference on Applications of Statisticsand Probability in Civil Engineering Tokyo (CDROM)
Vrouwenvelder T and Faber M H (2007) Practical methods of structural reli-ability In Proceedings of the 10th International Conference on Applications ofStatistics and Probability in Civil Engineering Tokyo (CDROM)
Winterstein S R Ude T C and Kleiven G (1994) Springing and slow driftresponses predicted extremes and fatigue vs simulation Proceedings Behaviourof Offshore Structures Vol 3 Elsevier Cambridge pp 1ndash15
Chapter 2
Spatial variability andgeotechnical reliability
Gregory B Baecher and John T Christian
Quantitative measurement of soil properties differentiated the new disciplineof soil mechanics in the early 1900s from the engineering of earth workspracticed since antiquity These measurements however uncovered a greatdeal of variability in soil properties not only from site to site and stratumto stratum but even within what seemed to be homogeneous deposits Wecontinue to grapple with this variability in current practice although newtools of both measurement and analysis are available for doing so Thischapter summarizes some of the things we know about the variability ofnatural soils and how that variability can be described and incorporated inreliability analysis
21 Variability of soil properties
Table 21 illustrates the extent to which soil property data vary accordingto Phoon and Kulhawy (1996) who have compiled coefficients of variationfor a variety of soil properties The coefficient of variation is the standarddeviation divided by the mean Similar data have been reported by Lumb(1966 1974) Lee et al (1983) and Lacasse and Nadim (1996) amongothers The ranges of these reported values are wide and are only suggestiveof conditions at a specific site
It is convenient to think about the impact of variability on safety byformulating the reliability index
β = E[MS]SD[MS]or
E[FS]minus 1SD[FS] (21)
in which β = reliability index MS = margin of safety (resistance minusload) FS = factor of safety (resistance divided by load) E[middot] = expectationand SD[middot] = standard deviation It should be noted that the two definitionsof β are not identical unless MS = 0 or FS = 1 Equation 21 expressesthe number of standard deviations separating expected performance from afailure state
Spatial variability and geotechnical reliability 77
Table 21 Coefficient of variation for some common field measurements (Phoon andKulhawy 1996)
Test type Property Soil type Mean Units Cov()
qT Clay 05ndash25 MNm2 lt 20CPT qc Clay 05ndash2 MNm2 20ndash40
qc Sand 05ndash30 MNm2 20ndash60VST su Clay 5ndash400 kNm2 10ndash40SPT N Clay and sand 10ndash70 blowsft 25ndash50
A reading Clay 100ndash450 kNm2 10ndash35A reading Sand 60ndash1300 kNm2 20ndash50B reading Clay 500ndash880 kNm2 10ndash35
DMT B Reading Sand 350ndash2400 kNm2 20ndash50ID Sand 1ndash8 20ndash60KD Sand 2ndash30 20ndash60ED Sand 10ndash50 MNm2 15ndash65pL Clay 400ndash2800 kNm2 10ndash35
PMT pL Sand 1600ndash3500 kNm2 20ndash50EPMT Sand 5ndash15 MNm2 15ndash65wn Clay and silt 13ndash100 8ndash30WL Clay and silt 30ndash90 6ndash30WP Clay and silt 15ndash15 6ndash30
Lab Index PI Clay and silt 10ndash40 _a
LI Clay and silt 10 _a
γ γd Clay and silt 13ndash20 KNm3 lt 10Dr Sand 30ndash70 10ndash40
50ndash70b
NotesaCOV = (3ndash12)meanbThe first range of variables gives the total variability for the direct method of determination and thesecond range of values gives the total variability for the indirect determination using SPT values
The important thing to note in Table 21 is how large are the reportedcoefficients of variations of soil property measurements Most are tens ofpercent implying reliability indices between one and two even for conser-vative designs Probabilities of failure corresponding to reliability indiceswithin this range ndash shown in Figure 21 for a variety of common distribu-tional assumptions ndash are not reflected in observed rates of failure of earthstructures and foundations We seldom observe failure rates this high
The inconsistency between the high variability of soil property data andthe relatively low rate of failure of prototype structures is usually attributedto two things spatial averaging and measurement noise Spatial averagingmeans that if one is concerned about average properties within some volumeof soil (eg average shear strength or total compression) then high spotsbalance low spots so that the variance of the average goes down as thatvolume of mobilized soil becomes larger Averaging reduces uncertainty1
78 Gregory B Baecher and John T Christian
1
01
Probability of Failure for Different Distributions
001
0001
00001
Pro
babi
lity
000001
00000010 1 2 3
Beta
Normal
Lognormal COV = 01
Triangular
Lognormal COV = 005
Lognormal COV = 015
4 5
Figure 21 Probability of failure as a function of reliability index for a variety of commonprobability distribution forms
Measurement noise means that the variability in soil property data reflectstwo things real variability and random errors introduced by the processof measurement Random errors reduce the precision with which estimatesof average soil properties can be made but they do not affect the in-fieldvariation of actual properties so the variability apparent in measurementsis larger ndash possibly substantially so ndash than actual in situ variability2
211 Spatial variation
Spatial variation in a soil deposit can be characterized in detail but only witha great number of observations which normally are not available Thus itis common to model spatial variation by a smooth deterministic trend com-bined with residuals about that trend which are described probabilisticallyThis model is
z(x) = t(x) + u(x) (22)
in which z(x) is the actual soil property at location x (in one or more dimen-sions) t(x) is a smooth trend at x and u(x) is residual deviation from thetrend The residuals are characterized as a random variable of zero-meanand some variance
Var(u) = E[z(x) minus t(x)2] (23)
Spatial variability and geotechnical reliability 79
in which Var(x) is the variance The residuals are characterized as randombecause there are too few data to do otherwise This does not presume thatsoil properties actually are random The variance of the residuals reflectsuncertainty about the difference between the fitted trend and the actual valueof soil properties at particular locations Spatial variation is modeled stochas-tically not because soil properties are random but because information islimited
212 Trend analysis
Trends are estimated by fitting lines curves or surfaces to spatially refer-enced data The easiest way to do this is by regression analysis For exampleFigure 22 shows maximum past pressure measurements as a function ofdepth in a deposit of Gulf of Mexico clay The deposit appears homogeneous
Maximum Past Pressure σvm (KSF)
Ele
vatio
n (f
t)
0 2 4 6 8
minus60
minus50
minus40
minus30
minus20
minus10
0
σvo
mean σvm
standard deviations
Figure 22 Maximum past pressure measurements as a function of depth in Gulf of Mexicoclays Mobile Alabama
80 Gregory B Baecher and John T Christian
and mostly normally consolidated The increase of maximum past pressurewith depth might be expected to be linear Data from an over-consolidateddesiccated crust are not shown The trend for the maximum past pressuredata σ prime
vm with depth x is
σ primevm = t(x) + u(x) = α0 +α1x = u (24)
in which t(x) is the trend of maximum past pressure with depth x α0and α1 are regression coefficients and u is residual variation about thetrend taken to be constant with depth (ie it is not a function of x)Applying standard least squares analysis the regression coefficients mini-mizing Var[u] are α0 = 3 ksf (014 KPa) and α1 = 006 ksfft (14 times 10minus3
KPam) yielding Var(u) = 10 ksf (005 KPa) for which the correspondingtrend line is shown The trend t(x) = 3 + 006x is the best estimate or meanof the maximum past pressure as a function of depth (NB ksf = kip persquare foot)
The analysis can be made of data in higher dimensions which in matrixnotation becomes
z = Xα + u (25)
in which z is the vector of the n observations z=z1 hellip zn X=x1 x2 isthe 2 x n matrix of location coordinates corresponding to the observationsα = αα1 hellip αn is the vector of trend parameters and u is the vectorof residuals corresponding to the observations Minimizing the variance ofthe residuals u(x) over α gives the best-fitting trend surface in a frequentistsense which is the common regression surface
The trend surface can be made more flexible for example in the quadraticcase the linear expression is replaced by
z = α0 +α1x +α2x2 + u (26)
and the calculation for α performed the same way Because the quadraticsurface is more flexible than the planar surface it fits the observed datamore closely and the residual variations about it are smaller On the otherhand the more flexible the trend surface the more regression parameters thatneed to be estimated from a fixed number of data so the fewer the degrees offreedom and the greater the statistical error in the surface Examples of theuse of trend surfaces in the geotechnical literature are given by Wu (1974)Ang and Tang (1975) and many others
Historically it has been common for trend analysis in geotechnicalengineering to be performed using frequentist methods Although this istheoretically improper because frequentist methods yield confidence inter-vals rather than probability distributions on parameters the numericalerror is negligible The Bayesian approach begins with the same model
Spatial variability and geotechnical reliability 81
However rather than defining an estimator such as the least squarescoefficients the Bayesian approach specifies an a priori probability distri-bution on the coefficients of the model and uses Bayesrsquos Theorem to updatethat distribution in light of the observed data
The following summarizes Bayesian results from Zellner (1971) for theone-dimensional linear case of Equation (24) Let Var(u)= σ 2 so thatVar(u) = Iσ 2 in which I is the identity matrix The prior probability densityfunction (pdf) of the parameters ασ is represented as f (ασ ) Given a set ofobservations z = z1 hellip zn the updated or posterior pdf of ασ is foundfrom Bayesrsquos Theorem as f (ασ |z) prop f (ασ )L(ασ |z) in which L(ασ |z) isthe Likelihood of the data (ie the conditional probability of the observeddata for various values of the parameters) If variations about the trend lineor surface are jointly Normal the likelihood function is
L(ασ |z) = MN(z|ασ )prop expminus(z minus Xα)primeminus1(z minus Xα) (27)
in which MN() is the Multivariate-Normal distribution having mean Xα
and covariance matrix = Iσ Using a non-informative prior f (α σ ) prop σminus1 and measurements y made
at depths x the posterior pdf of the regression parameters is
f (α0α1σ |xy) prop 1σ n+1 exp
[minus 1
2σ 2
sumn
i=1(y1 minus (α0 +α1x1))2
](28)
The marginal distributions are
f (α0α1|xy) prop [νs2 + n(α0 minusα0) + 2(α0 minusα0)(α1 minusα1)xi
+ (α1 minusα1)2x21]minusn2
f (α0|xy) prop [ν + (xi minus x)2
s2x2i n
(α0 minusα0)2]minus(νminus1)2 (29)
f (α1|xy) prop [ν + (xi minus x)2
s2 (α1 minusα1)2]minus(νminus1)2
f (σ |xy) prop 1σνminus1 exp
(minus νs2
2σ 2
)
in which
ν = n minus 2
α0 = y minusα1x α1 =[sum(
xi minus x)(
yi minus y)][sum(
xi minus x)]
s2 = νminus1sum(
yi minusα0 minusα1xi
)2
82 Gregory B Baecher and John T Christian
y = nminus1sum
yi
x = nminus1sum
xi
The joint and marginal pdfrsquos of the regression coefficients are Student-tdistributed
213 Autocorrelation
In fitting trends to data as noted above the decision is made to dividethe total variability of the data into two parts one part explained by thetrend and the other as variation about the trend Residual variations notaccounted for by the trend are characterized by a residual variance Forexample the overall variance of the blow count data of Figure 23 is 45 bpf2
(475 bpm2) Removing a linear trend reduces this total to a residual varianceof about 11 bpf2(116 bpm2) The trend explains 33 bpf2 (349 bpm2)or about 75 of the spatial variation and 25 is unexplained by thetrend
The spatial structure remaining after a trend is removed usually displayscorrelations among the residuals That is the residuals off the trend are notstatistically independent of one another Positive residuals tend to clumptogether as do negative residuals Thus the probability of encountering a
570
575
580
585
590
595
600
605
610
0 10 20 30 40
STP Blowcount
50
Figure 23 Spatial variation of SPT blow count data in a silty sand (data from Hilldale 1971)
Spatial variability and geotechnical reliability 83
minus25
minus2
minus15
minus1
minus05
0
05
1
15
2
Distance along transect
Figure 24 Residual variations of SPT blow counts
continuous zone of weakness or high compressibility is greater than wouldbe predicted if the residuals were independent
Figure 24 shows residual variations of SPT blow counts measured at thesame elevation every 20 m beneath a horizontal transect at a site The dataare normalized to zero mean and unit standard deviation The dark line is asmooth curve drawn through the observed data The light line is a smoothcurve drawn through artificially simulated data having the same meanand same standard deviation but probabilistically independent Inspectionshows the natural data to be smoothly varying whereas the artificial dataare much more erratic
The remaining spatial structure of variation not accounted for by the trendcan be described by its spatial correlation called autocorrelation Formallyautocorrelation is the property that residuals off the mean trend are notprobabilistically independent but display a degree of association amongthemselves that is a function of their separation in space This degree ofassociation can be measured by a correlation coefficient taken as a functionof separation distance
Correlation is the property that on average two variables are linearlyassociated with one another Knowing the value of one provides informa-tion on the probable value of the other The strength of this association ismeasured by a correlation coefficient ρ that ranges between ndash1 and +1 Fortwo scalar variables z1 and z2 the correlation coefficient is defined as
ρ = Cov(z1z2)radicVar(z1)Var(z2)
= 1σz1
σz2
E[(z1 minusmicroz1)(z2 minusmicroz2
)] (210)
84 Gregory B Baecher and John T Christian
in which Cov(z1z2) is the covariance Var(zi) is the variance σ is thestandard deviation and micro is the mean
The two variables might be of different but related types for examplez1 might be water content and z2 might be undrained strength or the twovariables might be the same property at different locations for examplez1 might be the water content at one place on the site and z2 the watercontent at another place A correlation coefficient ρ = +1 means that tworesiduals vary together exactly When one is a standard deviation above itstrend the other is a standard deviation above its trend too A correlationcoefficient ρ = minus1 means that two residuals vary inversely When one is astandard deviation above its trend the other is a standard deviation belowits trend A correlation coefficient ρ = 0 means that the two residuals areunrelated In the case where the covariance and correlation are calculated asfunctions of the separation distance the results are called the autocovarianceand autocorrelation respectively
The locations at which the blow count data of Figure 23 were measuredare shown in Figure 25 In Figure 26 these data are used to estimate autoco-variance functions for blow count The data pairs at close separation exhibita high degree of correlation for example those separated by 20 m have acorrelation coefficient of 067 As separation distance increases correlationdrops although at large separations where the numbers of data pairs are
A
B
C
D E F G H I
Cprime
Bprime
Aprime
Dprime Eprime Fprime Gprime Hprime
3225
33
21
26
27
2220
37
8
3
210
15
14
169
1
19ST 40033 31 34
302924
23
181265
4 711 13
17
36
ST 40044
28
35
Iprime
Figure 25 Boring locations of blow count data used to describe the site (TW Lambe andAssociates 1982 Earthquake Risk at Patio 4 and Site 400 Longboat Key FLreproduced by permission of TW Lambe)
Spatial variability and geotechnical reliability 85
40
30
20
10
0
minus10
minus20
minus30
minus40
Aut
ocov
aria
nce
0 100 200 300 400 500 600
Separation (m)
8ndash10m
10ndash12m
6ndash8m
Figure 26 Autocorrelation functions for SPT data at at Site 400 by depth interval(TW Lambe and Associates 1982 Earthquake Risk at Patio 4 and Site 400Longboat Key FL reproduced by permission of TW Lambe)
smaller there is much statistical fluctuation For zero separation distancethe correlation coefficient must equal 10 For large separation distancesthe correlation coefficient approaches zero In between the autocorrelationusually falls monotonically from 10 to zero
An important point to note is that the division of spatial variation intoa trend and residuals about the trend is an assumption of the analysis it isnot a property of reality By changing the trend model ndash for example byreplacing a linear trend with a polynomial trend ndash both the variance of theresiduals and their autocorrelation function are changed As the flexibility ofthe trend increases the variance of the residuals goes down and in generalthe extent of correlation is reduced From a practical point of view theselection of a trend line or curve is in effect a decision on how much of thedata scatter to model as a deterministic function of space and how much totreat probabilistically
As a rule of thumb trend surfaces should be kept as simple as possiblewithout doing injustice to a set of data or ignoring the geologic settingThe problem with using trend surfaces that are very flexible (eg high-order polynomials) is that the number of data from which the parameters ofthose equations are estimated is limited The sampling variance of the trendcoefficients is inversely proportional to the degrees of freedom involvedν = (n minus k minus 1) in which n is the number of observations and k is the num-ber of parameters in the trend The more parameter estimates that a trend
86 Gregory B Baecher and John T Christian
surface requires the more uncertainty there is in the numerical values of thoseestimates Uncertainty in regression coefficient estimates increases rapidly asthe flexibility of the trend equation increases
If z(xi) = t(xi)+u(xi) is a continuous variable and the soil deposit is zonallyhomogeneous then at locations i and j which are close together the residualsui and uj should be expected to be similar That is the variations reflectedin u(xi) and u(xj) are associated with one another When the locations areclose together the association is usually strong As the locations becomemore widely separated the association usually decreases As the separationbetween two locations i and j approaches zero u(xi) and u(xj) become thesame the association becomes perfect Conversely as the separation becomeslarge u(xi) and u(xj) become independent the association becomes zeroThis is the behavior observed in Figure 26 for the Standard Peretrationtest(SPT) data
This spatial association of residuals off the trend t(xi) is summarized bya mathematical function describing the correlation of u(xi) and u(xj) asseparation distance increases This description is called the autocorrelationfunction Mathematically the autocorrelation function is
Rz(δ) = 1Varu(x)E[u(xi)u(xi+δ)] (211)
in which Rz(δ) is the autocorrelation function Var[u(x)] is the variance ofthe residuals across the site and E[u(xi)u(xi+δ)]=Cov[u(xi)u(xi+δ)] is thecovariance of the residuals spaced at separation distance δ By definitionthe autocorrelation at zero separation is Rz(0) = 10 and empirically formost geotechnical data autocorrelation decreases to zero as δ increases
If Rz(δ) is multiplied by the variance of the residuals Var[u(x)] theautocovariance function Cz(δ) is obtained
Cz(δ) = E[u(xi)u(xi+δ)] (212)
The relationship between the autocorrelation function and the autocovari-ance function is the same as that between the correlation coefficient and thecovariance except that autocorrelation and autocovariance are functions ofseparation distance δ
214 Example TONEN refinery Kawasaki Japan
The SPT data shown earlier come from a site overlying hydraulic bay fillin Kawasaki (Japan) The SPT data were taken in a silty fine sand betweenelevations +3 and minus7 m and show little if any trend horizontally so aconstant horizontal trend at the mean of the data was assumed Figure 27shows the means and variability of the SPT data with depth Figure 26 shows
Spatial variability and geotechnical reliability 87
Upper Fill
Blow Counts N - BlowsFt
0 5 10 15 20 25 350
2
4
6
8
10
12
14
Dep
th m
Sandy Fill
SandyAlluvium
+σminusσ
mea
n
Figure 27 Soil model and the scatter of blow count data (TW Lambe and Associates1982 Earthquake Risk at Patio 4 and Site 400 Longboat Key FL reproducedby permission of TW Lambe)
autocovariance functions in the horizontal direction estimated for threeintervals of elevation At short separation distances the data show distinctassociation ie correlation At large separation distances the data exhibitessentially no correlation
In natural deposits correlations in the vertical direction tend to have muchshorter distances than in the horizontal direction A ratio of about 1 to 10 forthese correlation distances is common Horizontally autocorrelation may beisotropic (ie Rz(δ) in the northing direction is the same as Rz(δ) in the eastingdirection) or anisotropic depending on geologic history However in prac-tice isotropy is often assumed Also autocorrelation is typically assumed tobe the same everywhere within a deposit This assumption called station-arity to which we will return is equivalent to assuming that the deposit isstatistically homogeneous
It is important to emphasize again that the autocorrelation function is anartifact of the way soil variability is separated between trend and residualsSince there is nothing innate about the chosen trend and since changingthe trend changes Rz(δ) the autocorrelation function reflects a modelingdecision The influence of changing trends on Rz(δ) is illustrated in data
88 Gregory B Baecher and John T Christian
Figure 28 Study area for San Francisco Bay Mud consolidation measurements (Javete1983) (reproduced with the authorrsquos permission)
analyzed by Javete (1983) (Figure 28) Figure 29 shows autocorrelationsof water content in San Francisco Bay Mud within an interval of 3 ft (1 m)Figure 210 shows the autocorrelation function when the entire site is con-sidered The difference comes from the fact that in the first figure the meantrend is taken locally within the 3 ft (1 m) interval and in the latter the meantrend is taken globally across the site
Autocorrelation can be found in almost all spatial data that are analyzedusing a model of the form of Equation (25) For example Figure 211shows the autocorrelation of rock fracture density in a copper porphyrydeposit Figure 212 shows autocorrelation of cone penetration resistancein North Sea Clay and Figure 213 shows autocorrelation of water con-tent in the compacted clay core of a rock-fill dam An interesting aspectof the last data is that the autocorrelations they reflect are more a func-tion of the construction process through which the core of the damwas placed than simply of space per se The time stream of borrowmaterials weather and working conditions at the time the core was
Spatial variability and geotechnical reliability 89
10
05
0rk
minus05
minus100 5 10 15 20
lag k
Figure 29 Autocorrelations of water content in San Francisco Bay Mud within an intervalof 3 ft (1 m) ( Javete 1983) (reproduced with the authorrsquos permission)
0 5 10
Lak k
15 20
rk
10
05
00
minus05
minus10
Figure 210 Autocorrelations of water content in San Francisco Bay Mud within entire siteexpressed in lag intervals of 25 ft ( Javete 1983) (reproduced with the authorrsquospermission)
placed led to trends in the resulting physical properties of the compactedmaterial
For purposes of modeling and analysis it is usually convenient to approx-imate the autocorrelation structure of the residuals by a smooth functionFor example a commonly used function is the exponential
Rz(δ) = exp(minusδδ0) (213)
in which δ0 is a constant having units of length Other functions commonlyused to represent autocorrelation are shown in Table 22 The distance atwhich Rz(δ) decays to 1e (here δ0) is sometimes called the autocorrelation(or autocovariance) distance
90 Gregory B Baecher and John T Christian
minus200
minus100
0
100
200
300
400
500
50
SEPARATION (M)
CO
VA
RIA
NC
E
10 15
Figure 211 Autocorrelation of rock fracture density in a copper porphyry deposit(Baecher 1980)
Distance of Separation (m)
ρ =exp [ minus rb2]
r = 30 m
At Depth 3 m
0 20 40 60 800 20 40 60 80
10
08
06
04
02
00Cor
rela
tion
Coe
ffici
ent10
08
06
04
02
00Cor
rela
tion
Coe
ffici
ent
Distance of Separation (m)
r = 30 m
At Depth 36 m
Figure 212 Autocorrelation of cone penetration resistance in North Sea Clay (Tang1979)
215 Measurement noise
Random measurement error is that part of data scatter attributable toinstrument- or operator-induced variations from one test to another Thisvariability may sometimes increase or decrease a measurement but itseffect on any one specific measurement is unknown As a first approxima-tion instrument and operator effects on measured properties of soils canbe represented by a frequency diagram In repeated testing ndash presumingthat repeated testing is possible on the same specimen ndash measured val-ues differ Sometimes the measurement is higher than the real value ofthe property sometimes it is lower and on average it may systematically
Spatial variability and geotechnical reliability 91
0 20 40 60 80 100
Lag Distance (test number)
050
025
000
minus025
minus050
Aut
ocov
aria
nce
Figure 213 Autocorrelation of water content in the compacted clay core of a rock-filldam (Beacher 1987)
Table 22 One-dimensional autocorrelation models
Model Equation Limits of validity(dimension of relevant space)
White noise Rx(δ) =
1 if δ = 00 otherwise Rn
Linear Rx(δ) =
1 minus|δ|δ0 if δ le δ00 otherwise R1
Exponential Rx(δ) = exp(minusδδ0) R1
Squared exponential(Gaussian)
Rx(δ) = exp2(minusδδ0) Rd
Power Cz(δ) = σ 21(|δ|2δ20)minusβ Rd β gt 0
differ from the real value This is usually represented by a simple model ofthe form
z = bx + e (214)
in which z is a measured value b is a bias term x is the actual propertyand e is a zero-mean independent and identically distributed (IID) errorThe systematic difference between the real value and the average of themeasurements is said to be measurement bias while the variability of themeasurements about their mean is said to be random measurement errorThus the error terms are b and e The bias is often assumed to be uncertainwith mean microb and standard deviation σb The IID random perturbation is
92 Gregory B Baecher and John T Christian
usually assumed to be Normally distributed with zero mean and standarddeviation σe
Random errors enter measurements of soil properties through a varietyof sources related to the personnel and instruments used in soil investiga-tions or laboratory testing Operator or personnel errors arise in many typesof measurements where it is necessary to read scales personal judgment isneeded or operators affect the mechanical operation of a piece of testingequipment (eg SPT hammers) In each of these cases operator differenceshave systematic and random components One person for example mayconsistently read a gage too high another too low If required to make aseries of replicate measurements a single individual may report numbersthat vary one from the other over the series
Instrumental error arises from variations in the way tests are set uploads are delivered or soil response is sensed The separation of mea-surement errors between operator and instrumental causes is not onlyindistinct but also unimportant for most purposes In triaxial tests soilsamples may be positioned differently with respect to loading platens insucceeding tests Handling and trimming may cause differing amounts ofdisturbance from one specimen to the next Piston friction may vary slightlyfrom one movement to another or temperature changes may affect fluidsand solids The aggregate result of all these variables is a number of dif-ferences between measurements that are unrelated to the soil properties ofinterest
Assignable causes of minor variation are always present because a verylarge number of variables affect any measurement One attempts to controlthose that have important effects but this leaves uncontrolled a large numberthat individually have only small effects on a measurement If not identifiedthese assignable causes of variation may influence the precision and possiblythe accuracy of measurements by biasing the results For example hammerefficiency in the SPT test strongly affects measured blow counts Efficiencywith the same hammer can vary by 50 or more from one blow to the nextHammer efficiency can be controlled but only at some cost If uncontrolledit becomes a source of random measurement error and increases the scatterin SPT data
Bias error in measurement arises from a number of reasonably well-understood mechanisms Sample disturbance is among the more importantof these mechanisms usually causing a systematic degradation of average soilproperties along with a broadening of dispersion The second major contrib-utor to measurement bias is the phenomenological model used to interpretthe measurements made in testing and especially the simplifying assump-tions made in that model For example the physical response of the testedsoil element might be assumed linear when in fact this is only an approxima-tion the reversal of principal stress direction might be ignored intermediateprincipal stresses might be assumed other than they really are and so forth
Spatial variability and geotechnical reliability 93
The list of possible discrepancies between model assumptions and the realtest conditions is long
Model bias is usually estimated empirically by comparing predictionsmade from measured values of soil engineering parameters against observedperformance Obviously such calibrations encompass a good deal more thanjust the measurement technique they incorporate the models used to makepredictions of field performance inaccuracies in site characterization and ahost of other things
Bjerrumrsquos (1972 1973) calibration of field vein test results for theundrained strength su of clay is a good example of how measurementbias can be estimated in practice This calibration compares values of sumeasured with a field vane against back-calculated values of su from large-scale failures in the field In principle this calibration is a regression analysisof back-calculated su against field vane su which yields a mean trend plusresidual variance about the trend The mean trend provides an estimate ofmicrob while the residual variance provides an estimate of σb The residual vari-ance is usually taken to be the same regardless of the value of x a commonassumption in regression analysis
Random measurement error can be estimated in a variety of wayssome direct and some indirect As a general rule the direct techniques aredifficult to apply to the soil measurements of interest to geotechnical engi-neers because soil tests are destructive Indirect methods for estimating Veusually involve correlations of the property in question either with otherproperties such as index values or with itself through the autocorrelationfunction
The easiest and most powerful methods involve the autocorrelationfunction The autocovariance of z after the trend has been removed becomes
Cz(δ) = Cx(δ) + Ce(δ) (215)
in which Cx(δ) is from Equation (212) and Cx(δ) is the autocovariancefunction of e However since ei and ej are independent except when i = jthe autocovariance function of e is a spike at δ = 0 and zero elsewhereThus Cx(δ) is composed of two functions By extrapolating the observedautocovariance function to the origin an estimate is obtained of the fractionof data scatter that comes from random error In the ldquogeostatisticsrdquo literaturethis is called the nugget effect
216 Example Settlement of shallow footings on sandIndiana (USA)
The importance of random measurement errors is illustrated by a case involv-ing a large number of shallow footings placed on approximately 10 mof uniform sand (Hilldale 1971) The site was characterized by Standard
94 Gregory B Baecher and John T Christian
Penetration blow count measurements predictions were made of settlementand settlements were subsequently measured
Inspection of the SPT data and subsequent settlements reveals an interest-ing discrepancy Since footing settlements on sand tend to be proportional tothe inverse of average blow count beneath the footing it would be expectedthat the coefficient of variation of the settlements equaled approximatelythat of the vertically averaged blow counts Mathematically settlement ispredicted by a formula of the form ρ prop qNc in which ρ = settlementq = net applied stress at the base of the footing and Nc = average cor-rected blow count (Lambe and Whitman 1979) Being multiplicative thecoefficient of variation of ρ should be the same as that of Nc
In fact the coefficient of variation of the vertically averaged blow countsis about Nc
= 045 while the observed values of total settlements for268 footings have mean 035 inches and standard deviation 012 inchesso ρ = (012035) = 034 Why the difference The explanation maybe found in estimates of the measurement noise in the blow count dataFigure 214 shows the horizontal autocorrelation function for the blowcount data Extrapolating this function to the origin indicates that the noise(or small scale) content of the variability is about 50 of the data scattervariance Thus the actual variability of the vertically averaged blow counts is
aboutradic
12
2N =radic
12 (045)2 = 032 which is close to the observed variability
minus04
minus02
00
02
04
06
08
10
0 200 400 600 800 1000 1200
Separation (ft)
Figure 214 Autocorrelation function for SPT blow count in sand (Adapted from Hilldale1971)
Spatial variability and geotechnical reliability 95
of the footing settlements Measurement noise of 50 or even more of theobserved scatter of in situ test data particularly the SPT has been noted onseveral projects
While random measurement error exhibits itself in the autocorrelation orautocovariance function as a spike at δ = 0 real variability of the soil at ascale smaller than the minimum boring spacing cannot be distinguished frommeasurement error when using the extrapolation technique For this reasonthe ldquonoiserdquo component estimated in the horizontal direction may not be thesame as that estimated in the vertical direction
For many but not all applications the distinction between measure-ment error and small-scale variability is unimportant For any engineeringapplication in which average properties within some volume of soil areimportant the small-scale variability averages quickly and therefore has lit-tle effect on predicted performance Thus for practical purposes it can betreated as if it were a measurement error On the other hand if perfor-mance depends on extreme properties ndash no matter their geometric scale ndashthe distinction between measurement error and small scale is importantSome engineers think that piping (internal erosion) in dams is such a phe-nomenon However few physical mechanisms of performance easily cometo mind that are strongly affected by small-scale spatial variability unlessthose anomalous features are continuous over a large extent in at least onedimension
22 Second-moment soil profiles
Natural variability is one source of uncertainty in soil properties the otherimportant source is limited knowledge Increasingly these are referred to asaleatory and epistemic uncertainty respectively (Hartford 1995)3 Limitedknowledge usually causes systematic errors For example limited numbersof tests lead to statistical errors in estimating a mean trend and if there isan error in average soil strength it does not average out In geotechnical reli-ability the most common sources of knowledge uncertainty are model andparameter selection (Figure 215) Aleatory and epistemic uncertainties canbe combined and represented in a second-moment soil profile The second-moment profile shows means and standard deviations of soil properties withdepth in a formation The standard deviation at depth has two componentsnatural variation and systematic error
221 Example SHANSEP analysis of soft claysAlabama (USA)
In the early 1980s Ideal Basic Industries Inc (IDEAL) constructed a cementmanufacturing facility 11 miles south of Mobile Alabama abutting a shipchannel running into Mobile Bay (Baecher et al 1997) A gantry crane at the
96 Gregory B Baecher and John T Christian
Temporal
Spatial
Natural Variability
Model
Parameters
Knowledge Uncertainty
Objectives
Values
Time Preferences
Decision ModelUncertainty
Risk Analysis
Figure 215 Aleatory epistemic and decision model uncertainty in geotechnical reliabilityanalysis
facility unloaded limestone ore from barges moored at a relieving platformand place the ore in a reserve storage area adjacent to the channel As the sitewas underlain by thick deposits of medium to soft plastic deltaic clay con-crete pile foundations were used to support all facilities of the plant exceptfor the reserve limestone storage area This 220 ft (68 m) wide by 750 ft(230 m) long area provides limestone capacity over periods of interrupteddelivery Although the clay underlying the site was too weak to supportthe planned 50 ft (15 m) high stockpile the cost of a pile supported matfoundation for the storage area was prohibitive
To solve the problem a foundation stabilization scheme was conceivedin which limestone ore would be placed in stages leading to consolidationand strengthening of the clay and this consolidation would be hastened byvertical drains However given large scatter in engineering property data forthe clay combined with low factors of safety against embankment stabilityfield monitoring was essential
The uncertainty in soil property estimates was divided between that causedby data scatter and that caused by systematic errors (Figure 216) These wereseparated into four components
bull spatial variability of the soil depositbull random measurement noisebull statistical estimation error andbull measurement or model bias
The contributions were mathematically combined by noting that the vari-ances of these nearly independent contributions are approximately additive
V[x] asymp Vspatial[x]+ Vnoise[x]+ Vstatistical[x]+ Vbias[x] (216)
Spatial variability and geotechnical reliability 97
Uncertainty
Data Scatter
SpatialVariability
MeasurementNoise
SystematicError
ModelBias
StatisticalError
Figure 216 Sources of uncertainty in geotechnical reliability analysis
E-6400
minus180 minus100
minus20
minus40
minus60
minus80
0
20
0
ELE
VAT
ION
(ft)
100 200 280
E-6500
DRAINED ZONE
DENSE SAND
CLCL
CL
CLCL
CL
SANDY MATERIAL
Edge of primary limestone storage areaand stacker
of Westaccess road
MEDIUM-SOFT TO MEDIUMCL-CH GRAY SILTY CLAY
of drainageditch
of 20ft berm Stockplie of 20ft berm
I-4P
SPS-2I-12I-3
N-5381 RLSANORTH
of East occess road
EAST COORDINATE (ft)
DISTANCE FROM THE STOCKPILE CENTERLINE (ft)
E-6600 E-6700 E-6800
Figure 217 West-east cross-section prior to loading
in which V[x] = variance of total uncertainty in the property x Vspatial[x] =variance of the spatial variation of x Vnoise[x]= variance of the measurementnoise in x Vstatistical[x] = variance of the statistical error in the expectedvalue of x and Vbias[x] = variance of the measurement or model bias in xIt is easiest to think of spatial variation as scatter around the mean trendof the soil property and systematic error as uncertainty in the mean trenditself The first reflects soil variability after random measurement error hasbeen removed the second reflects statistical error plus measurement biasassociated with the mean value
Initial vertical effective stresses σ vo were computed using total unitweights and initial pore pressures Figure 22 shows a simplified profile priorto loading (Figure 217) The expected value σ vm profile versus elevation
98 Gregory B Baecher and John T Christian
was obtained by linear regression The straight short-dashed lines showthe standard deviation of the σ vm profile reflecting data scatter about theexpected value The curved long-dashed lines show the standard deviation ofthe expected value trend itself The observed data scatter about the expectedvalue σ vm profile reflects inherent spatial variability of the clay plus ran-dom measurement error in the determination of σ vm from any one testThe standard deviation about the expected value σ vm profile is about 1 ksf(005 MPa) corresponding to a standard deviation in over-consolidationratio (OCR) from 08 to 02 The standard deviation of the expected valueranges from 02 to 05 ksf (001 to 0024 MPa)
Ten CKoUDSS tests were performed on undisturbed clay samples to deter-mine undrained stressndashstrainndashstrength parameters to be used in the stresshistory and normalized soil engineering properties (SHANSEP) procedure ofLadd and Foott (1974) Reconsolidation beyond the in situ σ vm was used tominimize the influence of sample disturbance Eight specimens were shearedin a normally consolidated state to assess variation in the parameter s withhorizontal and vertical locations The last two specimens were subjected to asecond shear to evaluate the effect of OCR The direct simple shear (DSS) testprogram also provided undrained stressndashstrain parameters for use in finiteelement undrained deformation analyses
Since there was no apparent trend with elevation expected valueand standard deviation values were computed by averaging all data toyield
su = σ vos(σ vm
σ vo
)m
(217)
in which s =(0213 plusmn 0028) and m =(085 plusmn 005) As a first approximationit was assumed that 50 of the variation in s was spatial and 50 wasnoise The uncertainty in m estimated from data on other clays is primarilydue to variability from one clay type to another and hence was assumedpurely systematic It was assumed that the uncertainty in m estimated fromonly two tests on the storage area clay resulted from random measurementerror
The SHANSEP su profile was computed using Equation (217) If σ vo σ vms and m are independent and σ vo is deterministic (ie there is no uncertaintyin σ vo) first-order second-moment error analysis leads to the expressions
E[su] = σ voE[s](
E[σ vo]σ vo
)E[m](218)
2[su] = 2[s]+ E2[m]2[σ vm]+ 1n2(
E[σ vm]σ vo
)V[m] (219)
Spatial variability and geotechnical reliability 99
in which E[X] = expected value of X V[X] = variance of X and [X] =radicV[X]E[X]= coefficient of variation of X The total coefficient of variation
of su is divided between spatial and systematic uncertainty such that
2[su] = 2sp[su]+2
sy[su] (220)
Figure 218 shows the expected value su profile and the standard deviationof su divided into spatial and systematic components
Stability during initial undrained loading was evaluated using two-dimensional (2D) circular arc analyses with SHANSEP DSS undrainedshear strength profiles Since these analyses were restricted to the east andwest slopes of the stockpile 2D analyses assuming plane strain conditionsappeared justified Azzouz et al (1983) have shown that this simplifiedapproach yields factors of safety that are conservative by 10ndash15 for similarloading geometries
Because of differences in shear strain at failure for different modes offailure along a failure arc ldquopeakrdquo shear strengths are not mobilized simul-taneously all along the entire failure surface Ladd (1975) has proposed aprocedure accounting for strain compatibility that determines an averageshear strength to be used in undrained stability analyses Fuleihan and Ladd(1976) showed that in the case of the normally consolidated AtchafalayaClay the CKoUDSS SHANSEP strength was in agreement with the averageshear strength computed using the above procedure All the 2D analyses usedthe Modified Bishop method
To assess the importance of variability in su to undrained stability it isessential to consider the volume of soil of importance to the performanceprediction At one extreme if the volume of soil involved in a failure wereinfinite spatial uncertainty would completely average out and the system-atic component uncertainty would become the total uncertainty At the otherextreme if the volume of soil involved were infinitesimal spatial and sys-tematic uncertainties would both contribute fully to total uncertainty Theuncertainty for intermediate volumes of soil depends on the character ofspatial variability in the deposit specifically on the rapidity with which soilproperties fluctuate from one point to another across the site A convenientindex expressing this scale of variation is the autocorrelation distance δ0which measures the distance to which fluctuations of soil properties abouttheir expected value are strongly associated
Too few data were available to estimate autocorrelation distance for thestorage area thus bounding calculations were made for two extreme casesin the 2D analyses Lδ0 rarr 0 (ie ldquosmallrdquo failure surface) and Lδ0 rarr infin(ie ldquolargerdquo failure surface) in which L is the length of the potential fail-ure surface Undrained shear strength values corresponding to significantaveraging were used to evaluate uncertainty in the factor of safety forlarge failure surfaces and values corresponding to little averaging for small
100 Gregory B Baecher and John T Christian
Symbol
20
08
10
12
14
20 25 30
20 25
=E [FS] minus 10
SD[FS]
FILL HEIGHT Hf (ft)30
Mean minusone standarddeviation (SD)
Mean FS = E [FS]
35
35
RE
LIA
BIL
ITY
IND
EX
bFA
CTO
R O
F S
AF
ET
Y F
S
15
10
05
0
Case Effect of uncertainty in
clay undrained shearstrength su
Lro
Lro
Lro
Lro
Lro
Lro
fill total unit weight gfclay undrained shear strengthsu and fill total unit weight gf
(I)
b
infin
infin
infin
0
0
0
Figure 218 Expected value su profile and the standard deviation of su divided into spatialand systematic components
failure surface The results of the 2D stability analysis were plotted as afunction of embankment height
Uncertainty in the FS was estimated by performing stability analyses usingthe procedure of Christian et al (1994) with expected value and expectedvalue minus standard deviation values of soil properties For a given expected
Spatial variability and geotechnical reliability 101
value of FS the larger the standard deviation of FS the higher the chancethat the realized FS is less than unity and thus the lower the actual safety ofthe facility The second-moment reliability index [Equation (21)] was usedto combine E[FS] and SD[FS] in a single measure of safety and related to aldquonominalrdquo probability of failure by assuming FS Normally distributed
23 Estimating autocovariance
Estimating autocovariance from sample data is the same as making anyother statistical estimate Sample data differ from one set of observa-tions to another and thus the estimates of autocovariance differ Theimportant questions are how much do these estimates differ and howmuch might one be in error in drawing inferences There are two broadapproaches Frequentist and Bayesian The Frequentist approach is morecommon in geotechnical practice For discussion of Bayesian approaches toestimating autocorrelation see Zellner (1971) Cressie (1991) or Berger et al(2001)
In either case a mathematical function of the sample observations is usedas an estimate of the true population parameters θ One wishes to determineθ = g(z1 zn) in which z 1 hellip zn is the set of sample observations and θ which can be a scalar vector or matrix For example the sample mean mightbe used as an estimator of the true population mean The realized value of θfor a particular sample z1 hellip zn is an estimate As the probabilistic proper-ties of the z1hellip zn are assumed the corresponding probabilistic propertiesof θ can be calculated as functions of the true population parameters This iscalled the sampling distribution of θ The standard deviation of the samplingdistribution is called the standard error
The quality of the estimate obtained in this way depends on how variablethe estimator θ is about the true value θ The sampling distribution andhence the goodness of an estimate has to do with how the estimate mighthave come out if another sample and therefore another set of observationshad been made Inferences made in this way do not admit of a probabilitydistribution directly on the true population parameter Put another way theFrequentist approach presumes the state of nature θ to be a constant andyields a probability that one would observe those data that actually wereobserved The probability distribution is on the data not on θ Of coursethe engineer or analyst wants the reverse the probability of θ given the dataFor further discussion see Hartford and Baecher (2004)
Bayesian estimation works in a different way Bayesian theory allowsprobabilities to be assigned directly to states of nature such as θ ThusBayesian methods start with an a priori probability distribution f (θ ) whichis updated by the likelihood of observing the sample using Bayesrsquos Theorem
f (θ |z1 zn)infin f (θ )L(θ |z1 zn) (221)
102 Gregory B Baecher and John T Christian
in which f (θ |z1hellip zn) is the a posteriori pdf of θ conditioned on the obser-vations and L(θ |z1 hellip zn) is the likelihood of θ which is the conditionalprobability of z1 hellip zn as a function of θ Note the Fisherian conceptof a maximum likelihood estimator is mathematically related to Bayesianestimation in that both adopt the likelihood principle that all informationin the sample relevant to making an estimate is contained in the Likelihoodfunction however the maximum likelihood approach still ends up with aprobability statement on the variability of the estimator and not on the stateof nature which is an important distinction
231 Moment estimation
The most common (Frequentist) method of estimating autocovariance func-tions for soil and rock properties is the method of moments This uses thestatistical moments of the observations (eg sample means variances andcovariances) as estimators of the corresponding moments of the populationbeing sampled
Given the measurements z1hellip zn made at equally spaced locationsx1hellip xn along a line as for example in a boring the sample autocovarianceof the measurements for separation is
Cz(δ) = 1(n minus δ)
nminusδsumi=1
[z(xi) minus t(xi)z(xi+δ) minus t(xi+δ)] (222)
in which Cz(δ) is the estimator of the autocovariance function at δ (nminus δ) isthe number of data pairs having separation distance δ and t(xi) is the trendremoved from the data at location xi
Often t(xi) is simply replaced by the spatial mean estimated by the meanof the sample The corresponding moment estimator of the autocorrelationR(δ) is obtained by dividing both sides by the sample variance
Rz(δ) = 1s2z (n minus δ)
nminusδsumi=1
[z(xi) minus t(xi)z(xi+δ) minus t(xi+δ)] (223)
in which sz is the sample standard deviation Computationally this simplyreduces to taking all data pairs of common separation distance d calcu-lating the correlation coefficient of that set then plotting the result againstseparation distance
In the general case measurements are seldom uniformly spaced at least inthe horizontal plane and seldom lie on a line For such situations the sampleautocovariance can still be used as an estimator but with some modificationThe most common way to accommodate non-uniformly placed measure-ments is by dividing separation distances into bands and then taking theaverages within those bands
Spatial variability and geotechnical reliability 103
The moment estimator of the autocovariance function requires no assump-tions about the shape of the autocovariance function except that secondmoments exist The moment estimator is consistent in that as the samplesize becomes large E[(θ minus θ )2]rarr 0 On the other hand the moment estima-tor is only asymptotically unbiased Unbiasedness means that the expectedvalue of the estimator over all ways the sample might have been taken equalsthe actual value of the function being estimated For finite sample sizes theexpected values of the sample autocovariance can differ significantly from theactual values yielding negative values beyond the autocovariance distance(Weinstock 1963)
It is well known that the sampling properties of the moment estima-tor of autocorrelation are complicated and that large sampling variances(and thus poor confidence) are associated with estimates at large sep-aration distances Phoon and Fenton (2004) and Phoon (2006a) haveexperimented with bootstrapping approaches to estimate autocorrelationfunctions with promising success These and similar approaches fromstatistical signal processing should be exploited more thoroughly in thefuture
232 Example James Bay
The results of Figure 219 were obtained from the James Bay data ofChristian et al (1994) using this moment estimator The data are froman investigation into the stability of dykes on a soft marine clay at theJames Bay Project Queacutebec (Ladd et al 1983) The marine clay at the site is
Separation Distance
Aut
ocov
aria
nce
30
0
10
20Maximum Likelihood estimate(curve)
Moment estimates
150m100500
Figure 219 Autocovariance of field vane clay strength data James Bay Project (Christianet al 1994 reproduced with the permission of the American Society of CivilEngineers)
104 Gregory B Baecher and John T Christian
approximately 8 m thick and overlies a lacustrine clay The depth-averagedresults of field vane tests conducted in 35 borings were used for the cor-relation analysis Nine of the borings were concentrated in one location(Figures 220 and 221)
First a constant mean was removed from the data Then the product ofeach pair of residuals was calculated and plotted against separation distanceA moving average of these products was used to obtain the estimated pointsNote the drop in covariance in the neighborhood of the origin and alsothe negative sample moments in the vicinity of 50ndash100 m separation Notealso the large scatter in the sample moments at large separation distanceFrom these estimates a simple exponential curve was fitted by inspectionintersecting the ordinate at about 60 of the sample variance This yieldsan autocovariance function of the form
Cz(δ) =
22 kPa2 for δ = 0
13expminusδ23 m for δgt0(224)
in which variance is in kPa2 and distance in m Figure 222 shows variancecomponents for the factor of safety for various size failures
INDEX PROPERTIES FIELD VANE STRESS HISTORY
DE
PT
H Z
(m
)
0
SOILPROFILE
0
0 1 2 3 0 20 6040 0 100 200 300
10 20 30
2
4
6
8
10
12
14
16
18
20
Mean from8 FV
Selectedcu Profile
Ip
IL
Ip () cu (FV) (kPa) s primevo and s primep (kPa)
IL
Selecteds primep Profile
BlockTube
Till
Lacustrineclay
Marineclay
Crust
s primep
s primevo
gb=90kNm3
gb=90kNm3
gb=105kNm3
Figure 220 Soil property data summary James Bay (Christian et al 1994 reproduced withthe permission of the American Society of Civil Engineers)
160
140
120
100
Y a
xis
(m)
80
200 180 160 140 120 100 80 60 40 20 0 minus20
X axis (m)
Fill properties = 20kNmsup3rsquo= 30
Note All slopes are 3 horizontal to1 vertical
Y 123 m
Stage 2Berm 2Berm 1
Foundation clay Critical wedge - stage 2
Critical circles - single stage
Till Limit of vertical drains
6030
Stage 1
Figure 221 Assumed failure geometries for embankments of three heights (Christianet al 1994 reproduced with the permission of the American Society of CivilEngineers)
Embankment Height
6 m 12 m 23 m0
002
004
006
008
Var
ianc
e of
Fac
tor
of S
afet
y
01Rest of Sp VarAve Spatial VarSystematic Error
F = 1500R = 07
F = 1453R = 02
F = 1427R = 007
Figure 222 Variance components of the factor of safety for three embankment heights(Christian et al 1994 reproduced with the permission of the American Societyof Civil Engineers)
106 Gregory B Baecher and John T Christian
233 Maximum likelihood estimation
Maximum likelihood estimation takes as the estimator that value of theparameter(s) θ leading to the greatest probability of observing the dataz1 hellip zn actually observed This is found by maximizing the likelihoodfunction L(θ |z1 hellip zn) Maximum likelihood estimation is parametricbecause the distributional form of the pdf f (z1 hellip zn|θ ) must be specifiedIn practice the estimate is usually found by maximizing the log-likelihoodwhich because it deals with a sum rather than a product and becausemany common probability distributions involve exponential terms is moreconvenient
The appeal of the maximum likelihood estimator is that it possessesmany desirable sampling properties Among others it has minimum vari-ance (although not necessarily unbiased) is consistent and asymptoticallyNormal The asymptotic variance of θML is
limnrarrinfin Var[θML]=Iz(θ ) = nE[minusδ2LLpartθ2] (225)
in which Iz(θ ) is Fisherrsquos Information (Barnett 1982) and LL is the log-likelihood
Figure 223 shows the results of simulated sampling experiments in whichspatial fields were generated from a multivariate Gaussian pdf with speci-fied mean trend and autocovariance function Samples of sizes n = 36 64and 100 were taken from these simulated fields and maximum likelihoodestimators used to obtain estimates of the parameters of the mean trend andautocovariance function The smooth curves show the respective asymp-totic sampling distributions which in this case conform well with the actualestimates (DeGroot and Baecher 1993)
An advantage of the maximum likelihood estimator over moment esti-mates in dealing with spatial data is that it allows simultaneous estimation ofthe spatial trend and autocovariance function of the residuals Mardia andMarshall (1984) provide an algorithmic procedure finding the maximumDeGroot and Baecher used the Mardia and Marshall approach in analyzingthe James Bay data First they removed a constant mean from the data andestimated the autocovariance function of the residuals as
Cz(δ) =
23forδ = 0133 exp minus δ214 for δ gt 0
(226)
in which variance is in kPa2 and distance is in m Then using estimating thetrend implicitly
β0 = 407 kPa
β1 = minus20 times 10minus3 kPam
Spatial variability and geotechnical reliability 107
20
15
Num
ber
of D
ata
Num
ber
of D
ata
10
5
0
15
10
5
0
20
Num
ber
of D
ata
15
10
5
0
20
04 06 08 10 12 14 16
04 06 08 10 12 14 16
04 06 08 10Variance
12 14 16
0
4
3
2
1
0
1
2 f(x)
f(x)
3
4(a) n=36
(b) n=64
Asymptotic
0
1
2 f(x)
3
4(c) n=100
Asymptotic
Asymptotic
Figure 223 Simulated sampling experiments in which spatial fields were generated froma multivariate Gaussian pdf with specified mean trend and autocovariancefunction (DeGroot and Baecher 1993 reproduced with the permission of theAmerican Society of Civil Engineers)
β2 = minus59 times 10minus3 kPam (227)
Cz(δ) =
23 kPa2 forδ = 0133 kPa2 expminus(δ214 m) forδ gt 0
The small values of β1 and β2 suggest that the assumption of constantmean is reasonable Substituting a squared-exponential model for the
108 Gregory B Baecher and John T Christian
autocovariance results in
β0 = 408 kPa
β1 = minus21 times 10minus3 kPam
β2 = minus61 times 10minus3 kPam (228)
Cz(δ) =
229 kPa2 forδ = 0127 kPa2 expminus(δ373 m)2 forδ gt 0
The exponential model is superimposed on the moment estimates ofFigure 219
The data presented in this case suggest that a sound approach to estimatingautocovariance should involve both the method of moments and maximumlikelihood The method of moments gives a plot of autocovariance versusseparation providing an important graphical summary of the data whichcan be used as a means of determining if the data suggest correlation and forselecting an autocovariance model This provides valuable information forthe maximum likelihood method which then can be used to obtain estimatesof both autocovariance parameters and trend coefficients
234 Bayesian estimation
Bayesian inference for autocorrelation has not been widely used in geotechni-cal and geostatistical applications and it is less well developed than momentestimates This is true despite the fact that Bayesian inference yields theprobability associated with the parameters given the data rather than theconfidence in the data given the probabilistic model An intriguing aspect ofBayesian inference of spatial trends and autocovariance functions is that formany of the non-informative prior distributions one might choose to reflectlittle or no prior information about process parameters (eg the Jeffreysprior the Laplace prior truncated parameter spaces) the posterior pdfrsquoscalculated through Bayesian theorem are themselves improper usually inthe sense that they do not converge toward zero at infinity and thus thetotal probability or area under the posterior pdf is infinite
Following Berger (1993) Boger et al (2001) and Kitanidis (1985 1997)the spatial model is typically written as a Multinomial random process
z(x) =ksum
i=1
fi(x)β + ε(x) (229)
in which fi(x) are unknown deterministic functions of the spatial loca-tions x and ε(x) is a zero-mean spatial random function The random
Spatial variability and geotechnical reliability 109
term is spatially correlated with an isotropic autocovariance function Theautocovariance function is assumed to be non-negative and to decreasemonotonically with distance to zero at infinite separation These assumptionsfit most common autocovariance functions in geotechnical applications TheLikelihood of a set of observations z = z1 hellip zn is then
L(βσ |z) = (2πσ 2)minusn2|Rθ |minus12 expminus 1
2σ 2 (z minus Xβ)tRminus1θ (z minus Xβ)
(230)
in which X is the (n times k) matrix defined by Xij=fj(Xi) Rθ is the matrix ofcorrelations among the observations dependent on the parameters and |Rθ |is the determinant of the correlation matrix of the observations
In the usual fashion a prior non-informative distribution on the param-eters (β σ θ ) might be represented as f (β σ θ )infin(σ 2)minusaf (θ ) for variouschoices of the parameter a and of the marginal pdf f (θ ) The obvious choicesmight be a = 1 f (θ ) = 1 a = 1 f (θ ) = 1θ or a = 1 f (θ ) = 1 but eachof these leads to an improper posterior pdf as does the well-known Jeffreysprior A proper informative prior does not share this difficulty but it iscorrespondingly hard to assess from usually subjective opinion Given thisproblem Berger et al (2001) suggest the reference non-informative prior
f (βσθ ) prop 1σ 2
(|W2
θ |minus |W2θ |
(n minus k)
)12
(231)
in which
W2θ = partRθ
partθRminus1
θ I minus X(XprimeRminus1θ X)minus1X
primeRminus1
θ (232)
This does lead to a proper posterior The posterior pdf is usually evaluatednumerically although depending on the choice of autocovariance functionmodel and the extent to which certain of the parameters of that model areknown closed-form solutions can be obtained Berger et al (2001) presenta numerically calculated example using terrain data from Davis (1986)
235 Variograms
In mining the importance of autocorrelation for estimating ore reserveshas been recognized for many years In mining geostatistics a functionrelated to the autocovariance called the variogram (Matheron 1971) iscommonly used to express the spatial structure of data The variogramrequires a less-restrictive statistical assumption on stationarity than doesthe autocovariance function and it is therefore sometimes preferred for
110 Gregory B Baecher and John T Christian
inference problems On the other hand the variogram is more difficult touse in spatial interpolation and engineering analysis and thus for geotech-nical purposes the autocovariance is used more commonly In practice thetwo ways of characterizing spatial structure are closely related
Whereas the autocovariance is the expected value of the product oftwo observations the variogram 2γ is the expected value of the squareddifference
2γ = E[z(xi) minus z(xj)2] = Var[z(xi) minus z(xj)] (233)
which is a function of only the increments of the spatial properties nottheir absolute values Cressie (1991) points out that in fact the commondefinition of the variogram as the mean squared difference ndash rather thanas the variance of the difference ndash limits applicability to a more restric-tive class of processes than necessary and thus the latter definition is tobe preferred None the less one finds the former definition more commonlyreferred to in the literature The term γ is referred to as the semivariogramalthough caution must be exercised because different authors interchangethe terms The concept of average mean-square difference has been usedin many applications including turbulence (Kolmogorov 1941) and timeseries analysis (Jowett 1952) and is alluded to in the work of Mateacutern(1960)
The principal advantage of the variogram over the autocovariance is thatit makes less restrictive assumptions on the stationarity of the spatial prop-erties being sampled specifically only that their increment and not theirmean is stationary Furthermore the use of geostatistical techniques hasexpanded broadly so that a great deal of experience has been accumulatedwith variogram analysis not only in mining applications but also in environ-mental monitoring hydrology and even geotechnical engineering (Chiassonet al 1995 Soulie and Favre 1983 Soulie et al 1990)
For spatial variables with stationary means and autocovariances (iesecond-order stationary processes) the variogram and autocovariance func-tion are directly related by
γ (δ) = Cz(0) minus Cz(δ) (234)
Common analytical forms for one-dimensional variograms are given inTable 23
For a stationary process as |δ rarr infinCz(δ) rarr 0 thus γ (δ) rarr Cz(0) =Var(z(x)) This value at which the variogram levels off 2Cz(δ) is called thesill value The distance at which the variogram approaches the sill is calledthe range The sampling properties of the variogram are summarized byCressie (1991)
Spatial variability and geotechnical reliability 111
Table 23 One-dimensional variogram models
Model Equation Limits of validity
Nugget g(δ) =
0 if δ = 01 otherwise Rn
Linear g(δ) =
0 if δ = 0c0 + b||δ|| otherwise R1
Spherical g(δ) =
(15)(δa) minus (12)(δa)3 if δ = 01 otherwise
Rn
Exponential g(δ) = 1 minus exp(minus3δa) R1
Gaussian g(δ) = 1 minus exp(minus3δ2a2) Rn
Power g(δ) = hω Rn 0lt δ lt2
24 Random fields
The application of random field theory to spatial variation is based on theassumption that the property of concern z(x) is the realization of a randomprocess When this process is defined over the space x isin S the variable z(x)is said to be a stochastic process In this chapter when S has dimensiongreater than one z(x) is said to be a random field This usage is more or lessconsistent across civil engineering although the geostatistics literature usesa vocabulary all of its own to which the geotechnical literature occasionallyrefers
A random field is defined as the joint probability distribution
Fx1middotmiddotmiddot xn
(z1 zn
) = Pz(x1)le z1 z
(xn
)le zn
(235)
This joint probability distribution describes the simultaneous variation of thevariables z within a space Sx Let E[z(x)] = micro(x) be the mean or trend of z(x)and let Var[z(x)] = σ 2(x) be the variance The covariances of z(x1) hellip z(xn)are defined as
Cov[z(xi)z(xj)] = E[(z(xi) minusmicro(xi)) middot (z(xj) minusmicro(xj))] (236)
A random field is said to be second-order stantionary (weak or wide-sensestationary) if E[z(x)]= micro for all x and Cov[z(xi)z(xj)] depends only onvector separation of xi and xj and not on location Cov[z(xi)z(xj)] =Cz(xi minus xj) in which Cz(xindashxj) is the autocovariance function The randomfield is said to be stationary (strong or strict stationarity) if the complete prob-ability distribution Fx1xn
(z1 zn) is independent of absolute location
112 Gregory B Baecher and John T Christian
depending only on vector separations among the xihellipxn Strong stationarityimplies second-order stationarity In the geotechnical literature stationar-ity is sometimes referred to as statistical homogeneity If the autocovariancefunction depends only on the absolute separation distance and not directionthe random field is said to be isotropic
Ergodicity is a concept originating in the study of time series in whichone observes individual time series of data and wishes to infer proper-ties of an ensemble of all possible time series The meaning and practicalimportance of ergodicity to the inference of unique realizations of spatialrandom fields is less well-defined and is debated in the literature Simplyergodicity means that the probabilistic properties of a random process(field) can be completely estimated from observing one realization of thatprocess For example the stochastic time series z(t) = ν + ε(t) in which ν
is discrete random variable and ε(t) is an autocorrelated random processof time is non-ergotic In one realization of the process there is but onevalue of ν and thus the probability distribution of ν cannot be estimatedOne would need to observe many realizations of zt in order to have suf-ficiently many observations of ν to estimate Fν(ν) Another non-ergoticprocess of more relevance to the spatial processes in geotechnical engi-neering is z(x) = ν + ε(x) in which the mean of z(x) varies linearly withlocation x In this case z(x) is non-stationary Var[z(x)] increases with-out limit as the window within which z(x) is observed increases andthe mean mz of the sample of z(x) is a function of the location of thewindow
The meaning of ergodicity for spatial fields of the sort encountered ingeotechnical engineering is less clear and has not been widely discussed inthe literature An assumption weaker than full erogodicity which none theless should apply for spatial fields is that the observed sample mean mz andsample autocovariance function Cz(δ) converge in mean-squared error to therespective random field mean and autocovariance function as the volumeof space within which they are sampled increases This means that as thevolume of space increases E[(mz minusmicro)2]rarr0 and E[(Cz(δ) minus Cz(δ)2] rarr 0Soong and Grigoriu (1993) provide conditions for checking ergodicity inmean and autocorrelation
When the joint probability distribution Fx1xn(z1 zn) is multivari-
ate Normal (Gaussian) the process z(x) is said to be a Gaussian randomfield A sufficient condition for ergodicity of a Gaussian random field isthat lim|δ|rarrinfinCz(δ) = 0 This can be checked empirically by inspecting thesample moments of the autocovariance function to ensure they convergeto 0 Cressie (1991) notes limitations of this procedure Essentially all theanalytical autocovariance functions common in the geotechnical literatureobey this condition and few practitioners appear concerned about verifyingergodicity Christakos (1992) suggests that in practical situations it isdifficult or impossible to verify ergodicity for spatial fields
Spatial variability and geotechnical reliability 113
A random field that does not meet the conditions of stationarity is saidto be non-stationary Loosely speaking a non-stationary field is statisticallyheterogeneous It can be heterogeneous in a number of ways In the sim-plest case the mean may be a function of location for example if there is aspatial trend that has not been removed In a more complex case the vari-ance or autocovariance function may vary in space Depending on the wayin which the random field is non-stationary sometimes a transformation ofvariables can convert a non-stationary field to a stationary or nearly station-ary field For example if the mean varies with location perhaps a trend canbe removed
In the field of geostatistics a weaker assumption is made on stationaritythan that described above Geostatisticians usually assume only that incre-ments of a spatial process are stationary (ie differences |z1 minus z2|) and thenoperate on the probabilistic properties of those increments This leads to theuse of the variogram rather than the autocovariance function Stationarityof the autocovariance function implies stationarity of the variogram but thereverse is not true
Like most things in the natural sciences stationarity is an assumption ofthe model and may only be approximately true in the world Also station-arity usually depends on scale Within a small region soil properties maybehave as if drawn from a stationary process whereas the same propertiesover a larger region may not be so well behaved
241 Permissible autocovariance functions
By definition the autocovariance function is symmetric meaning
Cz(δ) = Cz(minusδ) (237)
and bounded meaning
Cz(δ) le Cz(0) = σ 2z (238)
In the limit as distance becomes large
lim |δ|rarrinfinCz(δ)
|δ|minus(nminus1)2 = 0 (239)
In general in order for Cz(δ) to be a permissible autocovariance functionit is necessary and sufficient that a continuous mathematical expression ofthe form
msumi=1
msumj=1
kikjCz(δ) ge 0 (240)
114 Gregory B Baecher and John T Christian
be non-negative-definite for all integers m scalar coefficients k1hellipkmand δ This condition follows from the requirement that variances of linearcombinations of the z(xi) of the form
var
[msum
i=1
kiz(xi)
]=
msumi=1
msumj=1
kikjCz(δ) ge 0 (241)
be non-negative ie the matrix is positive definite (Cressie 1991) Christakos(1992) discusses the mathematical implications of this condition on selectingpermissible forms for the autocovariance Suffice it to say that analyticalmodels of autocovariance common in the geotechnical literature usuallysatisfy the condition
Autocovariance functions valid in a space of dimension d are validin spaces of lower dimension but the reverse is not necessarily trueThat is a valid autocovariance function in 1D is not necessarily validin 2D or 3D Christakos (1992) gives the example of the linearly decliningautcovariance
Cz(δ) =σ 2(1 minus δδ0) for 0 le δ le δ0
0 for δ gt δ0(242)
which is valid in 1D but not in higher dimensionsLinear sums of valid autocovariance functions are also valid This means
that if Cz1(δ) and Cz2(δ) are valid then the sum Cz1(δ)+Cz2(δ) is also a validautocovariance function Similarly if Cz(δ) is valid then the product with ascalar αCz(δ) is also valid
An autocovariance function in d-dimensional space is separable if
Cz(δ) =dprod
i=1
Czi(δi) (243)
in which δ is the d-dimensioned vector of orthogonal separation distancesδ1hellipδd and Ci(δi) is the one-dimensional autocovariance function indirection i For example the autocovariance function
Cz(δ) = σ 2 expminusa2|δ|2= σ 2 expminusa2(δ1
2 +middotmiddot middot+ δ2d) (244)
= σ 2dprod
i=1
expminusa2δ2i
is separable into its one-dimensional components
Spatial variability and geotechnical reliability 115
The function is partially separable if
Cz(δ) = Cz(δi)Cz(δj =i) (245)
in which Cz(δj =i) is a (d minus 1) dimension autocovariance function implyingthat the function can be expressed as a product of autocovariance func-tions of lower dimension fields The importance of partial separability togeotechnical applications as noted by VanMarcke (1983) is the 3D case ofseparating autocorrelation in the horizontal plane from that with depth
Cz(δ1δ2δ3) = Cz(δ1δ2)Cz(δ3) (246)
in which δ1 δ2 are horizontal distances and δ3 is depth
242 Gaussian random fields
The Gaussian random field is an important special case because it is widelyapplicable due to the Central Limit Theorem has mathematically conve-nient properties and is widely used in practice The probability densitydistribution of the Gaussian or Normal variable is
fz(z) = minus 1radic2πσ
exp
minus1
2
(x minusmicro
σ
)2
(247)
for minusinfin le z le infin The mean is E[z] = micro and variance Var[z] = σ 2 For themultivariate case of vector z of dimension n the correponding pdf is
fz(z) = (2π )minusn2||minus12 expminus1
2(z minusmicro)primeminus1(z minusmicro)
(248)
in which micro is the mean vector and the covariance matrix
ij =Cov[zi(x)zj(x)
](249)
Gaussian random fields have the following convenient properties (Adler1981) (1) they are completely characterized by the first- and second-ordermoments the mean and autocovarinace function for the univariate caseand mean vector and autocovariance matrix (function) for the multivari-ate case (2) any subset of variables of the vector is also jointly Gaussian(3) the conditional probability distributions of any two variables or vectorsare also Gaussian distributed (4) if two variables z1 and z2 are bivariateGaussian and if their covariance Cov[z1 z2] is zero then the variables areindependent
116 Gregory B Baecher and John T Christian
243 Interpolating random fields
A problem common in site characterization is interpolating among spa-tial observations to estimate soil or rock properties at specific locationswhere they have not been observed The sample observations themselvesmay have been taken under any number of sampling plans random sys-tematic cluster and so forth What differentiates this spatial estimationquestion from the sampling theory estimates in preceding sections of thischapter is that the observations display spatial correlation Thus the assump-tion of IID observations underlying the estimator results is violated in animportant way This question of spatial interpolation is also a problemcommon to the natural resources industries such as forestry (Mateacutern 1986)and mining (Matheron 1971) but also geohydrology (Kitanidis 1997) andenvironmental monitoring (Switzer 1995)
Consider the case for which the observations are sampled from aspatial population with constant mean micro and autocovariance functionCz(δ) = E[z(xi)z(xi+δ)] The set of observations z=zihellip zn thereforehas mean vector m in which all the terms are equal and covariancematrix
=⎡⎢⎣
Var(z1) middot middot middot Cov(z1zn)
Cov(znz1) middot middot middot Var(zn)
⎤⎥⎦ (250)
in which the terms z(xi) are replaced by zi for convenience These termsare found from the autocovariance function as Cov(z(xi)z(xj)) = Cz (δij) inwhich δij is the (vector) separation between locations xi and xj
In principle we would like to estimate the full distribution of z(x0) atan unobserved location x0 but in general this is computationally intensiveif a large grid of points is to be interpolated Instead the most commonapproach is to construct a simple linear unbiased estimator based on theobservations
z(x0) =nsum
i=1
wiz(xi) (251)
in which the weights w=w1hellip wn are scalar values chosen to makethe estimate in some way optimal Usually the criteria of optimality areunbiasedness and minimum variance and the result is sometimes called thebest linear unbiased estimator (BLUE)
The BLUE estimator weights are found by expressing the variance of theestimate z(x0) using a first-order second-moment formulation and minimiz-ing the variance over w using a Lagrange multiplier approach subject to the
Spatial variability and geotechnical reliability 117
condition that the sum of the weights equals one The solution in matrixform is
w = Gminus1h (252)
in which w is the vector of optimal weights and the matrices G and h relatethe covariance matrix of the observations and the vector of covariances of theobservations to the value of the spatial variable at the interpolated locationx0 respectively
G =
⎡⎢⎢⎢⎣
Var(z1) middot middot middot Cov(z1zn) 1
1
Cov(znz1) middot middot middot Var(zn) 11 1 1 0
⎤⎥⎥⎥⎦
h =
⎡⎢⎢⎢⎣
Cov(z1z0)
Cov(znz0)1
⎤⎥⎥⎥⎦
(253)
The resulting estimator variance is
Var(z0) = E[(z0 minus z0)2]
= Var(z0) minusnsum
i=1
wiCov(z0zi) minusλ (254)
in which λ is the Lagrange multiplier resulting from the optimization Thisis a surprisingly simple and convenient result and forms the basis of theincreasingly vast literature on the subject of so-called kriging in the field ofgeostatistics For regular grids of observations such as a grid of borings analgorithm can be established for the points within an individual grid celland then replicated for all cells to form an interpolated map of the largersite or region (Journel and Huijbregts 1978) In the mining industry andincreasingly in other applications it has become common to replace the auto-covariance function as a measure of spatial association with the variogram
244 Functions of random fields
Thus far we have considered the properties of random fields themselvesIn this section we consider the extension to properties of functions of ran-dom fields Spatial averaging of random fields is among the most importantconsiderations for geotechnical engineering Limiting equilibrium stabil-ity of slopes depends on the average strength across the failure surface
118 Gregory B Baecher and John T Christian
Settlements beneath foundations depend on the average compressibility ofthe subsurface soils Indeed many modes of geotechnical performance ofinterest to the engineer involve spatial averages ndash or differences amongspatial averages ndash of soil and rock properties Spatial averages also playa significant role in mining geostatistics where average ore grades withinblocks of rock have important implications for planning As a result thereis a rich literature on the subject of averages of random fields only a smallpart of which can be reviewed here
Consider the one-dimensional case of a continuous scalar stochastic pro-cess (1D random field) z(x) in which x is location and z(x) is a stochasticvariable with mean microz assumed to be constant and autocovariance functionCz(r) in which r is separation distance r = (x1 minusx2) The spatial average ormean of the process within the interval [0X] is
MXz(x) = 1X
int X
0z(x)dx (255)
The integral is defined in the common way as a limiting sum of z(x) valueswithin infinitesimal intervals of x as the number of intervals increasesWe assume that z(x) converges in a mean square sense implying the existenceof the first two moments of z(x) The weaker assumption of convergence inprobability which does not imply existence of the moments could be madeif necessary (see Parzen 1964 1992) for more detailed discussion)
If we think of MXz(x) as a sample observation within one interval of theprocess z(x) then over the set of possible intervals that we might observeMXz(x) becomes a random variable with mean variance and possiblyother moments Consider first the integral of z(x) within intervals of length XParzen (1964) shows that the first two moments of
int X0 z(x)dx are
E
[int X
0z(x)dx
]=int X
0micro(x)dx = microX (256)
Var
[int X
0z(x)dx
]=int X
0
int X
0Cz(xi minus xj)dxidxj = 2
int X
0(X minus r)Cz(r)dr
(257)
and that the autocovariance function of the integralint X
0 z(x)dx as the interval[0 X] is allowed to translate along dimension x is (VanMarcke 1983)
Cint X0 z(x)dx(r) = Cov
[int X
0z(x)dx
int r+X
rz(x)dx
]
=int X
0
int X
0Cz(r + xi minus xj)dxidxj (258)
Spatial variability and geotechnical reliability 119
The corresponding moments of the spatial mean MXz(x) are
E[MXz(x)]= E
[1X
int X
0z(x)dx
]=int X
0
1Xmicro(x)dx = micro (259)
Var[MXz(x)]= Var
[1X
int X
0z(x)dx
]= 2
X2
int X
0(X minus r)Cz(r)dr (260)
CMXz(x)(r) = Cov
[1X
int X
0z(x)dx
1X
int r+X
rz(x)dx
]
= 1X2
int X
0
int X
0Cz(r + xi minus xj)dxidxj (261)
The effect of spatial averaging is to smooth the process The varianceof the averaged process is smaller than that of the original process z(x)and the autocorrelation of the averaged process is wider Indeed averaging issometimes referred to as smoothing (Gelb and Analytic Sciences CorporationTechnical Staff 1974)
The reduction in variance from z(x) to the averaged process MXz(x) canbe represented in a variance reduction function γ (X)
γ (X) = Var[MXz(x)]
var[z(x)] (262)
The variance reduction function is 10 for X = 0 and decays to zero as Xbecomes large γ (X) can be calculated from the autocovariance function ofz(x) as
γ (X) = 2x
xint0
(1 minus r
X
)R2(r)dr (263)
in which Rz(r) is the autocorrelation function of z(x) Note that the squareroot of γ (X) gives the corresponding reduction of the standard deviationof z(x) Table 24 gives one-dimensional variance reduction functions forcommon autocovariance functions It is interesting to note that each of thesefunctions is asymptotically proportional to 1X Based on this observationVanMarcke (1983) proposed a scale of fluctuation θ such that
θ = limXrarrinfin
X γ (X) (264)
or γ (X) = θ X as X rarr infin that is θ X is the asymptote of the variancereduction function as the averaging window expands The function γ (X)
Tabl
e2
4V
aria
nce
redu
ctio
nfu
nctio
nsfo
rco
mm
on1D
auto
cova
rian
ces
(aft
erV
anM
arck
e19
83)
Mod
elAu
toco
rrel
atio
nVa
rianc
ere
duct
ion
func
tion
Scal
eof
fluct
uatio
n
Whi
teno
ise
R x(δ
)= 1
ifδ=
00
othe
rwise
γ(X
)= 1
ifX
=0
0ot
herw
ise0
Line
arR x
(δ)= 1
minus|δ|
δ n
ifδ
leδ 0
0ot
herw
iseγ
(X)= 1
minusX3δ
0if
Xle
δ 0(δ
0X
)[ 1minusδ 03X] ot
herw
iseδ 0
Expo
nent
ial
R x(δ
)=ex
p(minusδ
δ 0
)γ
(X)=
2(δ 0X
)2( X δ 0
minus1+
exp2
(minusXδ 0
))4δ
0
Squa
red
expo
nent
ial
(Gau
ssia
n)R x
(δ)=
exp2
(minus|δ|
δ 0
)γ
(X)=
(δ0
X)2[ radic π
X δ 0
(minusXδ 0
)+ex
p2(minus
Xδ 0
)minus1]
inw
hich
is
the
erro
rfu
nctio
n
radic πδ 0
Spatial variability and geotechnical reliability 121
converges rapidly to this asymptote as X increases For θ to exist it isnecessary that Rz(r) rarr0 as r rarr infin that is that the autocorrelation func-tion decreases faster than 1r In this case θ can be found from the integralof the autocorrelation function (the moment of Rz(r) about the origin)
θ = 2int infin
0Rz(r)dr =
int infin
minusinfinRz(r)dr (265)
This concept of summarizing the spatial or temporal scale of autocorrelationin a single number typically the first moment of Rz(r) is used by a variety ofother workers and in many fields Taylor (1921) in hydrodynamics calledit the diffusion constant (Papoulis and Pillai 2002) Christakos (1992) ingeoscience calls θ 2 the correlation radius Gelhar (1993) in groundwaterhydrology calls θ the integral scale
In two dimensions the equivalent expressions for the mean and varianceof the planar integral
int X0
int X0 z(x)dx are
E
[int X
0z(x)dx
]=int X
0micro(x)dx = microX (266)
Var
[int X
0z(x)dx
]=int X
0
int X
0Cz(xi minus xj)dxidxj = 2
int X
0(X minus r)Cz(r)dr
(267)
Papoulis and Pillai (2002) discuss averaging in higher dimensions as doElishakoff (1999) and VanMarcke (1983)
245 Stochastic differentiation
The continuity and differentiability of a random field depend on the con-vergence of sequences of random variables z(xa)z(xb) in which xa xb aretwo locations with (vector) separation r = |xa minusxb| The random field is saidto be continuous in mean square at xa if for every sequence z(xa)z(xb)E2[z(xa)minusz(xb)] rarr0 as rrarr0 The random field is said to be continuousin mean square throughout if it is continuous in mean square at every xaGiven this condition the random field z(x) is mean square differentiablewith partial derivative
partz(x)partxi
= lim|r|rarr0
z(x + rδi) minus z(x)r
(268)
in which the delta function is a vector of all zeros except the ith termwhich is unity While stronger or at least different convergence propertiescould be invoked mean square convergence is often the most natural form in
122 Gregory B Baecher and John T Christian
practice because we usually wish to use a second-moment representation ofthe autocovariance function as the vehicle for determining differentiability
A random field is mean square continuous if and only if its autocovariancefunction Cz(r) is continuous at |r| = 0 For this to be true the first derivativesof the autocovariance function at |r|=0 must vanish
partCz(r)partxi
= 0 for all i (269)
If the second derivative of the autocovariance function exists and is finite at|r| = 0 then the field is mean square differentiable and the autocovariancefunction of the derivative field is
Cpartzpartxi(r) = part2Cz(r)partx2
i (270)
The variance of the derivative field can then be found by evaluating the auto-covariance Cpartzpartxi
(r) at |r| = 0 Similarly the autocovariance of the secondderivative field is
Cpart2zpartxipartxj(r) = part4Cz(r)partx2
i partx2j (271)
The cross covariance function of the derivatives with respect to xi and xj inseparate directions is
Cpartzpartxipartzpartxj(r) = minuspart2Cz(r)partxipartxj (272)
Importantly for the case of homogeneous random fields the field itself z(x)and its derivative field are uncorrelated (VanMarcke 1983)
So the behavior of the autocovariance function in the neighborhood ofthe origin is the determining factor for mean-square local properties of thefield such as continuity and differentiability (Crameacuter and Leadbetter 1967)Unfortunately the properties of the derivative fields are sensitive to thisbehavior of Cz(r) near the origin which in turn is sensitive to the choice ofautocovariance model Empirical verification of the behavior of Cz(r) nearthe origin is exceptionally difficult Soong and Grigoriu (1993) discuss themean square calculus of stochastic processes
246 Linear functions of random fields
Assume that the random field z(x) is transformed by a deterministicfunction g() such that
y(x) = g[z(x)] (273)
In this equation g[z(x0)] is a function of z alone that is not of x0 and not ofthe value of z(x) at any x other than x0 Also we assume that the transforma-tion does not depend on the value of x that is the transformation is space- or
Spatial variability and geotechnical reliability 123
time-invariant y(z + δ) = g[z(x + δ)] Thus the random variable y(x) is adeterministic transformation of the random variable z(x) and its probabilitydistribution can be obtained from derived distribution methods Similarlythe joint distribution of the sequence of random variables y(x1) hellip y(xn)can be determined from the joint distribution of the sequence of randomvariables x(x1) hellip x(xn) The mean of y(x) is then
E[y(x)] =infinint
minusinfing(z)fz(z(x))dz (274)
and the autocorrelation function is
Ry(y1y2) = E[y(x1)y(x2)] =infinint
minusinfin
infinintminusinfin
g(z1)g(z2)fz(z(x1)z(x2))dz1dz2
(275)
Papoulis and Pillai(2002) show that the process y(x) is (strictly) stationary ifz(x) is (strictly) stationary Phoon (2006b) discusses limitations and practicalmethods of solving this equation Among the limitations is that such non-Gaussian fields may not have positive definite covariance matrices
The solutions for nonlinear transformations are difficult but for linearfunctions general results are available The mean of y(x) for linear g(z) isfound by transforming the expected value of z(x) through the function
E[y(x)] = g(E[z(x)]) (276)
The autocorrelation of y(x) is found in a two-step process
Ryy(x1x2) = Lx1[Lx2
[Rzz(x1x2)]] (277)
in which Lx1is the transformation applied with respect to the first variable
z(x1) with the second variable treated as a parameter and Lx2is the trans-
formation applied with respect to the second variable z(x2) with the firstvariable treated as a parameter
247 Excursions (level crossings)
A number of applications arise in geotechnical practice for which one isinterested not in the integrals (averages) or differentials of a stochastic pro-cess but in the probability that the process exceeds some threshold eitherpositive or negative For example we might be interested in the probabilitythat a stochastically varying water inflow into a reservoir exceeds some rate
124 Gregory B Baecher and John T Christian
or in the properties of the weakest interval or seam in a spatially varyingsoil mass Such problems are said to involve excursions or level crossings ofa stochastic process The following discussion follows the work of Crameacuter(1967) Parzen (1964) and Papoulis and Pillai (2002)
To begin consider the zero-crossings of a random process the points xi atwhich z(xi) = 0 For the general case this turns out to be a surprisingly diffi-cult problem Yet for the continuous Normal case a number of statementsor approximations are possible Consider a process z(x) with zero mean andvariance σ 2 For the interval [xx+ δ] if the product
z(x)z(x + δ) lt 0 (278)
then there must be an odd number of zero-crossings within the interval forif this product is negative one of the values must lie above zero and the otherbeneath Papoulis and Pillai(2002) demonstrate that if the two (zero-mean)variables z(x) and z(x + δ) are jointly normal with correlation coefficient
r = E [z(x)z(x + δ)]σxσx+δ
(279)
then
p(z(x)z(x + δ) lt 0) = 12
minus arcsin(r)π
= arccos(r)π
p(z(x)z(x + δ) gt 0) = 12
+ arcsin(r)π
= π minus arccos(r)π
(280)
The correlation coefficient of course can be taken from the autocorrelationfunction Rz(δ) Thus
cos[πp(z(x)z(x + δ) lt 0)] = Rz(δ)
Rz(0)(281)
and the probability that the number of zero-crossings is positive is just thecomplement of this result
The probability of exactly one zero-crossing p1(δ) is approximatelyp1(δ) asymp p0(δ) and expanding the cosine in a Fourier series and truncating totwo terms
1 minus π2p21(δ)
2= Rz(δ)
Rz(0)(282)
or
p1(δ) asymp 1π
radic2[Rz(0) minus Rz(δ)]
Rz(0)(283)
Spatial variability and geotechnical reliability 125
In the case of a regular autocorrelation function for which the derivativedRz(0)dδ exists and is zero at the origin the probability of a zero-crossingis approximately
p1(δ) asymp δ
π
radicminusd2Rz(0)dδ2
Rz(0)(284)
The non-regular case for which the derivative at the origin is not zero (egdRz(δ) = exp(δδ0)) is discussed by Parzen (1964) Elishakoff (1999) andVanMarcke (1983) treat higher dimensional results The related probabilityof the process crossing an arbitrary level zlowast can be approximated by notingthat for small δ and thus r rarr1
P[z(x) minus zlowastz(x + δ) minus zlowast lt 0
]asymp P [z(x)z(x + δ) lt 0]eminusarcsin2(r)
2σ2
(285)
For small δ the correlation coefficient Rz(δ) is approximately 1 and thevariances of z(x) and z(x + δ) are approximately Rz(0) thus
p1zlowast (δ) asymp p1zlowast (δ) lt 0]e minusarcsin2(r)2Rz (0) (286)
and for the regular case
p1(δ) asymp δ
π
radicminusd2Rz(0)dδ2
Rz(0)e
minusarcsin2(r)2Rz (0) (287)
Many other results can be found for continuous Normal processes eg theaverage density of the number of crossings within an interval the probabilityof no crossings (ie drought) within an interval and so on A rich literature isavailable of these and related results (Yaglom 1962 Parzen 1964 Crameacuterand Leadbetter 1967 Gelb and Analytic Sciences Corporation TechnicalStaff 1974 Adler 1981 Cliff and Ord 1981 Cressie 1991 Christakos1992 2000 Christakos and Hristopulos 1998)
248 Example New Orleans hurricane protection systemLouisiana (USA)
In the aftermath of Hurricane Katrina reliability analyses were con-ducted on the reconstructed New Orleans hurricane protection system(HPS) to understand the risks faced in future storms A first-excursionor level crossing methodology was used to calculate the probability offailure in long embankment sections following the approach proposed
126 Gregory B Baecher and John T Christian
by VanMarcke (1977) This resulted in fragility curves for a reach oflevee The fragility curve gives the conditional probability of failure forknown hurricane loads (ie surge and wave heights) Uncertainties in thehurricane loads were convolved with these fragility curves in a systems riskmodel to generate unconditional probabilities and subsequently risk whenconsequences were included
As a first approximation engineering performance models and calcu-lations were adapted from the US Army Corps of Engineersrsquo DesignMemoranda describing the original design of individual levee reaches(USACE 1972) Engineering parameter and model uncertainties were prop-agated through those calculations to obtain approximate fragility curvesas a function of surge and wave loads These results were later calibratedagainst analyses which applied more sophisticated stability models and therisk assessments were updated
A typical design profile of the levee system is shown in Figure 224Four categories of uncertainty were included in the reliability analysisgeological and geotechnical uncertainties involving the spatial distribu-tion of soils and soil properties within and beneath the HPS geotechnicalstability modeling of levee performance erosion uncertainties involvingthe performance of levees and fills during overtopping and mechanicalequipment uncertainties including gates pumps and other operating sys-tems and human operator factors affecting the performance of mechanicalequipment
The principal uncertainty contributing to probability of failure of the leveesections in the reliability analysis was soil engineering properties specificallyundrained strength Su measured in Q-tests (UU tests) Uncertainties in soilengineering properties was presumed to be structured as in Figure 216
Figure 224 Typical design section from the USACE Design Memoranda for theNew Orleans Hurricane Protection System (USACE 1972)
Spatial variability and geotechnical reliability 127
and the variance of the uncertainty in soil properties was divided into fourterms
Var(Su) = Var(x) + Var(e) + Var(m) + Var(b) (288)
in which Var() is variance Su is measured undrained strength x is the soilproperty in situ e is measurement error (noise) m is the spatial mean (whichhas some error due to the statistical fluctuations of small sample sizes) andb is a model bias or calibration term caused by systematic errors in mea-suring the soil properties Measured undrained strength for one reach theNew Orleans East lakefront levees are shown as histograms in Figure 225Test values larger than 750 PCF (36 kPa) were assumed to be local effectsand removed from the statistics The spatial pattern of soil variability wascharacterized by autocovariance functions in each region of the system andfor each soil stratum (Figure 226) From the autocovariance analyses twoconclusions were drawn The measurement noise (or fine-scale variation)in the undrained strength data was estimated to be roughly 34 the totalvariance of the data (which was judged not unreasonable given the Q-testmethods) and the autocovariance distance in the horizontal direction forboth the clay and marsh was estimated to be on the order of 500 feetor more
The reliability analysis was based on limiting equilibrium calculationsFor levees the analysis was based on General Design Memorandum (GDM)calculations of factor of safety against wedge instability (USACE 1972)
0
1
2
3
4
5
6
7
8
9
10
0 250 500 750 1000 1250 1500
Undrained Shear Strength Q-test (PCF)
FillMarshDistributary Clay
Figure 225 Histogram of Q-test (UU) undrained soil strengths New Orleans Eastlakefront
128 Gregory B Baecher and John T Christian
C10minus2
29972
19928
9884
minus16
minus10204
minus202480 1025 205 3075 41
Distance h10minus2
5125 615 7175 82
Figure 226 Representative autocovariance function for inter-distributary clay undrainedstrength (Q test) Orleans Parish Louisiana
P(f
|Ele
vatio
n) P
roba
bilit
y of
Fai
lure 10
05
0
Authorization Basis
015 Fractile Level
085 Fractile Level
Median
Elevation (ft)
Range ofthe SystemReliability
Figure 227 Representative fragility curves for unit reach and long reach of levee
using the so-called method of planes The calculations are based onundrained failure conditions Uncertainties in undrained shear strength werepropagated through the calculations to estimate a coefficient of variationin the calculated factor of safety The factor of safety was assumed to beNormally distributed and a fragility curve approximated through threecalculation points (Figure 227)
The larger the failure surface relative to the autocorrelation of the soilproperties the more the variance of the local averages is reduced VanMarcke(1977) has shown that the variance of the spatial average for a unit-widthplain strain cross-section decreases approximately in proportion to (LrL)for Lgt rL in which L is the cross-sectional length of the failure surface and
Spatial variability and geotechnical reliability 129
rL is an equivalent autocovariance distance of the soil properties across thefailure surface weighted for the relative proportion of horizontal and verticalsegments of the surface For the wedge failure modes this is approximatelythe vertical autocovariance distance The variance across the full failure sur-face of width b along the axis of the levee is further reduced by averagingin the horizontal direction by an additional factor (brH) for b gt rH inwhich rH is the horizontal autocovariance distance At the same time thatthe variance of the average strength on the failure surface is reduced bythe averaging process so too the autocovariance function of this averagedprocess stretches out from that of the point-to-point variation
For a failure length of approximately 500 feet along the levee axis and30 feet deep typical of those actually observed with horizontal and ver-tical autocovariance distances of 500 feet and 10 feet respectively thecorresponding variance reduction factors are approximately 075 for aver-aging over the cross-sectional length L and between 073 and 085 foraveraging over the failure length b assuming either an exponential orsquared-exponential (Gaussian) autocovariance The corresponding reduc-tion to the COV of soil strength based on averaging over the failure plane isthe root of the product of these two factors or between 074 and 08
For a long levee the chance of at least one failure is equivalent to the chancethat the variations of the mean soil strength across the failure surface dropbelow that required for stability at least once along the length VanMarckedemonstrated that this can be determined by considering the first crossingsof a random process The approximation to the probability of at least onefailure as provided by VanMarcke was used in the present calculations toobtain probability of failure as a function of levee length
25 Concluding comments
In this chapter we have described the importance of spatial variation ingeotechnical properties and how such variation can be dealt with in aprobabilistic analysis Spatial variation consists essentially of two parts anunderlying trend and a random variation superimposed on it The distri-bution of the variability between the trend and the random variation is adecision made by the analyst and is not an invariant function of nature
The second-moment method is widely used to describe the spatial varia-tion of random variation Although not as commonly used in geotechnicalpractice Bayesian estimation has many advantages over moment-basedestimation One of them is that it yields an estimate of the probabilitiesassociated with the distribution parameters rather than a confidence intervalthat the data would be observed if the process were repeated
Spatially varying properties are generally described by random fieldsAlthough these can become extremely complicated relatively simple modelssuch as Gaussian random field have wide application The last portion of
130 Gregory B Baecher and John T Christian
the chapter demonstrates how they can be manipulated by differentiationby transformation through linear processes and by evaluation of excursionsbeyond specified levels
Notes
1 It is true that not all soil or rock mass behaviors of engineering importance aregoverned by averages for example block failures in rock slopes are governed by theleast favorably positioned and oriented joint They are extreme values processesNevertheless averaging soil or rock properties does reduce variance
2 In early applications of geotechnical reliability a great deal of work was focusedon appropriate distributional forms for soil property variability but this no longerseems a major topic First the number of measurements typical of site characteri-zation programs is usually too few to decide confidently whether one distributionalform provides a better fit than another and second much of practical geotechnicalreliability work uses second-moment characterizations rather than full distribu-tional analysis so distributional assumptions only come into play at the endof the work Second-moment characterizations use only means variances andcovariances to characterize variability not full distributions
3 In addition to aleatory and epistemic uncertainties there are also uncertainties thathave little to do with engineering properties and performance yet which affectdecisions Among these is the set of objectives and attributes considered importantin reaching a decision the value or utility function defined over these attributesand discounting for outcomes distributed in time These are outside the presentscope
References
Adler R J (1981) The Geometry of Random Fields Wiley ChichesterAng A H-S and Tang W H (1975) Probability Concepts in Engineering Planning
and Design Wiley New YorkAzzouz A Baligh M M and Ladd C C (1983) Corrected field vane strength
for embankment design Journal of Geotechnical Engineering ASCE 109(3)730ndash4
Baecher G B(1980) Progressively censored sampling of rock joint traces Journalof the International Association of Mathematical Geologists 12(1) 33ndash40
Baecher G B (1987) Statistical quality control for engineered fills EngineeringGuide US Army Corps of Engineers Waterways Experiment Station GL-87-2
Baecher G B Ladd C C Noiray L and Christian J T (1997) Formal obser-vational approach to staged loading Transportation Research Board AnnualMeeting Washington DC
Barnett V (1982) Comparative Statistical Inference John Wiley and Sons LondonBerger J O (1993) Statistical Decision Theory and Bayesian Analysis Springer
New YorkBerger J O De Oliveira V and Sanso B (2001) Objective Bayesian analysis of
spatially correlated data Journal of the American Statistical Association 96(456)1361ndash74
Bjerrum L (1972) Embankments on soft ground In ASCE Conference on Perfor-mance of Earth and Earth-Supported Structures Purdue pp 1ndash54
Spatial variability and geotechnical reliability 131
Bjerrum L (1973) Problems of soil mechanics and construction on soft clays InEighth International Conference on Soil Mechanics and Foundation EngineeringMoscow Spinger New York pp 111ndash59
Chiasson P Lafleur J Soulie M and Law K T (1995) Characterizing spatialvariability of a clay by geostatistics Canadian Geotechnical Journal 32 1ndash10
Christakos G (1992) Random Field Models in Earth Sciences Academic PressSan Diego
Christakos G (2000) Modern Spatiotemporal Geostatistics Oxford UniversityPress Oxford
Christakos G and Hristopulos D T (1998) Spatiotemporal Environmental HealthModelling A Tractatus Stochasticus Kluwer Academic Boston MA
Christian J T Ladd C C and Baecher G B (1994) Reliability applied toslope stability analysis Journal of Geotechnical Engineering ASCE 120(12)2180ndash207
Cliff A D and Ord J K (1981) Spatial Processes Models amp Applications PionLondon
Crameacuter H and Leadbetter M R (1967) Stationary and Related StochasticProcesses Sample Function Properties and their Applications Wiley New York
Cressie N A C (1991) Statistics for Spatial Data Wiley New YorkDavis J C (1986) Statistics and Data Analysis in Geology Wiley New YorkDeGroot D J and Baecher G B (1993) Estimating autocovariance of in situ soil
properties Journal Geotechnical Engineering Division ASCE 119(1) 147ndash66Elishakoff I (1999) Probabilistic Theory of Structures Dover Publications
Mineola NYFuleihan N F and Ladd C C (1976) Design and performance of Atchafalaya
flood control levees Research Report R76-24 Department of Civil EngineeringMassachusetts Institute of Technology Cambridge MA
Gelb A and Analytic Sciences Corporation Technical Staff (1974) Applied OptimalEstimation MIT Press Cambridge MA
Gelhar L W (1993) Stochastic Subsurface Hydrology Prentice-Hall EnglewoodCliffs NJ
Hartford D N D (1995) How safe is your dam Is it safe enough MEP11-5 BCHydro Burnaby BC
Hartford D N D and Baecher G B (2004) Risk and Uncertainty in Dam SafetyThomas Telford London
Hilldale C (1971) A probabilistic approach to estimating differential settle-ment MS thesis Civil Engineering Massachusetts Institute of TechnologyCambridge MA
Javete D F (1983) A simple statistical approach to differential settlements on clayPh D thesis Civil Engineering University of California Berkeley
Journel A G and Huijbregts C (1978) Mining Geostatistics Academic PressLondon
Jowett G H (1952) The accuracy of systematic simple from conveyer belts AppliedStatistics 1 50ndash9
Kitanidis P K (1985) Parameter uncertainty in estimation of spatial functionsBayesian analysis Water Resources Research 22(4) 499ndash507
Kitanidis P K (1997) Introduction to Geostatistics Applications to HydrogeologyCambridge University Press Cambridge
132 Gregory B Baecher and John T Christian
Kolmogorov A N (1941) The local structure of turbulence in an incompress-ible fluid at very large Reynolds number Doklady Adademii Nauk SSSR 30301ndash5
Lacasse S and Nadim F (1996) Uncertainties in characteristic soil properties InUncertainty in the Geological Environment ASCE specialty conference MadisonWI ASCE Reston VA pp 40ndash75
Ladd CC and Foott R (1974) A new design procedure for stability of soft claysJournal of Geotechnical Engineering 100(7) 763ndash86
Ladd C C (1975) Foundation design of embankments constructed on con-necticut valley varved clays Report R75-7 Department of Civil EngineeringMassachusetts Institute of Technology Cambridge MA
Ladd C C Dascal O Law K T Lefebrve G Lessard G Mesri G andTavenas F (1983) Report of the subcommittee on embankment stability__annexe II Committe of specialists on Sensitive Clays on the NBR Complex SocieteacutedrsquoEnergie de la Baie James Montreal
Lambe TW and Associates (1982) Earthquake risk to patio 4 and site 400 Reportto Tonen Oil Corporation Longboat Key FL
Lambe T W and Whitman R V (1979) Soil Mechanics SI Version WileyNew York
Lee I K White W and Ingles O G (1983) Geotechnical Engineering PitmanBoston
Lumb P (1966) The variability of natural soils Canadian Geotechnical Journal3 74ndash97
Lumb P (1974) Application of statistics in soil mechanics In Soil Mechanics NewHorizons Ed I K Lee Newnes-Butterworth London pp 44ndash112
Mardia K V and Marshall R J (1984) Maximum likelihood estimation of modelsfor residual covariance in spatial regression Biometrika 71(1) 135ndash46
Mateacutern B (1960) Spatial Variation 49(5) Meddelanden fran Statens Skogsforskn-ingsinstitut
Mateacutern B (1986) Spatial Variation Springer BerlinMatheron G (1971) The Theory of Regionalized Variables and Its Application ndash
Spatial Variabilities of Soil and Landforms Les Cahiers du Centre de Morphologiemathematique 8 Fontainebleau
Papoulis A and Pillai S U (2002) Probability Random Variables and StochasticProcesses 4th ed McGraw-Hill Boston MA
Parzen E (1964) Stochastic Processes Holden-Day San Francisco CAParzen E (1992) Modern Probability Theory and its Applications Wiley
New YorkPhoon K K (2006a) Bootstrap estimation of sample autocorrelation functions
In GeoCongress Atlanta ASCEPhoon K K (2006b) Modeling and simulation of stochastic data In GeoCongress
Atlanta ASCEPhoon K K and Fenton G A (2004) Estimating sample autocorrelation func-
tions using bootstrap In Proceedings Ninth ASCE Specialty Conference onProbabilistic Mechanics and Structural Reliability Albuquerque
Phoon K K and Kulhawy F H (1996) On quantifying inherent soil variabilityUncertainty in the Geologic Environment ASCE specialty conference MadisonWI ASCE Reston VA pp 326ndash40
Spatial variability and geotechnical reliability 133
Soong T T and Grigoriu M (1993) Random Vibration of Mechanical andStructural Systems Prentice Hall Englewood Cliffs NJ
Soulie M and Favre M (1983) Analyse geostatistique drsquoun noyau de barrage telque construit Canadian Geotechnical Journal 20 453ndash67
Soulie M Montes P and Silvestri V (1990) Modeling spatial variability of soilparameters Canadian Geotechnical Journal 27 617ndash30
Switzer P (1995) Spatial interpolation errors for monitoring data Journal of theAmerican Statistical Association 90(431) 853ndash61
Tang W H (1979) Probabilistic evaluation of penetration resistance Journal ofthe Geotechnical Engineering Division ASCE 105(10) 1173ndash91
Taylor G I (1921) Diffusion by continuous movements Proceeding of the LondonMathematical Society (2) 20 196ndash211
USACE (1972) New Orleans East Lakefront Levee Paris Road to South Point LakePontchartrain Barrier Plan DM 2 Supplement 5B USACE New Orleans DistrictNew Orleans
VanMarcke E (1983) Random Fields Analysis and Synthesis MIT PressCambridge MA
VanMarcke E H (1977) Reliability of earth slopes Journal of the GeotechnicalEngineering Division ASCE 103(GT11) 1247ndash65
Weinstock H (1963) The description of stationary random rate processes E-1377Massachusetts Institute of Technology Cambridge MA
Wu T H (1974) Uncertainty safety and decision in soil engineering JournalGeotechnical Engineering Division ASCE 100(3) 329ndash48
Yaglom A M (1962) An Introduction to the Theory of Random FunctionsPrentice-Hall Englewood Cliffs NJ
Zellner A (1971) An Introduction to Bayesian Inference in Econometrics WileyNew York
Chapter 3
Practical reliability approachusing spreadsheet
Bak Kong Low
31 Introduction
In a review on first-order second-moment reliability methods USACE (1999)rightly noted that a potential problem with both the Taylorrsquos series methodand the point estimate method is their lack of invariance for nonlinear perfor-mance functions The document suggested the more general HasoferndashLindreliability index (Hasofer and Lind 1974) as a better alternative but con-ceded that ldquomany published analyses of geotechnical problems have notused the HasoferndashLind method probably due to its complexity especiallyfor implicit functions such as those in slope stability analysisrdquo and that ldquothemost common method used in Corps practice is the Taylorrsquos series methodbased on a Taylorrsquos series expansion of the performance function about theexpected valuesrdquo
A survey of recent papers on geotechnical reliability analysis reinforces theabove USACE observation that although the HasoferndashLind index is perceivedto be more consistent than the Taylorrsquos series mean value method the latteris more often used
This chapter aims to overcome the computational and conceptual barriersof the HasoferndashLind index for correlated normal random variables and thefirst-order reliability method (FORM) for correlated nonnormals in thecontext of three conventional geotechnical design problems Specificallythe conventional bearing capacity model involving two random variablesis first illustrated to elucidate the procedures and concepts This is fol-lowed by a reliability-based design of an anchored sheet pile wall involvingsix random variables which are first treated as correlated normals thenas correlated nonnormals Finally reliability analysis with search for crit-ical noncircular slip surface based on a reformulated Spencer method ispresented This probabilistic slope stability example includes testing therobustness of search for noncircular critical slip surface modeling lognor-mal random variables deriving probability density functions from reliabilityindices and comparing results inferred from reliability indices with MonteCarlo simulations
Practical reliability approach 135
The expanding ellipsoidal perspective of the HasoferndashLind reliability indexand the practical reliability approach using object-oriented constrained opti-mization in the ubiquitous spreadsheet platform were described in Low andTang (1997a 1997b) and extended substantially in Low and Tang (2004)by testing robustness for various nonnormal distributions and complicatedperformance functions and by providing enhanced operational convenienceand versatility
Reasonable statistical properties are assumed for the illustrative cases pre-sented in this chapter actual determination of the statistical properties is notcovered Only parametric uncertainty is considered and model uncertaintyis not dealt with Hence this chapter is concerned with reliability methodand perspectives and not reliability in its widest sense The focus is onintroducing an efficient and rational design approach using the ubiquitousspreadsheet platform
The spreadsheet reliability procedures described herein can be appliedto stand-alone numerical (eg finite element) packages via the establishedresponse surface method (which itself is straightforward to implement inthe ubiquitous spreadsheet platform) Hence the applicability of the relia-bility approach is not confined to models which can be formulated in thespreadsheet environment
32 Reliability procedure in spreadsheet andexpanding ellipsoidal perspective
321 A simple hands-on reliability analysis
The proposed spreadsheet reliability evaluation approach will be illustratedfirst for a case with two random variables Readers who want a betterunderstanding of the procedure and deeper appreciation of the ellipsoidalperspective are encouraged to go through the procedure from scratch on ablank Excel worksheet After that some Excel files for hands-on and deeperappreciation can be downloaded from httpalummiteduwwwbklow
The example concerns the bearing capacity of a strip footing sustain-ing non-eccentric vertical load Extensions to higher dimensions and morecomplicated scenarios are straightforward
With respect to bearing capacity failure the performance function (PerFn)for a strip footing in its simplest form is
PerFn = qu minus q (31a)
where qu = cNc + poNq + B2γNγ (31b)
in which qu is the ultimate bearing capacity q the applied bearing pressurec the cohesion of soil po the effective overburden pressure at foundation
136 Bak Kong Low
level B the foundation width γ the unit weight of soil below the base offoundation and Nc Nq and Nγ are bearing capacity factors which areestablished functions of the friction angle (φ) of soil
Nq = eπ tanφ tan2(
45 + φ
2
)(32a)
Nc =(Nq minus 1
)cot (φ) (32b)
Nγ = 2(Nq + 1
)tanφ (32c)
Several expressions for Nγ exist The above Nγ is attributed to Vesic inBowles (1996)
The statistical parameters and correlation matrix of c and φ are shown inFigure 31 The other parameters in Equations (31a) and (31b) are assumedknown with values q = QvB = (200 kNm)B po = 18 kPa B = 12 m andγ = 20 kNm3 The parameters c and φ in Equations (31) and (32) readtheir values from the column labeled xlowast which were initially set equal tothe mean values These xlowast values and the functions dependent on them
qu = cNc + PoNq + B gNγ2
Units m kNm2kNm3 degrees as appropriate
X Qv q
Corr Matrix
PerFn
Solution procedure
nxnx =
xi ndash mi
C
Nq
NC
Ng
B Pogsm
f
b
63391463
3804
1074 1
00
= qu(x) - q
3268
1minus05
minus2732
minus0187
minus05
2507
2015
5 200 1667 12 20
Qv B
182
si
=si
xi ndash mi
si
xi ndash miT
[R ]minus1
1 Initially x values = mean values m
2 Invoke Solver to minimize b by changing x values subject to PerFn = 0 amp x values ge 0
Figure 31 A simple illustration of reliability analysis involving two correlated randomvariables which are normally distributed
Practical reliability approach 137
change during the optimization search for the most probable failure pointSubsequent steps are
1 The formula of the cell labeled β in Figure 31 is Equation (35b)in Section 323 ldquo=sqrt(mmult(transpose(nx) mmult(minverse(crmat)nx)))rdquo The arguments nx and crmat are entered by selecting thecorresponding numerical cells of the column vector
(xi minusmicroi
)σi and
the correlation matrix respectively This array formula is then enteredby pressing ldquoEnterrdquo while holding down the ldquoCtrlrdquoand ldquoShiftrdquo keysMicrosoft Excelrsquos built-in matrix functions mmult transpose andminverse have been used in this step Each of these functions containsprogram codes for matrix operations
2 The formula of the performance function is g (x) = qu minus q wherethe equation for qu is Equation (31b) and depends on the xlowastvalues
3 Microsoft Excelrsquos built-in constrained optimization program Solveris invoked (via ToolsSolver) to Minimize β By Changing the xvalues Subject To PerFn le 0 and x values ge 0 (If used for thefirst time Solver needs to be activated once via ToolsAdd-insSolverAdd-in)
The β value obtained is 3268 The spreadsheet approach is simple andintuitive because it works in the original space of the variables It does notinvolve the orthogonal transformation of the correlation matrix and iter-ative numerical partial derivatives are done automatically on spreadsheetobjects which may be implicit or contain codes
The following paragraphs briefly compares lumped factor of safetyapproach partial factors approach and FORM approach and provideinsights on the meaning of reliability index in the original space of therandom variables More details can be found in Low (1996 2005a) Lowand Tang (1997a 2004) and other documents at httpalummiteduwwwbklow
322 Comparing lumped safety factor and partial factorsapproaches with reliability approach
For the bearing capacity problem of Figure 31 a long-established determin-istic approach evaluates the lumped factor of safety (Fs) as
Fs = qu minus po
q minus po= f (cφ ) (33)
where the symbols are as defined earlier If c = 20 kPa and φ = 15 andwith the values of Qv B γ and po as shown in Figure 31 then the factor of
138 Bak Kong Low
45 Limit state surface boundary between safeand unsafe domains
40
35
30
25
20
Coh
esio
n c
15
10
5
0
SAFE
one-sigmadispersion ellipse
Mean-valuepointDesign point
Friction angle φ (degrees)0 10 20 30
UNSAFER
r
β-ellipse
b = R rsc
mc
sφ
mφ
Figure 32 The design point the mean-value point and expanding ellipsoidal perspectiveof the reliability index in the original space of the random variables
safety is Fs asymp 20 by Equation (33) In the two-dimensional space of c andφ (Figure 32) one can plot the Fs contours for different combinations ofc andφ including the Fs =10 curve which separates the unsafe combinationsfrom the safe combinations of c and φ The average point (c = 20 kPa andφ = 15) is situated on the contour (not plotted) of Fs = 20 Design by thelumped factor of safety approach is considered satisfactory with respect tobearing capacity failure if the factor of safety by Equation (33) is not smallerthan a certain value (eg when Fs ge 20)
A more recent and logical approach (eg Eurocode 7) applies partial fac-tors to the parameters in the evaluation of resistance and loadings Designis acceptable if
Bearing capacity (based on reduced c and φ) ge Applied pressure
(amplified) (34)
A third approach is reliability-based design where the uncertainties andcorrelation structure of the parameters are represented by a one-standard-deviation dispersion ellipsoid (Figure 32) centered at the mean-value pointand safety is gaged by a reliability index which is the shortest distance(measured in units of directional standard deviations Rr) from the safemean-value point to the most probable failure combination of parame-ters (ldquothe design pointrdquo) on the limit state surface (defined by Fs = 10for the problem in hand) Furthermore the probability of failure (Pf )
Practical reliability approach 139
can be estimated from the reliability index β using the established equa-tion Pf = 1 minus(β) = (minusβ) where is the cumulative distribution (CDF)of the standard normal variate The relationship is exact when the limitstate surface is planar and the parameters follow normal distributions andapproximate otherwise
The merits of a reliability-based design over the lumped factor-of-safetydesign is illustrated in Figure 33a in which case A and case B (with dif-ferent average values of effective shear strength parameters cprime and φprime) showthe same values of lumped factor of safety yet case A is clearly safer thancase B The higher reliability of case A over case B will correctly be revealedwhen the reliability indices are computed On the other hand a slope mayhave a computed lumped factor of safety of 15 and a particular foun-dation (with certain geometry and loadings) in the same soil may have acomputed lumped factor of safety of 25 as in case C of Figure 33(b)Yet a reliability analysis may show that they both have similar levels ofreliability
The design point (Figure 32) is the most probable failure combination ofparametric values The ratios of the respective parametric values at the cen-ter of the dispersion ellipsoid (corresponding to the mean values) to those atthe design point are similar to the partial factors in limit state design exceptthat these factored values at the design point are arrived at automatically(and as by-products) via spreadsheet-based constrained optimization Thereliability-based approach is thus able to reflect varying parametric sensitiv-ities from case to case in the same design problem (Figure 33a) and acrossdifferent design realms (Figure 33b)
one-standard-deviationdispersion ellipsoidFs = 14
Fs = 10
Unsafe Fs lt 10
Fs = 12
Fs = 30
Fs = 10
Fs = 10
Slope
(Slope)Unsafe
(Foundation)
Foundation
Safe
A
B
20 15
12C
cprimecprime
φprime φprime(a) (b)
Figure 33 Schematic scenarios showing possible limitations of lumped factor of safety(a) Cases A and B have the same lumped Fs = 14 but Case A is clearly morereliable than Case B (b) Case C may have Fs = 15 for a slope and Fs = 25 fora foundation and yet have similar levels of reliability
140 Bak Kong Low
323 HasoferndashLind index reinterpreted via expandingellipsoid perspective
The matrix formulation (Veneziano 1974 Ditlevsen 1981) of the HasoferndashLind index β is
β = minxisinF
radic(x minusmicro)T Cminus1 (x minusmicro) (35a)
or equivalently
β = minxisinF
radic[xi minusmicroi
σi
]T
[R]minus1[
xi minusmicroi
σi
](35b)
where x is a vector representing the set of random variables xi micro the vec-tor of mean values microi C the covariance matrix R the correlation matrixσi the standard deviation and F the failure domain Low and Tang (1997b2004) used Equation (35b) in preference to Equation (35a) because the cor-relation matrix R is easier to set up and conveys the correlation structuremore explicitly than the covariance matrix C Equation (35b) was enteredin step (1) above
The ldquoxlowastrdquo values obtained in Figure 31 represent the most probable failurepoint on the limit state surface It is the point of tangency (Figure 32) of theexpanding dispersion ellipsoid with the bearing capacity limit state surfaceThe following may be noted
(a) The xlowast values shown in Figure 31 render Equation (31a) (PerFn) equalto zero Hence the point represented by these xlowast values lies on the bearingcapacity limit state surface which separates the safe domain from theunsafe domain The one-standard-deviation ellipse and the β-ellipse inFigure 32 are tilted because the correlation coefficient between c and φ isminus05 in Figure 31 The design point in Figure 32 is where the expandingdispersion ellipse touches the limit state surface at the point representedby the xlowast values of Figure 31
(b) As a multivariate normal dispersion ellipsoid expands its expandingsurfaces are contours of decreasing probability values according tothe established probability density function of the multivariate normaldistribution
f (x) = 1
(2π )n2 |C|05
exp[minus1
2(x minusmicro)T Cminus1 (x minusmicro)
](36a)
= 1
(2π )n2 |C|05
exp[minus1
2β2]
(36b)
Practical reliability approach 141
where β is defined by Equation (35a) or (35b) without the ldquominrdquoHence to minimize β (or β2 in the above multivariate normal distribu-tion) is to maximize the value of the multivariate normal probabilitydensity function and to find the smallest ellipsoid tangent to thelimit state surface is equivalent to finding the most probable failurepoint (the design point) This intuitive and visual understanding ofthe design point is consistent with the more mathematical approachin Shinozuka (1983 equations 4 41 and associated figure) in whichall variables were transformed into their standardized forms and thelimit state equation had also to be written in terms of the standard-ized variables The differences between the present original space versusShinozukarsquos standardized space of variables will be further discussed in(h) below
(c) Therefore the design point being the first point of contact betweenthe expanding ellipsoid and the limit state surface in Figure 32is the most probable failure point with respect to the safe mean-value point at the centre of the expanding ellipsoid where Fs asymp 20against bearing capacity failure The reliability index β is the axisratio (Rr) of the ellipse that touches the limit state surface and theone-standard-deviation dispersion ellipse By geometrical properties ofellipses this co-directional axis ratio is the same along any ldquoradialrdquodirection
(d) For each parameter the ratio of the mean value to the xlowast value is similarin nature to the partial factors in limit state design (eg Eurocode 7)However in a reliability-based design one does not specify the par-tial factors The design point values (xlowast) are determined automaticallyand reflect sensitivities standard deviations correlation structure andprobability distributions in a way that prescribed partial factors cannotreflect
(e) In Figure 31 the mean value point at 20 kPa and 15 is safe againstbearing capacity failure but bearing capacity failure occurs when thec and φ values are decreased to the values shown (6339 1463) Thedistance from the safe mean-value point to this most probable failurecombination of parameters in units of directional standard deviationsis the reliability index β equal to 3268 in this case
(f) The probability of failure (Pf ) can be estimated from the reliabilityindex β Microsoft Excelrsquos built-in function NormSDist() can be usedto compute () and hence Pf Thus for the bearing capacity problemof Figure 31 Pf = NormSDist(minus3268) = 0054 This value com-pares remarkably well with the range of values 0051minus0060 obtainedfrom several Monte Carlo simulations each with 800000 trials usingthe commercial simulation software RISK (httpwwwpalisadecom)The correlation matrix was accounted for in the simulation The excel-lent agreement between 0054 from reliability index and the range
142 Bak Kong Low
0051minus0060 from Monte Carlo simulation is hardly surprisinggiven the almost linear limit state surface and normal variates shownin Figure 32 However for the anchored wall shown in the nextsection where six random variables are involved and nonnormal dis-tributions are used the six-dimensional equivalent hyperellispoid andthe limit state hypersurface can only be perceived in the mindrsquos eyeNevertheless the probabilities of failure inferred from reliability indicesare again in close agreement with Monte Carlo simulations Com-puting the reliability index and Pf = (minusβ) by the present approachtakes only a few seconds In contrast the time needed to obtain theprobability of failure by Monte Carlo simulation is several ordersof magnitude longer particularly when the probability of failure issmall and many trials are needed It is also a simple matter to inves-tigate sensitivities by re-computing the reliability index β (and Pf ) fordifferent mean values and standard deviations in numerous what-ifscenarios
(g) The probability of failure as used here means the probability that inthe presence of parametric uncertainties in c and φ the factor of safetyEquation (33) will be le 10 or equivalently the probability that theperformance function Equation (31a) will be le 0
(h) Figure 32 defines the reliability index β as the dimensionless ratio Rrin the direction from the mean-value point to the design point This isthe axis ratio of the β-ellipsoid (tangential to the limit state surface) tothe one-standard-deviation dispersion ellipsoid This axis ratio is dimen-sionless and independent of orientation when R and r are co-directionalThis axis-ratio interpretation in the original space of the variables over-comes a drawback in Shinozukarsquos (1983) standardized variable spacethat ldquothe interpretation of β as the shortest distance between the ori-gin (of the standardized space) and the (transformed) limit state surfaceis no longer validrdquo if the random variables are correlated A furtheradvantage of the original space apart from its intuitive transparency isthat it renders feasible and efficient the two computational approachesinvolving nonnormals as presented in Low amp Tang (2004) and (2007)respectively
33 Reliability-based design of an anchored wall
This section illustrates reliability-based design of anchored sheet pile walldrawing material from Low (2005a 2005b) The analytical formulationsin a deterministic anchored wall design are the basis of the performancefunction in a probabilistic-based design Hence it is appropriate to brieflydescribe the deterministic approach prior to extending it to a probabilistic-based design An alternative design approach is described in BS8002(1994)
Practical reliability approach 143
331 Deterministic anchored wall design based on lumpedfactor and partial factors
The deterministic geotechnical design of anchored walls based on the freeearth support analytical model was lucidly presented in Craig (1997) Anexample is the case in Figure 34 where the relevant soil properties are theeffective angle of shearing resistance φprime and the interface friction angle δ
between the retained soil and the wall The characteristic values are cprime = 0φprime = 36 and δ = frac12φprime The water table is the same on both sides of thewall The bulk unit weight of the soil is 17 kNm3 above the water table and20 kNm3 below the water table A surcharge pressure qs = 10 kNm2 actsat the top of the retained soil The tie rods act horizontally at a depth 15 mbelow the top of the wall
In Figure 34 the active earth pressure coefficient Ka is based on theCoulomb-wedge closed-form equation which is practically the same asthe KeriselndashAbsi active earth pressure coefficient (Kerisel and Absi 1990)The passive earth pressure coefficient Kp is based on polynomial equations
17
20
10qs
36 (Characteristic value)
Surcharge qs = 10 kNm2
18
200
M5
M1 4
02
Ka Kah Kp Kph
0225 6966 662 329
Boxed cells contain equations
15 m
Water table
TA
d Fs
d
γsat
γ φprime δ
γ
γsat
64 m
24 m
45
1 -272
Forces Lever arm
(kNm) (m) (kN-mm)
Moments
-782
-139
-371
3657
4545
2767
7745
8693
949
-123
-216
-1077
-322
3472
1733
2
3
4
5
3
1
2
δ
φprime
Figure 34 Deterministic design of embedment depth d based on a lumped factor ofsafety of 20
144 Bak Kong Low
(Figure 35) fitted to the values of KeriselndashAbsi (1990) for a wall with avertical back and a horizontal retained soil surface
The required embedment depth d of 329 m in Figure 34 ndash for alumped factor of safety of 20 against rotational failure around anchor pointA ndash agrees with Example 69 of Craig (1997)
If one were to use the limit state design approach with a partial factor of12 for the characteristic shear strength one enters tanminus1(tanφprime12) = 31in the cell of φprime and changes the embedment depth d until the summationof moments is zero A required embedment depth d of 283 m is obtainedin agreement with Craig (1997)
Function KpKeriseIAbsi(phi del)lsquoPassive pressure coefficient Kp for vertical wall back and horizontal retained filllsquoBased on Tables in Kerisel amp Absi (1990) for beta = 0 lamda = 0 andlsquo delphi = 0033 05 066 100x = del phiKp100=000007776 phi ^ 4 - 0006608 phi ^ 3 + 02107 phi ^ 2 - 2714 phi + 1363Kp66=000002611 phi ^ 4 - 0002113 phi ^ 3 + 006843 phi ^ 2 - 08512 phi + 5142Kp50 = 000001559 phi ^ 4 - 0001215 phi ^ 3 + 003886 phi ^ 2 - 04473 phi + 3208Kp33 = 0000007318 phi ^ 4 - 00005195 phi ^ 3 + 00164 phi ^ 2 - 01483 phi + 1798Kp0 = 0000002636 phi ^ 4 - 00002201 phi ^ 3 + 0008267 phi ^ 2 - 00714 phi + 1507Select Case xCase 066 To 1 Kp = Kp66 + (x - 066) 1 - 066) (Kp100 - Kp66)Case 05 To 066 Kp = Kp50 + (x - 05) (066 - 05) (Kp66 - Kp50)Case 033 To 05 Kp = Kp33 + (x - 033) (05 - 033) (Kp50 - Kp33)Case 0 To 033 Kp = Kp0 + x 033 (Kp33 - Kp0)End SelectKpKeriselAbsi = KpEnd Function
Value from Kerisel-Absi tables
Polynomial curves equations as givenin the above Excel VBA code
40
35
30
25
20
Pas
sive
ear
th p
ress
ure
coef
ficie
nt K
p
15
10
5
00 10 20
Friction angle φ30
0
033
050
066
40 50
d f = 10
Figure 35 User-created Excel VBA function for Kp
Practical reliability approach 145
The partial factors in limit state design are applied to the characteristicvalues which are themselves conservative estimates and not the most proba-ble or average values Hence there is a two-tier nested safety first during theconservative estimate of the characteristic values and then when the partialfactors are applied to the characteristic values This is evident in Eurocode 7where Section 243 clause (5) states that the characteristic value of a soil orrock parameter shall be selected as a cautious estimate of the value affectingthe occurrence of the limit state Clause (7) further states that character-istic values may be lower values which are less than the most probablevalues or upper values which are greater and that for each calculation themost unfavorable combination of lower and upper values for independentparameters shall be used
The above Eurocode 7 recommendations imply that the characteristicvalue of φprime (36) in Figure 34 is lower than the mean value of φprime Hence inthe reliability-based design of the next section the mean value of φprime adoptedis higher than the characteristic value of Figure 34
While characteristic values and partial factors are used in limit state designmean values (not characteristic values) are used with standard deviations andcorrelation matrix in a reliability-based design
332 From deterministic to reliability-based anchoredwall design
The anchored sheet pile wall will be designed based on reliability analy-sis (Figure 36) As mentioned earlier the mean value of φprime in Figure 36is larger ndash 38 is assumed ndash than the characteristic value of Figure 34In total there are six normally distributed random variables with meanand standard deviations as shown Some correlations among parametersare assumed as shown in the correlation matrix For example it is judgedlogical that the unit weights γ and γsat should be positively correlatedand that each is also positively correlated to the angle of friction φprime sinceγ prime = γsat minus γw
The analytical formulations based on force and moment equilibrium in thedeterministic analysis of Figure 34 are also required in a reliability analysisbut are expressed as limit state functions or performance functions ldquo= Sum(Moments1rarr5)rdquo
The array formula in cell β of Figure 36 is as described in step 1 of thebearing capacity example earlier in this chapter
Given the uncertainties and correlation structure in Figure 36 wewish to find the required total wall height H so as to achieve a relia-bility index of 30 against rotational failure about point ldquoArdquo Initiallythe column xlowast was given the mean values Microsoft Excelrsquos built-inconstrained optimization tool Solver was then used to minimize β bychanging (automatically) the xlowast column subject to the constraint that
146 Bak Kong Low
Surcharge qs
03
Ka Kah Kp Kph
0249 5904 5637 279 000
d
d = H ndash 64 ndash z
1215
Boxed cells containequations
15 m
Water table
TA
H PerFn
crmatrix (Correlation matrix)
nX
d
γsat
γ φprime δ64 m
300
b
z
45
1 -311
Forces Lever arm
(kNm) (m) (kN-mm)
Moments
-828
-149
-356
1892
4575
2767
7775
8733
972
-142
-229
-1156
-311
1839
Normal
Normal
Normal
Normal
Normal
Normal
1844
1028
162
3351
1729
2963
20
10
17
Mean StDev
38
19
24
1
1
05
0
05
0
0
05
1
0
05
0
0
0
0
1
0
0
0
05
05
0
1
08
0
0
0
0
08
1
0
0
0
0
0
0
1
2
085 -09363
-15572
013807
-22431
-17077
187735
2
1
03
γsat
qs
γ
x
φprime
δ
z
γsat
γsat
qs
qs
γ
γ
φprime
φprime
δ
δ
z
z
2
3
4
5 3
1
2
Dredge level
Figure 36 Design total wall height for a reliability index of 30 against rotational failureDredge level and hence z and d are random variables
the cell PerFn be equal to zero The solution (Figure 36) indicates thata total height of 1215 m would give a reliability index of 30 againstrotational failure With this wall height the mean-value point is safeagainst rotational failure but rotational failure occurs when the mean val-ues descendascend to the values indicated under the xlowast column Thesexlowast values denote the design point on the limit state surface and rep-resent the most likely combination of parametric values that will causefailure The distance between the mean-value point and the design pointin units of directional standard deviations is the HasoferndashLind reliabilityindex
Practical reliability approach 147
As noted in Simpson and Driscoll (1998 81 158) clause 8321 ofEurocode 7 requires that an ldquooverdigrdquo allowance shall be made for wallswhich rely on passive resistance This is an allowance ldquofor the unforeseenactivities of nature or humans who have no technical appreciation of thestability requirements of the wallrdquo For the case in hand the reliability anal-ysis in Figure 36 accounts for uncertainty in z requiring only the mean valueand the standard deviation of z (and its distribution type if not normal) to bespecified The expected embedment depth is d = 1215minus64minusmicroz = 335 mAt the failure combination of parametric values the design value of z iszlowast = 29632 and dlowast = 1215 minus 64 minus zlowast = 279 m This corresponds to anldquooverdigrdquo allowance of 056 m Unlike Eurocode 7 this ldquooverdigrdquo is deter-mined automatically and reflects uncertainties and sensitivities from case tocase in a way that specified ldquooverdigrdquo cannot Low (2005a) illustrates anddiscusses this automatic probabilistic overdig allowance in a reliability-baseddesign
The nx column indicates that for the given mean values and uncer-tainties rotational stability is not surprisingly most sensitive to φprime andthe dredge level (which affects z and d and hence the passive resistance)It is least sensitive to uncertainties in the surcharge qs because the aver-age value of surcharge (10 kNm2) is relatively small when compared withthe over 10 m thick retained fill Under a different scenario where thesurcharge is a significant player its sensitivity scale could conceivably bedifferent It is also interesting to note that at the design point where thesix-dimensional dispersion ellipsoid touches the limit state surface bothunit weights γ and γsat (1620 and 1844 respectively) are lower thantheir corresponding mean values contrary to the expectation that higherunit weights will increase active pressure and hence greater instability Thisapparent paradox is resolved if one notes that smaller γsat will (via smaller γ prime)reduce passive resistance smaller φprime will cause greater active pressure andsmaller passive pressure and that γ γsat and φprime are logically positivelycorrelated
In a reliability-based design (such as the case in Figure 36) one doesnot prescribe the ratios meanxlowast ndash such ratios or ratios of (characteristicvalues)xlowast are prescribed in limit state design ndash but leave it to the expand-ing dispersion ellipsoid to seek the most probable failure point on the limitstate surface a process which automatically reflects the sensitivities of theparameters The ability to seek the most-probable design point without pre-suming any partial factors and to automatically reflect sensitivities from caseto case is a desirable feature of the reliability-based design approach Thesensitivity measures of parameters may not always be obvious from a pri-ori reasoning A case in point is the strut with complex supports analyzedin Low and Tang (2004 85) where the mid-span spring stiffness k3 andthe rotational stiffness λ1 at the other end both turn out to have surpris-ingly negligible sensitivity weights this sensitivity conclusion was confirmed
148 Bak Kong Low
by previous elaborate deterministic parametric plots In contrast reliabilityanalysis achieved the same conclusion relatively effortlessly
The spreadsheet-based reliability-based design approach illustrated inFigure 36 is a more practical and relatively transparent intuitive approachthat obtains the same solution as the classical HasoferndashLind method forcorrelated normals and FORM for correlated nonnormals (shown below)Unlike the classical computational approaches the present approach doesnot need to rotate the frame of reference or to transform the coordinatespace
333 Positive reliability index only if mean-value point isin safe domain
In Figure 36 if a trial H value of 10 m is used and the entire ldquoxlowastrdquo columngiven the values equal to the ldquomeanrdquo column values the performance func-tion PerFn exhibits a value of minus4485 meaning that the mean value point isalready inside the unsafe domain Upon Solver optimization with constraintPerFn = 0 a β index of 134 is obtained which should be regarded as anegative index ie minus134 meaning that the unsafe mean value point is atsome distance from the nearest safe point on the limit state surface that sep-arates the safe and unsafe domains In other words the computed β indexcan be regarded as positive only if the PerFn value is positive at the meanvalue point For the case in Figure 36 the mean value point (prior to Solveroptimization) yields a positive PerFn for H gt 106 m The computed β indexincreases from 0 (equivalent to a lumped factor of safety equal to 10 ie onthe verge of failure) when H is 106 m to 30 when H is 1215 m as shownin Figure 37
4
3
2
1
0
8 9 10 11 12 13
Total wall height H
Rel
iabi
lity
inde
x b
minus1
minus2
minus3
minus4
minus5
Figure 37 Reliability index is 300 when H = 1215 m For H smaller than 106 m themean-value point is in the unsafe domain for which the reliability indices arenegative
Practical reliability approach 149
334 Reliability-based design involving correlatednonnormals
The two-parameter normal distribution is symmetrical and theoretically hasa range from minusinfin to +infin For a parameter that admits only positive valuesthe probability of encroaching into the negative realm is extremely remoteif the coefficient of variation (Standard deviationMean) of the parameter is020 or smaller as for the case in hand Alternatively the lognormal dis-tribution has often been suggested in lieu of the normal distribution sinceit excludes negative values and affords some convenience in mathematicalderivations Figure 38 shows an efficient reliability-based design when therandom variables are correlated and follow lognormal distributions Thetwo columns labeled microN and σN contain the formulae ldquo=EqvN(hellip 1)rdquoand ldquoEqvN(hellip 2)rdquo respectively which invoke the user-created functionsshown in Figure 38 to perform the equivalent normal transformation(when variates are lognormals) based on the following RackwitzndashFiesslertwo-parameter equivalent normal transformation (Rackwitz and Fiessler1978)
Equivalent normal standard deviation σN = φminus1 [F (x)]
f (x)
(37a)
Equivalent normal mean microN = x minusσN timesminus1 [F (x)] (37b)
where x is the original nonnormal variateminus1[] is the inverse of the cumula-tive probability (CDF) of a standard normal distribution F(x) is the originalnonnormal CDF evaluated at x φ is the probability density function (PDF)of the standard normal distribution and f (x) is the original nonnormalprobability density ordinates at x
For lognormals closed form equivalent normal transformation is avail-able and has been used in the VBA code of Figure 38 Efficient Excel VBAcodes for equivalent normal transformations of other nonnormal distribu-tions (including Gumbel uniform exponential gamma Weibull triangularand beta) are given in Low and Tang (2004 2007) where it is shown thatreliability analysis can be performed with various distributions merely byentering ldquoNormalrdquo ldquoLognormalrdquo ldquoExponentialrdquo ldquoGammardquo hellip in thefirst column of Figure 38 and distribution parameters in the columns to theleft of the xlowast column In this way the required RackwitzndashFiessler equivalentnormal evaluations (for microN and σN) are conveniently relegated to functionscreated in the VBA programming environment of Microsoft Excel
Therefore the single spreadsheet cell object β in Figure 38 contains severalExcel function objects and substantial program codes
For correlated nonnormals the ellipsoid perspective (Figure 32) and theconstrained optimization approach still apply in the original coordinate
150 Bak Kong Low
Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
Lognormal
03
Ka Kah Kp Kph
0239 6307 601 261 000
d
d = H ndash 64 ndash z
122
Boxed cells containequations
H PerFn
Correlation matrix
300
b
1 -293
Forces Lever arm
(kNm) (m) (kN-mm)
Moments
-802
-145
-36
1835
4595
2767
7795
8760
982
-135
-222
-1131
-311
1802
Mean
17 085 1636 1697
1994
9803
3779
1893
226
0817
0938
1988
1814
0927
0396
-074
-124
0119
-181
-139
2327
1878
1004
345
1763
3182
1
2
2
1
03
20
10
38
19
24
StDev
1
05
0
05
0
0
05
1
0
05
0
0
0
0
1
0
0
0
05
05
0
1
08
0
0
0
0
08
1
0
0
0
0
0
0
1
γsat
qs
γ
x
φprime
δ
z
γsat
γsat
qs
qs
γ
γ
φprime
φprime
δ
δsNmN
z
z
2
3
4
5
nX
The microN and σN column invoke theuser-defined function EqvN_LNbelow to obtain the equivalentnormal mean microN and equivalentnormal standard deviation σN of thelognormal variates
More nonnormal options availablein Low amp Tang (2004 2007)
Function EqvN_LN(mean StDev x code)lsquoReturns the equivalent mean of the lognormal variateif if code is 1lsquoReturns the equivalent standard deviation of the lognormal variates if code is 2del = 00001 lsquovariable lower limitIf x lt del Then x = dellamda = Log(mean) - 05 Log(1 + (StDev mean) ^ 2)If code = 1 Then EqvN_LN = x (1 - Log(x) + lamda)If code = 2 Then EqvN_LN = x Sqr(Log(1 + (StDev mean) ^ 2))End Function
Figure 38 Reliability-based design of anchored wall correlated lognormals
system except that the nonnormal distributions are replaced by an equiv-alent normal ellipsoid centered not at the original mean of the nonnormaldistributions but at an equivalent normal mean microN
β = minxisinF
radicradicradicradic[xi minusmicroNi
σNi
]T
[R]minus1
[xi minusmicroN
i
σNi
](38)
Practical reliability approach 151
as explained in Low and Tang (2004 2007) One Excel file associated withLow (2005a) for reliability analysis of anchored sheet pile wall involving cor-related nonnormal variates is available for download at httpalummiteduwwwbklow
For the case in hand the required total wall height H is practically thesame whether the random variables are normally distributed (Figure 36)or lognormally distributed (Figure 38) Such insensitivity of the design tothe underlying probability distributions may not always be expected par-ticularly when the coefficient of variation (standard deviationmean) or theskewness of the probability distribution is large
If desired the original correlation matrix (ρij) of the nonnormalscan be modified to ρprime
ij in line with the equivalent normal transforma-tion as suggested in Der Kiureghian and Liu (1986) Some tables ofthe ratio ρprime
ijρij are given in Appendix B2 of Melchers (1999) includ-ing a closed-form solution for the special case of lognormals For thecases illustrated herein the correlation matrix thus modified differs onlyslightly from the original correlation matrix Hence for simplicity theexamples of this chapter retain their original unmodified correlationmatrices
This section has illustrated an efficient reliability-based design approachfor an anchored wall The correlation structure of the six variables wasdefined in a correlation matrix Normal distributions and lognormal dis-tributions were considered in turn (Figures 36 and 38) to investigatethe implication of different probability distributions The procedure is ableto incorporate and reflect the uncertainty of the passive soil surface ele-vation Reliability index is the shortest distance between the mean-valuepoint and the limit state surface ndash the boundary separating safe and unsafecombinations of parameters ndash measured in units of directional standarddeviations It is important to check whether the mean-value point is inthe safe domain or unsafe domain before performing reliability analysisThis is done by noting the sign of the performance function (PerFn) inFigures 36 and 38 when the xlowast columns were initially assigned the meanvalues If the mean value point is safe the computed reliability index ispositive if the mean-value point is already in the unsafe domain the com-puted reliability index should be considered a negative entity as illustratedin Figure 37
The differences between reliability-based design and design based on spec-ified partial factors were briefly discussed The merits of reliability-baseddesign are thought to lie in its ability to explicitly reflect correlation struc-ture standard deviations probability distributions and sensitivities and toautomatically seek the most probable failure combination of parametric val-ues case by case without relying on fixed partial factors Corresponding toeach desired value of reliability index there is also a reasonably accuratesimple estimate of the probability of failure
152 Bak Kong Low
34 Practical probabilistic slope stability analysisbased on reformulated Spencer equations
This section presents a practical procedure for implementing Spencer methodreformulated for a computer age first deterministically then probabilisti-cally in the ubiquitous spreadsheet platform The material is drawn fromLow (2001 2003 both available at the authorrsquos website) and includes test-ing the robustness of search for noncircular critical slip surface modelinglognormal random variables deriving probability density functions fromreliability indices and comparing results inferred from reliability indices withMonte Carlo simulations The deterministic modeling is described first as itunderlies the limit state function (ie performance function) of the reliabilityanalysis
341 Deterministic Spencer method reformulated
Using the notations in Nash (1987) the sketch at the top of Figure 39(below columns I and J) shows the forces acting on a slice (slice i) thatforms part of the potential sliding soil mass The notations are weight Wibase length li base inclination angle αi total normal force Pi at the baseof slice i mobilized shearing resistance Ti at the base of slice i horizontaland vertical components (Ei Eiminus1 λiEi λiminus1Eiminus1) of side force resultantsat the left and right vertical interfaces of slice i where λiminus1 and λi are thetangents of the side force inclination angles (with respect to horizontal) atthe vertical interfaces Adopting the same assumptions as Spencer (1973)but reformulated for spreadsheet-based constrained optimization approachone can derive the following from MohrndashCoulomb criterion and equilibriumconsiderations
Ti = [cprimeili +(Pi minus uili
)tanφprime
i
]F (MohrndashCoulomb criteria) (39)
Pi cosαi = Wi minusλiEi +λiminus1Eiminus1 minus Ti sinαi (vertical equilibrium)(310)
Ei = Eiminus1 + Pi sinαi minus Ti cosαi (horizontal equilibrium) (311)
Pi =
[Wi minus
(λi minusλiminus1
)Eiminus1
minus1F
(cprime
ili minus uili tanφprimei
)(sinαi minusλi cosαi
)]
[λi sinαi + cosαi
+1F
tanφprimei
(sinαi minusλi cosαi
)]
(from above three equations) (312)
Undrained shear strength profile of soft clay Embankment
depth
cu
0
40
15
28
3
20
5
20
7
26
100 25 50
Cm
10(kPa) (deg) (kNm3)
30 20
φm γm
γclay
16 (kNm3)
0
2
4
6Dep
th
cu (kPa)
8
10
37
(m)
(kPa)
1 minus10
Xn
γw
Pw
Mw X
cX
minX
max
(XiminusX
0)
(XnminusX
0)
Xn
X0
yc
Rru
X0
minus5 0 5 10
ytop
Ω
Ω H
27
X ybot
ytop
γave C φ W α
rad u l P T E λ σprime Lx
Ly
Units m kNm3 kNm2 kNor other consistent set of units
5 173 10 15 61518 02
hc
ybot
slope angle
Soft clay
10
5
0
minus5
minus10
15 20
A B C D E F G H I J K L M N O P Q R
234567
8
9
10
111213
14
15
1617181920212223242526272829303132333435363738394041424344
454647484950
Embankment
Reformulated Spencer method
λiE
i
λ
λiminus1Eiminus1
αi
Eiminus1
Ei
F
0105
Array formulas
λ
Varying side-force angle λ
000 000
1287 1287
DumEq
Varying λ
λ=λ sin π
TRUE
ΣM ΣForces
Ti
Wi
l
Pi
491 minus589795 1340
Center of Rotation
Framed cells contain quations
17467
0 1747 327 500 15 0000
1 1658 137 500 2000 1000 30 474 1136 1073 210 4972 2850 481 0012 1298 1212 563
2 1570 000 500 2000 1000 30 763 0997 1727 163 8096 3636 963 0025 3249 1124 726
3 1472 minus118 500 1958 3526 0 1074 0879 2189 154 11201 4214 1557 0038 5094 1030 854
4 1374 minus214 500 1900 2715 0 1242 0770 2531 137 13712 2884 2305 0050 7497 932 961
5 1276 minus292 500 1866 2252 0 1378 0673 2809 125 14925 2195 3063 0062 9090 834 1047
6 1178 minus356 500 1843 2000 0 1490 0582 3037 117 15635 1825 3770 0072 10275 735 1119
7 1079 minus410 500 1827 2000 0 1583 0496 3225 112 16090 1734 4384 0082 11192 638 1178
8 981 minus453 500 1815 2000 0 1658 0415 3380 107 16506 1666 4896 0090 12014 540 1226
9 883 minus487 450 1801 2000 0 1670 0336 3404 104 16403 1615 5285 0096 12377 442 1264
10 785 minus513 400 1784 2000 0 1619 0259 3300 102 15816 1578 5538 0101 12279 343 1295
11 687 minus531 350 1767 2066 0 1556 0184 3171 100 15237 1603 5659 0103 12093 245 1317
12 589 minus542 300 1751 2110 0 1481 0110 3018 099 14654 1619 5659 0105 11826 147 1331
13 491 minus546 250 1734 2132 0 1394 0037 2841 098 14044 1627 5548 0104 11461 049 1339
14 393 minus542 200 1717 2132 0 1296 minus0037 2641 098 13380 1627 5337 0101 10985 minus049 1339
15 294 minus531 150 1698 2110 0 1186 minus0110 2418 099 12636 1619 5037 0097 10381 minus147 1331
16 196 minus513 100 1677 2066 0 1065 minus0184 2171 100 11786 1603 4663 0091 9636 minus245 1317
17 098 minus487 050 1652 2000 0 932 minus0259 1900 102 10799 1578 4234 0083 8737 minus343 1295
18 000 minus453 000 1620 2000 0 787 minus0336 1604 104 9674 1615 3763 0074 7703 minus442 1264
19 minus098 minus410 000 1600 2000 0 677 minus0415 1380 107 8912 1666 3251 0064 6932 minus540 1226
20 minus196 minus356 000 1600 2000 0 601 minus0496 1225 112 8534 1734 2692 0053 6422 minus638 1178
21 minus294 minus292 000 1600 2000 0 509 minus0582 1037 117 7974 1825 2101 0040 5752 minus736 1119
22 minus393 minus214 000 1600 2252 0 397 minus0673 809 125 7392 2195 1469 0027 5084 minus834 1047
23 minus491 minus118 000 1600 2715 0 261 minus0770 531 137 6839 2884 786 0014 4470 minus932 961
24 minus589 000 000 1600 3526 0 93 minus0879 189 154 6712 4214 00 0000 4175 minus1030 854
Figure 39 Deterministic analysis of a 5 m high embankment on soft ground with depth-dependent undrained shear strength The limit equilibrium method of slices isbased on reformulated Spencer method with half-sine variation of side forceinclination
154 Bak Kong Low
sum[Ti cosαi minus Pi sinαi
]minus Pw = 0 (overall horizontal equilibrium)
(313)
sum[(Ti sinαi + Pi cosαi minus Wi
)lowast Lxi+(Ti cosαi minus Pi sinαi
)lowast Lyi
]minus Mw = 0
(overall moment equilibrium) (314)
Lxi = 05(xi + ximinus1
)minus xc (horizontal lever arm of slice i) (315)
Lyi = yc minus 05(yi + yiminus1
)(vertical lever arm of slice i) (316)
where ci φi and ui are cohesion friction angle and pore water pressurerespectively at the base of slice i Pw is the water thrust in a water-filledvertical tension crack (at x0) of depth hc and Mw the overturning momentdue to Pw Equations (315) and (316) required for noncircular slip surfacegive the lever arms with respect to an arbitrary center The use of both λiand λiminus1 in Equation (312) allows for the fact that the right-most slice(slice 1) has a side that is adjacent to a water-filled tension crack henceλ0 = 0 (ie the direction of water thrust is horizontal) and for differentλ values (either constant or varying) on the other vertical interfaces
The algebraic manipulation that results in Equation (312) involves open-ing the term (Pi minus uili)tanφprime of Equation (39) an action legitimate only if(Pi minus uili) ge 0 or equivalently if the effective normal stress σ prime
i (= Pili minus ui)at the base of a slice is nonnegative Hence after obtaining the critical slipsurface in the section to follow one needs to check that σ prime
i ge 0 at the baseof all slices and Ei ge 0 at all the slice interfaces Otherwise one should con-sider modeling tension cracks for slices near the upper exit end of the slipsurface
Figure 39 shows the spreadsheet set-up for deterministic stability analysisof a 5 m high embankment on soft ground The undrained shear strength pro-file of the soft ground is defined in rows 44 and 45 The subscript m in cellsP44R44 denotes embankment Formulas need be entered only in the firstor second cell (row 16 or 17) of each column followed by autofilling downto row 40 The columns labeled ytop γave and c invoke the functions shownin Figure 310 created via ToolsMacroVisualBasicEditorInsertModuleon the Excel worksheet menu The dummy equation in cell P2 is equalto Flowast1 This cell unlike cell O2 can be minimized because it contains aformula
Initially xc = 6 yc = 8 R = 12 in cells I11K11 and λprime = 0 F = 1in cells N2O2 Microsoft Excelrsquos built-in Solver was then invoked to settarget and constraints as shown in Figure 311 The Solver option ldquoUseAutomatic Scalingrdquo was also activated The critical slip circle and factor ofsafety F = 1287 shown in Figure 39 were obtained automatically withinseconds by Solver via cell-object oriented constrained optimization
Practical reliability approach 155
Function Slice_c(ybmid dmax dv cuv cm)lsquocomment dv = depth vectorlsquocuv = cu vectorIf ybmid gt 0 Then Slice_c = cm Exit FunctionEnd Ifybmid = Abs(ybmid)If ybmid gt dmax Then lsquoundefined domain Slice_c = 300 lsquohence assume hard stratum Exit FunctionEnd IfFor j = 2 T o dvCount lsquoarray size=dvCount If dv(j) gt= ybmid Then interp = (ybmid - dv(j - 1)) (dv(j) - dv(j - 1)) Slice_c = cuv(j - 1) + (cuv(j) - cuv(j - 1)) interp Exit For End IfNext jEnd Function
Function ytop(x omega H)grad = Tan(omega 314159 180)If x lt 0 Then ytop = 0If x gt=0 And x lt H grad Then ytop = x gradIf x gt=H grad Then ytop = HEnd Function
Function AveGamma(ytmid ybmid gm gclay)If ybmid lt 0 Then Sum = (ytmid gm + Abs(ybmid) gclay) AveGamma = Sum (ytmid - ybmid) Else AveGamma = gmEnd IfEnd Function
Figure 310 User-defined VBA functions called by columns ytop γave and c of Figure 39
Noncircular critical slip surface can also be searched using Solver asin Figure 311 except that ldquoBy Changing Cellsrdquo are N2O2 B16 B18B40 C17 and C19C39 and with the following additional cell constraintsB16 ge B11tan(radians(A11)) B16 ge B18 B40 le 0 C19C39 le D19D39O2 ge 01 and P17P40 ge 0
Figure 312 tests the robustness of the search for noncircular critical sur-face Starting from four arbitrary initial circles the final noncircular criticalsurfaces (solid curves each with 25 degrees of freedom) are close enoughto each other though not identical Perhaps more pertinent their factors ofsafety vary narrowly within 1253 ndash 1257 This compares with the minimumfactor of safety 1287 of the critical circular surface of Figure 39
Figure 311 Excel Solver settings to obtain the solution of Figure 39
8
4
0
-4
-8
-12
-20 -10 0 10
Circular initial
Noncircular critical
20
FS =125-126
Figure 312 Testing the robustness of search for the deterministic critical noncircular slipsurface
Practical reliability approach 157
342 Reformulated Spencer method extendedprobabilistically
Reliability analyses were performed for the embankment of Figure 39as shown in Figure 313 The coupling between Figures 39 and 313 isbrought about simply by entering formulas in cells C45H45 and P45R45of Figure 39 to read values from column vi of Figure 313 The matrixform of the HasoferndashLind index Equation (35b) will be used except thatthe symbol vi is used to denote random variables to distinguish it from thesymbol xi (for x-coordinate values) used in Figure 39
Spatial correlation in the soft ground is modeled by assuming an autocor-relation distance (δ) of 3 m in the following established negative exponentialmodel
ρij = eminus∣∣Depth(i) minus Depth(j)
∣∣δ (317)
mean StDev Correlation matrixldquocrmatrdquo
Dist Name vi microi σi microiN σi
N (vi-microiN)σi
N
lognormal Cm 104815 10 150 9872 1563 0390
lognormal φm 3097643 30 300 29830 3090 0371
lognormal γm 2077315 20 100 19959 1038 0784
lognormal Cu1 3426899 40 600 39187 5112 -0962
lognormal Cu2 2287912 28 420 27246 3413 -1279
lognormal Cu3 1598835 20 300 19390 2385 -1426
lognormal Cu4 1583775 20 300 19357 2362 -1490
lognormal Cu5 2207656 26 390 25442 3293 -1022
lognormal Cu6 3459457 37 555 36535 5160 -0376
nv
EqvN(1)
Array formula Ctrl+Shift then enter 1961PrFailβ
00249 =normsdist(-β)
Probability of failure
EqvN(2)
Function EqvN(Distributions Name v mean StDev code)Select Case UCase(Trim(Distribution Name)) lsquotrim leadingtrailing spaces amp convert to uppercaseCase ldquoNORMALrdquo If code = 1 Then EqvN = mean If code = 2 Then EqvN = StDevCase ldquoLOGNORMALrdquo If vlt0000001 Then v=0000001 lamda = Log(mean) - 05 Log(1+(StDevmean)and2) If code=1 Then EqvN=v(1-Log(v) +lamda) If code=2 Then EqvN=vSqr(Log(1+(StDevmean)and2))End selectEnd Function
1
-03
05
0
0
0
0
0
0
-03
1
05
0
0
0
0
0
0
05
05
1
0
0
0
0
0
0
0
0
0
1
06065
03679
01889
0097
00357
0
0
0
06065
1
06065
03114
01599
00588
0
0
0
03679
06065
1
05134
02636
0097
0
0
0
01889
03114
05134
1
05134
01889
0
0
0
0097
01599
02636
05134
1
03679
0
Cm -φm γm Cu1 Cu2 Cu3 Cu4 Cu4 Cu1
0
0
00357
00588
0097
01889
03679
1
0 15 3 5 7 10
0
15
3
5
7
10
depth
Equation 16 autofilled
=SQRT(MMULT(TRANSPOSE(nv) MMULT(MINVERSE(crmat)nv)))
Figure 313 Reliability analysis of the case in Figure 39 accounting for spatial variation ofundrained shear strength The two templates are easily coupled by replacingthe values cm φm γm and the cu values of Figure 39 with formulas that pointto the values of the vi column of this figure
158 Bak Kong Low
(Reliability analysis involving horizontal spatial variations of undrainedshear strengths and unit weights is illustrated in Low et al 2007 for a clayslope in Norway)
Only normals and lognormals are illustrated Unlike a widely documentedclassical approach there is no need to diagonalize the correlation matrix(which is equivalent to rotating the frame of reference) Also the iterativecomputations of the equivalent normal mean (microN) and equivalent normalstandard deviation (σN) for each trial design point are automatic during theconstrained optimization search
Starting with a deterministic critical noncircular slip surface of Figure 312the λprime and F values of Figure 39 were reset to 0 and 1 respectively The vivalues in Figure 313 were initially assigned their respective mean valuesThe Solver routine in Microsoft Excel was then invoked to
bull Minimize the quadratic form (a 9-dimensional hyperellipsoid in originalspace) ie cell ldquoβrdquo
bull By changing the nine random variables vi the λprime in Figure 39 and the 25coordinate values x0 x2 x24 yb1 yb3 yb23 of the slip surface (F remainsat 1 ie at failure or at limit state)
bull Subject to the constraints minus1 le λprime le 1 x0 ge Htan(radians()) x0 ge x2x24 le 0 yb3 yb23 le yt3 yt23
sumForces = 0
sumM = 0 and σ prime
1 σ prime24 ge 0
The Solver option ldquoUse automatic scalingrdquo was also activated
The β index is 1961 for the case with lognormal variates (Figure 313)The corresponding probability of failure based on the hyperplane assump-tion is 249 The reliability-based noncircular slip surface is shown inFigure 314 The nine vi values in Figure 313 define the most probablefailure point where the equivalent dispersion hyperellipsoid is tangent tothe limit state surface (F = 1) of the reliability-based critical slip surfaceAt this tangent point the values of cu are as expected smaller thantheir mean values but the cm and φm values are slightly higher thantheir respective mean values This peculiarity is attributable to cm and φmbeing positively correlated to the unit weight of the embankment γm andalso reflects the dependency of tension crack depth hc (an adverse effect)on cm since the equation hc = 2cm (γm
radicKa) is part of the model in
Figure 39By replacing ldquolognormalrdquo with ldquonormalrdquo in the first column of
Figure 313 re-initializing column vi to mean values and invoking Solverone obtains a β index of 1857 with a corresponding probability of fail-ure equal to 316 compared with 249 for the case with lognormalvariates The reliability-based noncircular critical slip surface of the casewith normal variates is practically indistinguishable from the case with log-normal variates Both are however somewhat different (Figure 314) fromthe deterministic critical noncircular surface from which they evolved via
Practical reliability approach 159
(m)
(m)
Deterministic critical
64
20
-2 0 5 10 15 20-5-10-4-6-8
Figure 314 Comparison of reliability-based critical noncircular slip surfaces (the two uppercurves for normal variates and lognormal variates) with the deterministiccritical noncircular slip surface (the lower dotted curve)
Table 31 Reliability indices for the example case
[Lognormal variates] [Normal variates]
Reliability index 1961 1857(1971)lowast (1872)lowast
Prob of failure 249 316(244)lowast (306)lowast
lowast at deterministic critical noncircular surface
Table 32 Solverrsquos computing time on a computer
F for specified slip surface asymp 03 sF search for noncircular surface 15 sβ for specified slip surface 2 sβ search for noncircular surface 20 s
25 degrees of freedom during Solverrsquos constrained optimization search Thisdifference in slip surface geometry matters little however for the followingreason If the deterministic critical noncircular slip surface is used for relia-bility analysis the β index obtained is 1971 for the case with lognormalsand 1872 for the case with normals These β values are only slightlyhigher (lt1) than the β values of 1961 and 1857 obtained earlier with aldquofloatingrdquo surface Hence for the case in hand performing reliability anal-ysis based on the fixed deterministic critical noncircular slip surface willyield practically the same reliability index as the reliability-based critical slipsurface Table 31 summarizes the reliability indices for the case in handTable 32 compares the computing time
160 Bak Kong Low
343 Deriving probability density functions fromreliability indices and comparison withMonte Carlo simulations
The reliability index β in Figure 313 has been obtained with respect tothe limit state surface defined by F = 10 The mean value point repre-sented by column microi is located on the 1253 factor-of-safety contour ascalculated in the deterministic noncircular slip surface search earlier A dis-persion hyperellipsoid expanding from the mean-value point will touchthe limit state surface F = 10 at the design point represented by thevalues in the column labeled vi The probability of failure 249 approx-imates the integration of probability density in the failure domain F le 10If one defines another limit state surface F = 08 a higher value of β isobtained with a correspondingly smaller probability that F le 08 reflect-ing the fact that the limit state surface defined by F = 08 is further awayfrom the mean-value point (which is on the factor-of-safety contour 1253)In this way 42 values of reliability indices (from positive to negative) cor-responding to different specified limit states (from F = 08 to F = 18)were obtained promptly using a short VBA code that automatically invokesSolver for each specified limit state F The series of β indices yields thecumulative distribution function (CDF) of the factor of safety based onPr[F le FLimitState] asymp (minusβ) where () is the standard normal cumula-tive distribution function The probability density function (PDF) plots inFigure 315 (solid curves) were then obtained readily by applying cubicspline interpolation (eg Kreyszig 1988) to the CDF during which pro-cess the derivatives (namely the PDF) emerged from the tridiagonal splinematrix The whole process is achieved easily using standard spreadsheetmatrix functions Another alternative was given in Lacasse and Nadim(1996) which examined pile axial capacity and approximated the CDFby appropriate probability functions
For comparison direct Monte Carlo simulations with 20000 realiza-tions were performed using the commercial software RISK with LatinHypercube sampling The random variables were first assumed to be nor-mally distributed then lognormally distributed The mean values standarddeviations and correlation structure were as in Figure 313
The solid PDF curves derived from the β indices agree remarkably wellwith the Monte Carlo PDF curves (the dashed curves in Figure 315)
Using existing state-of-the-art personal computer the time taken by Solverto determine either the factor of safety F or the reliability index β of theembankment in hand is as shown in Table 32
The computation time for 20000 realizations in Monte Carlo simula-tions would be prohibitive (20000times15 s or 83 h) if a search for thecritical noncircular slip surface is carried out for each random set (cm φmγm and the cu values) generated Hence the PDF plots of Monte Carlo
Practical reliability approach 161
3
from β indices Normalvariates
Factor of safety
from Monte Carlosimulation20000 trials
PD
F
25
2
15
1
05
006 08 1 12 14 16 18
3
from β indices Lognormalvariates
Factor of safety
from simulation20000 trials
25
2
15
1
05
006 08 1 12 14 16 18
3
Normal variates
Lognormals
Factor of safety
Both curves basedon β indices
PD
F
25
2
15
1
05
006 08 1 12 14 16 18
3
Normals
Lognormals
Factor of safety
Both curves fromMonte Carlosimulationseach 20000 trials
25
2
15
1
05
006 08 1 12 14 16 18
Figure 315 Comparing the probability density functions (PDF) obtained from reliabilityindices with those obtained from Monte Carlo simulations ldquoNormalsrdquo meansthe nine input variables are correlated normal variates and ldquoLognormalsrdquomeans they are correlated lognormal variates
simulations shown in Figure 315 have been obtained by staying put atthe reliability-based critical noncircular slip surface found in the previoussection In contrast computing 42 values of β indices (with fresh non-circular surface search for each limit state) will take only about 14 minNevertheless for consistency of comparison with the Monte Carlo PDFplots the 42 β values underlying each solid PDF curve of Figure 315have also been computed based on the same reliability-based critical non-circular slip surface shown in Figure 314 Thus it took about 14 min(42times2 s) to produce each reliability indices-based PDF curve but about70 times as long (asymp 20000 times 03 s) to generate each simulation-basedPDF curve What-if scenarios can be investigated more efficiently using thepresent approach of first-order reliability method than using Monte Carlosimulations
The probabilities of failure (F le 1) from Monte Carlo simulations with20000 realizations are 241plusmn for lognormal variates and 328plusmn fornormal variates These compare well with the 249 and 316 shown inTable 31 Note that only one β index value is needed to compute each prob-ability of failure in Table 31 Only for PDF does one need about 30 or 40β values of different limit states in order to define the CDF from which thePDF are obtained
162 Bak Kong Low
35 On the need for a positive definitecorrelation matrix
Although the correlation coefficient between two random variables has arange minus1 le ρij le 1 one is not totally free in assigning any values within thisrange for the correlation matrix This was explained in Ditlevsen (1981 85)where a 3times3 correlation matrix with ρ12 = ρ13 = ρ23 = minus1 implies a contra-diction and is hence inconsistent A full-rank (ie no redundant variables)correlation matrix has to be positive definite One may note that HasoferndashLind index is defined as the square-root of a quadratic form (Equation (35a)or (35b)) and unless the quadratic form and its covariance matrix and cor-relation matrix are positive-definite one is likely to encounter the square rootof a negative entity which is meaningless whether in the dispersion ellipsoidperspective or in the perspective of minimum distance in the reduced variatespace
Mathematically a symmetric matrix R is positive definite if and only ifall the eigenvalues of R are positive For present purposes one can createthe equation QForm = uTRminus1u side by side with β = radic
(QForm) DuringSolverrsquos searches if the cell Qform shows negative value it means that thequadratic form is not positive definite and hence the correlation matrix isinconsistent
The Monte Carlo simulation program RISK (httpwwwpalisadecom)will also display a warning of ldquoInvalid correlation matrixrdquo and offer tocorrect the matrix if simulation is run involving an inconsistent correlationmatrix
That a full-rank correlation or covariance matrix should be positive def-inite is an established theoretical requirement although not widely knownThis positive-definiteness requirement is therefore not a limitation of thepaperrsquos proposed procedure
The joint PDF of an n-dimensional multivariate normal distributionis given by Equation (36a) The PDF has a peak at mean-value pointand decreasing ellipsoidal probability contours surrounding micro only if thequadratic form and hence the covariance matrix and its inverse are positivedefinite (By lemmas of matrix theory C positive definite implies Cminus1 is pos-itive definite and also means the correlation matrix R and its inverse arepositive definite)
Without assuming any particular probability distributions it is docu-mented in books on multivariate statistical analysis that a covariance matrix(and hence a correlation matrix) cannot be allowed to be indefinite becausethis would imply that a certain weighted sum of its variables has a negativevariance which is inadmissible since all real-valued variables must havenonnegative variances (The classical proof can be viewed as reduction adabsurdum ie a method of proving the falsity of a premise by showing thatthe logical consequence is absurd)
Practical reliability approach 163
An internet search using the Google search engine with ldquopositive defi-nite correlation matrixrdquo as search terms also revealed thousands of relevantweb pages showing that many statistical computer programs (in addi-tion to the RISK simulation software) will check for positive definitenessof correlation matrix and offer to generate the closest valid matrix toreplace the entered invalid one For example one such web site titledldquoNot Positive Definite Matrices ndash Causes and Curesrdquo at httpwwwgsuedusimmkteernpdmatrihtml refers to the work of Wothke (1993)
36 Finite element reliability analysis via responsesurface methodology
Programs can be written in spreadsheet to handle implicit limit state func-tions (eg Low et al 1998 Low 2003 and Low and Tang 2004 87)However there are situations where serviceability limit states can only beevaluated using stand-alone finite element or finite difference programs Inthese circumstances reliability analysis and reliability-based design by thepresent approach can still be performed provided one first obtains a responsesurface function (via the established response surface methodology) whichclosely approximates the outcome of the stand-alone finite element or finitedifference programs Once the closed-form response functions have beenobtained performing reliability-based design for a target reliability index isstraightforward and fast Performing Monte-Carlo simulation on the closedform approximate response surface function also takes little time Xu andLow (2006) illustrate the finite element reliability analysis of embankmentson soft ground via the response surface methodology
37 Limitations and flexibilities of object-orientedconstrained optimization in spreadsheet
If equivalent normal transformation is used for first-order reliability methodinvolving nonnormal distributions other than the lognormal it is neces-sary to compute the inverse of the standard normal cumulative distributions(Equations (37a) and (37b)) There is no closed form equation for this pur-pose Numerical procedures such as that incorporated in Excelrsquos NormSInvis necessary In Excel 2000 and earlier versions NormSInv(CDF) returnscorrect values only for CDF between about 000000035 and 099999965ie within about plusmn5 standard deviations from the mean There are unde-sirable and idiosyncratic sharp kinks at both ends of the validity rangeIn EXCEL XP (2002) and later versions the NormSInv (CDF) functionhas been refined and returns correct values for a much broader band ofCDF 10minus16 lt CDF lt (1 minus 10minus16) or about plusmn8 standard deviations fromthe mean In general robustness and computation efficiency will be bet-ter with the broader validity band of NormSInv(CDF) in Excel XP This is
164 Bak Kong Low
because during Solverrsquos trial solutions using the generalized reduced gradientmethod the trial xlowast values may occasionally stray beyond 5 standard devia-tions from the mean as dictated by the directional search before returningto the real design point which could be fewer than 5 standard deviationsfrom the mean
Despite its simple appearance the single spreadsheet cell that contains theformula for reliability index β is in fact a conglomerate of program routinesIt contains Excelrsquos built-in matrix operating routines ldquommultrdquo ldquotransposerdquoand ldquominverserdquo which can operate on up to 50times50 matrices The membersof the ldquonxrdquo vector in the formula for β are each functions of their respec-tive VBA codes in the user-created function EqvN (Figure 313) Similarlythe single spreadsheet cell for the performance function g(x) can also con-tain many nested functions (built-in or user-created) In addition up to 100constraints can be specified in the standard Excel Solver Regardless of theircomplexities and interconnections they (β index g(x) constraints) appear asindividual cell objects on which Excelrsquos built-in Solver optimization programperforms numerical derivatives and iterative directional search using the gen-eralized reduced gradient method as implemented in Lasdon and WarenrsquosGRG2 code (httpwwwsolvercomtechnology4htm)
The examples of Low et al (1998 2001) with implicit performance func-tions and numerical methods coded in Excelrsquos programming environmentinvolved correlated normals only and have no need for the Excel NormSInvfunction and hence are not affected by the limitations of NormSInv in Excel2000 and earlier versions
38 Summary
This chapter has elaborated on a practical object-oriented constrained opti-mization approach in the ubiquitous spreadsheet platform based on thework of Low and Tang (1997a 1997b 2004) Three common geotechni-cal problems have been used as illustrations namely a bearing capacityproblem an anchored sheet pile wall embedment depth design and areformulated Spencer method Slope reliability analyses involving spatiallycorrelated normal and lognormal variates were demonstrated with searchfor the critical noncircular slip surface The HasoferndashLind reliability indexwas re-interpreted using the perspective of an expanding equivalent dis-persion ellipsoid centered at the mean in the original space of the randomvariables When nonnormals are involved the perspective is one of equiva-lent ellipsoids The probabilities of failure inferred from reliability indices arein good agreement with those from Monte Carlo simulations for the exam-ples in hand The probability density functions (of the factor of safety) werealso derived from reliability indices using simple spreadsheet-based cubicspline interpolation and found to agree well with those generated by the farmore time-consuming Monte Carlo simulation method
Practical reliability approach 165
The coupling of spreadsheet-automated cell-object oriented constrainedoptimization and minimal macro programming in the ubiquitous spreadsheetplatform together with the intuitive expanding ellipsoid perspective ren-ders a hitherto complicated reliability analysis problem conceptually moretransparent and operationally more accessible to practicing engineers Thecomputational procedure presented herein achieves the same reliability indexas the classical first order reliability method (FORM) ndash well-documented inAng and Tang (1984) Madsen et al (1986) Haldar and Mahadevan (1999)Baecher and Christian (2003) for example ndash but without involving the con-cepts of eigenvalues eigenvectors and transformed space which are requiredin a classical approach It also affords possibilities yet unexplored or difficulthitherto Although bearing capacity anchored wall and slope stability aredealt with in this chapter the perspectives and techniques presented couldbe useful in many other engineering problems
Although only correlated normals and lognormals are illustrated the VBAcode shown in Figure 313 can be extended (Low and Tang 2004 and 2007)to deal with the triangular the exponential the gamma the Gumbel andthe beta distributions for example
In reliability-based design it is important to note whether the mean-valuepoint is in the safe domain or in the unsafe domain When the mean-valuepoint is in the safe domain the distance from the safe mean-value pointto the most probable failure combination of parametric values (the designpoint) on the limit state surface in units of directional standard deviationsis a positive reliability index When the mean-value point is in the unsafedomain (due to insufficient embedment depth of sheet pile wall for example)the distance from the unsafe mean-value point to the most probable safecombination of parametric values on the limit state surface in units of direc-tional standard deviations is a negative reliability index This was shownin Figure 37 for the anchored sheet pile wall example It is also importantto appreciate that the correlation matrix to be consistent must be positivedefinite
The meaning of the computed reliability index and the inferred probabilityof failure is only as good as the analytical model underlying the performancefunction Nevertheless even this restrictive sense of reliability or failure ismuch more useful than the lumped factor of safety approach or the partialfactors approach both of which are also only as good as their analyticalmodels
In a reliability-based design the design point reflects sensitivities standarddeviations correlation structure and probability distributions in a way thatprescribed partial factors cannot
The spreadsheet-based reliability approach presented in this chaptercan operate on stand-alone numerical packages (eg finite element) viathe response surface method which is itself readily implementable inspreadsheet
166 Bak Kong Low
The present chapter does not constitute a comprehensive risk assessmentapproach but it may contribute component blocks necessary for such afinal edifice It may also help to overcome a language barrier that hamperswider adoption of the more consistent HasoferndashLind reliability index andreliability-based design Among the issues not covered in this chapter aremodel uncertainty human uncertainty and estimation of statistical parame-ters These and other important issues were discussed in VanMarcke (1977)Christian et al (1994) Whitman (1984 1996) Morgenstern (1995) Baecherand Christian (2003) among others
Acknowledgment
This chapter was written at the beginning of the authorrsquos four-monthsabbatical leave at NGI Norway summer 2006
References
Ang H S and Tang W H (1984) Probability Concepts in Engineering Planningand Design Vol 2 ndash Decision Risk and Reliability John Wiley New York
Baecher G B and Christian J T (2003) Reliability and Statistics in GeotechnicalEngineering John Wiley Chichester UK
Bowles J E (1996) Foundation Analysis and Design 5th ed McGraw-HillNew York
BS 8002 (1994) Code of Practice for Earth Retaining Structures British StandardsInstitution London
Christian J T Ladd C C and Baecher G B (1994) Reliability applied toslope stability analysis Journal of Geotechnical Engineering ASCE 120(12)2180ndash207
Craig R F (1997) Soil Mechanics 6th ed Chapman amp Hall LondonDer Kiureghian A and Liu P L (1986) Structural reliability under incom-
plete probability information Journal of Engineering Mechanics ASCE 112(1)85ndash104
Ditlevsen O (1981) Uncertainty Modeling With Applications to MultidimensionalCivil Engineering Systems McGraw-Hill New York
ENV 1997-1 (1994) Eurocode 7 Geotechnical Design Part 1 General RulesCENEuropean Committee for Standardization Brussels
Haldar A and Mahadevan S (1999) Probability Reliability and StatisticalMethods in Engineering Design John Wiley New York
Hasofer A M and Lind N C (1974) Exact and invariant second-moment codeformat Journal of Engineering Mechanics ASCE 100 111ndash21
Kerisel J and Absi E (1990) Active and Passive Earth Pressure Tables 3rd edAA Balkema Rotterdam
Kreyszig E (1988) Advanced Engineering Mathematics 6th ed John Wiley amp SonsNew York pp 972ndash3
Lacasse S and Nadim F (1996) Model uncertainty in pile axial capacity cal-culations Proceedings of the 28th Offshore Technology Conference Texaspp 369ndash80
Practical reliability approach 167
Low B K (1996) Practical probabilistic approach using spreadsheet Geotech-nical Special Publication No 58 Proceedings Uncertainty in the GeologicEnvironment From Theory to Practice ASCE USA Vol 2 pp 1284ndash302
Low B K (2001) Probabilistic slope analysis involving generalized slip sur-face in stochastic soil medium Proceedings 14th Southeast Asian Geotech-nical Conference December 2001 Hong Kong AA Balkema Rotterdampp 825ndash30
Low B K (2003) Practical probabilistic slope stability analysis In Proceedings Soiland Rock America MIT Cambridge MA June 2003 Verlag Gluumlckauf GmbHEssen Germany Vol 2 pp 2777ndash84
Low B K (2005a) Reliability-based design applied to retaining walls Geotech-nique 55(1) 63ndash75
Low B K (2005b) Probabilistic design of anchored sheet pile wall Proceedings16th International Conference on Soil Mechanics and Geotechnical Engineering12ndash16 September 2005 Osaka Japan Millpress pp 2825ndash8
Low B K and Tang W H (1997a) Efficient reliability evaluation using spread-sheet Journal of Engineering Mechanics ASCE 123(7) 749ndash52
Low B K and Tang W H (1997b) Reliability analysis of reinforced embankmentson soft ground Canadian Geotechnical Journal 34(5) 672ndash85
Low B K and Tang W H (2004) Reliability analysis using object-orientedconstrained optimization Structural Safety 26(1) 69ndash89
Low B K and Tang W H (2007) Efficient spreadsheet algorithm for first-orderreliability method Journal of Engineering Mechanics ASCE 133(12) 1378ndash87
Low B K Lacasse S and Nadim F (2007) Slope reliability analysis accountingfor spatial variation Georisk Assesment and Management of Risk for EngineeringSystems and Geohazards Taylor amp Francis 1(4) 177ndash89
Low B K Gilbert R B and Wright S G (1998) Slope reliability analysis usinggeneralized method of slices Journal of Geotechnical and GeoenvironmentalEngineering ASCE 124(4) 350ndash62
Low B K Teh C I and Tang W H (2001) Stochastic nonlinear pndashy analysisof laterally loaded piles Proceedings of the Eight International Conference onStructural Safety and Reliability ICOSSAR lsquo01 Newport Beach California 17ndash22June 2001 8 pages AA Balkema Rotterdam
Madsen H O Krenk S and Lind N C (1986) Methods of Structural SafetyPrentice Hall Englewood Cliffs NJ
Melchers R E (1999) Structural Reliability Analysis and Prediction 2nd edJohn Wiley New York
Morgenstern N R (1995) Managing risk in geotechnical engineering ProceedingsPan American Conference ISSMFE
Nash D (1987) A comparative review of limit equilibrium methods of stability anal-ysis In Slope Stability Eds M G Anderson and K S Richards Wiley New Yorkpp 11ndash75
Rackwitz R and Fiessler B (1978) Structural reliability under combined randomload sequences Computers and Structures 9 484ndash94
Shinozuka M (1983) Basic analysis of structural safety Journal of StructuralEngineering ASCE 109(3) 721ndash40
Simpson B and Driscoll R (1998) Eurocode 7 A Commentary ARUPBREConstruction Research Communications Ltd London
168 Bak Kong Low
Spencer E (1973) Thrust line criterion in embankment stability analysis Geotech-nique 23 85ndash100
US Army Corps of Engineers (1999) ETL 1110-2-556 Risk-based analysis ingeotechnical engineering for support of planning studies Appendix A pagesA11 and A12 (httpwwwusacearmymilpublicationseng-tech-ltrsetl1110-2-556tochtml)
VanMarcke E H (1977) Reliability of earth slopes Journal of GeotechnicalEngineering ASCE 103(11) 1247ndash66
Veneziano D (1974) Contributions to Second Moment Reliability Research ReportNo R74ndash33 Department of Civil Engineering MIT Cambridge MA
Whitman R V (1984) Evaluating calculated risk in geotechnical engineeringJournal of Geotechnical Engineering ASCE 110(2) 145ndash88
Whitman R V (1996) Organizing and evaluating uncertainnty in geotechnicalengineering Proceedings Uncertainty in Geologic Environment From Theoryto Practice ASCE Geotechnical Special Publication 58 ASCE New York V1pp 1ndash28
Wothke W (1993) Nonpositive definite matrices in structural modeling In TestingStructural Equation Models Eds KA Bollen and J S Long Sage Newbury ParkCA pp 257ndash93
Xu B and Low B K (2006) Probabilistic stability analyses of embankmentsbased on finite-element method Journal of Geotechnical and GeoenvironmentalEngineering ASCE 132(11) 1444ndash54
Chapter 4
Monte Carlo simulationin reliability analysis
Yusuke Honjo
41 Introduction
In this chapter Monte Carlo Simulation (MCS) techniques are described inthe context of reliability analysis of structures Special emphasis is placedon the recent development of MCS for reliability analysis In this contextgeneration of random numbers by low-discrepancy sequences (LDS) andsubset Markov Chain Monte Carlo (MCMC) techniques are explained insome detail The background to the introduction of such new techniques isalso described in order for the readers to make understanding of the contentseasier
42 Random number generation
In Monte Carlo Simulation (MCS) it is necessary to generate random num-bers that follow arbitrary probability density functions (PDF) It is howeverrelatively easy to generate such random numbers once a sequence of uniformrandom numbers in (00 10) is given (this will be described in section 23)
Because of this the generation of a sequence of random numbers in (0010) is discussed first in this section Based on the historical developmentsthe methods can be classified to pseudo random number generation andlow-discrepancy sequences (or quasi random number) generation
421 Pseudo random numbers
Pseudo random numbers are a sequence of numbers that look like randomnumbers by deterministic calculations This is a suitable method for acomputer to generate a large number of reproducible random numbers andis used commonly today (Rubinstein 1981 Tsuda 1995)
170 Yusuke Honjo
Pseudo random numbers are usually assumed to satisfy the followingconditions
1 A great number of random numbers can be generated instantaneously2 If there is a cycle in the generation (which usually exists) it should be
long enough to be able to generate a large number of random numbers3 The random numbers should be reproducible4 The random numbers should have appropriate statistical properties
which can usually be examined by statistical testings for example Chisquare goodness of fit test
The typical methods of pseudo random number generation include themiddle-square method and linear recurrence relations (sometimes called thecongruential method)
The origin of the middle-square method goes back to Von Neumann whoproposed generating a random number of 2a digits from xn of 2a digitsby squaring xn to obtain 4a digits number and then cut the first and lasta digits of this number to obtain xn+1 This method however was foundto have relatively short cycle and also does not satisfy satisfactory statisticalcondition This method is rarely used today
On the other hand the general form of linear recurrence relations can begiven by the following equation
xn+1 = a0xn + a1xnminus1 + + ajxnminusj + b (mod P) (41)
where mod P implies that xn+1 = a0xn + a1xnminus1 + + ajxnminusj + b minus Pknand kn = [(xn+1 = a0xn + a1xnminus1 + + ajxnminusj + b)P] denotes the largestpositive integer in (xn+1 = a0xn + a1xnminus1 + + ajxnminusj + b)P
The simplest form of Equation (41) may be given as
xn+1 = xn + xnminus1 (mod P) (42)
This is called the Fibonacci method If the generated random numbers are2 digits integer and suppose the initial two numbers are x1 = 11 and x2 = 36and P = 100 the generated numbers are 4783301443
The modified versions of Equation (41) are multiplicative congruentialmethod and mixed congruential method which are still popularly used andprograms are easily available (see for example Koyanagi 1989)
422 Low-discrepancy sequences (quasi random numbers)
One of the most important applications of MCS is to solve multi-dimensionalintegration problems which includes reliability analysis The problem can be
Monte Carlo simulation 171
generally described as a multi-dimensional integration in a unit hyper cube
I =int 1
0
int 1
0f (x)dx [x = (x1x2 xk)] (43)
The problem solved by MCS can be described as the following equation
S(N) = 1N
Nsumi=1
f (xi) (44)
where xirsquos are N randomly generated points in the unit hyper cube
The error of |S(N) minus I| is proportional to O(Nminus 12 ) which is related to
the sample number N Therefore if one wants to improve the accuracy ofthe integration by one digit one needs to increase the sample point number100 times
Around 1960 some mathematicians showed that if one can generate thepoints sequences with some special conditions the accuracy can be improvedby O(Nminus1) or even O(Nminus2) Such points sequences were termed quasirandom numbers (Tsuda 1995)
However it was found later that quasi random numbers are not so effec-tive as the dimensions of the integration increase to more than 50 On theother hand there was much demand especially in financial engineeringto solve integrations of more than 1000 dimensions which expeditedresearch in this area in 1990s
One of the triggers that accelerated the research in this area was theKoksmandashHlawka theorem which related the convergence of the integrationand discrepancy of numbers generated in the unit hyper cube According tothis theorem the convergence of integration is faster as the discrepancy ofthe points sequences is smaller Based on this theorem the main objectiveof the research in the 1990s was to generate point sequences in the veryhigh dimensional hyper cube with as low discrepancy as possible In otherwords how to generate point sequences with maximum possible uniformitybecame the main target of the research Such point sequences were termedlow-discrepancy sequences (LDS) In fact LDS are point sequences that allowthe accuracy of the integration in k dimensions to improve in proportion to
O(
(logN)k
N
)or less (Tezuka 1995 2003)
Based on this development the term LDS is more often used than quasirandom numbers in order to avoid confusion between pseudo and quasirandom numbers
The discrepancy is defined as follows (Tezuka 1995 2003)
DN(k) = supyisin[01]k
∣∣∣∣∣∣([0y)N)
Nminus
kprodi=1
yi
∣∣∣∣∣∣ (45)
172 Yusuke Honjo
10
0000
10y1
y2
Figure 41 Concept of discrepancy in 2-dimensional space
where ([0y]N) denotes number of points in [0y) Therefore DN(k)
indicates discrepancy between the distribution of N points xn in thek dimensional hyper cube [01]k from ideally uniform points distributions
In Figure 41 discrepancy in 2-dimensional space is illustrated conceptu-ally where N = 10 If the number of the points in a certain area is always inproportion to the area rate DN
(k) is 0 otherwise DN(k) becomes larger as
points distribution is biasedA points sequences xi is said to be LDS when it satisfies the conditions
below
|S(N) minus I| le VfDN(k) (46)
DN(k) = c(k)
(logN)k
N
where Vf indicates the fluctuation of function f (x) within the domain of theintegration and c(k) is a constant which only depends on dimension k
Halton Faure and Sobolrsquo sequences are known to be LSD Variousalgorithms are developed to generate LSD (Tezuka 1995 2003)
In Figure 42 1000 samples generated by the mixed congruential methodand LSD are compared It is understood intuitively from the figure that pointsgenerated by LSD are more uniformly distributed
423 Random number generation following arbitrary PDF
In this section generation of random numbers following arbitrary PDF isdiscussed The most popular method to carry out this is to employ inversetransformation method
Monte Carlo simulation 173
1
08
06
04
02
00 02 04 06 08 1
1
08
06
04
02
00 02 04 06 08 1
(a) (b)
Figure 42 (a) Pseudo random numbers by mixed congruential method and (b) pointsequences by LDS 1000 generated sample points in 2-dimensional unit space
Fu (u ) Fx (x )1
0 11u xu1 u3 u2 u4
u4
u2
u3
u1
x4 x2x3x1
Figure 43 Concept of generation of random number xi by uniform random number uibased on the inverse transformation
Suppose X is a random variable whose CDF is FX(x) = pFX(x) is a non-decreasing function and 0 le p le 10 Thus the following
inverse transformation always exists for any p as (Rubinstein 1981)
Fminus1X (p) = min
[x FX(x) ge p
]0 le p le 10 (47)
By using this inverse transformation any random variable X that followsCDF FX(x) can be generated based on uniform random numbers in (01)(Figure 43)
X = Fminus1X (U) (48)
174 Yusuke Honjo
Some of the frequently used inverse transformations are summarizedbelow (Rubinstein 1981 Ang and Tang 1985 Reiss and Thomas 1997)Note that U is a uniform random variable in (01)
Exponential distribution
FX(x) = 1 minus exp[minusλx] (x ge 0)
X = Fminus1X (U) = minus1
λln(1 minus U)
The above equation can be practically rewritten as follows
X = Fminus1X (U) = minus1
λln(U)
Cauchy distribution
fX(x) = α
πα2 + (x minusλ)2 α gt 0λ gt 0minusinfin lt x lt infin
FX(x) = 12
+πminus1tanminus1(
x minusλ
α
)
X = Fminus1X (U) = λ+α tan
[π
(U minus 1
2
)]
Gumbel distribution
FX(x) = exp[minusexpminusa(x minus b)] (minusinfin lt x lt infin)
X = Fminus1X (U) = minus1
aln(minusln(U)
)+ b
Frechet distribution
FX(x) = exp
[minus(
ν
x minus ε
)k]
(ε lt x lt infin)
X = Fminus1X (U) = ν
(minusln(U))minus1k + ε
Monte Carlo simulation 175
Weibull distribution
FX(x) = exp
[minus(ωminus xωminus ν
)k]
(minusinfin lt x ltω)
X = Fminus1X (U) = (ωminus ν)
(ln(U)
)1k +ω
Generalized Pareto distribution
FX(x) = 1 minus(
1 + γx minusmicro
σ
)minus1γ
(micro le x)
X = Fminus1X (U) = σ
γ
[(1 minus U)minusγ minus 1
]+micro
The normal random numbers which are considered to be used mostfrequently next to the uniform random numbers do not have inverse trans-formation in an explicit form One of the most common ways to generate thenormal random numbers is to employ the Box and Muller method which isdescribed below (see Rubinstein 1981 86ndash7 for more details)
U1 and U2 are two independent uniform random variables Based on thesevariables two independent standard normal random numbers Z1 and Z2can be generated by using following equations
Z1 = (minus2 ln(U1))12 cos(2πU2) (49)
Z2 = (minus2 ln(U1))12 sin(2πU2)
A program based on this method can be obtained easily from standardsubroutine libraries (eg Koyanagi 1989)
424 Applications of LDS
In Yoshida and Sato (2005a) reliability analyses by MCS are presentedwhere results by LDS and by OMCS (ordinary Monte Carlo simulation) arecompared The performance function they employed is
g(xS) =5sum
i=1
xi minus(1 + x1
2500
)S (410)
where xi are independent random variables which follow identically a lognormal distribution with mean 250 and COV 10
The convergence of the failure probability with the number of simulationruns are presented in Figure 44 for LSD and 5 OMCS runs (which are
176 Yusuke Honjo
10minus2
10minus3
Failu
re P
roba
bilit
y
10minus4
100 101 102
Number103
LDSCrude
Figure 44 Relationship between calculated Pf and number of simulation runs (Yoshida andSato 2005a)
indicated by ldquoCruderdquo in the figure) It is observed that LDS has fasterconvergence than OMCS
43 Some methods to improve the efficiencyof MCS
431 Accuracy of ordinary Monte Carlo Simulation(OMCS)
We consider the integration of the following type in this section
Pf =int
intI [g (x) le 0] fX (x)dx (411)
where fX is the PDF of basic variables X and I is an indicator functiondefined below
I (x) = 1 g(x) le 0
0 g(x) gt 0
The indicator function identifies the domain of integration Function g(x)is a performance function in reliability analysis thus the integration provides
Monte Carlo simulation 177
the failure probability Pf Note that the integration is carried out on all thedomains where basic variables are defined
When this integration is evaluated by MCS Pf can be evaluated as follows
Pf asymp 1N
Nsumj=1
I[g(xj
)le 0]
(412)
where xj is j-th sample point generated based on PDF fx()One of the important aspects of OMCS is to find the necessary number of
simulation runs to evaluate Pf of required accuracy This can be generallydone in the following way (Rubinstein 1981 115ndash18)
Let N be the total number of runs and Nf be the ones which indicatedthe failure Then the failure probability Pf is estimated as
Pf = Nf
N(413)
The mean and variance of Pf can be evaluated as
E[Pf
]= 1
NE[NF
]= NPf
N= Pf (414)
Var[Pf
]= Var
[Nf
N
]= 1
N2 Var[Nf]
= 1N2 NPf
(1 minus Pf
)= 1N
Pf(1 minus Pf
)(415)
Note that here these values are obtained by recognizing the fact that Nffollows binominal distribution of N trials where the mean and the varianceare given as NPf and NPf (1 minus Pf ) respectively
Suppose one wants to obtain a necessary number of simulation runs for
probability∣∣∣Pf minus Pf
∣∣∣le ε to be more than 100α
Prob[∣∣∣Pf minus Pf
∣∣∣ le ε]
ge α (416)
On the other hand the following relationship is given based onChebyshevrsquos inequality formula
Prob[∣∣∣Pf minus Pf
∣∣∣lt ε]
ge 1 minusVar[Pf
]ε2 (417)
178 Yusuke Honjo
From Equations (416) and (417) and using Equation (415) thefollowing relationship is obtained
α le 1 minusVar[Pf
]ε2 = 1 minus 1
Nε2 Pf(1 minus Pf
)1 minusα ge 1
Nε2 Pf(1 minus Pf
)
N ge Pf(1 minus Pf
)(1 minusα)ε2 (418)
Based on Equation (418) the necessary number of simulation runs N canbe obtained for any given ε and α Let ε = δPf and one obtains for N
N ge Pf(1 minus Pf
)(1 minusα)δ2P2
f
= (1Pf ) minus 1(1 minusα)δ2 asymp 1
(1 minusα)δ2Pf(419)
The necessary numbers of simulation runs N are calculated in Table 41for ε = 01Pf ie δ = 01 and α = 095 In other words N obtained inthis table ensures the error of Pf to be within plusmn10 for 95 confidenceprobability
The evaluation shown in Table 41 is considered to be very conservativebecause it is based on the very general Chebyshevrsquos inequality relationshipBroding (Broding et al 1964) gives necessary simulation runs as below(Melchers 1999)
N gtminus ln(1 minusα)
Pf(420)
Table 41 Necessary number of simulation runs to evaluate Pf of required accuracy
δ α Pf N by Chebyshev N by Bording
01 095 100E-02 200000 300100E-04 20000000 29957100E-06 2000000000 2995732
05 095 100E-02 8000 300100E-04 800000 29957100E-06 80000000 2995732
10 095 100E-02 2000 300100E-04 200000 29957100E-06 20000000 2995732
50 095 100E-02 80 300100E-04 8000 29957100E-06 800000 2995732
Monte Carlo simulation 179
where N is necessary number of simulation runs for one variable α is givenconfidence level and Pf is the failure probability For example if α = 095and Pf = 10minus3 the necessary N is by Equation (420) about 3000 In thecase of more than one independent random variable one should multiplythis number by the number of variables N evaluated by Equation (420) isalso presented in Table 41
432 Importance sampling (IMS)
As can be understood from Table 41 the required number of simulation runsin OMCS to obtain Pf of reasonable confidence level is not small especiallyfor cases of smaller failure probability In most practical reliability analysesthe order of Pf is of 10minus4 or less thus this problem becomes quite seriousOne of the MCS methods to overcome this difficulty is the ImportanceSampling (IMS) technique which is discussed in this section
Theory
Equation (411) to calculate Pf can be rewritten as follows
Pf =int
intI[g(x) le 0] fX(x)
hV (x)hV (x)dx (421)
where v is a random variable following PDF hV
Pf asymp 1N
⎧⎨⎩
Nsumj=1
I[g(vj) le 0] fX(vj)
hV (vj)
⎫⎬⎭ (422)
where hV is termed the Importance Sampling function and acts as a con-trolling function for sampling Furthermore the optimum hV is known tobe the function below (Rubinstein 1981 122ndash4)
hV (v) = I [g(v) le 0] fX(v)Pf
(423)
This function hV (v) in the case of a single random variable is illustratedin Figure 45
It can be shown that the estimation variance of Pf is null for the samplingfunction hV (v) given in Equation (423) In other words Pf can be obtained
180 Yusuke Honjo
fx (x ) hv (v )
G(x ) = x minus a
apf
x v
Figure 45 An illustration of the optimum Importance Function in the case of a singlevariable
exactly by using this sample function as demonstrated below
Pf =int int
intI [g(x) le 0]
fX(x)hV (x)
hV (x)dx
=int int
intI [g(x) le 0] fX(x)
[I [g(x) le 0] fX(x)
Pf
]minus1
hV (x)dx
= Pf
int int
inthV (x)dx = Pf
Therefore Pf can be obtained without carrying out MCS if the importancefunction of Equation (423) is employed
However Pf is unknown in practical situations and the result obtainedhere does not have an actual practical implication Nevertheless Equation(423) provides a guideline on how to select effective important samplingfunctions a sampling function that has a closer form to Equation (423) maybe more effective It is also known that if an inappropriate sampling func-tion is chosen the evaluation becomes ineffective and estimation varianceincreases
Selection of importance sampling functions in practice
The important region in reliability analyses is the domain where g(x) le 0The point which has the maximum fX(x) in this domain is known as thedesign point which is denoted by xlowast One of the most popular ways to sethV is to set the mean value of this function at this design point (Melcher1999) Figure 46 conceptually illustrates this situation
Monte Carlo simulation 181
Failure
Z (x)=0
Limit State Surface
Safe
0
hv (x )fx (x )
x2
x1
Figure 46 A conceptual illustration of importance sampling function set at the design point
If g(x) is a linear function and the space is standardized normal space thesample points generated by this hV fall in the failure region by about 50which is considered to give much higher convergence compared to OMCSwith a smaller number of sample points
However if the characteristic of a performance function becomes morecomplex and also if the number of basic variables increases it is not a sim-ple task to select an appropriate sample function Various so-called adaptivemethods have been proposed but most of them are problem-dependentIt is true to say that there is no comprehensive way to select an appro-priate sample function at present that can guarantee better results in allcases
433 Subset MCMC (Markov Chain Monte Carlo) method
Introduction to the subset MCMC method
The subset MCMC method is a reliability analysis method by MCS proposedby Au and Beck (2003) They combined the subset concept with the MCMCmethod which had been known as a very flexible MCS technique to generaterandom numbers following any arbitrary PDF (see for example Gilks et al1996) They combined these two concepts to develop a very effective MCStechnique for structural reliability analysis
The basic concept of this technique is explained in this subsection whereasa more detailed calculation procedure is presented in the next subsectionSome complementary mathematical proofs are presented in the appendix ofthis chapter and some simple numerical examples in subsection 433
182 Yusuke Honjo
Let F be a set indicating the failure region Also let F0 be the total set andFi(I = 1 middot middot middotm) be subsets that satisfy the relationship below
F0 sup F1 sup F2 sup sup Fm sup F (424)
where Fm is the last subset assumed in the Pf calculation Note that thecondition Fm sup F is always satisfied
The failure probability can be calculated by using these subsets as follows
Pf = P(F) = P(F | Fm)P(Fm | Fmminus1) P(F1 | F0) (425)
Procedure to calculate Pf by subset MCMC method
(1) Renewal of subsets
The actual procedure to set a subset is as follows where zi = g(xi) is thecalculated value of a given performance function at sample point xi Nt isthe total number of generated points in each subset Ns are number of pointsselected among Nt points from the smaller zi values and Nf is number offailure points (ie zi is negative) in Nt Also note that x used in this sectioncan be a random variable or a random vector that follows a specified PDF
Step 1 In the first cycle Nt points are generated by MCS that follow thegiven PDF
Step 2 Order the generated samples in ascending order by the value of ziand select smaller Ns(lt Nt) points A new subset Fk+1 is defined bythe equation below
Fk+1 =
x | z(x) le zs + zs+1
2
(426)
Note that zirsquos in this equation is assembled in ascending order Theprobability that a point generated in subset Fk+1 for points in subsetFk is calculated to be P(Fk+1 | Fk) = NsNt
Step 3 By repeating Step 2 for a range of subsets Equation (425) can beevaluatedNote that xi (i = 1 Ns) are used as seed points in MCMC togenerate the next Nt samples which implies that NtNs samplesare generated from each of these seed points In other words NtNspoints are generated from each seed point by MCMCFurthermore the calculation is ceased when Nf reaches certainnumber with respect to Nt
Monte Carlo simulation 183
(2) Samples generation by MCMC
By MCMC method Nt samples are generated in subset Fk+1 This method-ology is described in this subsection
Let π (x) be a PDF of x that follows PDF fX(x) and yet conditioned to bewithin the subset Fk PDF π (x) is defined as follows
π (x) = c middot fX(x) middot ID(x) (427)
where
ID(x) =
1 if (x sub Fk)0 otherwise
where c is a constant for PDF π to satisfy the condition of a PDFPoints following this PDF π (x) can be generated by MCMC by the
following procedure (Gilks et al 1996 Yoshida and Sato 2005)
Step 1 Determine any arbitrary PDF q(xprime | xi) which is termed proposaldensity In this research a uniform distribution having mean xiand appropriate variance is employed (The author tends to usethe variance of the original PDF to this uniform distribution Thesmaller variance tends to give higher acceptance rate (ie Equation(429)) but requires more runs due to slow movement of the subsetboundary)
Step 2 The acceptance probability α(xprimexi) is calculated Then set xi+1 = xprimewith probability α or xi+1 = xi with probability 1 minus α Theacceptance probability α is defined as the following equation
α(xprimexi) = min
10q(xi | xprime) middotπ (xprime)q(xprime | xi)π (xi)
= min
10q(xi | xprime) middot fX(xprime) middot ID(xprime)
q(xprime | xi)fX(xi)
(428)
Equation (428) can be rewritten as below considering the fact thatq(xprime | xi) is a uniform distribution and point xi is certainly in thesubset under consideration
α = min
10fX(xprime) middot ID(xprime)
fX(xi)
(429)
Step 3 Repeat Step 2 for necessary cycles
184 Yusuke Honjo
s
Failure Region
Performance Function
Stable Region
R
Z = R minus S
Design Point
Figure 47 A conceptual illustration of the subset MCMC method
(3) Evaluation of the failure probability Pf
In this study the calculation is stopped when Nf reaches certain number withrespect to Nt otherwise the calculation in Step 2 is repeated for a renewedsubset which is smaller than the previous one
Some of more application examples of the subset MCMC to geotechnicalreliability problems can be seen in Yoshida and Sato (2005b)
The failure probability can be calculated as below
Pf = P (z lt 0) =
Ns
Nt
mminus1 Nf
Nt(430)
The mathematical proof of this algorithm is presented in the nextsection
The conceptual illustration of this procedure is presented in Figure 47It is understood that by setting the subsets more samples are generated tothe closer region to the performance function which is supposed to improvethe efficiency of MCS considerably compared to OMCS
Examples of subset MCMC
Two simple reliability analyses are presented here to demonstrate the natureof subset MCMC method
Monte Carlo simulation 185
8
6
4S
R
2
00 2 4 6 8 10
k = 1
k = 2
k = 3
Z = 0
Zsub_1
Zsub_2
Figure 48 A calculation example of subset MCMC method when Z = R minus S
(1) An example on a linear performance function
The first example employs the simplest performance function Z = R ndash SThe resistance R follows a normal distribution N(7010) whereas theexternal force S a normal distribution N(3010) The true value of thefailure probability of this problem is Pf = 234E minus 03
Figure 48 exhibits a calculation example of this reliability analysis In thiscalculation Nt = 100 and two subsets are formed before the calculation isceased It is observed that more samples are generated closer to the limitstate line as the calculation proceeds
In order to examine the performance of subset MCMC method the num-ber of selected sample points to proceed to the next subset Ns and stoppingcriteria of the calculation are altered Actually Ns is set to 251020 or 50The stopping criteria are set based on the rate of Nf to Nt as Nf ge050020010 or 005Nt Also a criterion Nf ge Ns is added Nt is setto 100 in all cases
Table 42 lists the results of the calculation 1000 calculations are madefor each case where mean and COV of log10(calculated Pf ) log10(truePf )are presented
In almost all cases the means are close to 10 which exhibit very littlebias in the present calculation for evaluating Pf COV seems to be somewhatsmaller for case Ns is larger
186 Yusuke Honjo
Table 42 Accuracy of Pf estimation for Z = R minus S
Cutoff criterion Ns = 2 Ns = 5 Ns = 10 Ns = 20 Ns = 50
➀Nf 002N Mean 0992 0995 0997 0976 0943Cov 0426 0448 0443 0390 0351
➁Nf 005N Mean 0995 0993 0989 0989 0971Cov 0428 0378 0349 0371 0292
➂Nf 010N Mean 1033 0994 0991 0985 0985Cov 0508 0384 0367 0312 0242
➃Nf 020N Mean 1056 1074 1006 0993 0990Cov 0587 0489 0348 0305 0165
➄Nf 050N Mean 1268 1111 1056 1020 0996Cov 0948 0243 0462 0319 0162
Table 43 Number of performance function calls for Z = R minus S example
Cutoff criterion Ns = 2 Ns = 5 Ns = 10 Ns = 20 Ns = 50
➀Nf 002N Mean 1719 1881 2077 2329 3432sd 251 371 438 561 1092
➁Nf 005N Mean 1858 2090 2291 2716 4400sd 313 562 445 540 882
➂Nf 010N Mean 2029 2218 2518 2990 5066sd 324 616 396 532 754
➃Nf 020N Mean 2129 2406 2706 3265 5694sd 312 486 389 480 681
➄Nf 050N Mean 2260 2580 2930 3611 6547sd 257 371 426 484 627
Table 43 presents a number of performance function calls in each caseThe mean and sd (standard deviation) of these numbers are shown It isobvious from this table that the numbers of the function calls increase as Nsincreases It is understandable because as Ns increases the speed of subsetsconverging to the limit state line decreases which makes the number of callslarger
As far as the present calculation is concerned the better combinationof Ns and the stopping criterion seems to be around Ns = 20sim50 andNf ge 020sim050 Nt However the number of performance function callsalso increases considerably for these cases
(2) An example on a parabolic performance function
The second example employs a parabolic performance function Z =(R minus 110)2 minus (S minus 6) The resistance R follows a normal distribution
Monte Carlo simulation 187
10
8
6
k = 1
k = 3
k = 4
k = 5
Z = 0
k = 2
Zsub_1
Zsub_2
Zsub_3
Zsub_4S
4
24 6 8 10
R12 14
Figure 49 A calculation example of subset MCMC method when Z = (Rminus110)2 minus (Sminus6)
N(8507072) whereas the external force S follows a normal distributionN(5007072) The true value of the failure probability of this problemwhich is obtained by one million runs of OMCS is Pf = 320E minus 04
Figure 49 exhibits a calculation example of this reliability analysis In thiscalculation Nt = 100 and four subsets were formed before the calculationwas ceased It is observed that more samples are generated closer to thelimit state line as the calculation proceeds which is the same as for the firstexample
All the cases that have been calculated in the previous example are repeatedin this example as well where Nt = 100
Table 44 exhibits the same information as in the previous example Themean value is practically 10 in most of the cases which suggests there islittle bias in the evaluation On the other hand COV tends to be smaller forlarger Ns
Table 45 also provides the same information as the previous examplesThe number of function calls increases as Ns is set larger It is difficult toidentify the optimum Ns and the stopping criterion
44 Final remarks
Some of the recent developments in Monte Carlo simulation (MCS) tech-niques namely low-discrepancy sequences (LDS) and the subset MCMC
188 Yusuke Honjo
Table 44 Accuracy of Pf estimation for Z = (R minus 110)2 minus (S minus 6)
Cutoff criterion Ns = 2 Ns = 5 Ns = 10 Ns = 20 Ns = 50
➀Nf 002N Mean 1097 1101 1078 1069 0997Cov 1027 0984 0839 0847 0755
➁Nf 005N Mean 1090 1092 1071 1056 1002Cov 0975 0924 0865 0844 0622
➂Nf 010N Mean 1135 1083 1055 1051 1024Cov 1056 0888 0857 0737 0520
➃Nf 020N Mean 1172 1077 1089 1054 1013Cov 1129 0706 0930 0777 0387
➄Nf 050N Mean 1309 1149 1160 1100 0997Cov 1217 0862 1015 0786 0366
Table 45 Number of performance function calls for Z = (R minus 110)2 minus (S minus 6) example
Cutoff criterion Ns = 2 Ns = 5 Ns = 10 Ns = 20 Ns = 50
➀Nf 002N Mean 2173 2501 2825 3416 5515sd 709 735 789 830 1254
➁Nf 005N Mean 2343 2703 3091 3707 6388sd 519 628 794 678 1113
➂Nf 010N Mean 2457 2866 3235 3999 7124sd 361 775 696 656 926
➃Nf 020N Mean 2627 2997 3475 4308 7727sd 621 848 732 908 880
➄Nf 050N Mean 2679 3187 3703 4673 8427sd 392 816 716 851 923
method that are considered to be useful in structural reliability problemsare highlighted in this chapter
There are many other topics that need to be introduced for MCS in struc-tural reliability analysis These include generation of a three-dimensionalrandom field MCS technique for code calibration and partial factorsdetermination
It is the authorrsquos observation that classic reliability analysis tools such asFORM are now facing difficulties when applied to recent design methodswhere the calculations become more and more complex and sophisticatedOn the other hand our computational capabilities have progressed tremen-dously compared to the time when these classic methods were developedWe need to develop more flexible and more user-friendly reliability analysistools and MCMS is one of the strong candidate methods to be used forthis purpose More development however is still necessary to fulfill theserequirements
Monte Carlo simulation 189
Appendix Mathematical proof of MCMC byMetropolisndashHastings algorithm
(1) Generation of random numbers by MCMC basedon MetropolisndashHastings algorithm
Let us generate x that follows PDF π (x) Note that x used in this sectioncan be a random variable or a random vector that follows the specifiedPDF π (x)
Let us consider a Markov chain process x(t) rarr x(t+1) The procedure togenerate this chain is described as follows (Gilks et al 1996)
Step 1 Fix an appropriate proposal density which can generally bedescribed as
q(
xprime∣∣x(t))
(431)
Step 2 Calculate the acceptance probability which is defined by the equa-tion below
α(xprimex(t)
)= min
1
q(x(t)∣∣xprime)π (xprime)
q(xprime|x(t)
)π(x(t))
(432)
α is the acceptance probability in (0010)Step 3 x(t) is advanced to x(t+1) by the condition below based on the
acceptance probability calculated in the previous step
x(t+1) =
xprime with probability α
x(t) with probability 1 minusα
x(t+1) generated by this procedure surely depends on the previous stepvalue x(t)
(2) Proof of the MetropolisndashHastings algorithm
It is proved in this subsection that x generated following the proceduredescribed above actually follows PDF π (x) The proof is given in the threesteps
1 First the transition density which is to show the probability to moveform x to xprime in this Markov chain is defined
p(xprime∣∣x)= q
(xprime∣∣x)α (xprimex
)(433)
190 Yusuke Honjo
2 It is shown next that the transition density used in the MetropolisndashHastings algorithm satisfies so-called detailed balance This implies thatp(xprime | x) is reversible with respect to time
p(xprime∣∣x)π (x) = p
(x|xprime)π (xprime) (434)
3 Finally it is proved that generated sequences of numbers by this Markovchain follow a stationary distribution π (x) which is expressed by thefollowing relationshipint
p(xprime∣∣x)π (x)dx = π
(xprime) (435)
It is proved here that the transition density used in MetropolisndashHastingsalgorism satisfies the detailed balance
It is obvious that Equation (434) is satisfied when x = xprimeWhen x = xprime
p(xprime∣∣x)π (x) = q
(xprime∣∣x)α (xprimex
)π (x)
= q(xprime∣∣x)min
1
q (x|xprime)π (xprime)q (xprime|x)π (x)
π (x)
= minq(xprime∣∣x)π (x) q
(x|xprime)π (xprime) (436)
Therefore p (xprime|x)π (x) is symmetric with respect to x and xprime Thus
p(xprime∣∣x)π (x) = q
(xprime∣∣x)π (xprime) (437)
From this relationship the detailed balance of Equation (434) is provedFinally the stationarity of generated x and xprime are proved based on the
detailed balance of Equation (434) This can be proved if the results beloware obtained
x sim π (x)
xprime sim π(xprime) (438)
This is equivalent to prove the stationarity of the Markov chainintp(xprime | x
)π (x)dx = π
(xprime) (439)
which is obtained as followsintp(xprime∣∣x)π (x)dx =
intp(x|xprime)π (xprime)dx
= π(xprime)int p
(x|xprime)dx
= π(xprime)
Monte Carlo simulation 191
Thus a sequence of numbers generated by the MetropolisndashHastings algo-rithm follow PDF π (x)
References
Ang A and Tang W H (1984) Probability Concepts in Engineering Planning andDesign Vol II Decision Risk and Reliability John Wiley amp Sons
Au S-K and Beck J L (2003) Subset Simulation and its Appplication to SeismicRisk Based Dynamic Analysis Journal of Engineering Mechanics ASCE 129 (8)901ndash17
Broding W C Diederich F W and Parker P S (1964) Structural optimizationand design based on a reliability design criterion J Spacecraft 1(1) 56ndash61
Gilks W R Richardson S and Spiegelhalter D J (1996) Markov Chain MonteCarlo in Practice Chapman amp HallCRC London
Koyanagi Y (1989) Random Numbers Numerical Calculation Softwares byFORTRAN 77 Eds T Watabe M Natori and T Oguni Maruzen Co Tokyopp 313ndash22 (in Japanese)
Melcher R E (1999) Structure Reliability Analysis and Prediction 2nd ed JohnWiley amp Sons Ltd England
Reiss R D and Thomas M (1997) Statistical Analysis of Extreme ValuesBirkhauser Basel
Rubinstein R Y (1981) Simulation and the Monte Carlo Method John Wiley ampSons Ltd New York
Tezuka S (1995) Uniform Random Numbers Theory and Practice Kluwer BostonTezuka S (2003) Mathematics in Low Discrepancy Sequences In Computational
Statistics I Iwanami Publisher Tokyo pp 65ndash120 (in Japanese)Tsuda T (1995) Monte Carlo Methods and Simulation 3rd ed Baifukan Tokyo
(in Japanese)Yoshida I and Sato T (2005a) A Fragility Analysis by Low Discrepancy Sequences
Journal of Structural Engineering 51A 351ndash6 (in Japanese)Yoshida I and Sato T (2005b) Effective Estimation Method of Low Failure Prob-
ability by using Markov Chain Monte Carlo Journal of Structural EngineeringNo794I-72m 43ndash53 (in Japanese)
Chapter 5
Practical application ofreliability-based designin decision-making
Robert B Gilbert Shadi S Najjar Young-Jae Choi andSamuel J Gambino
51 Introduction
A significant advantage of reliability-based design (RBD) approaches is tofacilitate and improve decision-making While the attention in RBD hasgenerally been focused on the format and calibration of design-checkingequations the basic premise of a reliability-based design is that the tar-get of the design should achieve an acceptable level of reliability With thispremise in mind designers and stakeholders can utilize a range of possiblealternatives in order to achieve the desired reliability The objective of thischapter is to present practical methods and information that can be used tothis end
This chapter begins with a discussion of decision trees and decisioncriteria Next the methods for conducting reliability analyses and the resultsfrom reliability analyses are considered in the context of decision-makingFinally the calibration of mathematical models with data is addressed forthe purposes of reliability analyses and ultimately decision-making
The emphasis in presenting the material in this chapter is on illustrat-ing basic concepts with practical examples These examples are all basedon actual applications where the concepts were used in solving real-worldproblems to achieve more reliable andor more cost-effective designs
In order to provide consistency throughout the chapter the examples areintentionally narrowed in scope to the design of foundations for offshorestructures For context two types of structures are considered herein fixedjacket platforms and floating production systems Fixed jackets consist ofa steel frame with legs supported at the sea floor with driven pipe piles(Figure 51) Fixed jackets are used in water depths up to several hundredmeters with driven piles that are approximately 1 m in diameter and 100 mlong Floating production systems consist of a steel hull moored to the seafloor with 8ndash16 mooring lines anchored by steel caissons that are jackedinto the soil using under-pressure and are called suction caissons Floatingsystems are used in water depths of 1000 m or more with suction caissonsthat are approximately 5 m in diameter and 30 m long Both structures
Practical application of RBD 193
Ocean Surface
Mudline
Mudline
Driven SteelPipe Pile
Foundation
SuctionCaisson
Foundation
Jacket
MooringLine
TopsidesFacilities
TopsidesFacilities
Hull
FixedJacket
Platform
FloatingProduction
System
Figure 51 Schematic of typical offshore structures
provide a platform for the production and processing of oil and gas offshoreThe structures typically cost between $100 million for jackets up to morethan $1 billion for floating systems and the foundation can cost anywherefrom 5 to 50 of the cost of the structure
52 Decision trees
Decision trees are a useful tool for structuring and making decisions(Benjamin and Cornell 1970 Ang and Tang 1984) They organize theinformation that is needed in making a decision provide for a repeatableand consistent process allow for an explicit consideration of uncertainty andrisk and facilitate communication both in eliciting input from stakeholdersto the decision and in presenting and explaining the basis for a decision
There are two basic components in a decision tree alternatives andoutcomes A basic decision tree in reliability-based design is shown inFigure 52 to illustrate these components There are two alternatives eg twodifferent pile lengths For each alternative there are two possible outcomesthe design functions adequately eg the settlements in the foundation donot distress the structure or it functions inadequately eg the settlementsare excessive In a decision tree the alternatives are represented by limbsthat are joined by a decision node while the outcomes are represented by
194 R B Gilbert et al
limbs that are joined by an outcome node (Figure 52) The sequence of thelimbs from left to right indicates the order of events in the decision processIn the example in Figure 52 the decision about the design alternative willbe made before knowing whether or not the foundation will function ade-quately Finally the outcomes are associated with consequences which areexpressed as costs in Figure 52
The following examples taken from actual projects illustrate a variety ofapplications for decision trees in RBD
Design B
Design A
FunctionsInadequately
Outcome Node
Decision Node
Inadequately
Adequately
Adequately
Functions
Functions
Functions
Consequences
Cost forConstruction
and Corrective Action
Cost forConstruction
and CorrectiveAction
Cost forConstruction
Cost forConstruction
Figure 52 Basic decision tree
Box 51 Example I ndash Decision tree for design of new facility
The combination of load and resistance factors to be used for the design ofa caisson foundation for a floating offshore structure was called into ques-tion due to the unique nature of the facility the loads on the foundation weredominated by the sustained buoyant load of the structure versus transientenvironmental loads and the capacity of the foundation was dominated byits weight versus the shear resistance of the soil The design-build contractorproposed to use a relatively small safety margin defined as the ratio of theload factor divided by the resistance factor due to the relatively small uncer-tainty in both the load and the capacity The owner wanted to consider thisproposed safety margin together with a typical value for more conventionaloffshore structures which would be higher and an intermediate value Thetwo considerations in this decision were the cost of the foundation whichwas primarily affected by needing to use larger vessels for installation as theweight and size of the caisson increased and the cost associated with a pull-out failure of the caisson if its capacity was not adequate The decision treein Figure 53 shows the structure of decision alternatives and outcomes thatwere considered
Practical application of RBD 195
Box 51 contrsquod
Relative Costs
Use Contractorrsquos ProposedSafety Margin (Small)
No Caisson Failure
Caisson Failure
Use Intermediate SafetyMargin (Medium)
No Caisson Failure
Caisson Failure
No Caisson FailureUse Conventional SafetyMargin (High)
None(Base Case)
Facility Damage
Larger Foundation
Larger Foundation + Facility Damage
Largest Foundation
Largest Foundation +Facility Damage Caisson Failure
Figure 53 Decision tree for Example I ndash Design of new facility
53 Decision criteria
Once the potential alternatives have been established through use of deci-sion trees or by some other means the criterion for making a decision amongpotential alternatives should logically depend on the consequences for possi-ble outcomes and their associated probabilities of happening A rational anddefensible criterion to compare alternatives when the outcomes are uncer-tain is the value of the consequence that is expected if that alternative isselected (Benjamin and Cornell 1970 Ang and Tang 1984) The expectedconsequence E(C) is obtained mathematically as follows
E (C) =sumallci
ciPC
(ci
)or
intallc
cfC (c)dc (51)
where C is a random variable representing the consequence and PC
(ci
)or
fC (c) is the probability distribution for c in discrete or continuous formrespectively
The consequences of a decision outcome generally include a variety ofattributes such as monetary value human health and safety environmentalimpact social and cultural impact and public perception Within thecontext of expected consequence (Equation (51)) a simple approach toaccommodate multiple attributes is to assign a single value such as a
Box 52 Example II ndash Decision tree for site investigation in mature field
The need for drilling geotechnical borings in order to design new structures ina mature offshore field where considerable geologic and geotechnical infor-mation was already available was questioned The trade-off for not drillinga site-specific boring was that an increased level of conservatism would beneeded in design to account for the additional uncertainty The decision tree inFigure 54 shows the structure of decision alternatives and outcomes that wereconsidered In this case the outcome of drilling the boring ie the geotech-nical design properties derived from the boring data is not known beforedrilling the boring In addition there is a range of possible design propertiesthat could be obtained from the boring and therefore a range of possible piledesigns that would be needed for a given set of geotechnical properties Thesemi-circles in Figure 54 indicate that there is a continuous versus a discreteset of possibilities at the outcome and decision nodes
Relative Costs
No Pile Failure New BoringDrill New Site-SpecificBoring
GeotechnicalProperties
DevelopConventional
Design
Pile Failure New Boring + Facility Damage
No Pile Failure Added Conservatism in Design
AddConservatism
to Design
Pile Failure
Use AvailableGeotechnicalInformation Added Conservatism in Design
+ Facility Damage
Figure 54 Decision tree for Example II ndash Site investigation in mature field
Box 53 Example III ndash Decision tree for pile monitoring in frontier field
The pile foundation for an offshore structure in a frontier area was designedbased on a preliminary analysis of the site investigation data The geotechnicalproperties of the site were treated in design as if the soil conditions were similarto other offshore areas where the experience base was large The steel wasthen ordered Subsequently a more detailed analysis of the geotechnical prop-erties showed that the soil conditions were rather unusual calling into question
Box 53 contrsquod
how the properties should be used in design and leading to relatively largeuncertainty in the estimated pile capacity The owner was faced with a seriesof decisions First should they stay with the original pile design or change itgiven the uncertainty in the pile capacity considering that changing the piledesign after the steel had been ordered would substantially impact the costand schedule of the project Second if they decided to stay with the originalpile design should they monitor the installation to confirm the capacity wasacceptable considering that this approach required a flexible contract wherethe pile design may need to be updated after the pile is installed The decisiontree in Figure 55 shows the structure of decision alternatives and outcomesthat were considered all of the possibilities are shown for completeness eventhough some such as keeping the design in spite of evidence that the capacityis unacceptable could intuitively be ruled out
Relative CostsNone (Base Case)No Pile Failure
Make Design More Conservative
Keep OriginalDesign
AcceptableCapacity
Do Not Monitor Pile Driving
MonitorPile Driving
UnacceptableCapacity
Facility Damage Pile Failure Keep
DesignNo Pile Failure Monitoring + Contract Flexibility
Monitoring + Contract Flexibility + Facility Damage
Pile Failure
Monitoring + Contract Flexibility+ Design Change
No Pile Failure
ChangeDesign Monitoring + Contract Flexibility +
Design Change + Facility DamagePile Failure No Pile Failure Monitoring + Contract FlexibilityKeep
Design Monitoring + ContractFlexibility + Facility Damage
Pile Failure
Monitoring + Contract Flexibility+ Design Change
No Pile FailureChangeDesign Monitoring + Contract Flexibility +
Design Change + Facility DamageDesign Change
Pile Failure No Pile Failure
Pile Failure Design Change + Facility Damage
Figure 55 Decision tree for Example III ndash Pile monitoring in frontier field
Box 54 Decision criterion for Example I ndash Design of new facility
The decision tree for the design example described in Box 51 (Figure 53)is completed in Figure 56 with information for the values and probabilitiesof the consequences The probabilities and costs for caisson failure corre-spond to the event that the caisson will pull out of the mudline at some timeduring the 20-year design life for the facility The costs are reported as negative
Continued
198 R B Gilbert et al
Box 54 contrsquod
values in order to denote monetary loss It is assumed that a caisson failurewill result in a loss of $500 million The cost of redesigning and installingthe caisson to achieve an intermediate safety margin is $1 million while thecost of redesigning and installing the caisson to achieve a high safety margin is$5 million The cost of the caisson increases significantly if the highest safetymargin is used because a more costly installation vessel will be required Inthis example the maximum expected consequence (or the minimum expectedcost) is obtained for the intermediate safety margin Note that the contributionof the expected cost of failure to the total expected cost becomes relativelyinsignificant for small probabilities of failure
Relative Costs (Millions $)
Use Contractorrsquos ProposedSafety Margin (Small)
No Caisson Failure 0(099)
(001)E(C) = 0(099) ndash 500(001) = minus$5 Millionminus500Caisson Failure
minus1No Caisson Failure(0999)
Use Intermediate SafetyMargin (Medium)
(0001)E(C) = minus1 ndash 500(0001) = minus$15 Million Caisson Failure minus(1 + 500)
No Caisson Failure(09999)
minus5Use Conventional SafetyMargin (High)
(00001)E(C) = minus5 ndash 500(00001) = minus$505 MillionCaisson Failure minus(5 + 500)
Figure 56 Completed decision tree for Example I ndash Design of new facility
monetary value that implicitly considers all of the possible attributes for anoutcome
A second approach to consider factors that cannot necessarily be directlyrelated to monetary value such as human fatalities is to establish tolerableprobabilities of occurrence for an event as a function of the consequencesThese tolerable probabilities implicitly account for all of the attributesassociated with the consequences A common method of expressing thisinformation is on a risk tolerance chart which depicts the annual probabil-ity of occurrence for an event versus the consequences expressed as humanfatalities These charts are sometimes referred to as FndashN charts where Fstands for probability or frequency of exceedance typically expressed on anannual basis and N stands for number of fatalities (Figure 57) These chartsestablish a level of risk expressed by an envelope or line which will betolerated by society in exchange for benefits such as economical energy
Practical application of RBD 199
1E-08
1E-07
1E-06
1E-05
1E-04
1E-03
1E-02
1E-01
1 10 100 1000
Number of Fatalities (N)
Lower Bound(Nuclear Power
Plants)
Target for OffshoreStructures
Ann
ual P
roba
bilit
y of
Inci
dent
s w
ith N
or
Mor
e F
atal
ities
Upper Bound(IndustrialFacilities)
Figure 57 Risk tolerance chart for an engineered facility
The risk is considered acceptable if the combination of probability and conse-quences falls below the envelope Excellent discussions concerning the basisfor applications of and limitations with these types of charts are provided inFischhoff et al (1981) Whitman (1984) Whipple (1985) ANCOLD (1998)Bowles (2001) USBR (2003) and Christian (2004)
The chart on Figure 57 includes an upper and lower bound for risktolerance based on a range of FndashN curves that have been published Thelower bound (most stringent criterion) on Figure 57 corresponds to a pub-lished curve estimating the risk tolerated by society for nuclear power plants(USNRC 1975) For comparison purposes government agencies in both theUnited States and Australia have established risk tolerance curves for damsthat are one to two orders of magnitude above the lower bound on Figure 57(ANCOLD 1998 and USBR 2003) The upper bound (least stringent crite-rion) on Figure 57 envelopes the risks that are being tolerated for a varietyof industrial applications such as refineries and chemical plants based onpublished data (eg Whitman 1984) As an example the target levels of riskthat are used by the oil industry for offshore structures are shown as a shadedrectangular box on Figure 57 (Bea 1991 Stahl et al 1998 Goodwin et al2000)
The level of risk that is deemed acceptable for the same consequences onFigure 57 is not a constant and depends on the type of facility If expectedconsequences are calculated for an event which results in a given number
200 R B Gilbert et al
of fatalities the points along the upper bound on Figure 57 correspondto 01 fatalities per year and those along the lower bound correspond to1 times 10minus5 fatalities per year Facilities where the consequences affect thesurrounding population versus on-site workers ie the risks are imposedinvoluntarily versus voluntarily and where there is a potential for catas-trophic consequences such as nuclear power plants and dams are generallyheld to higher standards than other engineered facilities
The risk tolerance curve for an individual facility may have a steeper slopethan one to one which is the slope of the bounding curves on Figure 57A steeper slope indicates that the tolerable expected consequence in fatalitiesper year decreases as the consequence increases Steeper curves reflect anaversion to more catastrophic events
Expressing tolerable risk in terms of costs is helpful for facilities wherefatalities are not necessarily possible due to a failure The Economist (2004)puts the relationship between costs and fatalities at approximately in orderof magnitude terms $10000000 US per fatality For example a typicaltolerable risk for an offshore structure is 0001 to 01 fatalities per year(shaded box Figure 57) which corresponds approximately to a tolerableexpected consequence of failure between $10000 and $1000000 per yearA reasonable starting point for a target value in a preliminary evaluation ofa facility is $100000 per year
The effect of time is not shown explicitly in a typical risk tolerance curvesince the probability of failure is expressed on an annual basis Howevertime can be included if the annual probabilities on Figure 57 are assumed toapply for failure events that occur randomly with time such as explosionshurricanes and earthquakes Based on the applications for which these curveswere developed this assumption is reasonable In this way the tolerabilitycurve can be expressed in terms of probability of failure in the lifetime ofthe facility meaning that failure events that may or may not occur randomlywith time can be considered on a consistent basis
Box 55 Tolerable risk for Example I ndash Design of new facility
Revisiting the example for the design of a floating system in Boxes 51and 54 (Figures 53 and 56) the event of ldquocaisson failurerdquo does not leadto fatalities The primary consequences for the failure of this very large andexpensive facility are the economic damage to the operator and the associatednegative perception from the public In addition the event of ldquocaissonfailurerdquo is not an event that occurs randomly with time over the 20-yearlifetime of the facility The maximum load on the foundation occurs whena maintenance operation is performed which happens once every five yearsof operation The probability of ldquocaisson failurerdquo in Figure 56 corresponds
Practical application of RBD 201
Box 55 contrsquod
to the probability that the caisson will be overloaded in at least one of theseoperations over its 20-year life The information on Figure 57 has beenre-plotted on Figure 58 for this example by assuming that incidents occurrandomly according to a Poisson process eg an incident with an annualprobability of 0001 per year on Figure 57 corresponds to probability ofoccurrence of 1 minus eminus(0001year)(20 years) = 002 in 20 years Figure 58 showsthat an acceptable probability of failure in the 20-year lifetime for this facilityis between 0002 and 002 Therefore the first design alternative on Figure 56(use contractorrsquos proposed factor of safety (small) with a probability of caissonfailure equal to 001) would provide a level of risk that is marginal concerningwhat would generally be tolerated
1E-07
1E-06
1E-05
1E-04
1E-03
1E-02
1E-01
1E+00
10 100 1000 10000
Cost of Failure C ($ Millions)
Lower Bound(Nuclear Power
Plants)
Upper Bound(IndustrialFacilities)
Target for OffshoreStructures
Pro
babi
lity
of In
cide
nt w
ith C
or
Mor
e D
amag
e in
20
Yea
rs
Figure 58 Risk tolerance chart for Example I ndash Design of floating system
A third approach to consider factors in decision making that cannot nec-essarily be directly related to monetary value is to use multi-attribute utilitytheory (Kenney and Raiffa 1976) This approach provides a rational andsystematic means of combining consequences expressed in different mea-sures to be combined together into a single scale of measure a utility valueWhile flexible and general implementation of multi-attribute utility theoryis cumbersome and not commonly used in engineering practice
202 R B Gilbert et al
54 Reliability analysis
Reliability analyses are an integral part of the decision-making processbecause they establish the probabilities for outcomes in the decision trees(eg the probability of caisson failure in Figure 56) The following sectionsillustrate practical applications for relating reliability analyses to decisionmaking
541 Simplified first-order analyses
The probability of failure for a design can generally be expressed as the prob-ability that a load exceeds a capacity A useful starting point for a reliabilityanalysis is to assume that the load on the foundation and the capacity ofthe foundation are independent random variables with lognormal distribu-tions In this way the probability of failure can be analytically related tothe first two moments of the probability distributions for the load and thecapacity This simplified first-order analysis can be expressed in a convenientmathematical form as follows (eg Wu et al 1989)
P(Load gt Capacity
)sim= Φ
⎛⎜⎝minus ln
(FSmedian
)radicδ2
load + δ2capacity
⎞⎟⎠ (52)
where P(Load gt Capacity) is the probability that the load exceeds the capac-ity in the design life which is also referred to as the lifetime probability offailure FSmedian is the median factor of safety which is defined as the ratioof the median capacity to the median load δ is the coefficient of variation(cov) which is defined as the standard deviation divided by the mean valuefor that variable and Φ () is the cumulative distribution function for a stan-dard normal variable The median factor of safety in Equation (52) can berelated to the factor of safety used in design as follows
FSmedian = FSdesign times
(capacitymedian
capacitydesign
)(
loadmedian
loaddesign
) (53)
where the subscript ldquodesignrdquo indicates the value used to design thefoundation The ratios of the median to design values represent biasesbetween the median value (or the most likely value since the mode equals themedian for a lognormal distribution) in the design life and the value that isused in the design check with the factor of safety or with load and resistancefactors The coefficients of variation in Equation (52) represent uncertainty
Practical application of RBD 203
in the load and the capacity The denominator in Equation (52) is referredto as the total coefficient of variation
δtotal =radicδ2
load + δ2capacity (54)
An exact solution for Equation (52) is obtained if the denominator iethe total cov in Equation (54) is replaced by the following expressionradic
ln(1 + δ2
load
)+ ln(1 + δ2
capacity
) the approximation in Equation (52) is
reasonable for values of the individual coefficients of variation that are lessthan about 03
The relationship between the probability of failure and the median factorof safety and total cov is shown in Figure 59 An increase in the medianfactor of safety and a decrease in the total cov both reduce the probabilityof failure
542 Physical bounds
In many practical applications of reliability analyses there are physicalbounds on the maximum load that can be applied to the foundation orthe minimum capacity that will be available in response to the load It is
0
01
02
03
04
05
06
07
08
1 2 3 4 5 6 7 8 9 10
Tot
al c
ov
Median FS
0001
000001
00001
001Failure Probability = 01
0000001
FixedJackets
Floating Systems
FixedJackets
Floating Systems
Figure 59 Graphical solutions for simplified first-order reliability analysis
204 R B Gilbert et al
Box 56 Reliability benchmarks for Example I ndash Design of new facility
In evaluating the appropriate level of conservatism for the design of a newfloating system as described in Box 51 (Figure 53) it is important to con-sider the reliability levels achieved for similar types of structures Fixed steeljacket structures have been used for more than 50 years to produce oil and gasoffshore in relatively shallow water depths up to about 200 m In recent yearsfloating systems have been used in water depths ranging from 1000 to 3000 mTo a large extent the foundation design methods developed for driven pilesin fixed structures have been applied directly to the suction caissons that areused for floating systems
For a pile in a fixed jacket with a design life of 20 years the median factorof safety is typically between three and five and the total cov is typicallybetween 05 and 07 (Tang and Gilbert 1993) Hence the resulting probabil-ity of failure in the lifetime is on the order of 001 based on Figure 59 Notethat the event of foundation failure ie axial overload of a single pile in thefoundation does not necessarily lead to collapse of a jacket failure probabil-ities for the foundation system are ten to 100 times smaller than those for asingle pile (Tang and Gilbert 1993)
For comparison typical values were determined for the median factor ofsafety and the total cov for a suction caisson foundation in a floating pro-duction system with a 20-year design life (Gilbert et al 2005a) The medianfactor of safety ranges from three to eight The median factor of safety tendsto be higher for the floating versus fixed systems for two reasons First a newsource of conservatism was introduced for floating systems in that the foun-dations are checked for a design case where the structure is damaged (ieone line is removed from the mooring system) Second the factors of safetyfor suction caissons were generally increased above those for driven piles dueto the relatively small experience base with suction caissons In addition to ahigher median factor of safety the total cov value for foundations in floatingsystems tends to be smaller with values between 03 and 05 This decreasecompared to fixed jackets reflects that there is generally less uncertainty inthe load applied to a mooring system foundation compared to that applied toa jacket foundation The resulting probabilities of foundation failure tend tobe smaller for floating systems compared to fixed jackets by several orders ofmagnitude (Figure 59) From the perspective of decision making this simpli-fied reliability analysis highlights a potential lack of consistency in how thedesign code is applied to offshore foundations and provides insight into whysuch an inconsistency exists and how it might be addressed in practice
important to consider these physical bounds in decision making applicationbecause they can have a significant effect on the reliability
One implication of the significant role that a lower-bound capacity canhave on reliability is that information about the estimated lower-boundcapacity should possibly be included in design-checking equations for
Box 57 Lower-bound capacity for Example I ndash Design of new facility
In evaluating the appropriate level of conservatism for the design of a newfloating system as described in Box 51 (Figure 53) it is important to con-sider the range of possible values for the capacity of the caisson foundationsA reasonable estimate of a lower-bound on the capacity for a suction caissonin normally consolidated clay can be obtained using the remolded strengthof the clay to calculate side friction and end bearing Based on an analysisof load-test data for suction caissons Najjar (2005) found strong statisticalevidence for the existence of this physical lower bound and that the ratio oflower-bound capacities to measured capacities ranged from 025 to 10 withan average value of 06
The effect of a lower-bound capacity on the reliability of a foundationfor a floating production system is shown in Figure 510 The structure isanchored with suction caisson foundations in a water depth of 2000 m detailsof the analysis are provided in Gilbert et al (2005a) Design informationfor the caissons is as follows loadmedianloaddesign = 07 δload = 014capacitymediancapacitydesign = 13 and δcapacity = 03 The lognormaldistribution for capacity is assumed to be truncated at the lower-bound valueusing a mixed lognormal distribution where the probability that the capac-ity is equal to the lower bound is set equal to the probability that the capacity is
100E-07
100E-06
100E-05
100E-04
100E-03
100E-02
100E-01
100E+00
0 02 04 06 08
Ratio of Lower-Bound to Median Capacity
Pro
babi
lity
of F
ailu
re in
Life
time
FSdesign = 15
FSdesign = 20
FSdesign = 25
Figure 510 Effect of lower-bound capacity on foundation reliability for Example I ndashDesign of floating system
Continued
206 R B Gilbert et al
Box 57 contrsquod
less than or equal to the lower bound for the non-truncated distribution Theprobability of failure is calculated through numerical integration
The results on Figure 510 show the significant role that a lower-boundcapacity can have on the reliability For a lower-bound capacity that is 06times the median capacity the probability of failure is more than 1000 timessmaller with the lower-bound than without it (when the ratio of the lower-bound to median capacity is zero) for a design factor of safety of 15
RBD codes For example an alternative code format would be to have twodesign-checking equations
φcapacitycapacitydesign ge γloadloaddesign
or
φcapacitylower boundcapacitylower bound ge γloadloaddesign
(55)
where φcapacity and γload are the conventional resistance and load factorsrespectively and φcapacitylower bound
is an added resistance factor that is appliedto the lower-bound capacity Provided that one of the two equations is satis-fied the design would provide an acceptable level of reliability Gilbert et al(2005b) show that a conservative value of 075 for φcapacitylower bound
wouldcover a variety of typical conditions in foundation design
Box 58 Lower-bound capacity for Example III ndash Pile monitoring in frontier field
The example described in Box 53 (Figure 55) illustrates a practical applica-tion where a lower-bound value for the foundation capacity provides valuableinformation for decision-making Due to the relatively large uncertainty inthe pile capacity at this frontier location the reliability for this foundationwas marginal using the conventional design check even though a relativelylarge value had been used for the design factor of safety FSdesign = 225 Theprobability of foundation failure in the design life of 20 years was slightlygreater than the target value of 1 times 10minus3 therefore the alternative of usingthe existing design without additional information (the top design alternativeon Figure 55) was not acceptable However the alternative of supplementingthe existing design with information from pile installation monitoring wouldonly be economical if the probability that the design would need to be changedafter installation was small Otherwise the preferred alternative for the ownerwould be to modify the design before installation
Box 58 contrsquod
In general pile monitoring information for piles driven into normal toslightly overconsolidated marine clays cannot easily be related to the ultimatepile capacity due to the effects of set-up following installation The capacitymeasured at the time of installation may only be 20ndash30 of the capacity afterset-up However the pile capacity during installation does arguably providea lower-bound on the ultimate pile capacity In order to establish the proba-bility of needing to change the design after installation the reliability of thefoundation was related to the estimated value for the lower-bound capacitybased on pile driving as shown on Figure 511 In this analysis the pile capac-ity estimated at the time of driving was considered to be uncertain it wasmodeled as a normally distributed random variable with a mean equal to theestimated value and a coefficient of variation of 02 to account for errors inthe estimated value The probability of foundation failure in the design lifewas then calculated by integrating the lower-bound value over the range ofpossible values for a given estimate The result is shown in Figure 511 Notethat even though there is uncertainty in the lower-bound value it still can havea large effect on the reliability
The expected value of driving resistance to be encountered and mea-sured during installation is indicated by the arrow labeled ldquoinstallationrdquo in
1E-05
1E-04
1E-03
1E-020 01 02 03 04 05 06
Estimated Lower Bound of Pile Capacity based on DrivingResistance
(Expressed as Ratio to Median Capacity)
Installation
Pro
babi
lity
of F
ound
atio
n F
ailu
re in
Des
ign
Life
Re-Tap
Target
Figure 511 Reliability versus lower-bound capacity for Example III ndash Pile monitoringin frontier field
Continued
208 R B Gilbert et al
Box 58 contrsquod
Figure 511 While it was expected that the installation information would justbarely provide for an acceptable level of reliability (Figure 511) the conse-quence of having to change the design at that point was very costly In orderto provide greater assurance the owner decided to conduct a re-tap analysis 5days after driving The arrow labeled ldquore-taprdquo in Figure 511 shows that thisinformation was expected to provide the owner with significant confidencethat the pile design would be acceptable The benefits of conducting the re-tapanalysis were considered to justify the added cost
543 Systems versus components
Nearly all design codes whether reliability-based or not treat foundationdesign on the basis of individual components However from a decision-making perspective it is insightful to consider how the reliability of anindividual foundation component is related to the reliability of the overallfoundation and structural system
Box 59 System reliability analysis for Example I ndash Design of new facility
The mooring system for a floating offshore facility such as that describedin Box 51 (Figure 53) includes multiple lines and foundation anchors Theresults from a component reliability analysis are shown in Figure 512 fora facility moored in three different water depths (Choi et al 2006) Theseresults correspond to the most heavily loaded line in the system where theprimary source of loading is from hurricanes Each mooring line consists ofsegments of steel chain and wire rope or polyester The points labeled ldquorope ampchainrdquo in Figure 512 correspond to a failure anywhere within the segmentsof the mooring line the points labeled ldquoanchorrdquo correspond to a failure at thefoundation and the points labeled ldquototalrdquo correspond to a failure anywherewithin the line and foundation The probability that the foundation fails isabout three orders of magnitude smaller than that for the line
The reliability of the system of lines was also related to that for the mostheavily loaded line (Choi et al 2006) In this analysis the event that thesystem fails is related to that for an individual line in two ways First the loadswill be re-distributed to the remaining lines when a single line fails Secondthe occurrence of a single line failure indicates information about the loads onthe system if a line has failed it is more likely that the system is being exposedto a severe hurricane (although it is still possible that the loads are relativelysmall and the line capacity for that particular line was small) The results forthis analysis are shown in Figure 513 where the probability that the mooringsystem fails is expressed as a conditional probability for the event that a
Box 59 contrsquod
single line in the system fails during a hurricane For the mooring systems in2000 and 3000 m water depths the redundancy is significant in that there isless than a 10 chance the system will fail even if the most heavily loaded linefails during a hurricane Also the mooring system in the 1000 m water depthhas substantially less redundancy than those in deeper water
10E-08
10E-07
10E-06
10E-05
10E-04
10E-03
10E-02
Water Depth (m)
1000 2000 3000
Rope ampChain
Anchor
Total
Pro
babi
lity
of F
ailu
re in
Des
ign
Life
Figure 512 Comparison of component reliabilities for most heavily loaded line inoffshore mooring system
70
Pro
babi
lity
of S
yste
m F
ailu
regi
ven
Fai
lure
of S
ingl
e Li
ne
0
10
20
50
60
40
30
1000 2000 3000Water Depth (m)
Figure 513 System redundancy for offshore mooring system
Continued
210 R B Gilbert et al
Box 59 contrsquod
The results in Figures 512 and 513 raise a series of questions forconsideration by the writers of design codes
(1) Is it preferable to have a system fail in the line or at the foundation Thebenefit of a failure in the foundation is that the weight of the foundationattached to the line still provides a restoring force to the entire mooringsystem even after it has pulled out from the seafloor The cost of failurein the foundation is that the foundation may cause collateral damage toother facilities if it is dragged across the seafloor
(2) Would the design be more effective if the reliability of the foundationwere brought closer to that of the line components It is costly to designfoundations that are significantly more conservative than the line Inaddition to the cost a foundation design that is excessively conserva-tive may pose problems in constructability and installation due to itslarge size
(3) Should the reliability for individual lines and foundations be consistentfor different water depths
(4) Should the system reliability be consistent for the different waterdepths
(5) How much redundancy should be achieved in the mooring system
55 Model calibration
The results from a reliability analysis need to be realistic to be of practicalvalue in decision-making Therefore the calibration of mathematical modelswith real-world data is very important
A general framework for calibrating a model with a set of data is obtainedfrom the following assumptions
bull The variation for an individual measurement about its mean value isdescribed by a Hermite Polynomial transformation of a standard normaldistribution which can theoretically take on any possible shape as theorder of the transformation approaches infinity (Journel and Huijbregts1978 Wang 2002)
bull The relationship between an individual data point with other data pointsis described by a linear correlation
bull The shape of the Hermite Polynomial transformation is consistentbetween data points ie affine correlation as described in Journel andHuijbregts (1978)
Practical application of RBD 211
In mathematical terms the probability distribution to describe variationsbetween a single measurement yi and a model of that measurement Yi isdescribed by a mean value a standard deviation correlation coefficientsand a conditional probability distribution
microYi= gmicro
(xYi
θ1 θnmicro
)(56)
σYi= gσ
(xYi
microYiθnmicro+1 θnmicro+nσ
)(57)
ρYiYj= gρ
(xYi
xYjθnmicro+nσ+1
θnmicro+nσ+nρ
)(58)
FYi|y1yiminus1
(yi
∣∣y1 yiminus1)= gF
(xYi
microYi|y1yiminus1σYi|y1yiminus1
θnmicro+nσ+nρ+1 θnmicro+nσ+nρ+nF
)(59)
where microYiis the mean value and gmicro () is a model with nmicro model param-
eters θ1 θnmicro that relate the mean to attributes of the measurementxYi
σYiis the standard deviation and gσ () is a model with nσ model
parameters θnmicro+1 θnmicro+nσ that relate the standard deviation to attributes
and the mean value of the measurement ρYiYjis the correlation coeffi-
cient between measurements i and j and gρ () is a model with nρ modelparameters θnmicro+nσ+1 θnmicro+nσ+nρ
that relate the correlation coefficient to
the attributes for measurements i and j and FYi|y1yiminus1
(yi
∣∣y1 yiminus1)
isthe cumulative distribution function for measurement i conditioned on themeasurements y1 to yiminus1 and gF () is a model with nF model parametersθnmicro+nσ+nρ+1 θnmicro+nσ+nρ+nF
that relate the cumulative distribution func-tion to the data attributes and the conditional mean value microYi|y1yiminus1
andstandard deviation σYi|y1yiminus1
which are obtained from the mean valuesstandard deviations and correlations coefficients for measurements 1 to iand the measurements y1 to yiminus1 The nF model parameters in gF () are thecoefficients in the Hermite Polynomial transformation function where theorder of the transformation is nF + 1
gF
(xYi
microYi|y1yiminus1σYi|y1yiminus1
θnmicro+nσ+nρ+1 θnmicro+nσ+nρ+nF
)
=int
ϕ(u)leyi minusmicroYi|y1yiminus1
σYi|y1yiminus1
1radic2π
eminus 12 u2
du (510)
212 R B Gilbert et al
where
ϕ (u) =
[H1 (u) +
nF+1summ=2
ψmm Hm (u)
]
minusradic
1 +nF+1summ=2
ψ2m
m
=y minusmicroYi|y1yiminus1
σYi|y1yiminus1
(511)
in which ψm are polynomial coefficients that are expressed as modelparameters ψm = θnmicro+nσ+nρ+mminus1 and Hm (u) are Hermite polynomialsHm+1 (u) = minusuHm (u)minusmHmminus1 (u) with H0 (u) = 1 and H1 (u) = minusu Thepractical implementation of Equation (510) involves first finding all of
the intervals of u in Equation (511) where ϕ (u) le yiminusmicroYi|y1yiminus1σYi|y1yiminus1
and then
integrating the standard normal distribution over those intervals If thedata points are assumed to follow a normal distribution then the order
of the polynomial transformation is one nF = 0 ϕ (u) = u = yminusmicroYi|y1yiminus1σYi|y1yiminus1
and FYi|y1yiminus1
(yi
∣∣y1 yiminus1)= (u) where (u) is the standard normal
functionThe framework described above contains nθ = nmicro + nσ + nρ + nF model
parameters θ1 θnmicro+nσ+nρ+nF that need to be estimated or calibrated
based on available information It is important to bear in mind that thesemodel parameters have probability distributions of their own since they arenot known with certainty For a given set of measurements the calibrationof the various model parameters follows a Bayesian approach
f1nθ |y1yn
(θ1 θnθ
∣∣y1 yn
)(512)
=P(y1 yn
∣∣∣θ1 θnθ
)f1nθ
(θ1 θnθ
)P(y1 yn
)where f1nθ |y1yn
(θ1 θnθ
∣∣y1 yn
)is the calibrated or updated
probability distribution for the model parameters P(y1 yn
∣∣∣θ1 θnθ
)is
the probability of obtaining this specific set of measurements for a given set
of model parameters and f1nθ
(θ1 θnθ
)is the probability distribu-
tion for the model parameters based on any additional information that isindependent of the measurements The probability of obtaining the mea-
surements P(y1 yn
∣∣∣θ1 θnθ
) can include both point measurements
as well as interval measurements such as proof loads or censored values
Practical application of RBD 213
(Finley 2004) While approximations exist for solving Equation (512) (egGilbert 1999) some form of numerical integration is generally required
The information from model calibration can be used in decision makingin two ways First the calibrated distribution from Equation (512) providesmodels that are based on all available information at the time of making adecision uncertainty in the calibrated parameters can be included directlyin the decision since the model parameters are represented by a probabilitydistribution and not deterministic point estimates Second the effect andtherefore value of obtaining additional information before making a decision(eg Figure 54) can be assessed
Box 510 Probability distribution of capacity for Example I ndash Design of new facility
The probability distribution for axial pile capacity plays an important role inreliability-based design such as described in Box 51 because it dictates theprobability of foundation failure due to overloading A database of pile loadtests with 45 driven piles in clay was analyzed by Najjar (2005) to investi-gate the shape of the probability distribution First a conventional analysiswas conducted on the ratio of measured to predicted capacity where the dis-tribution for this ratio was assumed to be lognormal Individual data pointswere assumed to be statistically independent In terms of the framework pre-sented in Equations (56) through (511) the attribute for each data point xiis the predicted capacity the function gmicro () is microInYi
= ln(θ1xi) the function
gσ () is σlnYi=radic
ln(1 + θ2
2
) the data points are statistically independent ie
the correlation coefficients between data points are zero and the cumulativedistribution function for lnYi gF () is a normal distribution with mean microlnYiand standard deviation σlnYi
Therefore there are two model parameters tobe calibrated with the data θ1 and θ2
In calibrating the model parameters with Equation (512) the only infor-mation used was the measured data points in the load test database ief12
(θ1θ2
)is taken as a non-informative diffuse prior distribution The cal-
ibrated expected values for θ1 and θ2 from the updated probability distributionobtained with Equation (512) are 096 and 024 respectively The resultingprobability distribution is shown on Figure 514 and labeled ldquoconventionallognormalrdquo
In Boxes 57 and 58 the effect of a lower bound on the pile capacity wasshown to be significant In order to investigate the existence of a lower-boundvalue an estimate for a lower-bound capacity was established for each datapoint based on the site-specific soil properties and pile geometry Specificallyremolded undrained shear strengths were used to calculate a lower-boundcapacity for driven piles in normally consolidated to slightly overconsolidated
Continued
Box 510 contrsquod
00001
0001
001
01
1
050 151
Actual CapacityPredicted Capacity
Cum
ulat
ive
Dis
tribu
tion
Fun
ctio
n
Hermite PolynomialTransformation
ConventionalnormalLog
Lower Bound
Figure 514 Calibrated probability distributions for axial pile capacity
clays and residual drained shear strengths and at-rest conditions were used tocalculate a lower-bound capacity for driven piles in highly overconsolidatedclays (Najjar 2005) For all 45 data points the measured capacity was abovethe calculated value for the lower-bound capacity To put this result in per-spective the probability of obtaining this result if the distribution of capacityreally is a lognormal distribution (with a lower bound of zero) is essentiallyzero meaning that a conventional lognormal distribution is not plausible
The model for the probability distribution of the ratio of measured to pre-dicted capacity was therefore refined to account for a more complicated shapethan a simple lognormal distribution The modifications are the following anadditional attribute the predicted lower-bound capacity is included for eachdata point and a 7th-order Hermite Polynomial transformation is used as ageneral model for the shape of the probability distribution In terms of Equa-tions (56) through (511) the attributes for each data point are the predictedcapacity x1i and the calculated lower-bound capacity x2i the function gmicro ()is microYi
= θ1x1i the function gσ () is σYi= θ2microYi
the correlation coefficientsbetween data points were zero and the cumulative distribution function forlnYi gF () is obtained by combining Equations (511) and (512) as follows
FY(yi)=
⎧⎪⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎪⎩
0 yi lt x2i
Φ(ui)where
H1(ui)+ 7sum
m=2
θm+1m Hm
(ui)
minusradicradicradicradic1 +
7summ=2
θ2m+1m
= yi minusmicroYi
σYi
yi ge x2i
(513)
Box 510 contrsquod
Therefore there are now eight model parameters to be calibrated with thedata θ1 to θ8
The calibrated expected values for the model parameters from the updatedprobability distribution obtained with Equation (512) are θ1 = 096θ2 = 026 θ3 = minus0289 θ4 = 0049 θ5 = minus0049 θ6 = 0005 θ7 = 005 andθ8 = 0169 The resulting probability distribution is shown on Figure 514 andlabeled ldquoHermite polynomial transformationrdquo
There are two practical conclusions from this calibrated model First theleft-hand tail of the calibrated distribution with the 7th-order Hermite Poly-nomial transformation is substantially different than that for the conventionallognormal distribution (Figure 514) This difference is significant because itis the lower percentiles of the distribution for capacity say less than 10which govern the probability of failure Second truncating the conventionallognormal distribution at the calculated value for the lower-bound capacityprovides a reasonable and practical fit to the more complicated 7th-orderHermite Polynomial transformation
Box 511 Value of new soil boring for Example II ndash Site investigation in mature field
A necessary input to the decision about whether or not to drill an additional soilboring in a mature field as described in Box 52 and illustrated in Figure 54is a model describing spatial variability in pile capacity across the field Thegeology for the field in this example is relatively uniform marine clays thatare normally to slightly overconsolidated The available data are soil boringswhere the geotechnical properties were measured and a design capacity wasdetermined Two types of soil borings are available modern borings wherehigh-quality soil samples were obtained and older borings where soil sampleswere of lower quality Pile capacity for these types of piles is governed by sideshear capacity
The model for the spatial variability in design pile capacity expressed interms of the average unit side shear over a length of pile penetration is givenby Equations (514) through (517)
microYi=(θ1 + θ2x1i + θ3x2
1i
)ex2i θ4 (514)
where Yi is the average unit side shear which is the total capacity due to sideshear divided by the circumference of the pile multiplied by its length x1i isthe penetration of the pile and x2i is 0 for modern borings and 1 for older
Continued
Box 511 contrsquod
borings (the term ex2i θ4 accounts for biases due to the quality of soil samples)
σYi= eθ4ex2i θ5 (515)
where the term eθ4 accounts for a standard deviation that does not dependon pile penetration and that is non-negative and the term ex2i θ5 accounts forany effects on the spatial variability caused by the quality of the soil samplesincluding greater or smaller variability
ρYiYj=⎛⎝e
minusradic(
x3iminusx3j)2+(x4iminusx4j
)2(eθ6+x1ieθ7)⎞⎠e
minus∣∣x2iminusx2j∣∣eθ8
(516)
where x3i and x4i are the coordinates describing the horizontal position
of boring i the term eminusradic(
x3iminusx3j)2+(x4iminusx4j
)2(eθ6+x1ieθ7)
accounts for apositive correlation between nearby borings that is described by a positive cor-
relation distance of eθ6 +x1ieθ7 and the term eminus∣∣x2iminusx2j
∣∣eθ8 accounts for asmaller positive correlation between borings with different qualities of samplesthan those with the same qualities of samples and the probability distributionfor the average side friction is modeled by a normal distribution
FYi|y1yiminus1
(yi∣∣y1 yiminus1
)= Φ
(yi minusmicroYi|y1yiminus1
σYi|y1yiminus1
)(517)
Therefore there are eight parameters to be calibrated with the data θ1 to θ8The detailed results for this calibration are provided in Gambino and Gilbert(1999)
Before considering the decision at hand there are several practicalconclusions that can be obtained from the calibrated model in its own rightThe expected values for the mean and standard deviation of average unitside shear from modern borings are 98 kPa and 72 kPa respectively Thecoefficient of variation for pile capacity is then 7298 or 007 meaning thatthe magnitude of spatial variability in pile capacity across this field is relativelysmall The expected horizontal correlation distance is shown with respect topile penetration in Figure 515 it is on the order of thousands of meters andit increases with pile penetration The expected bias due to older soil boringsis equal to 091 (ex2iθ4 in Equation (514)) meaning that designs based onolder borings tend to underestimate the pile capacity that would be obtainedfrom a modern boring Also the expected effect on the spatial variability dueto older soil borings is to reduce the standard deviation by 086 (ex2iθ5 ) in
Box 511 contrsquod
0
20
40
60
80
100
120
140
160
0 1000 2000 3000 4000
Horizontal Correlation Distance (m)
Pile
Pen
etra
tion
(m)
Figure 515 Calibrated horizontal correlation distance for pile capacities betweenborings
Equation (515) meaning that the data from the older soil borings tends tomask some of the naturally occurring spatial variations that are picked upwith modern borings thereby reducing the apparent standard deviation Fur-thermore the expected correlation coefficient between an older boring and a
modern boring at the exact same location is 058 (eminus∣∣x2iminusx2j∣∣eθ8 in Equa-
tion (516) and notably less than the ideal value of 10 meaning that it is notpossible to deterministically predict the pile capacity obtained from a modernboring by using information from an older boring
The results for this calibrated model are illustrated in Figures 516a and516b for an example 4000 m times 4000 m block in this field Within this blockthere are two soil borings available a modern one and an older one at therespective locations shown in Figures 516a and 516b The three-dimensionalsurface shown in Figure 1516a denotes the expected value for the calculateddesign capacity at different locations throughout the block The actual designpile capacities at the two boring locations are both above the average for thefield (Figure 516a) Therefore the expected values for the design pile capac-ity at other locations of an offshore platform within this block are aboveaverage however as the location moves further away from the existing bor-ings the expected value tends toward the unconditional mean for the fieldapproximately 30 MN
The uncertainty in the pile capacity due to not having a site-specificmodern soil boring is expressed as a coefficient of variation on Figure 516b
Continued
Box 511 contrsquod0
500
1000
2000
2500
3500
4000
1500
3000 0
1500
2000
2500
3000
3500
4000
500
1000
26
27
28
29
30
31
32
33
34
x Coordinate (m)
Exp
ecte
d P
ile C
apac
ity (
MN
)
y Coordinate (m)
Unconditional Mean
Older BoringModern Boring
Figure 516a Expected value for estimated pile capacity versus location for a 100 mlong 1 m diameter steel pipe pile
At the location of the modern boring this cov is zero since a modern boringis available at that location and the design capacity is known However at thelocation of the older boring the cov is greater than zero since the correlationbetween older and modern borings is not perfect (Figure 516b) therefore eventhough the design capacity is known at this location the design capacity thatwould have been obtained based on a modern soil boring is not known Asthe platform location moves away from both borings the cov approachesthe unconditional value for the field approximately 0075
The final step is to put this information from the calibrated model for spatialvariability into the decision tree in Figure 54 The added uncertainty in nothaving a site-specific modern boring is included in the reliability analysis byadding uncertainty to the capacity the greater the uncertainty the lower thereliability for the same set of load and resistance factors in a reliability-baseddesign code In order to achieve the same reliability as if a modern soil boring
Box 511 contrsquod
0
500
1000
1500
2000
2500
3000
3500
4000
0500
1000
1500
2000
2500
3000
3500
40000000
0020
0040
0060
0080
x Coordinate (m) y Coordinate (m)
co
v in
Pile
Cap
acity
Older Boring
Modern Boring
Unconditional cov
Figure 516b Coefficient of variation for estimated pile capacity versus location fora 100 m long 1 m diameter steel pipe pile
were available the resistance factor needs to be smaller as expressed in thefollowing design check format
φspatial
(φcapacitycapacitydesign
)ge γloadloaddesign (517)
where φspatial is a partial resistance factor that depends on the magnitudeof spatial variability obtained from Figure 516b and the design capacityis obtained from Figure 516a Figure 517 shows the relationship betweenφspatial and the coefficient of variation due to spatial variability for thefollowing assumptions design capacity is normally distributed with a cov of03 when a site-specific boring is available load is normally distributed witha cov of 05 and the target probability of failure for the pile foundationis 6 times 10minus3 (Gilbert et al 1999) To use this chart a designer would first
Continued
Box 511 contrsquod
075
080
085
090
095
100
0 002 004 006 008 01
Coefficient of Variation due to Spatial Variability
Partial ResistanceFactor
due to SpatialVariabilityφspatial
Figure 517 Partial resistance factor versus coefficient of variation for spatialvariability
quantify the spatial variability as a cov value using the approach describedabove as shown in Figure 516b The designer would then read the corre-sponding partial resistance factor directly from Figure 517 as a functionof this cov due to spatial variability When there is not additional uncer-tainty because a modern soil boring is available at the platform location(ie the cov due to spatial variability is 0) the partial resistance factoris 10 and does not affect the design As the cov due to spatial variabilityincreases the partial resistance factor decreases and has a larger effect on thedesign
The information in Figures 516a 516b and 517 can be combined togetherto support the decision-making process For example consider a location in theblock where the cov due to spatial variability is 006 from Figure 516b Therequired partial resistance factor to account for this spatial variability is 095or 95 of the design capacity could be relied upon in this case comparedto what is expected if a site-specific boring is drilled Since pile capacity isapproximately proportional to pile length over a small change in length thisreduction in design capacity is roughly equivalent to a required increase in pilelength of about 5 Therefore the cost of drilling a site-specific boring couldbe compared to the cost associated with increasing the pile lengths by 5 inthe decision tree in Figure 54 to decide whether or not to drill an additionalsoil boring
Practical application of RBD 221
56 Summary
There can be considerable value in using the reliability-based designapproach to facilitate and improve decision-making This chapter hasdemonstrated the practical potential of this concept Several different levelsof decisions were addressed
bull project-specific decisions such as how long to make a pile foundationor how many soil borings to drill
bull design code decisions such as what resistance factors should be usedand
bull policy decisions such as what level of reliability should be achieved forcomponents and systems in different types of facilities
An emphasis has been placed on first establishing what is important in thecontext of the decisions that need to be made before developing and applyingmathematical tools An emphasis has also been placed on practical meth-ods to capture the realistic features that will influence decision-making It ishoped that the real-world examples described in this chapter motivate thereader to consider how the role of decision making can be utilized in thecontext of reliability-based design for a variety of different applications
Acknowledgments
We wish to acknowledge gratefully the following organizations for support-ing the research and consulting upon which this paper is based AmericanPetroleum Institute American Society of Civil Engineers BP ExxonMobilNational Science Foundation Offshore Technology Research Center ShellState of Texas Advanced Research and Technology Programs Unocal andUnited States Minerals Management Service In many cases these organiza-tions provided real-world problems and data as well as financial supportWe would also like to acknowledge the support of our colleagues andstudents at The University of Texas at Austin The views and opinionsexpressed herein are our own and do not reflect any of the organizationsparties or individuals with whom we have worked
References
ANCOLD (1998) Guidelines on Risk Assessment Working Group on Risk Assess-ment Australian National Committee on Large Dams Sydney New South WalesAustralia
Ang A A-S and Tang W H (1984) Probability Concepts in Engineering Plan-ning and Design Volume II ndash Decision Risk and Reliability John Wiley amp SonsNew York
222 R B Gilbert et al
Bea R G (1991) Offshore platform reliability acceptance criteria DrillingEngineering Society of Petroleum Engineers June 131ndash6
Benjamin J R and Cornell C A (1970) Probability Statistics and Decision forCivil Engineers McGraw-Hill New York
Bowles D S (2001) Evaluation and use of risk estimates in dam safetydecisionmaking In Proceedings Risk-Based Decision-Making in Water ResourcesASCE Santa Barbara California 17 pp
Choi Y J Gilbert R B Ding Y and Zhang J (2006) Reliability ofmooring systems for floating production systems Final Report for MineralsManagement Service Offshore Technology Research Center College StationTexas 90 pp
Christian J T (2004) Geotechnical engineering reliability how well do we knowwhat we are doing Journal of Geotechnical and Geoenvironmental EngineeringASCE 130 (10) 985ndash1003
Finley C A (2004) Designing and analyzing test programs with censored datafor civil engineering applications PhD dissertation The University of Texas atAustin
Fischhoff B Lichtenstein S Slovic P Derby S L and Keeney R L (1981)Acceptable Risk Cambridge University Press Cambridge
Gambino S J and Gilbert RB (1999) Modeling spatial variability in pile capacityfor reliability-based design Analysis Design Construction and Testing of DeepFoundations ASCE Geotechnical Special Publication No 88 135ndash49
Gilbert R B (1999) First-order second-moment bayesian method for data analysisin decision making Geotechnical Engineering Center Report Department of CivilEngineering The University of Texas at Austin Austin Texas 50 pp
Gilbert R B Gambino S J and Dupin R M (1999) Reliability-based approachfor foundation design without with-specific soil borings In Proceedings Off-shore Technology Conference Houston Texas OTC 10927 Society of PetroleumEngineers Richardson pp 631ndash40
Gilbert R B Choi Y J Dangyach S and Najjar S S (2005a) Reliability-baseddesign considerations for deepwater mooring system foundations In Proceed-ings ISFOG 2005 Frontiers in Offshore Geotechnics Perth Western AustraliaTaylor amp Francis London pp 317ndash24
Gilbert R B Najjar S S and Choi Y J (2005b) Incorporating lower-boundcapacities into LRFD codes for pile foundations In Proceedings Geo-FrontiersAustin TX ASCE Virginia pp 361ndash77
Goodwin P Ahilan R V Kavanagh K and Connaire A (2000) Integratedmooring and riser design target reliabilities and safety factors In ProceedingsConference on Offshore Mechanics and Arctic Engineering New Orleans TheAmerican Society of Mechanical Engineers (ASME) New York 785ndash92
Journel A G and Huijbregts Ch J (1978) Mining Geostatistics Academic PressSan Diego
Kenney R L and Raiffa H (1976) Decision with Multiple Objectives Preferencesand Value Tradeoffs John Wiley and Sons New York
Najjar S S (2005) The importance of lower-bound capacities in geotech-nical reliability assessments PhD dissertation The University of Texas atAustin
Practical application of RBD 223
Stahl B Aune S Gebara J M and Cornell C A (1998) Acceptance criteria foroffshore platforms In Proceedings Conference on Offshore Mechanics and ArcticEngineering OMAE98-1463
Tang W H and Gilbert R B (1993) Case study of offshore pile system reliabilityIn Proceedings Offshore Technology Conference OTC 7196 Houston Societyof Petroleum Engineers pp 677ndash83
The Economist (2004) The price of prudence A Survey of Risk 24 January 6ndash8USBR (2003) Guidelines for Achieving Public Protection in Dam Safety Decision
Making Dam Safety Office United States Bureau of Reclamation Denver COUSNRC (1975) Reactor safety study an assessment of accident risks in US com-
mercial nuclear power plants United States Nuclear Regulatory CommissionNUREG-75014 Washington D C
Wang D (2002) Development of a method for model calibration with non-normaldata PhD dissertation The University of Texas at Austin
Whipple C (1985) Approaches to acceptable risk In Proceedings Risk-Based Deci-sion Making in Water Resources Ed Y Y Haimes and EZ Stakhiv ASCENew York CA pp 31ndash45
Whitman R V (1984) Evaluating calculated risk in geotechnical engineeringJournal of Geotechnical Engineering ASCE 110(2) 145ndash88
Wu T H Tang W H Sangrey D A and Baecher G (1989) Reliability of offshorefoundations Journal of Geotechnical Engineering ASCE 115(2) 157ndash78
Chapter 6
Randomly heterogeneous soilsunder static and dynamicloads
Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
61 Introduction
The values of soil properties used in geotechnical engineering and geome-chanics involve a significant level of uncertainty arising from several sourcessuch as inherent random heterogeneity (also referred to as spatial variabil-ity) measurement errors statistical errors (due to small sample sizes) anduncertainty in transforming the index soil properties obtained from soil testsinto desired geomechanical properties (eg Phoon and Kulhawy 1999) Inthis chapter only the first source of uncertainty will be examined (inher-ent random heterogeneity) as it can have a dramatic effect on soil responseunder loading (in particular soil failure) as will be demonstrated in the fol-lowing As opposed to lithological heterogeneity (which can be describeddeterministically) the aforementioned random soil heterogeneity refers tothe natural spatial variability of soil properties within geologically distinctlayers Spatially varying soil properties are treated probabilistically and areusually modeled as random fields (also referred to as stochastic fields) In thelast decade or so the probabilistic characteristics of the spatial variabilityof soil properties have been quantified using results of various sets of in-situtests
The first effort in accounting for uncertainties in the soil mass in geotech-nical engineering and geomechanics problems involved the random variableapproach According to this approach each soil property is modeled bya random variable following a prescribed probability distribution function(PDF) Consequently the soil property is constant over the analysis domainThis approach allowed the use of established deterministic analysis meth-ods developed for uniform soils but neglected any effects of their spatialvariability It therefore reduced the natural spatial variation of soil prop-erties to an uncertainty in their mean values only This rather simplisticapproach was quickly followed by a more sophisticated one modeling thespatial heterogeneity of various soil properties as random fields This becamepossible from an abundance of data from in-situ tests that allowed the
Soils under static and dynamic loads 225
establishment of various probabilistic characteristics of the spatial variabilityof soil properties
In the last decade or so a series of papers have appeared in the litera-ture dealing with the effect of inherent random soil heterogeneity on themechanical behavior of various problems in geomechanics and geotechnicalengineering using random field theory A few representative papers are men-tioned here grouped by the problem examined (1) foundation settlementsBrzakala and Pula (1996) Paice et al (1996) Houy et al (2005) (2) soilliquefaction Popescu et al (1997 2005ab) Fenton and Vanmarke (1998)Koutsourelakis et al (2002) Elkateb et al (2003) Hicks and Onisiphorou(2005) (3) slope stability El Ramly et al (2002 2005) Griffiths and Fenton(2004) and (4) bearing capacity of shallow foundations Cherubini (2000)Nobahar and Popescu (2000) Griffiths et al (2002) Fenton and Griffiths(2003) Popescu et al (2005c) The methodology used in essentially all ofthese studies was Monte Carlo Simulation (MCS)
First-order reliability analysis approaches (see eg Christian (2004) fora review) as well as perturbationexpansion techniques postulate the exis-tence of an ldquoaverage responserdquo that depends on the average values of thesoil properties This ldquoaverage responserdquo is similar to that obtained froma corresponding uniform soil having properties equal to the average prop-erties of the randomly variable soil The inherent soil variability (randomheterogeneity) is considered to induce uncertainty in the computed responseonly as a random fluctuation around the ldquoaverage responserdquo For exam-ple in problems involving a failure surface it is quite common to assumethat the failure mechanism is deterministic depending only on the averagevalues of the soil properties More specifically in most reliability analysesof slopes involving spatially variable soil deterministic analysis methodsare first used to estimate the critical slip surface (corresponding to a mini-mum factor of safety and calculated assuming uniform soil properties) Afterthat the random variability of soil properties along this pre-determinedslip surface is used in the stochastic analysis (eg El Ramly et al 20022005)
In contrast to the type of work mentioned in the previous paragrapha number of recent MCS-based studies that did not impose any restric-tions on the type and geometry of the failure mechanism (eg Paice et al1996 for foundation settlements Popescu et al 1997 2005ab for soilliquefaction Griffiths and Fenton 2004 for slope stability and Popescuet al 2005c for bearing capacity of shallow foundations) observed thatthe inherent spatial variability of soil properties can significantly modifythe failure mechanism of spatially variable soils compared to the corre-sponding mechanism of uniform (deterministic) soils For example Fochtand Focht (2001) state that ldquothe actual failure surface can deviate fromits theoretical position to pass through weaker material so that the averagemobilized strength is less than the apparent average strengthrdquo The deviation
226 Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
of the failure surface from its deterministic configuration can be dramaticas there can be a complete change in the topology of the surface Further-more the average response of spatially variable soils (obtained from MCS)was observed to be significantly different from the response of the corre-sponding uniform soil For example Paice et al (1996) predicted an up to12 increase in average settlements for an elastic heterogeneous soil withcoefficient of variation CV = 42 compared to corresponding settlementsof a uniform soil deposit with the same mean soil properties Nobahar andPopescu (2000) and Griffiths et al (2002) found a 20ndash30 reduction in themean bearing capacity of heterogeneous soils with CV = 50 comparedto the corresponding bearing capacity of a uniform soil with the same aver-age properties Popescu et al (1997) predicted an increase of about 20 inthe amount of pore-water pressure build-up for a heterogeneous soil depositwith CV = 40 compared to the corresponding results of uniform soilwith the same mean properties It should be noted that the effects of spatialvariability were generally stronger for phenomena governed by highly non-linear constitutive laws (such as bearing capacity and soil liquefaction) thanfor phenomena governed by linear laws (eg settlements of foundations onelastic soil)
This chapter discusses the effects of inherent spatial variability (randomheterogeneity) of soil properties on two very different phenomena bearingcapacity failure of shallow foundations under static loads and seismicallyinduced soil liquefaction The first implies a limit equilibrium type of failureinvolving a large volume of soil while the second is based on a micro-scale(soil grain level) mechanism Based on results of numerical work previouslypublished by the authors it is shown that the end effects of spatial variabilityon limit loads are qualitatively similar for both phenomena
The probabilistic characteristics of the spatial variability of soils are brieflydiscussed in the first part of this chapter A general MCS approach forgeotechnical systems exhibiting random variation of their properties is alsopresented The second part focuses on the effects of inherent soil spatialvariability Based on examples dealing with the bearing capacity of shallowfoundations and seismically induced soil liquefaction it is shown that thespatial variability of soil properties can change the soil behavior from thewell known theoretical results obtained for uniform (deterministic) soils dra-matically The mechanisms of this intriguing behavior are analyzed in somedetail and interpreted using experimental and numerical evidence It is alsoshown that in some situations the resulting factors of safety and failureloads for randomly heterogeneous soils can be considerably lower than thecorresponding ones determined from some current reliability approachesused in geotechnical engineering with potentially significant implicationsfor design The concept of characteristic values or ldquoequivalent uniformrdquo soilproperties is discussed in the last part of the chapter as a means of accountingfor the effects of soil variability in practical applications
Soils under static and dynamic loads 227
62 Representing the random spatial variability ofsoil properties
The natural heterogeneity in a supposedly homogeneous soil layer may bedue to small-scale variations in mineral composition environmental condi-tions during deposition past stress history variations in moisture contentetc (eg Tang 1984 Lacasse and Nadim 1996) From a geotechnical designviewpoint this small-scale heterogeneity is manifested as spatial variation ofrelevant soil properties (such as relative density shear strength hydraulicconductivity etc)
An example of soil spatial variability is presented in Figure 61 in terms ofrecorded cone tip resistance values in a number of piezocone (CPTu) profilesperformed for core verification at one of the artificial islands (namely TarsiutP-45) constructed in the Canadian Beaufort Shelf and used as drilling plat-forms for oil exploration in shallow waters (Gulf Canada Resources 1984)The cone tip resistance qc is directly related to shear strength and to liq-uefaction potential of sandy soils The profiles shown in Figure 61 werelocated on a straight line at 9 m center-to-center distances Those recordsare part of a more extended soil investigation program consisting of 32piezocone tests providing an almost uniform coverage of the area of inter-est (72 mtimes72 m) The soil deposit at Tarsuit P-45 consists of two distinct
9m
MAC31
dept
h -
z(m
)
0minus5
minus10
minus15
minus20
0 10 20 0 10 20 0 10 20cone tip resistance (MPa)
averagevalues
example of loosesand pocket
example of densesand pocket
0 10 20 0 10 20 0 10 20
MAC04 MAC05 MAC06 MAC07 MAC08
9m 9m 9m 9m
Figure 61 Cone tip resistance recorded in the sand core at Tarsiut P-45 (after Popescuet al 1997)
228 Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
a50m
P3
P2 P5 P8
P6
P9
P12
clay
sandclay
sandclay
sand
clay
sandclay
sand
clay
sand
clay
sandclay
sandclay
sand
0 25 50 0 25
Stress normalized standard penetration resistance - (N1)60recorded in the dense sand layer
50 0 25 50 0 25 50
clay
sand
clay 25 50
25 50
250 50
b
P11
sand profile P10
finesandwithsilt
siltyclay
densesand
clay
sandde
pth
(m)
dept
h (m
)55
m
55 m
50m 50m
minus45
minus40
minus35
minus30
minus25
P1 P4
P10P7
minus45
minus40
minus35
minus30
minus25
dept
h (m
)
minus45
minus40
minus35
minus30
minus25
minus20
minus15
minus10
minus5
dept
h (m
)
minus45
minus40
minus35
minus30
minus25
(N1)60
Figure 62 Standard Penetration Test results from a 2D measurement array in the TokyoBay area Japan (after Popescu et al 1998a)
layers an upper layer termed the ldquocorerdquo and a base layer termed the ldquobermrdquo(see eg Popescu et al 1998a for more details) Selected qc records fromthe core are shown in Figure 61 Records from both layers are shown inFigure 63a
A second example is presented in Figure 62 in terms of StandardPenetration Test (SPT) results recorded at a site in the Tokyo Bay areaJapan where an extended in situ soil test program had been performed forliquefaction risk assessment A two-dimensional measurement array consist-ing of 24 standard penetration test profiles was performed in a soil depositformed of three distinct soil layers (Figure 62b) a fine sand with silt inclu-sions a silty clay layer and a dense to very dense sand layer The resultsshown in Figure 62a represent stress normalized SPT blowcounts (N1)60recorded in the dense sand layer in 12 of the profiles Results from all soillayers are shown in Figure 63b
Soils under static and dynamic loads 229
c Probability density functions fitted to normalized fluctuations of recorded data
Gaussian distribution(for comparison)
Beta distribution (Tarsiut - core)Beta distribution (Tarsiut - berm)
Lognormal distrib(Tokyo Bay - Fine sand) Gamma distribution(Tokyo Bay - Dense sand)
a CPT data from Tarsiut P-45 (hydraulic fill) b SPT data from Tokyo Bay (natural deposit)
0 5 10Cone tip resistance - qc (MPa) Normalized SPT blowcount - (N1)60
15 20 25 0
minus50
minus30
05
04
03
02
01
0
minus4 minus3 minus2 minus1 0Normalized data (zero-mean unit standard deviation)
1 2 3 4
minus25
minus20
minus15
minus10
minus50
minus40
minus30
SILTY CLAY
Dep
th -
z(m
)
Dep
th -
z(m
)
FINE SAND
minus20mminus16m dry
DENSESAND
minus20
saturated
CORE
BERM
minus10
0
20 40 60 80 100
Figure 63 In-situ penetration test results at two sites (after Popescu et al 1998a) and cor-responding fitted probability distribution functions for the normalized data (zeromean unit standard deviation) (a) cone penetration resistance from 8 CPTprofiles taken at an artificial island in the Beaufort Sea (Gulf Canada Resources1984) (b) normalized standard penetration data from 24 borings at a site in theTokyo Bay area involving a natural soil deposit (c) fitted PDFs to normalizedfluctuations of field data Thick straight lines in (a) and (b) represent averagevalues (spatial trends)
It can be easily observed from Figures 61 and 62 that soil shear strength ndashwhich is proportional to qc and (N1)60 ndash is not uniform over the test area butvaries from one location to another (both in the horizontal and vertical direc-tions) As deterministic descriptions of this spatial variability are practicallyimpossible ndash owing to the prohibitive cost of sampling and to uncertaintiesinduced by measurement errors (eg VanMarcke 1989) ndash it has becomestandard practice today in the academic community to model this smallscale random heterogeneity using stochastic field theory The probabilistic
230 Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
characteristics of a stochastic field describing spatial variability of soilproperties are briefly discussed in the following
621 Mean value
For analysis purposes the inherent spatial variability of soil propertiesis usually separated into an average value and fluctuations about it (egVanMarcke 1977 DeGroot and Baecher 1993) The average values of cer-tain index soil properties (eg qc) are usually assumed to vary linearly withdepth (eg Rackwitz 2000) reflecting the dependence of shear strength onthe effective confining stress They are usually called ldquospatial trendsrdquo andcan be estimated from a reasonable amount of in-situ soil data (using forexample a least square fit) In a 2D representation denoting the horizontaland vertical coordinates by x and z respectively the recorded qc values areexpressed as
qc(xz) = qavc (z) + uf (xz) (61)
where uf(xz) are the random fluctuations around the spatial trend qavc (z)
Assuming isotropy in the horizontal plane extension to 3D is straightfor-ward The spatial trends qav
c (z) are shown in Figures 61 and 63ab bythick straight lines representing a linear increase with depth of the aver-age recorded index values qc and (N1)60 It is mentioned that the thick linesin all plots in Figure 61 correspond to the same trend for all profiles (oneaverage calculated from all records) Several authors (eg Fenton 1999bEl-Ramly et al 2002) note the importance of correct identification of thespatial trends Phoon and Kulhawy (2001) state that ldquodetrending is as muchan art as a sciencerdquo and recommend using engineering judgment
622 Degree of variability
The degree of variability ndash or equivalently the magnitude of the randomfluctuations around the mean values ndash is quantified by the variance or thestandard deviation As the systematic trends can be identified and expresseddeterministically they are usually removed from the subsequent analysis anditrsquos only the local fluctuations that are modeled using random field theoryFor mathematical convenience the fluctuations uf(xz) in Equation (61) arenormalized by the standard deviation σ (z)
u(xz) = [qc(xz) minus qavc (z)]σ (z) (62)
so that the resulting normalized fluctuations u(xz) can be described by azero-mean unit standard deviation random field
As implied by Equation (62) the standard deviation σ (z) may be a func-tion of the depth z A depth-dependent standard deviation can be inferred
Soils under static and dynamic loads 231
by evaluating it in smaller intervals over which it is assumed to be constantrather than using the entire soil profile at one time For example the standarddeviation of the field data shown in Figure 61 was calculated over 1 m longmoving intervals and was found to vary between 1 and 4 MPa (Popescu1995)
The most commonly used parameter for quantifying the magnitude offluctuations in soil properties around their mean values is the coefficient ofvariation (CV) For the data in Figure 61 CV(z) = σ (z)qav
c (z) was found tovary between 20 and 60 The average CV of SPT data from the densesand layer in Figure 62 was about 40
It should be mentioned at this point that the CPT and SPT records shown inFigures 61 and 62 include measurement errors and therefore the degreesof variability discussed here may be larger than the actual variability ofsoil properties The measurement errors are relatively small for the caseof CPT records (eg for the electrical cone the average standard devia-tion of measurement errors is about 5 American Society for Testing andMaterials 1989) but they can be significant for SPT records which areaffected by many sources of errors (discussed by Schmertmann 1978 amongothers)
623 Correlation structure
The fluctuations of soil properties around their spatial trends exhibit in gen-eral some degree of coherencecorrelation as a function of depth as can beobserved in Figures 61 and 62 For example areas with qc values consis-tently larger than average indicate the presence of dense (stronger) pocketsin the soil mass while areas with qc values consistently smaller than aver-age indicate the presence of loose (less-resistant) soil Examples of denseand loose soil zones inferred from the fluctuations of qc are identified inFigure 61 Similarly the presence of a stronger soil zone at a depth of about35 m and spanning profiles P5 and P8ndashP12 can be easily observed from theSPT results presented in Figure 62
The similarity between fluctuations recorded at two points as a function ofthe distance between those two points is quantified by the correlation struc-ture The correlation between values of the same material property measuredat different locations is described by the auto-correlation function A signif-icant parameter associated with the auto-correlation function is called thescale of fluctuation (or correlation distance) and represents a length overwhich significant coherence is still manifested Owing to the geological soilformation processes for most natural soil deposits the correlation distancein the horizontal direction is significantly larger than the one in the verticaldirection This can be observed in Figure 61 by the shape of the dense andloose soil pockets In this respect separable correlation structure models (egVanMarcke 1983) seem appropriate to model different spatial variability
232 Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
characteristics in the vertical direction (normal to soil strata) and in thehorizontal one
There are several one-dimensional theoretical auto-correlation functionmodels that can be combined into separable correlation structures and usedto describe natural soil variability (eg VanMarcke 1983 Shinozuka andDeodatis 1988 Popescu 1995 Rackwitz 2000) Inferring the parame-ters of a theoretical correlation model from field records is straightfor-ward for the vertical (depth) direction where a large number of closelyspaced data are usually available (eg from CPT records) However it ismuch more challenging for the horizontal direction where sampling dis-tances are much larger DeGroot and Baecher (1993) and Fenton (1999a)review the most common methods for estimating the auto-correlationfunction of soil variability in the vertical direction Przewlocki (2000)and Jaksa (2007) present methods for estimating the two-dimensionalcorrelation of soil spatial variability based on close-spaced field measure-ments in both vertical and horizontal directions As close-spaced recordsin horizontal direction are seldom available in practice Popescu (1995)and Popescu et al (1998a) introduce a method for estimating the hor-izontal auto-correlation function based on a limited number of verticalprofiles
As a note of caution it is mentioned here that various different definitionsfor the correlation distance exist in the literature yielding different values forthis parameter (see eg Popescu et al 2005c for a brief discussion) Anotherproblem stems from the fact that when estimating correlation distances fromdiscrete data the resulting values are dependent on the sampling distance(eg Der Kiureghian and Ke 1988) and the length scale or the distance overwhich the measurements are taken (eg Fenton 1999b)
When two or more different soil properties can be recorded simultane-ously it is interesting to assess the spatial interdependence between themand how it affects soil behavior For example from piezocone test resultsone can infer another index property ndash besides qc ndash called soil classifica-tion index which is related to grain size and soil type (see eg Jefferies andDavies 1993 Robertson and Wride 1998) The spatial dependence betweentwo different properties measured at different locations is described by thecross-correlation function In this case the variability of multiple soil prop-erties is modeled by a multivariate random field (also known as a vectorfield) and the ensemble of auto-correlation and cross-correlation functionsform the cross-correlation matrix
As the field data u(xz) employed to estimate the correlation structurehave been detrended and normalized (refer to Equation (62)) the result-ing field is stationary It is always possible to model directly non-stationarycharacteristics in spatial correlation but this is very challenging because oflimited data and modeling difficulties (eg Rackwitz 2000) Consequentlystationarity is almost always assumed (at least in a piecewise manner)
Soils under static and dynamic loads 233
624 Probability distribution
Based on several studies reported in the literature soil properties can followdifferent probability distribution functions (PDFs) for different types of soilsand sites but due to physical reasons they always have to follow distribu-tions defined only for nonnegative values of the soil properties (excludingthus the Gaussian distribution) Unlike Gaussian stochastic fields wherethe first two moments provide complete probabilistic information non-Gaussian fields require knowledge of moments of all orders As it is extremelydifficult to estimate moments higher than order two from actual data (egLumb 1966 Phoon 2006) the modeling and subsequent simulation of soilproperties that are represented as homogeneous non-Gaussian stochasticfields are usually done using their cross-correlation matrix (or equiva-lently cross-spectral density matrix) and non-Gaussian marginal probabilitydistribution functions
While there is no clear evidence pointing to any specific non-Gaussianmodel for the PDF of soil properties one condition that has to be satisfiedis for the PDF to have a lower (nonnegative) bound The beta gamma andlognormal PDFs are commonly used for this purpose as they all satisfy thiscondition For prescribed values of the mean and standard deviation thegamma and lognormal PDFs are one-parameter distributions (eg the valueof the lower bound can be considered as this parameter) They are bothskewed to the left Similarly for prescribed values of the mean and standarddeviation the beta PDF is a two-parameter distribution and consequentlymore flexible in fitting in-situ data Moreover it can model data that aresymmetrically distributed or skewed to the right Based on penetration datafrom an artificial and a natural deposit Popescu et al (1998a) observed thatPDFs of soil strength in shallow layers are skewed to the left while for deepersoils the corresponding PDFrsquos tend to follow more symmetric distributions(refer to Figure 63)
As will be discussed later the degree of spatial variability (expressed bythe coefficient of variation) is a key factor influencing the behavior of het-erogeneous soils as compared to uniform ones It has also been determinedthat both the correlation structure and the marginal probability distributionfunctions of soil properties affect the response behavior of heterogeneoussoils to significant degrees (eg Popescu et al 1996 2005a Fenton andGriffiths 2003)
63 Analysis method
631 Monte Carlo Simulation approach
Though expensive computationally Monte Carlo Simulation (MCS) isthe only currently available universal methodology for accurately solving
234 Radu Popescu George Deodatis and Jean-Herveacute Preacutevost
problems in stochastic mechanics involving strong nonlinearities and largevariations of non-Gaussian uncertain system parameters as is the case ofspatially variable soils As it will be shown later in this chapter failuremechanisms of heterogeneous soils may be different from the theoreticalones (derived for assumed deterministic uniform soil) and can change signif-icantly from one realization (sample function) of the random soil propertiesto another This type of behavior prevents the use of stochastic analysismethods such as the perturbationexpansion ones that are based on theassumption of an average failure mechanism and variations (fluctuations)around it
MCS is a means of solving numerically problems in mathematics physicsand other sciences through sampling experiments and basically consists ofthree steps (eg Elishakoff 1983) (1) simulation of the random quanti-ties (variables fields) (2) solution of the resulting deterministic problemfor a large number of realizations in step 1 and (3) statistical analysis ofresults For the case of spatially variable soils the realizations of the ran-dom quantities in step 1 consist of simulated sample functions of stochasticfields with probabilistic characteristics estimated from field data as discussedin Section 62 Each such sample function represents a possible realizationof the relevant soil index properties over the analysis domain The deter-ministic problem in step 2 is usually solved by finite element analysis Onesuch analysis is performed for each realization of the (spatially variable) soilproperties
632 Simulation of homogeneous non-Gaussian stochasticfields
Stochastic fields modeling spatially variable soil properties are non-Gaussianmulti-dimensional (2D or 3D) and can be univariate (describing the variabil-ity of a single soil property) or multivariate (referring to the variability of sev-eral soil properties and their interdependence) Among the various methodsthat have been developed to simulate homogeneous non-Gaussian stochas-tic fields the following representative ones are mentioned here Yamazakiand Shinozuka (1988) Grigoriu (1995 1998) Gurley and Kareem (1998)Deodatis and Micaletti (2001) Puig et al (2002) Sakamoto and Ghanem(2002) Masters and Gurley (2003) Phoon et al (2005)
As discussed in Section 624 it is practically impossible to estimate non-Gaussian joint PDFs from actual soil data Therefore a full descriptionof the random field representing the spatial variability of soil propertiesat a given site is not achievable Under these circumstances the simu-lation methods considered generate sample functions matching a targetcross-correlation structure (CCS) ndash or equivalently cross-spectral densitymatrix (CSDM) in the wave number domain ndash and target marginal PDFsthat can be inferred from the available field information It is mentioned