dakota/uq: a toolkit for uncertainty quantification in a ...structural dynamics research, dept. 9124...
TRANSCRIPT
S iaC d StatesD 000.
DAKO uantification ina M mputational
24
g Group
andia is a multiprogram laboratory operated by Sandorporation, a Lockheed Martin Company,for the Uniteepartment of Energy under contract DE-AC04-94AL85
TA/UQ: A Toolkit For Uncertainty Qultiphysics, Massively Parallel Co
Environment
Steven F. [email protected]
Structural Dynamics Research, Dept. 91Sandia-Albuquerque
LANL Uncertainty Quantification Workin
October 4, 2001
Uncertainty Quantification
Real
• terials,
•
•
•
• y
Unce
Engineering Sciences Center
Physical Systems:
Display random and systematic variation- geometry, ma
boundary conditions, initial conditions, excitations
Vary from one realization to the next
Display behavior that cannot be precisely measured
rtainty occurs in various forms:
Irreducible, variability, aleatoric
Reducible, epistemic, subjective, model form uncertaint
Uncertainty Quantification
Usefu
requirements.
I Program Plan,hich a computerm the perspective
• An
• Mo
Engineering Sciences Center
l in:
alysis and Design
–To assess the reliability of physical systems.
–To establish designs that satisfy pre-established reliability
–To establish sensitivities to key uncertainties
del validation, certification, and accreditation
–As defined in the DOE Defense Programs (DOE/DP) ASCvalidation is the process of determining the degree to wmodel is an accurate representation of the real world froof the intended model applications.
–Convey confidence in predictions to decision makers
Uncertainty Quantification:Gen
Gene
s
and as
ually
f
Stati
• Mo(po
• Seob
Engineering Sciences Center
eral Framework
ral Description:
tical Approach:
del components of as Random Variables or Fields, ssibly) Random External Input
ek quantities such as . However, what is acttained are conditional statistics .
M . X,( )f U
: vector of uncertain parameters
f
X
U : output(s) of system
M: a deterministic mapping
: input(s) to system
x
E g U( )[ ]E g U( ) M[ ]
Probabilistic/Statistical Approach: EssentialEle
Conc
o
PD
rization of the
acterization ofters
Esse
• Ra
• Pr
• Ch
Genethe re
Engineering Sciences Center
ments of a Statistical Approach:
lusion: Need a Generalized Outlook.
ntial Elements of a Statistical Approach:
ndom External Inputs
pagation Techniques–Analytical Reliability Methods; Sampling; Response Surface
Approximations; Stochastic Finite Element Methods.
aracterization of Models–Verification and Validation.
E g U( )[ ] E E E g U( ) M X,[ ] M{ }⟨ ⟩=
ropagation Techniques (Note:ependency on M)
CharacteModel, M
Probabilistic charthe input parameral functions of
sponse, u
Random external loads
Anatomy of Global Uncertainty
E
Engineering Sciences Center⟨ ⟩
E M{ }
E g U( ) M X,[ ]
E g U( )[ ]=
Anatomy of Global Uncertainty
Engineering Sciences Center
E g U( ) M X,[ ]
- Random Vibration- Earthquake Engineering- Ocean Engineering- Weapons Applications: Launch Shocks/Re-entry Loads, Penetration Loads, Hostile Environments
Uncertainty due to External Loads: f( )
Anatomy of Global Uncertainty
Un
- E
Engineering Sciences Center
E M{ }
E g U( ) M X,[ ]
certainty Propagation:
ffects of parametric uncertainty:Intrinsic variabilities, Tolerances,Lack of repeatability
X( )
Anatomy of Global Uncertainty
E
Engineering Sciences Center⟨ ⟩
E M{ }
E g U( ) M X,[ ]
E g U( )[ ]=
Uncertainty Quantification at Sandia-NM
• DAel problems,e approximation,
el Uncertainty
• Po
• Ep
• Se
Engineering Sciences Center
KOTA (Design Analysis Kit for OpTimizAion)/UQ–Framework for multi-level, parallel computation: ASCI-lev
optimization, nondeterministic analysis, response surfacdesign of experiments, optimization under uncertainty
lynomial Chaos and Stochastic Finite Elements–Analysis of response of stochastic systems
istemic Uncertainty–Non-Probabilistic Approach, Probabilistic Approach, Mod
nsitivity Analysis
Objectives of Toolkit
Provi n a unifiedframe
s
e problems
n
t
non-gradient)
• Di
• AS
• Mi
• Fle
Why
• Ex
• Su
• Mu
• Ex
• Ex
Engineering Sciences Center
de uncertainty quantification tools to the analyst community iwork to be used in the design and certification processes.
cipline independent
CI (Accelerated Strategic Computing Initiative)-scal
imize number of function evaluations
xibility in uncertainty model
ie UQ tools to the DAKOTA framework?
isting, proven software framework
ccessfully linked with over 20 application codes
ltilevel parallelism
tensive optimization algorithm library (gradient and
tensive selection of approximation strategies
Answ• W
oo
eeo sSe
DAKOTA toolkitDes
Contours of GaMe3on Reacting Surface
Streamlines
Inlet:
GaMe3
AsH3
H2
Outlet:GaMe3
AsH3
H2
CH4
• H• H
Addi• R• L• N• A• B
Engineering Sciences Center
er fundamental engineering questions:hat is the best design?w safe is it?w much confidence in my answer
tional motivations:use tools and interfacesverage optimization, UQ, et al.nconvex, nonsmooth design spaces → state-of-the-art methodologieCI-scale applications and architectures → scalable parallelism a pathfinder in enabling M&S-based culture change at Sandia
DAKOTAOptimizationUncertainty Quant.Parameter Est.Sensitivity Analysis
DesignModelParameters Metrics
MERCURY
SALINASALEGRA
PRONTO
GOMAEAGLE
NG ion opticsNG power supplyAF&F subsystemlaydown, gas transfer
Coatings consortiumGoodyear CRADA
CTH explosives performance
ign optimization of engineering applications
Ove
Responses
Functionsobjectivesconstraints
Gradientsnumericalanalyticmixed
Hessians
least sq. termsgeneric
analytic
I
e
ainty
odelFormExtrap
DO
Strat
DDADopt
N
LHS/MAMFOR
Engineering Sciences Center
rview of DAKOTA framework
Iterator
Optimizer ParamStudy
SGOPTNPSOLDOT OPT++
Least Sq.E
InterfaceParameters
Model:
Designcontinuousdiscrete
Uncertainnormal/lognuniform/logu
histogramState
continuousdiscrete
Application
system call, fork
direct
Approximationpolynomialneural networksplines
synch & asynch
synch & asynch
terator
Model
gy: control of multiple iterators and models
Iterator
Model
Iterator
Model
Coordination:
Cascaded
Nested
ConcurrentAdaptive/Interactive
Hybrid
SeqApprox
OptUnderUnc
Branch&Bound/PICO
Strategy
Optimization Uncert
M
UncOfOptima
message passing
kriging
weibull
BCGN
Vector
Centered MultiD
List
CECCD/FF
onDeterm
C
SFEM
Taylor series
Parallelism:
Peer (static, distr. sched.)Master-slave (self-sched.)
4 nested levels
V+M
Boot/Imp
Layered (surrogates, hierarchical)
hierarchicalrSQP++
Optimization/UQ Projects
DAKO-7471
ary):
5482bina
P-7471
s):-747196
itus Leung
rch libraries):v
houcq,illiams
DAKO
SGO
PICO
OPT+
Engineering Sciences Center
TA project (optimization with engineering simulations):Sandia manager - David Womble, 9211, [email protected], 845PI - Mike Eldred, 9211, [email protected], 844-6479Team members - Tony Giunta, Bill Hart, Bart van Bloemen Waandershttp://endo.sandia.gov/DAKOTA/
TA/UQ project (analytic reliability, sampling, and SFE UQ librSandia manager - Martin Pilch, 9133, [email protected], 845-3047PI - Steve Wojtkiewicz, 9124, [email protected], 284-Team members - Mike Eldred, Rich Field, John Red-Horse, Angel Ur
T project (stochastic global optimization):Sandia manager - David Womble, 9211, [email protected], 845PI - Bill Hart, 9211, [email protected], 844-2217http://www.cs.sandia.gov/~wehart/main.html
project (mixed integer programming, scheduling and logisticSandia manager - David Womble, 9211, [email protected], 845PI - Cindy Phillips, 9211, [email protected], 845-72Team members - Bob Carr, Jonathan Eckstein (Rutgers), Bill Hart, Vhttp://www.cs.sandia.gov/~caphill/proj/pico.html
+/DDACE/APPS/IDEA projects (NLP, sampling, & pattern seaSandia manager - Chuck Hartwig(acting), 8950, [email protected] - Juan Meza, 8950, [email protected], 294-2425Team members - Paul Boggs, Patty Hough, Tamara Kolda, Leslea Le
Kevin Long, Monica Martinez-Canales, and Pam Whttp://csmr.ca.sandia.gov/~meza/research.html
Current Dakota/UQ Capabilities
Samp
r
Order Reliabil-
s
mize
• Ra
• St
Analy
• Me
• FOity
Robu
Stoch
Respo
• Apco
Engineering Sciences Center
ling Techniques:
ndom Sampling (Monte Carlo)
atified Sampling (LHS (Latin Hypercube Sampling)
tical Reliability Techniques:
an Value (MV), Advanced Mean Value (AMV/AMV+)
RM (First Order Reliability Method)/SORM (Second Method)
tness Analysis
astic Finite Element/ Polynomial Chaos Expansions
nse Surface Approximations:
plication of UQ tools to a surrogate function to minimputational expense.
Sampling Techniques
Mont
iscipline ofo
capabilities
)
• Gepr
• Ea
• Co
• Tw
• Un
Engineering Sciences Center
e Carlo-Style (Sampling-based) Analysis:
neral, simple to implement and robust to size and dblem being investigated
sily wrapped around current deterministic analysis
mputationally expensive (many function evaluations
o current options:–Traditional Monte Carlo–Latin Hypercube Sampling
der investigation:–Bootstrap Sampling–Importance Sampling Techniques–Quasi-Monte Carlo SImulation–Markov Chain Monte Carlo
Overview of Analytically BasedRel
• Inv d normal ran-do
ization problem distance func-ethod.
ble space.
ues about anp
• Na
• MVwhtio
• MV
• FO
• Eq“o
Engineering Sciences Center
iability Methods
olve a transformation to unit variance, uncorrelatem variable space.
taf Transformation used in DAKOTA/UQ.
, AMV/AMV+, FORM all solve a constrained optimere the objective function is always this minimumn with the constraint function depending on the m
and AMV/AMV+ work in the original random varia
RM/SORM work in the transformed space.
uivalent to Polynomial Response Surface Techniqtimally” selected expansion point
Probabilistic Robustness Analysis
• “G e of outputfun
and itsxU
• Po
such
wherlowe
Engineering Sciences Center
iven the bounds on the input parameters, what rangction is possible?”
se two global optimization problems:
that
e is the size of uncertain input vector, denoter and upper bounds, respectively.
gupper max g M f x,( )( )=x
glower min g M f x,( )( )=x
xi( )L xi xi( )U≤ ≤ i 1…N=∀
N xL
Probabilistic Robustness Analysis
Answ
m variables of SDM 2002)
• Reun
Engineering Sciences Center
er:
cently extended to mixed case of intervals and randoknown dependence (to appear in Wojtkiewicz, AIAA
g u( ) glower gupper,[ ]∈
SFEM/Polynomial Chaos Techniques
• Ap
p of randomar
e ransforma-io
s
o
• Ov
• Rt
• E
• C
Engineering Sciences Center
proximation of full stochastic representation
timal approximation in inner product spaces, spaceiables.
presents a more general alternative to the Rosenblatt tn
–avoid assuming full distribution when faced with limited input data
timating coefficients is the key issue–requires realizations of the function it replaces
nvergence issues–are there sufficient samples to compute coefficients?
–possibility of non-physical realizations
–mean square convergence
L2
SFEM/Polynomial Chaos Techniques
•
•
r+ )
uni x( )
δi x( )-----------=
Engineering Sciences Center
Consider PCE of general random process,
, where
– th order polynomial in , where
–function of underlying random variables
Solve for the Fourier coefficients,
can be solved in closed-form
u
u x Φ;( ) u x Φ;( )P( )
ui x( )Γ i ξ( )i 0=
P
∑≡≈ P1s!---- m(
r 0=
s 1–
∏
s 1=
q
∑=
q ξ ξ ξ1 ξ2 … ξm
T=
m
ui x( )
i x( )u x h ξ[ ],( ) Γ i ξ( )⟨ ⟩
Γ i2 ξ( )⟨ ⟩
-------------------------------------------
… u x h ξ[ ],( ) Γ i ξ( ) f ξ ξ( ) ξd
∞–
∞
∫∞–
∞
∫∞–
∞
∫
… Γ i2 ξ( ) f ξ ξ( ) ξd
∞–
∞
∫∞–
∞
∫∞–
∞
∫-----------------------------------------------------------------------------------= =
δi x( )
Epistemic Uncertainty
•
•
bases do
P
•
Non-
Engineering Sciences Center
Epistemic Uncertainty results from a lack of information.
Epistemic Uncertainty manifests itself in several ways
–Uncertainty in parameters for which statistically significant datanot exist
–The form of the model is not known exactly
robabilistic Approach
Variety of approaches investigated:
–Interval analysis–Possibility Theory–Evidence Theory (Dempster-Shafer)–Imprecise Probability–Probability Bounds–Interval-valued Probability Distributions–Convex Sets of Probability Distributions
The Penetrator Problem
• Pro
ent failure,
n with simplified
P f
u SRS ω( )=} )
codeslinear transient dynamicslinear soil mechanicsr filter
tain parametersangle-of-attack soil parameter
wind
Engineering Sciences Center
blem statement
–During the penetration event, predict the probability of compon
–Consider a nonlinear, full-body, 3D, coupled-physics simulatioprobabilistic properties.
α
-γ
D
CLv
f 0=M . ; α D{ ,(
M
∼ASCI FE
• structure - non• boundary - non• SRS - nonlinea
Uncer• α, • D,
The model, M: a complex, cascaded system
102
eq (Hz)
α (deg)h
f )
✓ Nodyn
✓ 50,
✓ Sp✓ Lo
sim
✓ 33 sim
Struc
Soil
Engineering Sciences Center
P
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09−400
−300
−200
−100
0
100
200
300
400
101
300
400
500
600
700
800
900
1000
Filters
Time (s)A
ccel
(g)
Pea
kA
ccel
(g)
Natural Fr
D (in)
Pea
kA
ccel
(g)
nlinear transientamics FEA000 DOF
erical cavity expansionads couple with mechanicsulation
LP, SRS
CPU hours perulation
✓ 49 total runs performed
✓ Performed simultaneouslyon network of Sun Ultra IIs
Component Response
tural Mechanics
Mechanics
Approximate RS Models
u(
Overview of the Shock Response Spectrum (SRS)
Why u
• potential
•
•
•
103
104
105
Frequency (Hz)
SRS
0−8000
−6000
−4000
−2000
0
2000
4000
6000
8000
Acc
el (
g)
Engineering Sciences Center
se SRS?
measure of shock severity; indicative of shock damage
frequency-domain representation of shock response
long history of use in weapon design; test-based spec
used for component qualification - compare to
0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01Time (sec)
101
102
100
101
102
103
104
105
Natural
Pea
k A
ccel
(g)
Time history
SRSref
UQ analysis of Penetrator System
• Tw
e model
• Us(res
Engineering Sciences Center
o design variables, :– , angle of attack is a normal random variable with mean 1
and standard deviation of 1.
– , soil depth is a lognormal random variable with mean 25and standard deviation 16.
ing the results from simulations, build a approximat
ponse surface approximation) for .
Xα
D
u mini
SRSref f i( ) SRS f i( )–( )=
Z g U( ) I U( )= =
P f P Z 0≤( ) 1 E g U( )[ ]–= =
u
UQ analysis of Penetrator System
• Ap :
LHS
0/0.02400
0/0.06767
0/0.05581
0/0.05071
Z 0( )
Engineering Sciences Centerply MC/LHS to these surrogate models to evaluate
and
Response SurfaceApproximation Method
MC
Kriging 0.02000/0.02300 0.0200
Splines 0.06900/0.06781 0.0672
Neural Net 0.05024/0.05588 0.0550
Quadratic Polynomial 0.04960/0.05077 0.0507
F
Ns 14×10= Ns 5
6×10=
Sum
Curre
n s
r
pling
-
Anal
Sam
Prob
Poly
Futu
Enha
Non
Engineering Sciences Center
mary
nt Capabilities
ytical Reliability Techniques:
•MV, AMV, AMV+, FORM/SORM
pling techniques:
•Pure Random Sampling (Monte Carlo)
•Stratified Sampling (LHS)
abilistic Robustness Analysis
omial Chaos Expansions/Stochastic Finite Element Technique
e Capabilities:
nced sampling methods:
•Importance Sampling, Bootstrap Sampling,
Quasi-Monte Carlo Sampling, Markov Chain Monte Carlo Sam
traditional uncertainty methodologies