transient 3d heat flo w analysis for integrated circuit ...tjs/atar/tlm.pdf · the tlm method is...
TRANSCRIPT
Transient 3D heatflow analysisfor integrated circuit devices
using the transmissionline matrix methodon a quad tr eemesh
T. Smy�, D. Walkey, S.K. Dew
�Departmentof Electronics
CarletonUniversity
Ottawa,ON, CanadaK1S5B6�Departmentof ElectricalandComputerEngineering
Universityof Alberta,Edmonton,AB, CanadaT6G2G7
Abstract
Thispaperpresentsa3D transmissionline matrix implementation(TLM)for thesolutionof transient
heatflow in integratedsemiconductordevices. The implementationusesa rectangulardiscontinuous
meshto allow for localmeshrefinement.Thisapproachis basedonaquadtreemeshingtechniquewhich
canhave a complex geometryusingblocksof varyingsizes.Eachsuchblock canhave a maximumof
two adjacentblocksonany verticalsideanda maximumof four blockson thetopor bottom.
TheTLM implementationis basedon a physicalextractionof a resistanceandcapacitancenetwork
andthenthecreationof theappropriateTLM matrix. Theformulationallowsfor temperaturedependent
materialparametersandanon-uniformtimestepping.
The simulatoris first testedusinga 2D exampleof a heatsourcein a rectangularregion. Using
this examplethenumericalerror is determinedandfoundto belessthan0.4%. Next, nonlinearitiesare
included,anda numberof non-uniformtimesteppingalgorithmsaretested.Then,a 3D problemis also
comparedto ananalyticalsolutionandagaintheerroris verysmall.Finally, anexampleof afull solution
of heatflow in a realisticSi trenchdevice is presented.
Keywords:Simulation,Device,Thermal,Transient,TLM
�Correspondingauthor
1
1 Intr oduction
Thesolutionof theheatdiffusionequationusingtransmissionline matrix (TLM) techniquesis well estab-
lished[1, 2]. A specificapplicationof TLM hasbeenthe solutionof steadystateand transientthermal
responsesof integratedcircuit (IC) devices[3, 4]. This problemis of substantialinterestdueto theneedto
know maximumoperatingtemperaturesfor reliability reasonsandto developtemperatureconsistentmodels
of thedevice.
Thebasicequationfor describingtheheatflow in IC devicesis[6, 7]:����� ��� ��������� �� � ���� � ����� �!�#"$� �#� (1)
where ��� ��is the temperature-dependent thermalconductivity, ��� thespecificheatof thematerial, � the
materialdensity, and� ����� �!�#"$� �#� describesthe local heatgeneration.A simplisticdescriptionof theprob-
lem to be solved is that of a small thin heatgenerationplateat the base/collectorjunction of the device
undersimulationwhich is placedin a large semiconductorsubstratecharacterizedby a constantthermal
conductivity, materialdensityandheatcapacity. Theboundaryconditionsaresomewhatproblematic,but a
fixedtemperatureon thebottomof thesemiconductorandtheassumptionof eitherdevice “tiling” (leading
to a reflective boundaryconditionat the edges)or of a semi-infinitesubstrateleadsto a problemthat can
be solved analyticallyusinga Green’s function approach.However, solving this equationfor realisticIC
devicesis complicatedby anumberof factors:
1. Theproblemmustbesolvedin 3D, becauseof thedominantroleof 3D heatspreadingfrom thesmall
sourceinto thesubstrate.
2. Theelectricaldevice equationleadsto non-uniformheatgenerationat thebase.
3. Thedevice metalizationmustbesimulated.This is particularlycrucialfor substrateswith poorther-
malconductivities suchasGaAs.
4. Temperature-dependentmaterialpropertiesmustbeincluded.BothGaAsandSi exhibit temperature-
dependentthermalconductivities andheatcapacities.ThismakesEq.1 anonlinearequation.
Accountingfor theeffectsdescribedabove necessitatestheuseof a3D numericalmodel.Someapplica-
tionssimply requirethesolutionof thesteadystateproblemandthentheextractionof aneffective thermal
2
resistance.%'& � which is definedasthesteadystatetemperatureabove ambientdividedby thepower dissi-
patedin thedevice. However, many problemsrequirea full solutionof thetransientresponseof thedevice.
TLM hasshown itself highly appropriatefor solving time dependenttransientheatflow problemsandis
particularlyattractive for multimaterialandnonlinearproblemsasany temperatureor materialdependence
of thethermalparameterscanbelocally accommodated.
Anotherconstrainton this typeof problemrelatesto theefficiency of themethodin the faceof a large
spatialdynamicrange.Theheatflow problemfor IC devicesis characterizedby verylargethermalgradients
localizedaroundthe active componentswhich are relatively small regions ( ( 1 ) m) containinga power
source.In theseregions,ahighdensityof meshelementsis requiredtoobtainanaccuratesolution.However,
it is alsonecessaryto solve for relatively large volumesdueto substratethicknesseswhich are typically
50-500 ) m. Thesetwo requirementsmake the efficient solutionof the heatequationfor 3D IC devices
only possiblewith a highdegreeof meshrefinementaroundthedevice junctionsandmuchlargerelements
towardstheedgesof thesimulationregion.
The TLM methodis traditionally performedusinga continuousrectangularmesh. Clearly, a uniform
grid spacingwould be inefficient becausevery many cellswould be requiredif forcedto usea fine mesh
throughout.A nonuniformmeshcanbereadilyimplementedusingTLM, but this is usuallystill continuous
sothatthefinemeshrequiredneartheactive devicesmustalsoextendthroughat leastpartof thesubstrate,
therebystill requiringan unnecessarilylarge numberof cells. Ideally, a multigrid or quadtreeapproach
shouldbe usedsuchthat local meshrefinementneednot extendbeyond whereit is absolutelyrequired,
therebyallowing theminimumnumberof cellsin thecalculation.
In suchmeshes,two or moresmallcellsmaybeadjacentalonga singlesideof a largercell. While this
allows the necessarylocal meshrefinement,it leavesa problemof how to connectall of the transmission
lines (TL’s) of thesmall cells to thatof the large one. So far, threeapproachesto this problemhave been
made,generallytreatingthe2D case.[8, 9, 10] Of these,Pulko[10] andAit-Sadi[8] optedto connectall of
thesmallcell TL’s togetherat thecell boundary. ThesemergedTL’sarethenconnectedto asingleTL from
thelargecell. In effect, this imposesa constanttemperaturecriterionover mostof the largecell boundary
which is aqualitative changeto theproblembeingconsidered.Wong[9] avoidsthisproblemby connecting
only theTL of themiddlesmallcell to thatof thelargeone.Dependentvoltagesourcesarethenconnected
to eachof theTL’s of all of theothersmallcellsto linearly interpolatethetemperaturealongthe largecell
boundary. This appearsto correctsomeof theinaccuraciesintroducedby theearliermethods.[9] However,
3
somepotentialproblemswith thisapproacharethatdependentsourcescanreducethenumericalstabilityof
themethodandalsothatsuchactive sourcesno longerexplicitly conserve energy. Theformersituationcan
leadto problemswhenutilizing dynamictime steppingto improve efficiency, andthe latter canintroduce
systematicerrorsinto thesolution. A preferableapproachwould strictly staywith passive componentsto
avoid thesedrawbacks.
This paperwill describethedevelopmentof a non-uniformmeshingtechniquebasedon a 3D quadtree
mesh(QTM) approachanda TLM implementationusingthis mesh.A distinct featureof this approachis
thatmultiple transmissionlinesareusedif necessaryto connectlargecells to eachof their smallerneigh-
bors. Suchan approachavoids imposingconstanttemperaturesalongthe cell boundaryandavoids using
unnecessarydependentsources.However, suchanapproachresultsin a variablenumberof TL’s for each
cell. Further, this TLM implementationwill beusedfor bothconstantandtemperature-dependent material
properties,andthetime stepwill ultimatelybeadjusteddynamicallyduringsimulationto reduceexecution
time. Resultsconfirmingtheaccuracy of this TLM approachwill bepresented,andfinally a full transient
simulationof a realisticSi-basedtrenchdevice will begiven.
2 Model Development
2.1 BasicTLM for Heat Flow
TLM modelingrepresentsa physicalmodelof heatflow asa sequenceof voltagepulsestravelling through
amatrixnetwork of transmissionlines.By computingthevoltage(temperature)at thenodesconnectingthe
transmissionlines,anexplicit, unconditionallystablesolutionto thetelegrapher’s equation canbedescribed
by asetof equationsthatareiteratedin astraightforwardmanner.
Thetelegrapher’s equationis givenby:
�+*-,.� %'/ � / � ,�� �10 / � / � * ,�� * (2)
where %�/ , 0 / and � / are respectively, the resistance,inductanceand capacitanceper unit length in the
transmissionline andV is thenodevoltage.Equations1 and2 have a similar structure.Replacing, with, setting %�/ � / � ����2 � , addinga currentsourceto the left handsideof Eq. 2 andrequiringthe term
involving thesecondtimederivative to benegligible resultsin acorrespondencebetweenthetwo equations.
Once %�/ and � / have beendeterminedby thephysicalproblem,a requirementthat thepulsestravel from
4
nodeto nodein a time step 3 �resultsin the complex transmissionline impedance4 �65 0 / 27� / being
equalto 3 � 2 � � /83 � �where 3 � is thelocaldistancebetweennodes.Therequirementthatthesecondterm
in Eq.2 beneglectablethenbecomesaconditionthat 3 �besufficiently small.
With theseconditions,the transientheatequationcanbe solved in a TLM framework by iteratingthe
following steps:
Nodal Temperatures: At eachnode9 , thetemperatureat thetimestep: , ; � 9 �, is calculatedfrom thelo-
calsourcesandtheincidenttemperaturepulses,; ,=<> � 9 �. Thesubscript? indicateswhichtransmission
line thepulseis on.
Scattered Pulses: The scatteredtemperaturepulses; ,A@> � 9 �, arecalculatedfrom the nodal temperatures
andincidentpulses.
Incident Pulses: The incidentpulsefor the subsequentiteration, ;CB � , <> � 9 �, arecomputedfrom the scat-
teredpulsesfrom the previous time step,accountingfor any time stepchangesanddifferencesin
nodalseparation.
A substantialamountof work hasbeendoneonTLM solutionsof heatflow, includingthedevelopmentof
3D and2D modelsbasedon continuousrectangularmeshes[3]. Oneadvantageof TLM over conventional
finite differencetechniquesis greaterstability. In particular, this allows the time stepto be increasedas
the simulationprogressestowardssteadystate,therebyacceleratingthe computationsubstantially. Incor-
poratingthis into the TLM equationssimply requiressomeadditionalcomputationsfor scatteringeffects
at cell boundariesdueto changesin the time step[10], althoughcareis requiredto ensurethesecondtime
derivative termin Equation2 remainssmall. Finally, by updatingresistanceandcapacitancevaluesduring
thesimulation,nonlineartemperature-dependent materialpropertiesareeasilyincorporatedinto theTLM
framework.
2.2 TLM on a Quad TreeMesh
A full descriptionof the 3D modelbuilding codeis describedelsewhere[11]. This modelbuilding code
will, givenanIC technologydescriptionandlayoutinformation,automaticallygeneratea 3D modelof the
device,discretizethemodel,andsolve for thetemperaturedistribution in thedevice. Thebasisof thiscode
is aquadtreemeshingtechniquethatcreatesacomplex meshgeometryusingrectangularblocksof varying
5
sizes.The modelis createdsuchthat a block andits neighborsform a relatively simplesetof topologies
wherebyeachblockcanhaveamaximumof two blocksonany verticalsideandamaximumof four blocks
on the top or bottom. Fromtheseblock topologies,a network of thermalresistancesandcapacitancescan
beextractedrepresentingthethermalequationto besolved.
Theessenceof themultigrid quadtreemesh(QTM) is shown in Fig. 1(a).Thebottomright cornershows
anareaof local refinement,anda reductionin meshdensityis obtainedby increasingtheblock sizeaswe
move away from it. Theprimaryrequirementof a 2D QTM is thattherecanbeat mosttwo smallerblocks
presentonany sideof agivenblock.Thisallowsfor averyquickexpansionof blocksize,but still minimizes
thenumberof localblock topologies.Theextensionof thisstyleof meshingto 3D canbedonein anumber
of ways,but perhapsthemostobviousis to allow thetopandbottomof ablock to haveatmostfour smaller
blocksadjacent.This resultsin block topologiessuchasshown in Fig. 1(b).
The fundamentalapproachto TLM modelingof the QTM is shown for a 2D casein Fig. 2(a). At the
centerof eachblock is a temperaturenode,anda transmissionline from this nodeto all adjacentnodesis
created.Eachtransmissionline consistsof two impedancesandtwo resistances,oneof eachassociatedwith
eachblock. This is shown in Fig. 2(b) for two nodesdenotedas 9 < and 9ED . Although the associationof%�/ � / with �$��2 � allows for a numberof choicesin choosingthevaluesof % and 4 of eachtransmission
line, a naturalchoiceis to usethe physicalvaluesof thermalresistanceandcapacitanceassociatedwith
nodalconnections.
Figure3(a) shows the basicgeometriesusedto calculatethe two resistorvaluesassociatedwith a link
betweenadjacentblocksin a 2D slab. Theresistance,% � , betweennode 9 � andthecommonboundaryis
simply given by % � 0 2 ���GFIH �where H is the thicknessof the slab,and F and
0arethe widths and
lengthsasshown in the figure. For cells which areof differentsizes,the sameformula holds; however,
multiple branchesareneededand the F usedin the calculationfor a given branchin the larger cell is
reducedto that of the correspondingsmallercell. This subdivision of the large cells resultsin several
parallelresistorswhoseequivalentresistanceis equalto thatwhichwouldhaveresultedhadtheentirewidth
of thecell beenusedin a singleresistor. Extendingthis to 3D just requirestreating H in thesamemanner
asis donefor F .
To calculatethethermalcapacitanceassociatedwith eachnode,thetotal capacitancein a block is deter-
minedby � &KJL& �M, &KJL& �$��� where , &KJN& is theblockvolume, � thematerialdensityand ��� thespecificheatof
6
thematerial.As shown in Fig. 3(b),thiscapacitancecanbeapportionedto eachtransmissionline according
to thearea(or volume)it correspondsto. In 2D, this is given simply as � &KJL& 2PO divided by thenumberof
adjacentcellsalongthecorrespondingedge.
Althoughthecalculationof theline resistancesandcapacitanceswasdevelopedfor a2D mesh,verysimi-
lar expressionsaredeterminedfor the3D case.Theaboveapproachto theextractionof thetransmissionline
resistancescorrespondsmathematicallyto theapproachdescribedin [9] with theappropriatereductionfor
aquadtreemeshandanextensionto 3D. Oncethecapacitanceshave beendetermined,theline impedances
canbecalculatedusing 4RQ � 3 � 27� Q for agiventimestep.
Theuseof this formulationin 2D or 3D QTM resultsin a network of nodes(at theblock centers)con-
nectedby transmissionlines.Unlike previousmultigrid[8, 9, 10] approaches,connectionsonly occurat the
nodes.We denoteeachline asa “branch” with thenumberof branchesat eachnode 9 being S � 9 �. Each
TLM nodein the problemcanbe describedashaving a nodetemperature; � 9 �at time step : andasso-
ciatedwith eachbranch? of thenodeis a resistance% > � 9 �, an impedance4 > � 9 �
, a scatteredtemperature
pulse; , @> � 9 �anda incidentpulse; , <> � 9 �
. Also associatedwith thenodeis acurrentsourceT � 9 �if heatis
beinggeneratedin theblock representedby thatnode.A generalizedTLM nodeis shown in Fig. 4.
UsingthestandardTLM approachandtheabovenomenclature,weobtainfor thenodetemperatureat the
timestep : : ; � 9 � � UVNWYX Q[Z\ > ] ; , <> � 9 �% > � 9 �^� 4 > � 9 � � T � 9 �`_acbd (3)
with, d � WYX Q7Z\ > b% > � 9 ��� 4 > � 9 � (4)
thesummationsareoverall branchespresentat thatnode.
Thescatteredpulseoneachbranchis thengivenby:
; , @> � 9 � � ; � 9 � 4 > � 9 �% > � 9 ��� 4 > � 9 � � ; , <> � 9 � % > � 9 ��e 4 > � 9 �% > � 9 �^� 4 > � 9 � (5)
Theequationconnectingthescatteredpulsesat thetimestep: andtheincidentpulsesat : � b musttake
into accountboth thescatteringat thediscontinuityin 4 at theblock boundariesandthescatteringdueto
timestepchanges.Takingtheseinto accountweobtaina relationfor eachbranchof thenode:
;fB � , <> � 9 � � ; , @>hg � 9^i �Lj >hg � ; , @> � 9 � � > (6)
7
where j >hg � � b �lkY� 4 > � 9 �4 > � 9 �^� 4 >hg � 9 i � (7)� � 4 >hg � 9 i �Yemk 4 > � 9 �4 > � 9 �^� 4 > g � 9 i � (8)
andk � ;n3 � 2 ;Po � 3 �
is theratio of thecurrenttime stepto thepreviousone.For theabove expressions9 iand ? i denoterespectively thenodeandreversebranchat theendof thetransmissionline.
The TLM solutionof the heatdiffusion problemrequiresonly the repeatedsolutionof Eq. 3, 5 and6
with appropriateinitial andboundaryconditions. Two basicboundaryconditionsareneeded:1) a fixed
temperatureon a block faceor 2) a zero heatflow condition acrossa block face. Implementingthese
boundaryconditionsrequiresreplacingEq.6 for thebranchesconnectedto a boundary. Thefirst condition
requirestheimplementationof thefollowing equationfor thebranchconnectedto theface:
;fB � , <> � 9 � �M,!p e ; , @> � 9 �(9)
where ,!p is thefixedvoltageon theface.Thezeroflow conditionresultsin:
;fB � , <> � 9 � � ; , @> � 9 �(10)
for any branchwith a reflective boundaryconditionat theend.
Theinitial conditionusedwasto assumethedevice wasat a uniform initial temperature,,Eq . To impose
thiscondition,all initial incidentpulsesweresetto ,Eq 2 ] .
Theabove formulationassumedconstantmaterialparameters.It is, however, straightforward to update
temperature-dependent materialparametersat eachiteration. An additionalrequirementfor an accurate
solutionis thatthesecondtermin Eq.2 involving a secondtime derivative is negligible. This is essentially
a requirementfor asufficiently smalltimestep.A significantadvantageof theTLM methodis theability to
increasethetimestepsizeasthesimulationprogressestowardssteadystate.Theimplicationsof increasing
thetimesteponsolutionaccuracy will bediscussedbelow.
3 Results
In orderto evaluatetheaccuracy of theTLM algorithmdiscussedabove, theTLM solutiononbotha QTM
meshandauniformmeshwerecomparedtoananalyticalsolutionfor anumberof representativegeometries.
8
3.1 2D slab
The first geometryconsideredwasa 2D slabwith a fixed temperatureon the left sideanda squareheat
generationregion in the centerof the slab. The otheredgesof theslabwereassumedto have a zeroheat
flow condition. The basicgeometryis shown for botha QTM anda uniform meshin Fig. 5(a). The slab
dimensionsare64 ) m by 64 ) m with a16 ) m squareuniformheatgenerationregion (100mF ). For these
initial simulations,the time stepwaskept uniform andwasat a valueof onetenththe smallestRC time
constantpresentin theslab. TheslabmaterialwasSi with a thermalconductivity of b[r ]ts bfu ov F 2 )^w - x ,
densityof ] rzy7y s bfu o � * � 2 )^w|{ andspecificheatof u�r~}�� 2 � - x wereused.
A time dependentanalyticalsolution for this problemcanbe developed[12] usinga Green’s function
approach.This leadsto thefollowing expressionfor thetemperature:
����� �!� �#� � �� � UV � * e � �� � �\Q[� � � * �������`� oE���P�� &�� ]R�C��� ��� Q � �� � � �#�K� ��� Q � * ��e �#�K� ��� Q � � �� Q � _a(11)UV �\� � � � { � ������� � oE�[� �� & � ]R�#�K� � k � � �� � � �C��� � k � � * �Ye �C��� � k � � � �k � � _a
where � and�
aretheslabdimensions,�
is thepower densityover theregion definedby thecornerpoints��� � � � � � and ��� * � � * � , � Q � 9�� 2P� ,k � � w|� 2 ] � and � ��� 27����� .
In Fig. 5, thebasicslabmeshgeometryis shown for boththeQTM anda uniform meshof 32 s 32. This
QTM meshshows an aggressive attemptat reducingthe numberof blocksandcontains220 blocks. The
characteristicQTM meshstructurecanbeseen.Theuniformmeshhas,of course,1024blocks.Also shown
in this figurearegrayscaledepictionsof the temperaturedistribution in theslabat timesof 0.1 ) s and15) s. As canbeseen,thetwo meshesproducevery similar distributions. In Fig. 6(a),a comparisonbetween
the two meshsolutionsandtheanalyticalsolutionis shown for the temperatureof a point at thecenterof
the slab. The temperatureis shown asa function of time andit canclearly be seenthat the solutionsfor
bothmeshescomparevery favorablywith theanalyticalsolution.However, theuniformmeshsolutiontook
considerablylongerto compute.
The residualerror (the absolutedifferencebetweenthe analyticalsolutionand the numericalsolution
normalizedto thefinal temperature)is alsoshown for anumberof casesin Fig. 6(b). Shown in thisplot are
threeTLM solutionsandtwo solutionsbasedon a simplefinite difference(FD) approach.(While bothFD
9
andTLM producevery similar resultsfor a given mesh,theTLM methodhasthepotentialto do sowith
considerablylesscalculationby stretchingthe time step. This is discussedfurther below.) The minimal
error is found for thehighestmeshdensityanda uniform grid of 1024elements.In this instance,theFD
solutioninitially shows a largererrorfor thetransientbut by 5 )� , theresidualerrorsarevery similar. The
low densityQTM with 220elementsshows a larger steadystateerrorby a factorof about4, with theFD
solutionagainhaving a larger error initially. An intermediateQTM meshof 350 elementsusinga TLM
solutionreducesthe error to only around2.5 timesthat of the high densityresult. It is importantto note
that for all casestheresidualerror is lessthan0.4%of thenormalizedtransientandlessthan0.5 x for an
absoluteerror.
To enableaneasyevaluationof theeffect of thequadtreemeshon thespatialtemperaturedistribution,
thetemperaturehasbeenplottedin Fig. 7 in thex andy directionsthroughthecenterof theslabfrom one
edgeto theother. Thetemperatureprofile is plottedfor threedifferenttimesduringtheevolution to steady
state.As canbeseen,bothmeshesproducea reasonablyaccuratesolutionthroughouttheregion.
Theresultsdescribedabove indicatethata2D QTM meshessentiallyproducesasaccurateasolutionfor
thetemperatureasauniformhighdensitymesh,with areductionin thenumberof blocksby afactorof four.
3.2 Non-uniform time stepping
Oneof themainadvantagesof TLM over FD simulationmethodsis theability to reduceexecutiontimeby
optimizingthetimestepduringthesimulation.In this section,weexploresomesimplemethodsof altering
or controllingthetimestepduringthesimulation.To characterizetheerrorpresentin thesimulationduring
thetransient,anerrorparameterwasevaluated[5]. Theparameterusedwascalculatedusing:
� �¢¡¡¡¡ ] 0 / � / � ; � 9 ��e ] ;Po � � 9 �^� ;8o * � 9 �;[3 �^� ;8o � 3 � � ¡¡¡¡ (12)
which wasaveragedover theheatgenerationregion to calculate�-£¥¤¥¦ . This parameteris a directestimation
of thesizeof thesecondderivative errortermin Eq.2.
A variety of algorithmsfor stepsizegrowth was investigatedusingthe QTM of 220 blocksshown in
Fig. 5. The mostobvious methodof increasingthe stepsize is a uniform growth rategiven by ;P3 � �� b � T � ;Po � 3 �whereG is a fixed andtypically lessthan5%. Fig. 8(a) shows the transienttemperature
responseat thecenterof theblockandtheaverageerror, �-£#¤§¦ , for threedifferentgrowth rates.Thesmallest
growth rateof 0.5%follows theanalyticalcurveverywell. Increasingthegrowth rateto 5%causesanerror
10
to develop in the steeppart of the transientwhich thendisappearsassteadystateis approached.Further
increasingthegrowth rateby a factorof 2 producesa largeovershootandthenoscillatorybehavior asthe
solutionmovestowardssteadystate.Theplot of averageerrorparameterclearlyshows the inaccuracy of
thesolutionathighergrowth rates.For T � u�rz¨ %, theerrorparameterstaysessentiallybelow bfu ov , andthe
solutionis well behaved. For the large growth rates,� £#¤¥¦ quickly becomesratherlarge introducingerrors
in the initial part of the transient.For the growth rateof 10%, the error parameteris large for the whole
simulation,rangingbetweenbfu o { and bfu o * .It was noticedthat for longer simulationtimes, the absolutesize of the time stepwould becometoo
large,producingerrorsandoscillationsin thesolution.An approachto minimizing bothsolutionerrorand
executiontime is to usea moderategrowth ratebut limit the time stepto a maximumvalue. In Fig. 8(b),
theresultsof this approacharepresented.A growth rateof 1% wasusedandthreesimulationsareshown
correspondingto amaximumtimestepof 20,50or 100timesthesmallestRC timeconstant(j � < Q ) present
in the slab. The plot of the temperaturetransientshows that whenlimited to 20 timesj � < Q the solution
closelyfollows theanalyticalresult.Using50or 100j � < Q producesincreasingerror;however, thesolution
doesstayrelatively closeto thecorrectanswerandapproachesthecorrectsteadystatevalue.Theeffect of
fixing thetimestepcanclearlybeseenin theplot of theaveragederrorparameterwith adampedoscillation
occurringin � £¥¤¥¦ from thepoint at which the time stepstopsgrowing. After this point, theerrorbecomes
progressively smaller, exceptin thecaseof 100j � < Q , indicatingthatthisfixedtimestepis too large.
Finally, a third approachto timestepcontrolwastriedwhich involvedmonitoringtheerrorparameter. In
thiscase,if �8£#¤¥¦ becamelargerthan ¨ s bfu ov , thegrowth ratewasheldconstantuntil it againdroppedbelow
this threshold,afterwhich it wasallowedto increaseagain.In Fig. 8(c), theresultsfrom this approachare
shown for four differentgrowth ratesof 1, 2.5, 5 and10%. As canbeseenfrom theplots, this methodis
quite successful.A high growth rateof even 2.5 or 5% canbeusedwith a goodresultobtained.A 10%
growth ratedoesproducesomeerror in the initial part of the transient,but even for this extremecasethe
erroris well behavedin thelatterpartof thetransient.
Of the above techniquesmonitoringthe errorparameterwasthe mostsuccessful.Of course,the point
of growing the time stepis to reducethenumberof iterationsusedto simulatethe transientresponseto a
particularvalueof�. In Fig. 9, the total numberof iterationsasa functionof
�is presentedfor four cases
which producean accuratesolution. Thesecasesare: a uniform time stepof u�rKb j � < Q , a uniform growth
rateof 0.5%,a growth rateof 1% anda maximumtime stepof ] u j � < Q , anda growth rateof 2.5% and
11
a constraintof ¨ s bfu ov on the error parameter. As can be seen,the last caseproducedthe minimum
numberof iterationsand thereforeminimal executiontime; however, all the non-uniformtime stepping
algorithmsproduceaverydramaticdecreasein thesimulationtime. Theabove algorithmsareby nomeans
eitheroptimalor comprehensive, but do indicatethelargeadvantageof usingTLM with non-uniformtime
stepping.
3.3 Nonlinear materials
To investigatethe accuracy of the simulationwith respectto nonlinearmaterialproperties,the 2D slab
problemdescribedabove wassimulatedwith a temperature-dependent thermalconductivity givenby:
�©��� q � b � :�ª � � «e¬ q ��� :�ª * � e¬ q � * � (13)
where �Eq � b[r O ¨ s bfu ov F 2 )^w - x , : ª � � e O rz¨ s bfu o { x o �, : ª * � bfu oE® x o * and
q � y[u7u x for
Si. In Fig. 10(a),the transientresponsein theslabcenterfor constantandtemperature-dependent thermal
conductivity is shown. The importanceof including the nonlinearitycanclearly be seen. Unfortunately,
thereis no analyticsolutionto the nonlineartransientproblem,so that numericaltechniquesarerequired
for even thesimplestgeometries.Thesecondplot in Fig. 10 shows the time evolution of the temperature
profile in the � -directionthroughthe slabcenter. Although it is not possibleto get ananalyticalsolution
of the transientresponseof the non-lineartemperature-dependent problem,it is possibleto get a steady
statesolutionusinga Kirchoff transformation[13]. This analyticalresultis alsoshown in thesecondplot,
andit canbeseenthatthesimulationsdo smoothlyapproachthesteadystateanalyticalsolution,providing
confidencein thenon-linearTLM solution.
4 3D block
Theadvantageof theQTM is mostdramaticallyillustratedfor a3 dimensionalproblem.In Fig. 11,aQTM
for a3D blockof heightof 50 ) m anda width andlengthof 32 ) m is shown. Thebottomfaceof theblock
wasfixedat 300K, anda thin heatsourceregion wassituatednearthetop of theblock. Theheatsourceof
25 mW is centeredat theblock, is 8 s 8 ) m in its lateraldimensions,hasa thicknessof 0.2 ) m andis 0.2) m from thetop face.Theblockmaterialis Si.
As with the2D slab,ananalyticaltransientsolutionof this problemis available(assumingtemperature
independentmaterialparameters),resultingin a solutionsimilar to Eq. 11, but with a triple sum. Figure
12
11presentsthetemperaturedistribution of theblock usinga QTM modelwith only 1974blocks.However,
if a uniform � - � meshwereto be used,the numberof nodeswould be 122,880,assumingthe samenon-
uniform meshingin the " direction. A comparisonof the time responseof the modelandthe analytical
solutionis shown in Fig.12. As canbeseenfrom boththeplot of thetemperaturetransientresponseandthe
residualerror, theQTM TLM solutionis veryaccurateover theentireresponsetime. A spatialtemperature
distribution is shown in thesecondfigure. Both theanalyticalandQTM TLM solutionfor thetemperature
at thecenterof theblock down throughthedepthof theblock arepresented.The agreementis excellent,
lendingconfidenceto the3D extensionsof theQTM TLM approach.The situationanalyzedhereis very
similar to thefinal device simulationsto be investigated,consistingof a block of Si with a backsidefixed
temperatureandasmallthinheatingplateatthetopof theblockrepresentingtheheatingatthebase/collector
junctionof amicroelectronicdevice.
5 Full Si tr enchsimulation
Finally, a modelof a realisticSi trenchdevice is shown in Fig. 13(a). The figureshows 1/4 of thedevice
which hasa central1 ) m x 20 ) m emitterwith a collectorcontacton eachside. The technologyhasa
PtSi/W plug contactstructureandtwo aluminummetallayerswereincludedin the modelwith a W plug
via. Thesimulationregion is 64 ) m squareandthewaferthicknesswas100 ) m. Thebacksidetemperature
was300K andthepower generatedin thedevice was1.75mW, uniformly generatedin a region underthe
emitter. Theendsof theemitterandcollectormetalizationwerealsofixedat300K. Thetotalmodelconsists
of 15,000blocks.
A simulationof thetransientresponseof this Si device is shown in Fig. 13(b). Thetemperatureof point
at thecenterof thebase/collectorjunctionis presentedasa functionof time. TheFig. 14shows thechange
in thetemperatureprofileat four differenttimes.
6 Conclusions
This paperpresents2D and3D TLM implementationsof thesolutionfor heatflow, ultimately in realistic
integratedsemiconductordevices.Theimplementationis basedona quadtreemeshwhich is veryefficient
at localizedmeshrefinement. Model building codewill, given an IC technologydescriptionand layout
information,automaticallygeneratea 3D modelof the device anddiscretizethe model. The basisof the
13
quadtreemeshingtechniqueis to createa modelconsistingof differing sizeblocks,suchthata block and
its neighborsform a relatively simplesetof topologies,wherebyeachblock canhave a maximumof two
blocksonany verticalsideandamaximumof four blockson thetopor bottomof a block.
From this physicalmodel,a generalizedTLM implementationis formulatedby extractinga physical
resistanceandcapacitancenetwork. Theformulationallowsfor non-uniformtimesteppingandtemperature-
dependentmaterialparameters.
The TLM implementationwas first testedfor a simple 2D problemfor which an analyticalsolution
is obtainable. A direct comparisonbetweenthe QTM TLM solution,uniform meshsolutionsandfinite
differencesolutionsshowed the QTM meshTLM solutionworks very well. A numberof algorithmsfor
optimizingthetimesteppingwereinvestigatedwith thebestfoundto beastraightforwardmonitoringof an
errorparameter. It wasfoundthat increasingthe time stepduringsimulationcouldreduceexecutiontime
by up to 100times.This is on top of thereductionfactorresultingfrom usingtheQTM over a continuous
mesh.
A simple3D situationwasalsoanalyzedandit wasfound thata QTM of 1974blocks(versus122,880
blocksfor a uniform � - � mesh)produceda very accuratesolutionwith lessthan0.5%error. Finally, an
exampleof a full simulationof a Si trenchdevice using15,000blockswaspresented,displayingthe full
capabilitiesof thesimulator.
Acknowledgement
This work wassupportedby Nortel Networks, theNaturalSciencesandEngineeringResearchCouncilof
Canada(NSERC)andtheNetworksof Centersof Excellence(Micronet).
14
References
[1] PeterB. Johns,A simpleexplicit andunconditionallystablenumericalroutinefor the solutionof the
diffusionequation,Int. J.Numer. Eng. 11, 1407-1328,(1977).
[2] M.Y. Al-Zeben,A.H. SalehandM. A. Al-Omar, TLM modellingof diffusion,drift andrecombination
of chargecarriersin semiconductors,Int. J.Numer. Model.: Electron.Netw. DevicesFields,5, 219-225,
(1992).
[3] Xiang Gui, Paul W. Webb,andGuang-boGao,Useof three-dimensionalTLM methodin the thermal
simulationanddesignof semiconductordevices,IEEETrans.Electr. Dev., 39(6) 1295–1302(1992).
[4] Xiang Gui, Guang-boGao,HadisMorkoc, Theeffectsof surfaceMetallizationtheThermalBehavior
of GaAsMicrowave PowerDevices,IEEETrans.Microwave Th. Techniques,bf 42(2)342–344,1994.
[5] PaulW. WebbandIanA. D. Russell,Applicationof theTLM methodto transientthermalsimulationof
microwave power transistors,IEEETrans.Electr. Dev., 42(4) 624–631(1995).
[6] J.Higgens,ThermalPropertiesof PowerHBT’s, IEEETrans.Electr. Dev., 40(12) 2171-2177(1993).
[7] H. Carslaw, Conduction of heat in solids, Ch.1,Oxford,1959.
[8] R. Ait-Sadi, A.J. Lowery and B. Tuck, Combinedfine-coarsemeshtransmissionline modelling for
diffusionproblems,Int. J.Numer. Model.: Electron.Netw. DevicesFields,3, 111-126,(1990).
[9] C.C.WongandH. Xiao, A combinedfine-coarsemeshmethodfor the transmissionline modellingof
diffusion,Int. J.Numer. Model.: Electron.Netw. DevicesFields,9 405-415(1996).
[10] S. H. Pulko, I.A. HalleronandC.P. Phizacklea,Substructuringof spaceandtime in TLM diffusion
applications,Int. J.Numer. Model.: Electron.Netw. DevicesFields,3 207-214(1990).
[11] T. Smy, D. Walkey andS.Dew, A 3D thermalsimulationtool for integrateddevices– Atar, Accepted
by IEEETransonComputerAidedDesign(Sept.2000).
[12] M. NecatiOzisik,Heat Conduction, Ch.2-4,JohnWiley andSons,1980.
[13] W.B.Joyce, Thermal resistanceof heatsinks with temperature-dependent conductivity, Solid-State
Electronics,18321-322(1975).
15
List of Figures
1 (a) A 2D quadtreemeshshowing refinementand(b) a single3D elementwith 4 blocks
above,2 on theright and1 onall othersides. . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 (a)TLM transmissionline links in aquadtreeand(b) detailsof asinglelink in thequadtree
mesh. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Geometryfor a 2D meshfor calculating(a) resistancesand(b) capacitances.. . . . . . . . . 18
4 A generalTLM nodefor aQTM mesh.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5 Meshgeometriesandtemperaturedistributionsfor a2D rectangularslab. . . . . . . . . . . 20
6 Transientresponseof the2D rectangularslabshowing (a) temperatureand(b) residualerror. 21
7 Spatial temperaturedistributions throughthe centerof the rectangularslab (a) in the x-
directionand(b) in they-direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
8 Transientresponseof theslabusinga QTM with differentstrategiesfor non-uniformtime
steppingsuchas(a) unconstrainedgrowth, (b) 1% growth constrainedto a maximumtime
stepand(c) growth constrainedby limiting �8£#¤¥¦ to ¨ s bfu ov . . . . . . . . . . . . . . . . . . 22
9 Totalnumberof iterationsasa functionof simulationtime. . . . . . . . . . . . . . . . . . . 23
10 Effect of a temperature-dependent heatconductivity on (a) transientresponsein the slab
centerand(b) temperatureprofile in thex-direction. . . . . . . . . . . . . . . . . . . . . . . 23
11 Grayscalerepresentationof thetemperaturedistribution at 100 ) s in a rectangular3D solid
usinga QTM mesh.Thedarkestregionscorrespondto a temperatureriseof 65 x . Dueto
thesymmetry, only onequadrantof theblock is shown. . . . . . . . . . . . . . . . . . . . . 24
12 Temperatureresponsewithin a 3D rectangularsolidshowing (a) transientresponseof max-
imumtemperatureand(b) temperatureprofileasa functionof depthnearthesteadystate.. . 24
13 (a)Detailsof aSi trenchdeviceshowing structureandmetalizationand(b) transientthermal
responseof thedevice. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
14 Four temperaturedistributionsin a Si trenchdevice at increasingtimes. . . . . . . . . . . . 26
16
(a) (b)
Figure1: (a) A 2D quadtreemeshshowing refinementand(b) a single3D elementwith 4 blocksabove, 2on theright and1 onall othersides.
Ri
ZiZj
Rj
nj
in
(a) (b)
Figure2: (a)TLM transmissionline links in aquadtreeand(b) detailsof asinglelink in thequadtreemesh.
17
(a)
n1
n2
WL
T1 T2
T T2
T
1
3
n1
n2
n3
L
W
(b)
16totC
16totC
16totC
16totC
tot8
Ctot8
C
4C
tot8
C
tot8
C
tot
Figure3: Geometryfor a2D meshfor calculating(a) resistancesand(b) capacitances.
18
RjZi
Z
n
n
n
n
Gi
i
j+1
j
j
Ri
j+2
Figure4: A generalTLM nodefor aQTM mesh.
19
Uniform Mesh QTM Mesh
model
z
x
y
� � u�rKb )� � £¥¯ � bfu x
� � b-¨ )� � £#¯ � O ] x
Figure5: Meshgeometriesandtemperaturedistributionsfor a2D rectangularslab.
20
(a)0°
50°
100°
150±Time (µs)
0
10
20
30
40
Tem
pera
ture
Ris
e (K
)
Analytical²Grid Mesh TLM³Quad Tree Mesh TLM´
(b)0µ
50µ
100µ
150¶Time (µs)
0
0.001
0.002
0.003
0.004
Nor
mal
ized
Err
or· FD uniform mesh (1024)FD QTM (220)TLM uniform mesh (1024)¸TLM QTM (350)¸TLM QTM (220)¸
Figure6: Transientresponseof the2D rectangularslabshowing (a) temperatureand(b) residualerror.
(a)0¹
20¹
40¹
60¹
x (µº m)
0
10
20
30
40
Tem
pera
ture
Ris
e (K
) Analytical »Uniform mesh TLMQTM TLM¼
t=15 ½
µs
t=1.0 ½
µs
t=0.1 ½
µs
(b)0¹
20¹
40¹
60¹
y (µº m)
0
10
20
30
40
Tem
pera
ture
Ris
e (K
)
Analytical»Uniform mesh TLMQTM TLM¼
t=15 ½
µs
t=1.0 ½
µs
t=0.1 ½
µs
Figure7: Spatialtemperaturedistributionsthroughthecenterof therectangularslab(a) in thex-directionand(b) in they-direction.
21
(a)0¾
50¾
100¾
150¿Time (µs)
0
10
20
30
40
50
Tem
pera
ture
Ris
e (K
)
AnalyticalÀG = 0.5%ÁG = 5%ÁG = 10%Á
0Â
50Â
100Â
150ÃTime (µs)
10−7
10−6
10−5
10−4
10−3
10−2
10−1
eÄG = 0.5%ÅG = 5%ÅG = 10%Å
(b)0¾
50¾
100¾
150¿Time (µs)
0
10
20
30
40
50
Tem
pera
ture
Ris
e (K
)
AnalyticalÀMax Æ
∆t = 100τmin
Max Æ
∆t = 50τmin
Max Æ
∆t = 20τmin
0Ç
50Ç
100Ç
150ÈTime (µs)
10−7
10−6
10−5
10−4
10−3
10−2
eÉMax Ê
∆t at 100τmin
Max Ê
∆t at 50τmin
Max Ê
∆t at 20τmin
(c)0¾
50¾
100¾
150¿Time (µs)
0
10
20
30
40
50
Tem
pera
ture
Ris
e (K
)
AnalyticalÀ1%2.5%Ë5%¿10%
0Â
50Â
100Â
150ÃTime (µs)
10−7
10−6
10−5
10−4
10−3
10−2
eÄ1%Ì2.5%Í5%Ã10%Ì
TemperatureResponse AverageErrorParameter
Figure8: Transientresponseof theslabusinga QTM with differentstrategiesfor non-uniformtime step-ping suchas(a) unconstrainedgrowth, (b) 1% growth constrainedto a maximumtime stepand(c) growthconstrainedby limiting �8£#¤¥¦ to ¨ s bfu ov .
22
0Î
50Î
100Î
150ÏTime (µs)
0
2000
4000
6000
8000
10000
Num
ber
of It
erat
ions
G = 0.5%ÐG = 1%, Max Ð
∆t of 20τmin
G = 2.5%, e < 5x10Ð −4
Uniform time steps at 0.1τmin
Figure9: Totalnumberof iterationsasa functionof simulationtime.
(a)0Ñ
50Ñ
100Ñ
150ÒTime (µs)
0
10
20
30
40
50
Tem
pera
ture
Ris
e (K
)
Nonconstant κConstant Ó
κ
(b)0Ô
20Ô
40Ô
60Ô
x (µÕ m)
0
10
20
30
40
50
Tem
pera
ture
Ris
e (K
)
Steady StateÖTLM at 15 ×
µsTLM at 5 ×
µsTLM at 1 ×
µsTLM at 0.1 ×
µs
Figure10: Effect of a temperature-dependent heatconductivity on (a) transientresponsein theslabcenterand(b) temperatureprofile in thex-direction.
23
Figure11: Grayscalerepresentationof thetemperaturedistributionat100 ) sin arectangular3D solidusingaQTM mesh.Thedarkestregionscorrespondto atemperatureriseof 65 x . Dueto thesymmetry, only onequadrantof theblock is shown.
(a)0Ø
20Ø
40Ø
60Ø
80Ø
100Ø
Time (µs)
0
20
40
60
80
Tem
pera
ture
Ris
e (K
)
AnalyticalÙQuad Tree TLMÚ
0Ø0.002Ø0.004Ø0.006Ø0.008Ø
Nor
mal
ized
Err
or
(b)0Û
10Û
20Û
30Û
40Û
50Û
Depth (µm)
0
20
40
60
80
Tem
pera
ture
Ris
e (K
)
AnalyticalÜQuad Tree TLMÝ
Figure12: Temperatureresponsewithin a3D rectangularsolidshowing (a) transientresponseof maximumtemperatureand(b) temperatureprofileasa functionof depthnearthesteadystate.
24
(a) (b)0Þ
5ß
10Þ
15ß
Time (µs)
0
2
4
6
8T
empe
ratu
re R
ise
(K)
Figure13: (a) Detailsof a Si trenchdevice showing structureandmetalizationand(b) transientthermalresponseof thedevice.
25
t = 0.05 )� , 3 � £#¯ � yàráu K t = 0.25 )� , 3 � £#¯ � yàrzâ K
t = 5.0 )� , 3 � £#¯ ��ã ráu�¨ K t = 15.0 )� , 3 � £¥¯ ��ã r ] â K
Figure14: Four temperaturedistributionsin aSi trenchdevice at increasingtimes.
26