Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

MULTIDISCIPLINARY DESIGN OPTIhIIZATION WITH SUPERELEMENTS I S ;CISC/NA4STRAN

The method usctl in hlS('/NASl'11AN for computing sensitivity coefficient,\ (or gratlirnts), called the scmianalytical rnetliod, is rnhancctl to inclutlc tllc cor~cept of supcrelcments. Tllc superc~li:nwirt. cq~~ilihrinrn ccluatiorr arc dilfcreritiatcd r~sirlg a finite diffwmce tccllniqui~ to i.o111~)11t~~ solut,ior~. The r c d ~ ~ i t tecllnique io~~ thc change i n thc s y s t e n ~ used in the analysis phase to reduce the s u p c ~ r ~ l ( ~ nstifflr~r~t ness. Illass, tlal~lpingslid load matrices is also used i l l the sellsit ivity colnputations. First ortlrr respolrw scnsitivitics S~~I(VII(\. arc i m ~ r l p n t d11bi11g a f'urward f i ~ ~ itlifFrrr~nc.c' tc J)esign \~iriahlrys(botli shalw ar~tlsizi~rg) can I)c tlc~tirretl n c . 1 . 0 5 ~x i ~ l ) r ~ . c l ( ~ locally for a s n l ~ e r c ~ l ( ~ r n or cn c~ a ~ r1ji. l i ~ l k d rrlcmt Ijoll~itlarics.'I'hc sizing varial)lvs ran I)o ally rtxal field specified on a propert,y bl~lk data entry. For sllape vari;ll)l(.s, the tjasis vector conccyt is utilizrd, wllcrc tlre slrnpc rrariablcs arc scalar rnr~lt ipliers of t hv hasis vwtors. 1)ilfbrvnt sllap~ vectors may bc xpccilicd. witir t h r assu~rlptiorr t.liat a s a lincar colr~l,iirntion the optirllal design can bc c~xprcsxd of t h e hasis vt~-tors. Iksign rwponscs [nay hc I'IQIII an! of the s l ~ l ) p o ~ t esoll~tion tl disciplines. u ~ t lrlall t~olistriri~~etl of pri~rlar?.) sufor I~otlr j)rirlrary a d wc.onrlary (i~~lagcls 1 Introduction pereler~icvts. First. tach solution discipline is proressctl iili:rcn~ent,i~lly ' 1 ' 1 1 ~ 1)a1)('r pres('nt5 t l ~ c t(di~~iq of' w ri~~rIt,idisci~)li~~i~ry tlein the analysis phases. Second, the design i.ons~r.aints nlltl t \~~pf~~~cle (or ~ nsulxt e n t sr ~ ~ i t ~ ~ ill rvs) sig~ropt i ~ ~ ~ i io11 za~ t v i 11 their sensitivities are co~llpuied i n c r e ~ n e ~ ~ t a for l l y t.acll de.\IS('/S.ZS.L I i ~ \ N .I 11rcapxljility lrlay used for csamplc. sign discipline. Third, thc t l c s i g ~ comt,ritints ~ and serrsit,ivi l l t \I(. aut o l ~ ~ a t rtlcsigrr tl of a large, structure rnotlr.let1 with ties for the various tlisciplines are nlergetl and c o ~ ~ v e r t tco~ l srrpewler~~er~t~s. for ~ n i n i m u m weight sltbject to a set of I)ean approximate tlesig~irnotlcl. Finally, the opti~uizermodliavio~~ral al~d sitlc, constraints in a n~~rltitlisciplinary design ifies the approximate design model. The t l c s i g ~process ~ is enviro~rn~rwt. Thr. a r l \ a ~ f a g rof , using a s ~ ~ p e r e l m w n niodrl t iterative in nature, with the design variables and constraints l e ~ ~ t ill tlesig~r opti~uizationis over it n o n - s l ~ p c r e l c ~ ~1rrod~1 for all superclenients considered sirnultancolrsly by thc opt,llat only superslernents affected by the design process have t irnizrr. to be rsgrneratetl during thc search for the optirnal design. D(-sign synthesis of large st,ruct,~~ral syst,en~s requir~s sigst ruc\Vhcreas, for t h s 1lol1-snperclerne11tmodel, t h e wl~ole nificant coniputer resource allocation. Furtherlnore, mantnrc has to he regeneratrd, which results in a si~bstant,ial agement of t h e enormous data creat,ed during matrix gc3n~ w n ally i l l cornpnter tirne. eration places an unnecessary hurclen on the analyst or rleDesign srmsitivity analysis and optimization first, introsigner. To alleviate this burden, an atlvar~ceddatabasc featl~~cccl in Version 66 of MS('/N/\Srl'R..\N, was originally limt~11.c of MSCINASTRAN called the multi-master teck~nic~uc~ ited t o t,hr residual structurt,. In Vrrsion 67, the design [2] is available, which provides the facility for scgmsnting thv sensitivity capability was extended to upstream supereletlatahase for effective utilization of disk space. This permits r~wnts[ I ] ant1 in Versior~ 68, the sensitivity capal~ilityfor a solution by processing groups of supt.relenients upstrcarn superelerner~ts was extended t o inclntle des-ign 01)i r ~ arr ir~depentlent manner arltl o f f ~ r spractical data mantirnization in a rnnltidisciplinary design er~vironmcnt,.T h e agtvllrnt for large rnotlcls. suIjportc:tl solntion disciplines are: aeroelasticity. bucklir~g. frequency response, transimt response, normal rrlocles and statics. \iersion 68 is in development at MSC; ant1 earlier versions are commerically available.

ABSTRACT

Copyright 8 1992 by the American Institute of Aeronautics, Inc All rights reserved.

Examples arc presented and the final designs are comp a r d to the optimal designs of non-superelement (i.e., residuid only) nlodcls. The first example, a Seventy-Two 1)ar truss [3] rnodelcd with two suprrelenierits, is optiniized for rnil~ilrir~m weight subject to a set of static response constraints. For thr: second example, the intermediate cornplvsity wing [ I ] : a rliiriimurl~ wcight tlesigri is sought subject, da e~ sig~ constrairits ~ (sthtics alltl to a set of r n u l t i d i s c i p l i ~ ~ cigcnvalrre). Finally, thc third exirrriple sch(~rriatically(I(,scriI)es use of thc. mnlti-maslcr terhniclue.

111 t,he superrlenicrit formulation ['i]. tile l)asi(. co11c~l)t ( O I . rris to divide t11c s t r ~ ~ c t uillto r e sewrill sr~pcrc~l(~~tlt=nth ~ t ~ ( l ( ~ f i l i ( ~as l t lir. g i o ~ ~ sTlie ) . int,erivr d e p e s of C r e f ~ l o(dofs) ' o ' o r otrlcltrl .ict , arc3 colltlenwtl to tllc c'strrior I)or~rr(lar\. ~ c1~ 1 l1 ~' r c ' . s r l . 'Slrii static. c.011tlegrres of Crertlo~n.t l ( ~ f i ~as tlei~satiorl pro(.(dure i:, r(y)(>itt cd for all the s ~ ~ p r ~ r f ~ l c ~ sn ir~ r c.ni the motl(11, Leading to a s n ~ a l l ( ~ set r of syxten~e q ~ ~ i l i l ) r i ~ r ~ i ~ equations to solvt,. i ll lit (,'or~sitlrl t hc cquilil)riuu~ cyuat,ion for a s ~ ~ p c w l r ~ n c terms of its interior and exterior (or \)oundary) dot's.

Related Work
[r)]

In standard b l S ( ~ / N A S 7 ' K A N not,atioli. a bar over a tlrvelopetl a tcchniql~rfor tlrsigrl optisymlml rrlearls tllc quantity shown is a part ition of i t k parent n~izationwith s ~ ~ b s t r u c t ~using ~ r e s adjoint variahl(,s. Iiirscli matrix. Thc symbol will he rcusetl withor~t the lmr when [(i] 113s tlrvnonst,ratc~(I the tlesigl~optimizatioli of s t , ~ . ~ i c t ~ ~ r e s terms arc added later. other is oprr~otleletlwith sul~structnrcs, where each sul)stn~cture T h e rcduced boundary eqt~ationis: tin~izctlintlrpentl(~nt1q.. This approach, which requires ilerII~CS until convergeliw. is inefficier~t ating hr't.tv(wi sl~ljstrr~ct arltl can lead to lion-optinial soIut,ior~s. The procctlure in MSC'/NASrrRAN, just as in refererice [5], treats the tlesigli variables and constraints si~r~ultaneously for all the supcrelements.
C' :I and . Arora

3
r

Problem Statement
antl the homclary transformation matrix is will

I,he prot)lem staternrnt is t,o seek the set of tlcsign variables

X,which

subject t o

The [I<,,,] matrix represents the stiffness of the superelernrnt ~neasurcd at it bo~~rltlary poirits. l'he {I',,} vector represents Imundary loads t,ransrnittrtl to the rerilaintlw of t,l~c, s t r ~ ~ c t , ~This ~ r e . static couderrsatio~l is rrpc~atetlfor all supc~.c~lerncnts i l l the model. dowrr to t h r r e s i d ~ ~ st a lrr~ctnrc.

where i=l,NL)V (number of independent cksign variables. which may he linked across superelcrrient bomdaries) antl j=l,NCON (total nnrriber of design colistraints). F ( X ) is the objective function, G ( x ) is the response constrairrt set ant1 .rl arid a" arc tlic side cor~straintson the tlesigrr variables. Note thc design variables may be linked acros. sl~prr(:lcrrlentbountlaries.

Static Response Sensitivity ('onsitler a l . Y (.lralrgc> resulting i l l a 1A' cllange i l l t,hc st,iffr~c%ss antl At' change, in t11r loatl vector of c.clr~ilil~ri~m eql~at ion ( 4 ) :

Derivation of Superelement Design Sensitivity


w ~ i s i ivit t 1.

For a cornplctcx tlvrivat ior~ of iuperrl~~rllcnt tleiiglr

t I I C r.ratlcr is tlircctc.tl t,o rcl'erencc [ I ] . :\ concise tlerivat i o ~ ~

for static a11cI c.igr~r\-aIuc r~xs~)ot~ sensitivit~. srs is l ) r ( w ~ r ~ t ( d herc to clarify tl~c.r~trtliotlology.

where the right hand side terms are known as pseudo loatls, resulting from t llc design va.riahle perturbations. The stiffness matrix is carried out perturbation of the load a ~ l t l ~lsinga finite difference method at the element level. The pseudo loads have NDV x NLC columns, where NDV is the number of independent design variables and NLC is the numbcr of load cases. Tlir change in the system solution. A[:, can he evalmted using the same reduction method used for analysis. First ortler response sensitivities are then evaluated using a forward finite difference scheme

the model, excluding external superelements. +k is the kth eigenvector of jth superelement. The default Guyan retlnction technique is exact for stiffness but approximate for mass. The assumption is that significant masses are assigned to the boundary points. Therefore, advanced reduction techniques such as Generalized Dynamic Reduction (GDR) [7] or ("omponent Mode Synthesis (C'MS) [7] are recon~~nenclecl to enhance the mass matrix reduction.

Design Optimization - Solution


200

where X- Design variable, GI- is the solntion vector change in the analysis solution vector due to a perturbation A.Y.

AM-

The design scwsit,ivity and opt.imization procedure in b I S ( ! / N A S r l 1 i A N Version 68 is available in SOL 200 - a st.ructur~tl solution sequence,. This is a multicliscilinary supwelenicnt s o l ~ ~ t i o sequence n supporting the following disciplines: aeroelast,ic, buckling, eigenvalue, frequency response (direct or rnodal), statics and transient response (modal only).

Eigenvalue Response Sensitivity The eigenval~leresponse is a function of the g1ol)al structure antl, therefore its sensitivity is computed at the residual superelenlent level only. Consider the following eigenvalue problern:

User Interface

cvhel~ I<- stiff~lessmatrix. i l -is the mass niatriz. Xis the square of the natural frequency of vilxation antl 4is thc- dynarnic rnocle shape corresponding to eigenvalue A. If we consider a AX c11angeresulting in a A l i ' change in stiffness and change in the mass, we obtain the cllange i l l XI

The design opt,i~nization bulk data entries illt.rotlucetl in bIS('/NASTKtZN [S] for the rlon-st~pereleme~it capa.l)ility are unmodified, with the exception of the DESOBJ a.nd DRESPl bulk cla.ta entries. The DESOBJ entry, which specifies the objective response, has a new field for specifying a snperelement ID. The DRESPl entry used for defining a design response, allows specification of superelement ID for volume and weight responses. The design entries are briefly described in Table 1.

Table 1. Bulk Data Entries

Expanding this equation as a first order variation, pre~llultiplyingboth sides by $* and using equation (12) we

DESOBJ DESVXR I)LINI<

I Design objective
/ Drsien variable

I)OPTF'RLI I O~tirnization~aranleters DKESP1 Response quantity DHF'\P:! / Svnthetic remonse

I Design variable linking


I

DSC'HEEN Note tllat equation (14) is valid only for the case of distinct eigenvalues. In the superelement context, eigenvalue sensitivity for the kth eigenvalue of the structure is simply the sum of the contribution that each superelement makes to the kth eigmvalue derivative:
I '

1 ('onstraint screening data

/ Tahlr ronstants I)C.(;RID I Design variable to grid relation I)VPRI<LI ( Drsign variable to linear property relation

where the superscript 'j' is the jth superelement, excluding external superelements, and subscript ' k ' is the kth eigenvalue. NSE is the total number of superelements in

Mult i-Master Technique

8.1

Seventy-Two Bar Truss

Tllc tlat,abase segmentation feature of MS('/NASrYR.4N is an extremely powerfnl and versat.ile procedure available for solvirrg very large, models which are subjrtct to colrstraints o n disk spa.ce and CPU availability. This feature enables a piece-meal solution by processing groups of ~ u ~ e r e l c n l e nin ts a stand a l o ~ w fashion. Normally, all processing information for s u p e ~ l e m e n t sis stored in the DBALL data base set, (dbset). The datablocks can be retlirectetl to different files by using a number of location PARAMeters. The two most comnlonly used are DBUP and DBDN.

This elementary problem from reference [5]is modeled with two superelements (see Figure 1). Elements 1-30 are assigned to superelement 1 and the remainder 31-72 are by default assigned to the residual superelement. The design objective is to minimize the weight of the structure subject to displacement and stress constraints. The design variables are the cross-sectio~lal areas of the bars, with members linked to enforce symmetry. The design data is given in the table below: Table 2. Iesign Data for Seventy-Two Bar Truss Weight -0.25~n Cl,2 0 . 2 5 ~ ~ ~ Displacement Constraints (Grids 1-4) Stress ( p s z ) -25E3 u,,,,,~ 25E3 (Elements 1-72) Minimum Area 0.01in2

DBUP - the location where 'upstream' processing information is stored a.ncl is only needed for data recovery for the supereleme~it. This dbset is typically very large.

I Objective

<

<

<

<

DBDN

- the location where 'downstream' processing is stored. The reduced clamping, mass, stiffness and load matrices are stored here. This dbset is small and is only needed when processing any superelement which is downstream of the current superelement.

Table 3. Final Design for Seventy-Two Bar Truss


ll12\

The database for each superelement or group of superelenients is partitioned into a upstream dbset (DBUP) containing data required for data recovery and a downstream dhset (DRDN) which contains the bare minimum of information for downstream processing. The individual databases are then attached to the downstream steps using the DBLOCATE capability [2].This enables incremental or even concnrrent processing of superelements since each group of superelements has its own independent database. For design sensitivity and optimization in SOL 200, the superelernents of interest can be processed at the user's tliscretion with the aid of the SEDV and SERESI' case control entries ['I. Thus only the necessary dbsets have to be online. Since t,his is a ~ n a n u a l procedure the optimization phase is limited to a sil~gletlesign cycle. The user has the option of rnanually updating the designed grids arid properties; or usir~g the DHJ,O(~'A'I'E feat,ure for restart,ing t,he tlrsig~i01)timizat,ion task

De5ign Variable 1 2 3

Elements

4
5 6 7 8 9 10 11 12 13 14 15 16 I Weight (11 Max. Constraint

Illustrative Examples
Com~arison of the final d e s i~n values in table 3 indicates that the final design variable values for the superelement model agree closely with the non-superelement model. This is to be expected since the accuracy of the static response values and their corresponding sensitivities are not affected (within numerical roundoff) by the static condensation procedure used in reducing the superelement matrices.

demonstrate the methodThe following example ology of designing structtlre modeled with superelements. All problc~ns were run on an IBM RS/6000 (Model 950). A fil~ite difference step size of 1.0E - 4 was used for t,he tlesign sellsit ivit y calculations.

~notlclc(l 11sing invariant corlc~rntratcilmasses. A ~ u i n i r n u n ~ \vtxight (lcsign is sought s l ~ l ~ j r to c t a set of static constrai~rth The design t,ask for this rnoclcl (originally an AS'PROS (vo~r-;llisc.s st rcw ant1 ply Hill-lailure illtliccs) ant1 the, second model) is t,o demonstrate the multidisciplinary design (statcigeuvalue (torsional) is constrailled to be above 40 IIz. The ics antl normal rnotles) of a wing stri~cture [6] rnodcled tlefar~lt(:uyan reduction procedure is used with the 1,ancwith superclen~ents.The structure is partitioned chordwise zos rigen\raluc extraction rnethod. Thc c o n v ~ linearization .~ into t,hrce supereler~rents (sec Figurc 2): with a cantilevered method for constrairlt approximation is s ~ l e c t k d for all the bountlary i~ontlitionat thi, root. All t h r w s i ~ p e r c ~ l e r ~ ~ e n t sresponse ttypes. Fiual designs for thc no~l-sujjerclcmentverh a w sizing variables and constraints. The structure is mods11s superele~rlrntmodels are as follows: clid usir~g 1-D (mass a11d rod) and 2-D (plate and shear panel) elcmmt ty11c.s. T h e graphitr-epoxy wing skins are for Intermed~nte ~nodi~lr with d c o r r ~ ~ ~ o splat,e i t e (~lenli~nts. antl t h i ~ t l~ickni,ss Table 1. Fmal Des~gns of' i,acl~ply gror~pfor each clement tlesignetl i n d e p c ~ ~ t l r ~ ~ t l y . ('ornplex~t y M lng Rot,h top antl hotton1 wing skins lor the same ply oriiwtatio~ ~~ I I Clir~kcatito c > ~ ~ f ' oi l) ~.c~ ' ~~nlc ant1 ~ t r t11v ~ . pl) o ~ . i c ~ l l i ~ wi~~ si~rfaceb g arc c . o n r ~ c ~ t with id tioils ilr(' 11i'liI l i x ~ d .711~' tlrr ~,otl\ a ~ ~I ItI P l cross-srctiorlal arras arc si,lectcd as tlr-iigu varial)lr~s.711e spars antl webs are nlotli~l(d using tllc shear parlols arrtl the thiclincss of t,he eleme~ltsare also part of the il(~sig~lcd proptrtir7s. 7 ' 1 1 ~ ~ is. a total of 153 i~ltlt-l)cndt t o t w o tliwt sizi~lg tlr>sigu variahlrs. 'I'he ~vingis s u l ~ j ~i c The final des~gnq fol the 5uperelernent models ngree stat ic loatl caws ri~prcsent at ivi, of a s u h o ~ ~ arlcl i c a slipcrclowl) w ~ t h the non-superelement model. Inspection of the sonic air loatl. The ~lolrstructuralmass of t,hc str11ctrn.e is

8.2

Intermediate Complexity Wing

2. Internlediate ('omplcxity Wing

initial frequency for the torsional mode (second) reveals the need to use advanced reduction techriiyues t o match tlie analysis results of the non-superelement case. The error increases with the n u n h e r of superelenlents and, as reported in [ I ] : even with a few superele~nerits,the eigenvalue sensitivity coefficients for the higher modes may be in error. Even tho~~g the h superelcrnent rriodels converged in less computer time compared t o the non-superelernent model, insufficient d a t a is available to make a general recommendation for t h e method.

8.3

Superelement Design Optimization using the Multi-Master technique

T'sc of 111~ mdti-nlaster t,ccl~niql~c is p ~ . c w ~ ~ r t iu . c ~ la s c l ~ e ~ l ~ afashion, tic with emphasis on the t c c h n i q ~ ~ rather c than on specifics. C'onsider the tlesig~ropt,i~rlizationof a large structure modeled with a multi-level snprwlelriel~ttree structurcl as illustrated in F i g ~ ~ 3. r e The superrlenrcnt t,rec c o ~ ~ s i sof t s three levels: the top level is called the 'tip level' (srtperelements I0 1, 102, '20 I , '20%and 300); the secoutl level is called the 'collectol~'level (superelen~ents100, 200) which

contain tlie condensed data from the tip (or upstream s11perelcrrrcnts) a ~ i t l the last Icvel is the resitl~~al (superelenic~nt 0). The system solution is computed at this level. To permit ir~crementalor eve11 concurrent processing, each superelement or group of superelements is processed in a stand alone nmnner. The databases can be segrrrcnted as ilh~stratedin Figure 4. This enables storing only m i n muin information required for clownstrearn processing. 'I'lrus liniitations on disk space can be overcome by attaching o ~ d y the required databases. Once the system solution has been comput,ed a t t h e residual level, the path can be reversed for performing d a t a recovery for the upstream supcrrlcments. For design sensitivity analysis and optimization, databases for superelements with design variables and/or constraints have t,o be online. To facilitate the processing. two case control entries are available:

SEDV

- Selective design variable processir~gfor a set of superelemer~ts

SI4;RESP - Selective response sensitivity matrix gencrat,ion for a set of superelernents.

3. hlult,i-Level S ~ ~ p f v l e n ~ Tree ent

Superelements 100, 101, 102 and 0 are identified for redesign, while superelements 200. '201, 202, and 300 are to be l~clclfixed in tlie tlesigli task. T11e first two runs create the sr~perelen~ent matrices for supert4emcmt groups A arid B. The tlowlistream information is directed to the DRDN dl~sets.The third run for superelement 0 (i.e., residual), assenihles the reduced matrices of the upstream superele~nents (contailled in the DBDN dbsets) and solves the system matrices. The nest run reverses the process ortlcr antl goes up the branch of group A to perforn~data recovery. Tlic final run for design optimization attaches all dbsets created for superelement groups A antl C. Superelement group B tlhsets are not required lxcause they do not particapte in the design process; i.e., these superelements do not have &sign variables and/or constraints, and are not downstrea.m of any designed superelements. By attaching only required dbsets, the available disk space is used efficiently.

Conclusions

h m r r u* .. L mSE Tnm

4. Split L)at.al)asc.and its Atlvantagc~s

Consider the run sequence illustrated in Figure 5 , where the sllperelements are grouped as per the following:

The multidisciplinary design optimization of structures modeled with superelements (or substructures) is demonstrated in MSCJNASTRAN via example problems. The first example, the Seventy-Two bar truss, is a relatively s~nall problem. The final optimal designs for the superelement model agrees closely with the non-superelement model. The second example is the design of the intermediate complexity wing model subject to multidisciplinary design constraints. Comparable final designs for the superelement models are obtained. Since the mass matrix reduction is approximate, the accuracy of the eigenvalue response and its sensitivity is impacted. A deficiency in the current implementation is that mode tracking is not available. Constrained modes may change their order of occurrence as the design is modified, and orthogonality checks need to be made to automatically constrain the user requested modes. For this example, the second mode was verified to be the torsional mode throughout the design psocess for both the superelement and non-superelement models. T h e efficiency of the design procedure can he improved by exploiting the concurrent and/or parallel processing capabilities of the coniputer hardware and software. With the aid of the multi-master method, concurrent processing is lxesently ava,ilable in MSCJNASTRAN.

Table 5. Superelenlent Grouping Group Superelement IDS A 1 100, 101, 102

MATRIX GENERATION

3.

MATRIX GENERATION & SYSTEM SOLUTION

DBUP

5.

DESIGN SENSITIVITY & OPTIMIZATION

10

Acknowledgments

11

References

1. Patel H. D., "Development of superelenient design sensitivity in MSCINASTRAN," Proceedings, 19!)1 hlSC World Users Conference, Los Angeles. C'A., Xlarch 1991.

2. "Int,roduction to V67 Seminar Notes," The MacNcalScliwentllcr rorporation. Los Angeles, (IA.. Septem1)cr 1991. 3. Haftka K. T.. Gnrdal Z. arid Iiarnat M. P.. Htmr rrfs Publisllof'5'2ruclr~rnlOptirr~iintion.Ii111werAcatle~r~ic f.13, 10!10.

5. Arora 1. S arid Go\ril A. Ii., "An efficient method for optirnal structural design by sul~st,ructuring,"C'orrrp u t ~ r sccrrrl .Strr~ctur.es.Vol. 7,p p . 507-51,j.

7. (;ochel M. .A,. "Hantll~oohf o ~ Supelelement Analy


515."

The AIac Neal Schnendle~Corpolat~on.1982.

8. hIoore G . .J., "MSCI/NASTRAN Design Sensitivity and Optimization User's Guide, Version 67," The ;\/lacNeal Schwendler Corporation, 1992.

You might also like