A direct approach to structural topology optimization

Date
2017-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: This dissertation addresses various topics that emerge from the unification of conventional structural optimization—based on ‘sequential approximate optimization’ (SAO)—with the alternative ‘direct’—or ‘simultaneous analysis and design’ (SAND)—formulation of the structural topology design problem. Structural topology optimization—in the form of a ‘material distribution problem’—is a generalisation of structural optimization, encompassing and simultaneously addressing al the aspects of structural design. In structural optimization, SAO techniques are preferred because the number of structural analyses—which are expensive, computationally speaking—are ostensibly minimised. However, particularly in the presence of local state-based constraints—e.g. local stress constraints—the sensitivity analyses which accompany traditional ‘nested analysis an design’ (NAND) methods require a prohibitive number of structural analysis runs per design iteration. In the alternative SAND setting, structural analysis is conducted approximately and sequentially: the finite element (FE) equilibrium equations are retained as a set of nonlinear equality constraints and the state variables—i.e. displacements—form part of the overall set of primal variables. Therefore, the FE equilibrium equations may only be satisfied at convergence of the optimization algorithm, and the complex and expensive sensitivity analyses associated with state-based constraints, simplify to the calculation of partial derivatives. Moreover, the equivalent of a single structural analysis only is required per design iteration, notwithstanding the imposition of a large number of state-based constraints. Based on a dual method in theory, we propose a separable and strictly convex quadratic Lagrange-Newton approximate subproblem for use in SAO of the SAND formulated topology design problem. In classical (simply-constrained) minimum compliance design, the dual statement of the subproblem is equivalent to the ever-popular optimality criteria (OC) approach—a class of NAND methods. This relates, in turn, to the known equivalence between dual SAONAND algorithms based on intervening variables and the OC method. Due to the presence of nonlinear equality constraints, the classical SAO procedure (exclusively geared, traditionally, for inequality constrained problems) is extended to a general, nonlinear and nonconvex, mathematical programming framework. It turns out that conventional techniques of enforced convergence and termination in traditional NAND-based SAO may be transplanted into the SAND setting with only minor complications. It is demonstrated that the compounded issues of existence of solutions, mesh-dependence, local minima, and macro-scale manufacturability, may be addressed in a computationally efficient manner by the imposition of so-called ‘slope constraints’—point-wise bounds on the gradient of the material distribution function. For global optimization, random multistart strategies may be pursued. A specialized version of ‘linear independence constraint qualification’ (LICQ) may hold in many practical situations, and because the global stiffness matrix is not inverted per se, material density variables are permitted a value of zero on the lower bound. Hence, singular local minima are feasible and available—in both simply-constrained and local stress-constrained problems—and may be converged to with standard gradient-based optimization methods without having to resort to any relaxation or perturbation techniques whatsoever.
AFRIKAANSE OPSOMMING: Hierdie proefskrif spreek verskeie aspekte aan wat volg op die samevoeging van konvensionele struktuur optimering—gebaseer op herhaalde benaderde optimering (sequential approximate optimization; SAO)—met die alternatiewe ‘direkte’—of ‘gelyktydige analise en ontwerp’ (simultaneous analysis and design; SAND)—formulering van die strukturele topologie ontwerp probleem. Strukturele topologie optimering is ’n veralgemening van struktuur optimering—in die vorm van ’n materiaal verspreidings probleem—en dus kan al die aspekte van struktuur ontwerp gelyktydig aangespreek word. In struktuur optimering word SAO tegnieke verkies, want die aantal struktuur analises—wat baie duur is, in terme van berekeninge—word o¨enskynlik geminimeer. Nietemin, wanneer lokale spanning begrensings bygevoeg word, vereis die sensitiwiteit analise wat gepaard gaan met tradisionele ‘geneste analise en ontwerp’ (nested analysis and design; NAND) metodes, ’n groot aantal struktuur analises. In die alternatiewe SAND raamwerk word struktuur analise benaderd en herhaaldelik uitgevoer: die eindige element (EE) ewewigs vergelykings word behou in die optimering probleem as ’n groot aantal gelykheidsbegrensings, en die verplasings vorm deel van die algehele stel primale veranderlikes. Dit wil sˆe, die EE ewewigs vergelykings sal net tevrede gestel word indien die optimerings algoritme konvergeer, en die duur sensitiwiteit analises wat gepaard gaan met begrensings in terme van verplasings—soos bv. spanning begrensings—vereenvoudig na die berekening van eenvoudige parsi¨ele afgeleides. Wat meer is, net ’n enkele struktuur analise word benodig per ontwerp iterasie, al word daar ’n groot aantal lokale spanning begrensings bygevoeg. Gebaseer op ’n duale metode in teorie, stel ons voor dat ’n skeibare en streng konvekse kwadratiese Lagrange-Newton subprobleem gebruik word vir SAO van die SAND geformuleerde probleem. In klassieke maksimum styfheid ontwerp kan aangetoon word dat die duale metode soortgelyk is aan die populˆere optimale kriteria (optimality criteria; OC) metode—wat ’n NAND formulering is. Dit is, op sy beurt, identies aan die ooreenstemming tussen duale SAO-NAND algoritmes gebaseer op tussentydse veranderlikes en die OC metode. Omdat die optimerings probleem nie-lineˆere gelykheidsbegrensings bevat, moet die SAO prosedure (wat gewoonlik net van toepassing is op nie-gelykheid begrensde probleme) toegepas word in ’n algehele, nie-lineˆere en nie-konvekse, wiskundige programmerings raamwerk. Inderdaad, konvensionele tegnieke wat konvergensie af dwing in tradisionele NAND-SAO, kan amper net so gebruik word in die SAND raamwerk. Dit word bewys dat die ooreenstemmende uitdagings wat gepaard gaan met die bestaan van oplossings, maas-onafhanklikheid, lokale minima en makro-skaal vervaardiging tegnieke, kan geadresseer word deur middel van helling-begrensings—wat puntsgewys die gradi¨ent van die materiaal verspreiding bind—solank dit ook gepaard gaan met ’n lukrake multi-begin strategie (vir globale optimering). ’n Gespesialiseerde soort ‘lineˆere onafhanklike begrensing kwalifikasie’ (linear independence constraint qualification; LICQ) mag geld in baie praktiese gevalle, en omdat die inverse van die globale styfheid matriks nie per se uitgewerk word nie, word ’n onderste grens van presies nul toegelaat vir die materiaal veranderlikes. Dit beteken, op sy beurt, dat die singulariteite wat gepaard gaan met die lokale minima in beide eenvoudig-begrensde probleme en lokale spanning-begrensde probleme, oorkom kan word met gewone gradi¨entgebaseerde optimerings metodes, sonder dat daar enigsins verslapping- of perturbasie-tegnieke gebruik moet word.
Description
Thesis (PhD)--Stellenbosch University, 2017.
Keywords
Structural design., SAO (Sequential approximation optimization), Convex programming, Design services, Structural optimization, UCTD
Citation