THESIS
2021
1 online resource (xvi, 118 pages) : illustrations (some color)
Abstract
Topology design of structures, layouts and metamaterials refers to the strategic distribution of
their constituent materials or components in a given design domain in order to achieve specified
properties or functionalities. Owing to its novelty, it can find wide engineering applications, such
as the design of lightweight airplane wings and automotive components. Conventional topology
design approaches, such as solid isotropic material with penalization, level-set based methods and
genetic algorithm, have achieved a great success in designing various structures, layouts and
materials. However, they suffer from high computational cost owing to the repetitive high-dimensional
physics-based simulations needed during the optimization process, which greatly
limits their application scope.
In...[
Read more ]
Topology design of structures, layouts and metamaterials refers to the strategic distribution of
their constituent materials or components in a given design domain in order to achieve specified
properties or functionalities. Owing to its novelty, it can find wide engineering applications, such
as the design of lightweight airplane wings and automotive components. Conventional topology
design approaches, such as solid isotropic material with penalization, level-set based methods and
genetic algorithm, have achieved a great success in designing various structures, layouts and
materials. However, they suffer from high computational cost owing to the repetitive high-dimensional
physics-based simulations needed during the optimization process, which greatly
limits their application scope.
In this thesis, two efficient machine learning based design methods are proposed for topology
design. The first one uses artificial neural network (ANN)-based surrogate models to replace the
high-dimensional simulations for the objective function/sensitivity evaluation in each design
iteration. A novel dual-model ANN-based surrogate model is proposed which can rapidly evaluate
the design performance as well as its sensitivity with respect to the design variables. The ANN-based
surrogate model is then integrated into conventional topology optimization (TO) algorithms to perform inverse design. Compared to conventional TO methods, the computational efficiency
has been improved by at least two orders of magnitude. Moreover, to save the computational cost
for training data generation, an efficient training data generation method is proposed by exploiting
the similarity of designs in later iterations. Three design problems, including elastic metamaterials
for wave deflection, cantilever beams with minimum compliance and metamaterials with negative
Poisson’s ratio are used to demonstrate the performance of the proposed methods.
The other proposed machine learning based design method combines generative adversarial
network, ANN-based surrogate model and genetic algorithm to perform inverse design. The
generative adversarial network produces design candidates satisfying geometric constraints, and
the ANN-based surrogate model evaluates design performance. Together they form the design
network on which the inverse design can be efficiently conducted. One main issue in this method
as well as other generative design methods is the need for a large number of labelled training data.
In this thesis, an adaptive learning and optimization strategy is proposed, which greatly reduces the
number of training data needed. The method is applied to solving two type of design problems.
The first problem is the design of heat source layout for thermal management. The other is to design
2D architectured composite materials for achieving high toughness, high stiffness near the
theoretical upper bound and prescribed bulk/shear moduli. In all cases, excellent design results are
obtained. Compared with the existing machine learning approaches which use tens of thousands to
millions of training data, the proposed method only needs a few hundred to thousand training data,
demonstrating the effectiveness and efficiency of the proposed adaptive generative design
approach.
Post a Comment