Files in this item

FilesDescriptionFormat

application/pdf

application/pdfGUO-DISSERTATION-2018.pdf (14MB)Restricted Access
(no description provided)PDF

Description

Title:On the use of machine learning with design optimization data for system topology design
Author(s):Guo, Tinghao
Director of Research:Allison, James T
Doctoral Committee Chair(s):Allison, James T
Doctoral Committee Member(s):Kim, Harrison; Wang, Pingfeng; Guest, Jeremy S
Department / Program:Industrial&Enterprise Sys Eng
Discipline:Industrial Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):network analysis, data mining, topic modeling, natural language processing, cluster analysis, design optimization data, system topology design, machine learning, deep learning, variational autoencoder, style transfer, topology optimization, heat spreader, design space coverage, indirect design representation, generative design, active learning, random forest, evaluation cost, generative adversarial network, circuit synthesis, design methodology, design formulation space
Abstract:In this dissertation, several machine learning strategies are presented to advance solution capabilities for homogeneous and heterogeneous system topology design. The core contribution is to begin bridging the gap between data science and design science. The key principle is to extract meaningful knowledge and insights from design data, and to build machine learning models that enable effective design exploration and lead to generalizable design insights. This work provides an alternative perspective for system topology design, leveraging design data instead of designer intuition derived from experience or established gradient-based topology optimization methods. As a preliminary study for this dissertation, the research literature for a relevant segment of the engineering design research community was analyzed using network analysis. This study was based on a collection of 1,668 articles published in the American Society of Mechanical Engineers (ASME) Design Automation Conference (DAC) from 2002-2015. Several methodologies were developed and used, and useful insights were provided. These analyses revealed several insights, including opportunities for strengthening the link between design automation methods, such as design optimization, and machine learning. This result serves as a basis and motivation for the remainder of the work presented here. The remainder of the dissertation concentrates on efforts to advance understanding of how to use machine learning effectively with design data generated using design automation methods. A novel design framework using deep learning was developed for homogeneous system topology optimization. The application chosen here involves heat conduction, with competing objective functions of temperature and power density. Existing methods can solve related heat conduction problems efficiently (e.g., thermal compliance), but not combined temperature and power density. The strategy presented here seeks to use data generated from related easy-to-solve thermal compliance problems to support efficient solution of the desired problem. An indirect design representation was constructed using a variational autoencoder (VAE), and was combined with a deep convolutional style transfer network to improve the quality of generated designs. The VAE maps the original large-dimensional design space onto a lower dimensional space (called the latent space). The heat conduction problem was solved by optimizing with respect to latent variables, and system performance was evaluated using full-dimension design representations. Several variants of the optimization formulation have been examined, and the Pareto-optimal solutions are presented. The method is shown to successfully navigate the design space and identify many non-dominated designs that outperform those found using conventional topology optimization. Topology optimization of heterogeneous systems, sometimes referred to as synthesis, requires different design representations and solution techniques. Here we considered two classes of synthesis problems where existing methods and recent advances, such as efficient enumeration, are limited in practical solution capability. The first class of synthesis problems (Case 1) considered is where efficient enumeration methods can be used to list all unique, feasible system design topologies, but not all topologies can be evaluated in a practical amount of time due to computational expense. Active learning is investigated here as a strategy to select subsets of topologies for evaluation with the goal of finding high-performance designs without the need to evaluate all candidates. Active learning is a semi-supervised learning technique that interactively improves predictive model accuracy with strategically selected training examples. The predictive model used here is an ensemble method called random forest. Several active learning strategies are considered and results indicate that active learning is a promising strategy for solving Case 1 synthesis problems. Case 2 synthesis problems involve systems where all topologies of interest can neither be enumerated nor evaluated in a practical time period. Here a new approach for Case 2 problems is introduced where machine learning techniques are used to generate topologies in a way that implicitly satisfies constraints and aids search for high-performance designs, in essence creating a targeted design space representation for efficient search. This eliminates the need to enumerate all unique, feasible topologies, and supports approximate solution of Case 2 synthesis problems. Generative adversarial networks (GANs) are investigated as the design representation tool for this design automation process. Experiments were conducted to explore capabilities of GAN-based synthesis methods for electronic circuit synthesis. Multiple GAN strategies are investigated. The numerical results demonstrate that the improved Wasserstein GAN is capable of generating feasible circuit topologies efficiently. The GAN-based design framework may also be extended to more general design synthesis tasks.
Issue Date:2018-07-12
Type:Thesis
URI:http://hdl.handle.net/2142/101819
Rights Information:Copyright 2018 Tinghao Guo
Date Available in IDEALS:2018-09-27
Date Deposited:2018-08


This item appears in the following Collection(s)

Item Statistics