Generation and Evaluation of Design Using Deep Neural Networks

Open Access
Author:
Dering, Matthew Lewis
Graduate Program:
Computer Science and Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
April 13, 2018
Committee Members:
  • Daniel Kifer, Dissertation Advisor
  • Conrad S Tucker, Committee Chair
  • Rebecca Jane Passonneau, Committee Member
  • Mohamed Khaled Almekkawy, Committee Member
  • Timothy Raymond Brick, Outside Member
Keywords:
  • Deep learning
  • design
  • generative models
  • prototyping
Abstract:
The objective of this dissertation is to reduce inefficiencies and manual tasks that are part of the product design lifecycle, using proposed deep neural network based methods. Computer Aided Design (CAD) and CAD software are an integral part of the product design lifecycle. However, many inefficiencies still exist in the design generation and evaluation process, which require manual work performed by human designers. This dissertation addresses several stages of the product design lifecycle where human intervention is still necessary. Four deep learning based methods are presented which help reduce inefficiencies in these processes. Deep learning methods use neural networks to learn data representations which are specific to the task at hand, and with minimal input from a model designer. These methods model problems as large non-linear systems, whose parameters are estimated during training to minimize a loss function. The contributions of this work are i) a Generative Adversarial Network which generates sketches during the ideation phase, to speed up the ideation process ii) a Variational AutoEncoder which generates high fidelity meshes of a given object class, which assists in the 3D concept generation process iii) a predictive neural network which validates 3D design concepts by predicting the object’s function, given its form, hereby reducing modeling time for the validation process iv) an autoencoder based unsupervised learning method which evaluates a user’s activities during physical prototyping, detecting errors in the prototyping process.