Degree Type

Thesis

Date of Award

2020

Degree Name

Doctor of Philosophy

Department

Mechanical Engineering

Major

Mechanical Engineering; Computer Engineering

First Advisor

Adarsh Krishnamurthy

Abstract

Computer-Aided Engineering (CAE) is necessary for fast and efficient product development in design and manufacturing. Historically, CAE played a vital role in the drastic improvement of production time and production cost over the past two to three decades. While it used to take several months to years to realize a product earlier, the same process takes only a few hours to days now, due to the availability of state-of-the-art CAE techniques. This thesis is a contribution to the CAE ecosystem with the motivation to integrate advanced ideas from machine learning and computational sciences with current CAE tools to improve the product development cycle. These ideas can be broadly classified into (i) data-driven/machine learning-based approaches, (ii) GPU acceleration for massively parallelizable tasks. These ideas have been adopted extensively and have become prevalent over the past five to eight years, with a substantial focus on other cyber-physical systems in the aerospace, automotive, agriculture, healthcare, and transportation sectors. However, applications to the manufacturing sector have not seen as much advancement compared to the other areas. In this work, we focus on our contributions to CAE, which is an integral part of the manufacturing cyber-physical system.

Over the past decade, transition to a new economy driven by automation and revolutionary changes in manufacturing technologies has enabled highly sophisticated, creative, and customizable products to be manufactured on demand by flexible robotic systems. Consequently, the demand for designing and customizing products for each user has grown exponentially. However, end-users who wish to customize their products or designs often do not have sufficient knowledge, experience, or expertise about manufacturing technologies and computational methods to analyze the design. This limitation requires CAE systems to be intelligent in making decisions with fewer interventions from the end-users (often termed as Industry 4.0). While several cyber-physical systems have embedded advanced data-driven tools such as deep learning and reinforcement learning into their workflow, their usage in the manufacturing systems is still very sparse. This dissertation is an attempt to fill this lacuna.

Another critical idea explored for integration to the state-of-the-art CAE tools is the development of algorithms accelerated with graphical processing units (GPUs). Almost a decade ago, GPUs were meant for accelerating the rendering pipeline of an application by computing a the geometric transformations of triangles and rendering them interactively (at a rate of more than 30 frames per second). However, the trend to use GPUs for general-purpose computations (GPGPUs) has been catching on. Many CAE applications have started to embrace the use of GPUs for general computations since they can perform certain typical computations in less than a minute, while the traditional computational methods take minutes/hours to do the same. Such low latency provides users with the opportunity to perform interactive custom designs, which are necessary for the current age of personalized products rather than mass production-based systems. In this dissertation, we leverage the speed of GPU-accelerated algorithms to accelerate data-driven CAE tools.

Specifically, we develop four CAE tools in this dissertation. First, we propose an intelligent decision-making system that can be applied in the product development process. This tool is useful in developing designs without iterative design reviews involving design or manufacturing engineers. Next, we develop a tool for performing quick design analysis to validate the designs. We then develop a tool for design exploration in situations where the computation of physics is trivial using CAE, but the inverse (finding the design satisfying the desired physical phenomenon) is an intractable problem. Third, we develop a tool for optimizing and obtaining designs with creativity while maintaining the design intent. Finally, we present frameworks developed for scaling deep learning to distributed manufacturing cyber-physical systems.

DOI

https://doi.org/10.31274/etd-20200624-61

Copyright Owner

Aditya Balu

Language

en

File Format

application/pdf

File Size

199 pages

Available for download on Tuesday, June 15, 2021

Share

COinS