Developing physics-informed machine learning models for complex engineering systems design

Xu, Yanwen

Permalink

https://hdl.handle.net/2142/121344 Copy

Description

Title

Developing physics-informed machine learning models for complex engineering systems design

Author(s)

Xu, Yanwen

Issue Date

2023-07-11

Director of Research (if dissertation) or Advisor (if thesis)

Wang, Pingfeng

Doctoral Committee Chair(s)

Wang, Pingfeng

Committee Member(s)

• Gardoni, Paolo
• Nagi, Rakesh
• Wang, Qiong

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Industrial Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Physics-informed machine learning
• design automation
• uncertainty modeling
• surrogate modeling
• active learning
• adaptive sampling
• high-dimensional problem
• reliability in design.

Abstract

Engineering systems design relies on quantitative and qualitative data to describe design-related engineering phenomena and prescribe improvements for design practice. Advanced computational methods that create mathematical models and develop data-driven techniques based on data are crucial for engineering design. Data-driven design refers to making engineering design decisions based on data. In a data-driven design framework, data plays a primary role. One notable aspect of data-driven design is the use of data to inform and guide the design process, relying on data collection and analysis to gain insights into system behavior. One of the challenges in data-driven design for complex systems is the uncertainty and scarcity of data, as well as the high dimensionality of available information. Inadequate data can hinder the development of accurate models, while the high dimensionality of data can make it difficult to extract meaningful insights and improve system design. The focus of this thesis is to explore and develop physics-informed machine learning methods that effectively analyze complex engineering systems under uncertainty while maintaining accuracy and efficiency in their predictions. Chapter 1 provides background information on engineering system design under uncertainty. Chapter 2 introduces a novel uncertainty quantification approach that efficiently estimates rare event probabilities with narrow estimation bounds simultaneously for high-dimensional problems and complex engineering systems. The theoretical proof of the developed estimator's asymptotic behavior is presented without imposing strong assumptions. An asymptotic confidence interval is established to quantify uncertainties for the developed estimator. This study offers important insights into robust estimations of reliability for complex engineering systems. Additionally, it highlights the need for further investigation and development of advanced surrogate modeling techniques for complex system design. In Chapter 3, the thesis outlines the development of physics-informed machine learning techniques, with a focus on recent milestones in the literature. The aim is to highlight the progress of research in surrogate modeling towards building a framework that integrates diverse knowledge sources to address challenges related to reliability and system safety. The goal is to mitigate issues arising from uncertainty, data scarcity, and high dimensionality, providing a comprehensive approach to improving reliability and safety in complex engineering systems. Chapter 4 proposes a novel physics-informed framework for engineering design under uncertainty, which integrates physics information and data from different sources across different scales to achieve holistic system analysis, uncertainty quantification, and design optimization goals. The results demonstrate that the proposed strategy enhances both design reconstruction and system performance prediction accuracy and can be easily extended to uncertainty quantification and reliability-based design applications. In Chapter 5, an in-depth exploration of active learning strategies is conducted to refine and improve the surrogate model after the initial model construction. A novel active learning criterion is proposed, considering both the missing pattern and information cost of partially observed data, to iteratively select new training sample points and refine the model. This approach effectively utilizes all available information, including both fully observed and partially observed data points, resulting in an accurate and cost-effective solution for developing adaptive surrogate models. The developed methodology offers a valuable means to leverage additional knowledge in the construction of advanced surrogate models. In Chapter 6, the knowledge and techniques discussed in Chapters 2-5 are synthesized and applied to analyze and design various engineering systems, including battery management systems, carbon capture and storage systems, and semiconductor designs. By leveraging the insights and methodologies developed in earlier chapters, comprehensive analyses and design optimizations are conducted to enhance the performance, reliability, and safety of these engineering systems. Chapter 7 concludes by summarizing the fundamental progress made in the modeling and design of complex engineering systems, providing insight and an outlook for the future. The author believes that the journey towards creating a transformable physics-informed modeling platform will undoubtedly result in remarkable tools for analyzing, designing, and optimizing complex engineering systems, as well as advancing our understanding of data-driven design methods.

Graduation Semester

2023-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/121344

Copyright and License Information

Copyright 2023 Yanwen Xu

Owning Collections