Physics-informed machine learning for the modeling and inverse design of microwave devices

Liu, Yanan

Permalink

https://hdl.handle.net/2142/125798 Copy

Description

Title

Physics-informed machine learning for the modeling and inverse design of microwave devices

Author(s)

Liu, Yanan

Issue Date

2024-07-10

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

Jin, Jian-Ming

Doctoral Committee Chair(s)

Jin, Jian-Ming

Committee Member(s)

• Goddard, Lynford L.
• Feng, Milton
• Schutt-Ainé, José E.

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• machine learning
• neural networks
• microwave devices
• design optimization

Abstract

This dissertation focuses on the development of physics-based machine learning models and their applications in the forward modeling and inverse design of microwave devices. There are two aspects of the research. First, we develop efficient surrogate modeling methods by incorporating physics knowledge into neural network models. Second, we propose a robust optimization technique utilizing the developed fast forward model. In the first part, we propose two neural network models which incorporate the second-order analytic extension of eigenvalues (AEE), and the second-order characteristic mode analysis (CMA). In these methods, the output nodes of neural networks are parameters of the physics models, instead of network parameters directly. This allows the models to learn a more reliable and generalizable mapping between the input and output. We introduce physics-based regularization to maximize information gain from data samples and advanced training strategies to further improve modeling efficiency. The proposed methods not only significantly reduce the cost for training data generation, but also demonstrate superior performance in terms of accuracy, superior data efficiency, and generalization capability. These models are differentiable, stackable, and fully compatible with system-level simulations. In the second part, a hybrid optimization framework combining genetic algorithms (GA) and gradient descent (grad-opt) is presented. In this approach, GA is used to start the optimization with random search and heuristics-based evolution. During the GA process, individual designs that satisfy a certain goodness criteria are seeded as the starting point for gradient-based updates, where an optimal solution can be reached quickly within only a few iterations. To facilitate the optimization process, we proposed the use of a neural network model that can speed up fitness evaluation in GA and gradient calculation in grad-opt. To further improve modeling efficiency and accelerate the design process when dealing with a large number of design variables, we adopt the divide-and-conquer strategy, which is fully compatible with the ML w/AEE model. We introduce a special neural network block called the fusion module to perform component cascading numerically, allowing the gradients to be passed from the objective function to design variables. We also propose a robust metric for global control, that involves the cost function value and its cumulative gradient. This control parameter allows us to achieve a good balance between efficiency and stability. We use two numerical examples to demonstrate the efficacy and strength of the proposed method. Hybrid optimization is shown to be much more efficient than standard GA and more robust than gradient-based methods.

Graduation Semester

2024-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/125798

Copyright and License Information

Copyright 2024 Yanan Liu

Owning Collections