ORCID

https://orcid.org/0000-0002-8051-7282

Date of Award

Spring 2025

Language

English

Embargo Period

4-15-2027

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

College/School/Department

Department of Electrical and Computer Engineering

Program

Electrical and Computer Engineering

First Advisor

Mohammed Agamy

Committee Members

Ming-Ching Chang, Gary Saulnier, Won Namgoong

Keywords

Electric circuits, Power Electronics, Graph Neural Networks (GNN), Machine Learning, Bond Graph, Heterogeneous Representation

Subject Categories

Electrical and Electronics | Other Electrical and Computer Engineering | Power and Energy | Signal Processing

Abstract

Traditional methods for electronic circuit analysis and design face significant challenges in efficiency, computational cost, and scalability, often encountering convergence difficulties. machine learning (ML) models provides a powerful alternative, since it can predict performance metrics, explore design trade-offs (Pareto fronts), and inherently manage variations in circuit structure, thus significantly speeding up design/analysis iterations, reducing computational burden, and facilitating the optimization of complex electronic systems. While ML models offer potential solutions, its application in power electronics has largely focused on control or component-level tasks using surrogate models, neglecting the fundamental representation of circuit topology and component interconnectivity. The problem stems from the existing ML-based circuit modeling approaches, since they lack a systematic means to encode circuit topology and component values effectively. This thesis introduces a novel, systematic framework for representing electric circuits as graphs, specifically designed to enable graph-based ML applications. The framework provides a method to construct graph representations based on the bond graph modeling approach, which captures both the topology and the dynamics of electric circuits components. Building upon this representation, Graph Neural Network (GNN) models are developed and tailored for various circuit analysis tasks. The thesis starts with the systematic bond graph-based framework for graph construction suitable for circuits with varying parameters and operating modes, followed by the design and applications of GNN models for classification tasks (demonstrated on resonant circuits and DC-DC converter topologies) and multi-variable regression tasks (predicting DC-DC converter output voltage and efficiency across different configurations and operating conditions, including conduction modes). Additionally, a detailed characterization of the computational requirements and scalability of the framework is provided, validated on complex circuits like three-phase DC-AC inverter under diverse conditions. Furthermore, the framework is enhanced by incorporating heterogeneous graph properties, utilizing distinct node and edge types derived from the bond graph formalism to improve physical fidelity and representation accuracy compared to homogeneous approaches. Finally, a heterogeneous Physics-Informed GNN (PIGNN) is proposed, integrating fundamental circuit laws (KCL/KVL) via a custom loss function to enhance model generalization beyond the training data distribution. This establishes a robust methodology for interfacing physical electrical structures with machine learning, enabling a wide range of ML tasks such as classification and regression, and providing a foundational step towards more advanced applications like automated power electronics circuit synthesis and design.

License

This work is licensed under the University at Albany Standard Author Agreement.

Download

Available for download on Thursday, April 15, 2027

Share

COinS