A Comparative Study of Surrogate Modeling Techniques for Parasitic Inductance Estimation in Power Modules
Year: 2026
Type: conference-paper
Abstract: This paper presents an automated finite element method (FEM)-driven surrogate modeling framework for rapid prediction of parasitic inductances in power module commutation loops. Parasitic inductance critically affects switching transients, voltage overshoot, and electromagnetic interference in high frequency converters. Although FEM based extraction provides high fidelity results, it becomes computationally prohibitive for large scale parametric design exploration. A parameterized half-bridge power module was developed in Ansys Q3D and automated through a Python based PyAEDT workflow for systematic dataset generation. Latin hypercube sampling was employed to explore geometric variations, and inductance matrices were extracted for multiple configurations. Linear regression, artificial neural networks, and Gaussian process regression (GPR) were evaluated as surrogate models. Across all investigated case studies, the optimized GPR model achieved an average mean absolute error of 0.0224 nH, a root mean square error of 0.0351 nH, and an average$R^{2}$of 0.983. Additionally, the surrogate framework provided approximately$4.0 \times 10^{5}$reduction in per-sample computation time relative to full wave analysis. The probabilistic nature of GPR further enables uncertainty aware prediction and provides a basis for future active learning based sample refinement.
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Keywords: parasitic inductance, power modules, surrogate models, Gaussian process regression, artificial neural networks, linear regression, half-bridge module, Latin hypercube sampling, uncertainty quantification, active learning (machine learning), electromagnetic interference, switching transients, high-frequency converters
Open Access status: hybrid
OpenAlex ID: W7224964489
