학술논문
A Multi-Objective, Machine-Learning-Based Optimization Method and its Application to a Power Card Package Design
Document Type
Conference
Source
2023 IEEE Transportation Electrification Conference & Expo (ITEC) Transportation Electrification Conference & Expo (ITEC), 2023 IEEE. :1-6 Jun, 2023
Subject
Language
ISSN
2473-7631
Abstract
Increasing the power density of traction inverters used in electric vehicles is a challenging design problem that requires trade-offs in volume, electrical parasitics, thermal performance and thermomechanical reliability. This paper presents a novel mixed-variable, multi-physics, and multi-objective Bayesian Optimization (BO) method allowing for rapid co-optimization of the power module package topology, layout, and materials. Compared to genetic algorithms often used to optimize power modules, the developed BO solution requires a reduced number of simulations, and can, therefore, interface directly with conventional finite element modeling (FEM) environments such as ANSYS for multi-physic simulations. Additionally, it can support 40+ continuous variables, e.g., geometrical parameters, along with 1000+ categorical variables, e.g., accounting for material and architecture options. An overview of the approach, optimization method, and visualization of the optimization results are presented for a double-sided half-bridge 50 kW SiC-based power module, as a proof-of-concept demonstration. The BO predictions were in good accordance with FEM results of selected optimized geometries, with less than 10 % deviation regardless of the considered physics. This new method provides an efficient path for designers to compare arbitrary package architectures and select solutions that meet future performance and reliability objectives.