Continuous and Categorical Bayesian Optimization for 3D Power Module Package & Cold Plate Design

As land, air, and sea-based transportation electrification expands, inverter technology is subject to a variety of domain-specific requirements. With such a diversity in requirements, prior learning may not transfer between domains to offer appropriate solutions. To keep pace with the quickly evolving power conversion landscape, efficient numerical methods are needed to assess the performance of a new technology and determine whether changes in the design paradigm are necessary. In this research, a categorical embedded Bayesian optimization approach is extended to generate Pareto-optimal solutions for both an electronics power module and associated cold plate using a sequential design approach. Each design problem is independently solved using three-dimansional (3D) multi-physics evaluation methods to measure the impact of both continuous and categorical design variable selections. It was observed that the probabilistic models accurately captured Pareto designs near the utopia point thus expanding the number of eligible designs for down-selection. Furthermore, such models showed improved convergence with a limited number of function calls motivating the use of Bayesian optimization approaches for the design acceleration of power modules and cold plates.