Reinforcement Learning Based Optimization of Power Modules

Designing layouts for power modules is a complex and time-consuming problem due to the degrees of freedom involved and the tradeoff induced by the metrics to minimize - usually parasitic inductance and maximum temperature. While this multi-objective optimization was initially solved manually by domain experts, in recent years, several methods have been proposed that optimize the layouts autonomously using Random Sampling and Genetic Algorithms. In this work, we re-utilize a graph-based representation of power modules (including all design constraints, component specifications, and their connectivity) and benchmark Random Sampling, Genetic Algorithms, and the proposed Reinforcement Learning based optimization methods at optimizing the power modules' layouts w.r.t. their electro-thermal Pareto-Front. We show that while Reinforcement Learning takes longer to converge, it remains competitive with both established optimization techniques for exemplary half-bridge power modules, demonstrating its potential as a promising first step toward solving the more complex optimization challenges posed by future power module designs.