Multi-Objective Automatic Design of Power Module Packaging Based on Artificial Neural Network and Deep Reinforcement Learning

Traditional design methods for power module packaging heavily rely on manual experiences. Although the computer-aided optimization (CAO) methods utilizing metaheuristic algorithms have accelerated the design process, they are still affected by lengthy iterative cycles when including 3D numerical simulations. Regarding these issues, this paper proposes a multi-objective automated design (MOAD) method for power modules based on machine learning, combining Artificial Neural Networks (ANN) and Deep Reinforcement Learning (DRL). The proposed ANN model can save massive time of iterative 3D numerical calculations in multi-objective optimization (MOO) design of the power module packaging, meanwhile avoiding the complex interaction of different software. The DRL algorithm can output the optimal design variables efficiently, even when design requirements change after the agent are well trained. A 1200V/300A double-sided cooled (DSC) module is used as a case study to validate the proposed method.