Physics-enhanced U-net and deep reinforcement learning for automated optimization of pin-fin heat sinks in electric vehicle power modules

Abstract

The use of pin-fin structures in compact energy devices, such as electric vehicle power modules, is a widely adopted thermal management strategy to enhance heat transfer efficiency. In this study, we present an innovative deep learning framework that integrates a physics-enhanced U-net architecture with a deep reinforcement learning agent to achieve autonomous optimal design of pin-fin arrays. The physics-enhanced U-net is trained to predict thermal-flow fields, while the integrated deep reinforcement learning agent autonomously optimizes pin-fin configurations to minimize both pressure drop and junction temperature. First, we generate a high-fidelity training dataset through an automated computational pipeline that integrates COMSOL Multiphysics for thermal-flow field simulations with a custom Matlab script for parametric generation of 1080 training samples. Subsequently, we train our physics-enhanced U-net architecture to predict the velocity, pressure and temperature fields from various pin-fin structure inputs. The proposed model demonstrates both high prediction accuracy and robustness, achieving mean-squared-errors on the order of 10⁻⁴ for all output fields. As a result, the trained U-net model achieves exceptional prediction accuracy, demonstrating 93.9 % precision for pressure drop and 99.5 % for junction temperature. Finally, we integrate the deep reinforcement learning agent with the trained U-net model to establish an automated optimization framework for pin-fin design, enabling intelligent exploration of design space. The proposed deep learning framework successfully automates the optimization of pin-fin heat sinks for a high power density module. The model demonstrates exceptional capability in generating optimal designs, with the optimized configuration achieving an 8.8 K reduction in junction temperature and 11.3 % decrease in pressure drop comparing to a baseline design. These improvements can be translated into approximately 10 % augmentation in power output, which validates both the effectiveness and robustness of our deep learning driven design approach.