AI-Driven Optimization of Microfluidic Cooling Design Using CNN-Guided Reinforcement Learning
Please login to view abstract download link
Cooling performance and reliability in high power electronic packages increasingly rely on embedded microfluidic cooling design, but exploring the enormous design space of channel topologies and dimensions using full order simulations is computationally prohibitive. This work develops an AI assisted design framework that combines a pre trained dual input convolutional neural network (CNN) surrogate with a Double Deep Q Network (DDQN) reinforcement learning agent to automatically discover high performance microfluidic patterns design under manufacturability constraints. The CNN surrogate, trained on Representative Volume Element (RVE)–based finite element simulations, provides instantaneous predictions of effective thermal conductivity in the vertical direction k33, which is used as the reward signal for the DDQN agent to navigate the discrete design space of pattern design, channel width, spacing, and depth. In the surrogate based optimization, the DDQN agent improves the best effective vertical thermal conductivity from 115 to 144 W/(m·K), corresponding to a 24% increase achieved by coordinated reconfiguration of the microfluidic pattern and channel dimensions. The top ranked designs identified by the RL framework are reconstructed in CAD and evaluated using three dimensional conjugate heat transfer simulations of an ASIC and HBM stack under realistic heat flux and inlet conditions to quantify junction temperature and flow rate trade offs for a subset of DDQN optimized patterns. These results demonstrate that integrating surrogate models with reinforcement learning provides a practical route to accelerate microfluidic cooling design without exhaustive high fidelity simulation of every candidate.
