Machine Learning-Based Optimization of a Mini-Channel Heatsink Geometry

The research on cooling electronics devices in workstations, servers, and other devices has become one of the rapidly growing technologies today, with the generation of high heat fluxes associated with component compactness and increased power consumption. Hence, the current study aims to enhance the thermohydraulic performance of a compact mini-channel heat removal system based on a machine learning-based optimization technique. The design variables considered for the optimization study are the channel's Reynolds number (Re) , fin thickness (t_f) , fin spacing (s_f), and fin height (h_f). Initially, the performance of the heat sink geometry with different fin configurations is computed using 3D-RANS simulations. The generated data are then used to train six regression techniques in machine learning to propose the best approach capable of accurately predicting the heat transfer coefficient and pressure drop data. The selected machine learning model is then coupled with the multi-objective genetic algorithm to find the optimal heat sink geometry. It is found that the multilayered perceptron approach re-designed to a deep neural network model efficiently predicts both the heat transfer coefficient and the pressure drop from the available data. The overall performance of the optimized channel geometry is found up to 2.1 times improved than the best available channel configuration in the data pool. Further, the optimized channel geometry's heat transfer coefficient was 14% higher, and the corresponding pressure drop was five times lower.