Machine learning-optimized jet-enhanced immersion liquid cooling for high-power data centers

Yongping Huang, Chendong Liu, Bao Ding, Zilong Deng, Wei Gao,
Machine learning-optimized jet-enhanced immersion liquid cooling for high-power data centers,
International Communications in Heat and Mass Transfer,
Volume 169, Part B,
2025,
109654,
ISSN 0735-1933,
https://doi.org/10.1016/j.icheatmasstransfer.2025.109654.
(https://www.sciencedirect.com/science/article/pii/S0735193325010802)
Abstract: To address the thermal management challenges posed by multi-scale heat sources and high-heat-flux in data centers, jet-enhanced immersion liquid cooling (ILC) technology has emerged as the preferred solution. This study innovatively applies machine learning to resolve the inherent conflict between heat convection enhancement and flow resistance reduction in jet-enhanced ILC systems. A three-dimensional steady-state numerical model is established to generate a dataset correlating input parameters (pin-fin height H, jet hole diameter d, and hole spacing l) with optimization objectives (maximum temperature Tmax, temperature uniformity index σT, and flow resistance Fp). Four machine learning models are employed for dataset modeling and prediction accuracy analysis. The results demonstrate that the Deep Neural Network (DNN) model achieves the highest prediction accuracy, with R2 values maintained at 0.93–0.97. H is the dominant factor determining thermal performance (Tmax and σT), and the parameter influence on the flow resistance (Fp) follows the order: d > l > H. Moreover, Tmax and σT exhibit an extremely strong positive correlation, while both show moderate negative correlations with Fp. Through comprehensive balancing the game mechanism, the optimal parameters are determined as d = 1.84 mm, l = 17.47 mm, and H = 30 mm, with the multi-objective prediction showing a maximum relative error of 0.68 %.
Keywords: Single-phase immersion liquid cooling; Jet; Machine learning