Machine Learning Models of Heatsink Performance in Single-Phase Immersion Cooling: Application to Design Acceleration

Abstract The rapid increase in computational requirements of modern computing systems has led to exponentially increasing thermal design powers (TDPs) from the processors. Traditional air-cooling methods have reached their limits, being unable to handle the component level heat fluxes and rack level power densities of next generation data centers. A transition to single-phase immersion cooling is an effective strategy to properly manage the heat fluxes from the next generation processors, all while offering improved energy efficiency, decreased water usage, and less datacenter whitespace. To fully embrace the benefits of immersion cooling, the heatsinks used on the processors must be well designed and optimized, necessitating rapid and robust performance models. This study presents the use of neural networks to develop models that predict performance characteristics of heatsinks in single-phase immersion cooling. The networks are trained on a synthetic dataset of 15,552 points, taking inputs of immersion fluid, flow conditions, and heatsink geometry. An analysis of the various architectural and training conditions is conducted, yielding networks with enhanced validation set accuracy. When inferenced in generalization, the networks exhibit accurate predictive capabilities (4.084% MAPE) with an average inference time of 3.685 µsec. The networks are deployed in multi-variable, multi-objective optimizations, completing various 10,000-point optimization analyses in under three seconds (average 2.21s), with only a 3.9% difference from CFD optimization. A time saving analysis is performed demonstrating the extreme order of magnitude reductions in design turn around time possible.