Optimization of heater configurations in single-phase immersion cooling systems using CFD-validated artificial neural networks

초록

The growing energy demands of data centers have underscored the need for efficient thermal management solutions. This study investigates the optimization of single-phase immersion cooling (SPIC) using a hybrid methodology that combines experimental analysis, computational fluid dynamics (CFD) modeling and artificial intelligence (AI) techniques. A laboratory-scale SPIC system using a hydrocarbon-based dielectric fluid (LP-75F) was tested under various thermal loads to collect empirical temperature data, which was then used to validate a two-dimensional CFD model employing SST k-omega turbulence closure. Following validation, a parametric study was conducted through 572 simulations to evaluate the effects of heater elevation, lateral spacing, and inlet coolant temperature. The resulting dataset was used to train a feedforward artificial neural network (ANN), which achieved strong predictive accuracy (R2 = 0.98). The ANN model further enabled identification of design configurations that reduced maximum chamber temperatures by up to 6 K, all without hardware changes. Unlike existing CFD-AI coupled studies that have primarily focused on standardized heat exchanger flow, this study presents an experimentally verified CFD-ANN framework specialized for buoyancy-driven single-phase immersion cooling (SPIC). This framework quantitatively captures the interaction between heater height, lateral clearance, and coolant inlet temperature to establish a physically interpretable design space that lowers Tmax by up to 6 K without hardware modifications.

키워드