Neural operator models for predicting physical fields in server electronic microchips doped with water-Al2O3 nanofluid coolant

Abstract

In response to the increasingly severe thermal challenges brought by high heat flow in server electronic microchips, this work utilizes water-Al2O3 nanofluid coolant and microchannel cooling device to study the internal physical field flow and further determine the heat transfer performance of the flow standard for improvement. In particular, the prediction of two-dimensional physical fields is designed as a regression task using a network of five neural operators, taking eight design variables into account. Using the integral solution method, the Nusselt number and Fanning friction coefficient from the predicted field are extracted. The prediction accuracy of FNO exceeds the worst model DeepONet by an order of magnitude and the training cost reduces by three to five times. It is noteworthy that the resolution range of FNO is one order of magnitude lower than that of the largest DeepONet. It exhibits significant grid resolution invariance and maintains consistent prediction accuracy regardless of resolution changes. Here, FNO demonstrates a state-of-the-art performance in terms of predictive power, computational efficiency, and grid resolution. Additionally, this research has deepened our understanding of enhanced heat transfer mechanisms. It guides experimental measurements in industrial applications and provides a solid foundation for advances in thermal management technology and electronic microchips design.

Keywords

• Neural operator models
• Physical fields prediction
• Electronic internal flow
• Water-Al2O3 nanofluid