Mask neural network for temperature field prediction for three-dimensional thermal design of electronic devices

Abstract

Accurate temperature field prediction (TFP) is critical for optimizing the thermal management in complex electronic systems. Although deep learning surrogate models have demonstrated high accuracy for Heat Source Layout (HSL) problems, they may overlook key three-dimensional (3D) features, such as the layered heat sinks, thickness-dependent properties of Thermal Interface Material (TIM) and the volumetric fan airflow. In this work we propose a novel approach by integrating a mask module into a U-Net-based generator neural network, further enhanced with a mask area loss function, allowing for precise temperature field predictions by capturing 3D design properties of the electronic devices. Our model was tested on vehicle domain control units (DCUs), and results demonstrated substantial improvements over existing methods, with a 47 % reduction in Maximum Absolute Error (MaxAE), a 52 % reduction in Mean Absolute Error (MeanAE), and an 81 % reduction in Mean Square Error (MSE). These findings underscore the importance of including vertical and volumetric design factors in thermal analysis and suggest that our approach can aid researchers and engineers in fields such as automotive electronics, data-center cooling, and consumer-device thermal management. By advancing the state-of-the-art in TFP, this model holds promise for guiding future 3D thermal design optimization in diverse electronic systems.