Physics-Informed Neural Networks for Microprocessor Thermal Management Model

Abstract

The cooling of microprocessors has emerged as a crucial challenge in enhancing performance. Various studies are being conducted to optimize the structural design for microprocessor cooling. However, conducting direct experiments after designing a structure incurs substantial costs. Furthermore, utilizing conventional numerical methods for simulating various structures results in high computational expenses and diminished design flexibility. In this study, we employ Physics-informed neural networks (PINNs) to develop a simulation for predicting the heat distribution of a microprocessor with various structures. We train the model using a thermal diffusion equation with a point source generated from the chip at the bottom of the microprocessor, which is transferred between three regions: chip, heat sink, and ambient environment. Furthermore, we successfully predict heat transfer simulations that vary with different heat sink structures and achieve analytical results that are consistent with actual experimental outcomes. Finally, we propose a novel PINN that allows for the free design of heat sinks for microprocessors to evaluate the cooling performance.

Keywords

• Physics-informed neural networks
• deep learning
• temperature model
• mathematical modeling