Session: K9-11 Phonon Modeling and Machine Learning for Thermal Transport

Paper Number: 114697

114697 - Physics-Informed Deep Learning for Modeling Multi-Scale Thermal Transport Using Boltzmann Transport Equation 

With the increased integration and miniaturization of devices and structures, effective thermal management has become more and more essential for ensuring optimal performance and reliability of contemporary electronics. Therefore, it is vital to conduct thermal analysis accurately and efficiently, especially in the micro-/nanoscale, to meet the increasing demand for enhanced heat dissipation and energy conversion. However, modern systems exhibit characteristic lengths that can be much smaller than the mean free path of thermal energy carriers (e.g., phonons, electrons). This leads to ballistic transport, and Fourier’s law can fail due to the non-diffusive transport effect. The Boltzmann Transport Equations (BTE), in contrast, can accurately model multi-scale thermal transport, taking into account non-diffusive thermal transport. However, numerically solving BTE is computationally complex due to its high dimensionality. This complexity can hinder its applications in fast thermal evaluation, particularly when phonon dispersion and time evolution are considered. Recognizing this limitation, we recently developed a physics-informed deep learning (PIDL) framework for solving the time-dependent mode-resolved phonon BTE by minimizing the residuals from the governing physical laws (i.e., the phonon BTE) and boundary/initial conditions. It is noted that no labeled training data is required, and training can be conducted in a parametric setting. This enables a trained model to quickly evaluate thermal transport of structures spanning multiple scales, from transistors to packaging levels. Furthermore, we extend our PIDL framework to solve the coupled electron-phonon BTEs, where electron-phonon interactions are considered. Once offline training is completed, our PIDL framework can be utilized for online evaluation of heat conduction problems ranging from 1D to 3D. The framework is capable of predicting temperature and heat flux distribution at any position and time within a few seconds. Compared to existing numerical solvers, the proposed method demonstrates high efficiency and accuracy, making it promising for practical implementation in various fields, including the thermal design and management of microelectronic devices.

Presenting Author: Jiahang Zhou University of Notre Dame