Key physical descriptors for predicting interfacial thermal resistance by machine learning

Abstract

Interfacial thermal resistance (ITR) plays a crucial role in the thermal management of micro-nano devices. Compared with traditional approaches, machine learning is a cost-effective, accurate, and efficient method for predicting material properties, but its limited interpretability hinders the exploration of underlying mechanisms. In this study, we have constructed machine learning models that can effectively predict ITR by comparing six algorithms, with the random forest (RF) model achieving the highest accuracy. Through feature engineering, 39 descriptors have been narrowed down to 7 descriptors while maintaining accuracy, among which 5 descriptors are all strongly related to the bonding strength and, thus, the vibrational properties of the two materials consisting of the interface. In particular, we have defined new descriptors corresponding to the matching degree of the sound velocity and applied it to obtain an improved RF model. We further employed this improved RF model to explore materials with a low ITR when in contact with Si, providing guidance for the search of chip heat dissipation materials. The key physical descriptors agree well with the physical picture of phonon transport across interfaces and enhance the interpretability of the machine learning models, which can accelerate the design and optimization of thermal interfaces in relevant applications.