Studying Heat Conduction Properties by Using Machine Learning Potentials: a brief review

The reliability of simulations of thermal properties in micro-/nano-scale structures critically depends on the accuracy of potential function. While empirical force fields offer limited precision, machine learning potentials effectively bridge this gap by combining the computational accuracy of density functional theory with the efficiency of empirical force fields. This review provides a systematic investigation of various frameworks of machine learning potentials, including descriptors, training methodologies, and accuracy. Furthermore, a comparison of their predictive performance for thermal conductivity across different material systems is presented. Finally, key challenges and future research directions for machine learning potentials are discussed, aiming to promote their applications in thermal transport studies and enhance the understanding of heat conduction mechanisms.