Characterization of thermal properties of 2 dimensional materials with machine learning

The accurate prediction of phonon interactions plays a significant role in thermal transport simulations of semiconductors and insulators. Generating harmonic and anharmonic force constants using density functional theory and iterative solution of Boltzmann transport equation is one of the reliable methods for predicting lattice thermal transport properties. However, the high computational cost of DFT is a limiting factor except for calculation on perfect crystals. While molecular dynamics (MD) simulations offer an alternative method, the solid functional form of classical force fields adversely affects the accuracy of calculations. Machine learning (ML) techniques propose an approach to gathering positive aspects of DFT and MD simulations, i.e. working on realistic systems with high accuracy and reasonable computational cost. For this purpose, we generated machine learning-based interatomic potentials (MLIP) using Gaussian Approximation Potential (GAP) for 2-dimensional graphene, silicene, and h-XN (X = B, Al, Ga). In order to test the accuracy of the GAP for all considered structures, we calculated phonon dispersion curves and lattice thermal conductivity via harmonic and anharmonic force constants, respectively, and compared them to DFT results. Additionally, we generated force constants for both DFT and GAP using the HIPHIVE Python library, aimed to reduce the computational costs of calculating anharmonic force constants. GAP predicted not only acoustic modes but also optical modes in DFT accuracy.