Insight into the effect of force error on the thermal conductivity from machine-learned potentials

Abstract

Machine-learned potentials (MLPs) have been extensively used to obtain the lattice thermal conductivity (κ) via atomistic simulations. However, the impact of force errors in various MLPs on thermal transport has not been widely recognized and remains to be fully understood. Here, we employ MLP-driven molecular dynamics (MD) and anharmonic lattice dynamics (LD) to systematically investigate how the calculated κ varies with the force errors, using boron arsenide as a prototypical material to emphasize the challenges associated with high thermal conductivity. We consistently observe an underestimation of κ in MD simulations with different MLPs including the neuroevolution potential, deep potential, and moment tensor potential (MTP). We propose a robust second-order extrapolation scheme based on controlled force noises via the Langevin thermostat to correct this underestimation. The corrected results achieve a good agreement with previous experimental measurements from 200 K to 600 K. In contrast, the κ values from LD calculations with MLPs readily align with the experimental data, which is attributed to the much smaller effects of the force errors on the force-constant calculations. Our findings provide deeper physical insight into the effect of the force errors in machine-learned potentials on thermal transport, and are particularly instrumental for simulating and seeking high-κ materials. In addition, we also make our modified version of the MLIP package publicly accessible in order to facilitate the accurate calculation of heat current in MTP-based MD simulations.