Atomic potential energy uncertainty in machine-learning interatomic potentials and thermal transport in solids with atomic diffusion

Thermal transport simulations have attracted wide attention in recent years, and one standard approach is to use the Green-Kubo method based on machine-learning interatomic potentials and equilibrium molecular dynamics (GK-MLIP-EMD). In this work, we focus on the lattice thermal conductivities & kappa;Ls for solids with atomic diffusion by taking & beta;-Cu2-xSe (0 x 0.05) as an example. Surprisingly, the GK-MLIP-EMD approach fails in the evaluation of & kappa;Ls for & beta;-Cu1.95Se, whereas the direct method based on nonequilibrium molecular dynamics reliably predicts these values instead. The failure of GK-MLIP-EMD for & beta;-Cu1.95Se could be attributed to the ambiguous projection of the local atomic potential energy Ui in MLIPs, exacerbated by the Cu diffusion at elevated temperatures. The Cu diffusion in & beta;-Cu1.95Se greatly increases the ratio of the convective term and the uncertainty of the conductive term. These influences are considered negligible in crystalline solids. Our findings imply that the ambiguous definition of Ui in MLIPs breaks down the applicability of the GK-MLIP-EMD approach to & kappa;L prediction for solids with severe atomic diffusion.

Automated summary

This work investigates the application of the Green-Kubo method based on machine-learning interatomic potentials and equilibrium molecular dynamics (GK-MLIP-EMD) in thermal transport simulations of solids. Using β-Cu2-xSe (0 x 0.05) as an example, it is found that the GK-MLIP-EMD approach fails to evaluate the lattice thermal conductivities (κLs) for β-Cu1.95Se, while a direct method based on nonequilibrium molecular dynamics reliably predicts these values. The failure of GK-MLIP-EMD for β-Cu1.95Se is attributed to the ambiguous projection of the local atomic potential energy Ui in MLIPs.