Simulating lattice thermal conductivity in semiconducting materials using high-dimensional neural network potential

TL;DR

This paper introduces a high-dimensional neural network potential that accurately predicts lattice thermal conductivity in semiconductors, matching DFT results and enabling efficient simulations for materials like Si and GaN.

Contribution

The study develops and validates a high-dimensional neural network potential that achieves DFT-level accuracy in predicting thermal conductivity of semiconductors.

Findings

HDNNP predicts thermal conductivity within 1% of DFT for Si.

HDNNP predicts thermal conductivity within 5.4% of DFT for GaN.

Root mean square error of forces is less than 40 meV/Å.

Abstract

We demonstrate that a high-dimensional neural network potential (HDNNP) can predict the lattice thermal conductivity of semiconducting materials with an accuracy comparable to that of density functional theory (DFT) calculation. After a training procedure based on the force, the root mean square error between the forces predicted by the HDNNP and DFT is less than 40 meV/{\AA}. As typical examples, we present the results for Si and GaN bulk crystals. The deviation from the thermal conductivity calculated using DFT is within 1% at 200 to 500 K for Si and within 5.4% at 200 to 1000 K for GaN.