Equivariant graph neural network interatomic potential for Green-Kubo thermal conductivity in phase change materials

TL;DR

This paper introduces an $E(3)$-equivariant neural network interatomic potential combined with Green-Kubo formalism to accurately compute thermal conductivity in phase change materials like GeTe, capturing phase transitions and anharmonic effects.

Contribution

It presents a novel application of equivariant neural networks within Green-Kubo to evaluate thermal conductivity, effectively modeling phase transitions and anharmonicity in GeTe.

Findings

Accurately predicts phase transition temperatures in GeTe.

Captures anharmonic effects leading to precise thermal conductivity calculations.

Outperforms Boltzmann transport equation in crystalline phases.

Abstract

Thermal conductivity is a fundamental material property that plays an essential role in technology, but its accurate evaluation presents a challenge for theory. In this work, we demonstrate the application of $E(3)$-equivariant neutral network interatomic potentials within Green-Kubo formalism to determine the lattice thermal conductivity in amorphous and crystalline materials. We apply this method to study the thermal conductivity of germanium telluride (GeTe) as a prototypical phase change material. A single deep learning interatomic potential is able to describe the phase transitions between the amorphous, rhombohedral and cubic phases, with critical temperatures in good agreement with experiments. Furthermore, this approach accurately captures the pronounced anharmonicity that is present in GeTe, enabling precise calculations of the thermal conductivity. In contrast, the Boltzmann transport equation including only three-phonon processes tends to overestimate the thermal conductivity by approximately a factor of 2 in the crystalline phases.