Towards Accurate Prediction of Configurational Disorder Properties in Materials using Graph Neural Networks

TL;DR

This paper introduces a graph neural network-based framework to accurately predict configurational disorder properties in complex materials, surpassing traditional methods and aligning well with experimental data.

Contribution

It presents a novel GNN-based workflow combined with Monte Carlo simulations for predicting disorder properties in complex materials systems.

Findings

GNNs can predict phase transition temperatures close to experimental values.

Energy deviation variance influences prediction accuracy.

The framework offers a new data-driven approach for complex material design.

Abstract

The prediction of configurational disorder properties, such as configurational entropy and order-disorder phase transition temperature, of compound materials relies on efficient and accurate evaluations of configurational energies. Previous cluster expansion methods are not applicable to configurationally-complex material systems, including those with atomic distortions and long-range orders. In this work, we propose to leverage the versatile expressive capabilities of graph neural networks (GNNs) for efficient evaluations of configurational energies and present a workflow combining attention-based GNNs and Monte Carlo simulations to calculate the disorder properties. Using the dataset of face-centered tetragonal gold copper without and with local atomic distortions as an example, we demonstrate that the proposed data-driven framework enables the prediction of phase transition temperatures close to experimental values. We also elucidate that the variance of the energy deviations among configurations controls the prediction accuracy of disorder properties and can be used as the target loss function when training and selecting the GNN models. The work serves as a fundamental step toward a new data-driven paradigm for the accelerated design of configurationally-complex functional material systems.