Virtual Node Graph Neural Network for Full Phonon Prediction

TL;DR

This paper introduces a virtual node graph neural network that predicts full phonon spectra and bandstructures directly from atomic coordinates, enabling rapid materials design with improved phonon property predictions.

Contribution

The paper develops three types of virtual node approaches to enhance graph neural networks for predicting complex phonon properties in materials.

Findings

Successfully predicts phonon bandstructures for various alloys

Builds a large phonon database with over 146,000 materials

Enables rapid, high-quality phonon spectra prediction for materials design

Abstract

The structure-property relationship plays a central role in materials science. Understanding the structure-property relationship in solid-state materials is crucial for structure design with optimized properties. The past few years witnessed remarkable progress in correlating structures with properties in crystalline materials, such as machine learning methods and particularly graph neural networks as a natural representation of crystal structures. However, significant challenges remain, including predicting properties with complex unit cells input and material-dependent, variable-length output. Here we present the virtual node graph neural network to address the challenges. By developing three types of virtual node approaches - the vector, matrix, and momentum-dependent matrix virtual nodes, we achieve direct prediction of $\Gamma$-phonon spectra and full dispersion only using atomic coordinates as input. We validate the phonon bandstructures on various alloy systems, and further build a $\Gamma$-phonon database containing over 146,000 materials in the Materials Project. Our work provides an avenue for rapid and high-quality prediction of phonon spectra and bandstructures in complex materials, and enables materials design with superior phonon properties for energy applications. The virtual node augmentation of graph neural networks also sheds light on designing other functional properties with a new level of flexibility.