Density of states prediction for materials discovery via contrastive learning from probabilistic embeddings

Shufeng Kong,Francesco Ricci,Jeffrey B Neaton,Dan Guevarra,Carla P Gomes,John M Gregoire

doi:10.1038/s41467-022-28543-x

Abstract

Machine learning for materials discovery has largely focused on predicting an individual scalar rather than multiple related properties, where spectral properties are an important example. Fundamental spectral properties include the phonon density of states (phDOS) and the electronic density of states (eDOS), which individually or collectively are the origins of a breadth of materials observables and functions. Building upon the success of graph attention networks for encoding crystalline materials, we introduce a probabilistic embedding generator specifically tailored to the prediction of spectral properties. Coupled with supervised contrastive learning, our materials-to-spectrum (Mat2Spec) model outperforms state-of-the-art methods for predicting ab initio phDOS and eDOS for crystalline materials. We demonstrate Mat2Spec’s ability to identify eDOS gaps below the Fermi energy, validating predictions with ab initio calculations and thereby discovering candidate thermoelectrics and transparent conductors. Mat2Spec is an exemplar framework for predicting spectral properties of materials via strategically incorporated machine learning techniques.

Highlights

Machine learning for materials discovery has largely focused on predicting an individual scalar rather than multiple related properties, where spectral properties are an important example
We introduce Mat2Spec, which builds upon the concepts of H-CLMP to address the open challenge of spectral property prediction with a graph neural networks (GNN) materials encoder coupled to probabilistic embedding generation and contrastive learning
Given the computational expense of generating these spectra, the ability to predict these spectra, even in an approximate way, represents a valuable tool to perform materials screening to guide ab initio computation. We demonstrate such acceleration of materials discovery with a use case based on the identification of band gaps below but near the Fermi energy in metallic systems, which have been shown to be pertinent to thermoelectrics and transparent conductors[35,36]

Summary

Introduction

Machine learning for materials discovery has largely focused on predicting an individual scalar rather than multiple related properties, where spectral properties are an important example. The recent advent of machine learning prediction of materials properties has introduced the possibility of even higher throughput primary screening due to the minuscule expense of making a prediction for a candidate material using an already-trained model[7–10]. Toward this vision, we introduce the materials to spectrum (Mat2Spec) framework for predicting spectral properties of crystalline materials, demonstrated for the prediction of the ab initio phonon and electronic densities of state. For a complete review of target properties predicted via ML, see ref

Methods

Results

Conclusion