Deep learning of spectra: Predicting the dielectric function of semiconductors

TL;DR

This paper introduces a machine learning approach using graph attention neural networks to accurately predict the dielectric function and spectra of semiconductors from crystal structures, based on an extensive ab initio database.

Contribution

It creates a large dielectric tensor database and develops the OptiMate GAT models for spectral prediction, demonstrating high accuracy without explicit spectral constraints.

Findings

OptiMate models accurately predict dielectric spectra.

The approach produces smooth, artifact-free spectral curves.

The database enables data-driven materials property prediction.

Abstract

Predicting spectra and related properties such as the dielectric function of crystalline materials based on machine learning has a huge, hitherto unexplored, technological potential. For this reason, we create an ab initio database of 9915 dielectric tensors of semiconductors and insulators calculated in the independent-particle approximation (IPA). In addition, we present the OptiMate family of machine learning models, a series of graph attention neural networks (GAT) trained to predict the dielectric function and refractive index. OptiMate yields accurate prediction of spectra of semiconductors using only their crystal structure. Smooth, artifact-free curves are obtained without these properties being enforced by penalties.