Incorporating quasiparticle and excitonic properties into material discovery

TL;DR

This paper develops machine learning models trained on a new database of quasiparticle and excitonic properties, enabling rapid prediction and discovery of light-absorbing materials with high accuracy and reduced computational cost.

Contribution

The work introduces a high-throughput GW-BSE workflow and a machine learning approach to predict quasiparticle and excitonic properties, accelerating material discovery.

Findings

ML models predict quasiparticle gaps with RMSE of 0.36 eV

ML models classify excitonic binding energies with 90% accuracy

Application to Materials Project database identified new photoabsorbers

Abstract

In recent years, GW-BSE has been proven to be extremely successful in studying the quasiparticle (QP) bandstructures and excitonic effects in the optical properties of materials. However, the massive computational cost associated with such calculations restricts their applicability in high-throughput material discovery studies. Recently, we developed a Python workflow package, $py$GWBSE, to perform high-throughput GW-BSE simulations. In this work, using $py$GWBSE we create a database of various QP properties and excitonic properties of over 350 chemically and structurally diverse materials. Despite the relatively small size of the dataset, we obtain highly accurate supervised machine learning (ML) models via the dataset. The models predict the quasiparticle gap with an RMSE of 0.36 eV, exciton binding energies of materials with an RMSE of 0.29 eV, and classify materials as high or low excitonic binding energy materials with classification accuracy of 90%. We exemplify the application of these ML models in the discovery of 159 visible-light and 203 ultraviolet-light photoabsorber materials utilizing the Materials Project database.