High-Throughput GW Calculations via Machine Learning

TL;DR

This paper introduces a machine learning framework that predicts $G_0W_0$ quasiparticle energies from DFT data, enabling fast and accurate excited-state calculations across molecular dynamics trajectories.

Contribution

The novel ML approach predicts quasiparticle energies efficiently using only DFT inputs, reducing computational costs for dynamic systems.

Findings

Achieves RMSE below 0.1 eV on test data

Reproduces band structures and density of states accurately

Effective even for unseen BN polymorphs

Abstract

We present a machine learning (ML) framework that predicts $G_0W_0$ quasiparticle energies across molecular dynamics (MD) trajectories with high accuracy and efficiency. Using only DFT-derived mean-field eigenvalues and exchange-correlation potentials, the model is trained on 25\% of MD snapshots and achieves RMSEs below 0.1 eV. It accurately reproduces k-resolved quasiparticle band structures and density of states, even for BN polymorphs excluded from the training data. This approach bypasses the computational bottlenecks of $G_0W_0$ simulations over dynamic configurations, offering a scalable route to excited-state electronic structure simulations with many-body accuracy.