High-Throughput GW Calculations via Machine Learning

Ragab Ahmed Ragab Abdelghany^1*, Chih-En Hsu², Hung-Chung Hsueh², Yuan-Hong Tsai³, Ming-Chiang Chung¹

¹Physics, NCHU, Taichung, Taiwan
²Physics, Tamkang University, New Taipei, Taiwan
³Physics, AI Foundation, Taipei, Taiwan

* Presenter:Ragab Ahmed Ragab Abdelghany, email:abdelghany.ra@gmail.com

We present a machine learning framework for predicting $G_0W_0$ quasiparticle energies across molecular dynamics 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 root-mean-square errors below 0.1 eV. It accurately reproduces k-resolved quasiparticle band structures and density of states, even for unseen BN polymorphs. This approach bypasses the computational bottlenecks of traditional GW simulations, offering a scalable route to excited-state electronic structure simulations with many-body accuracy.

Keywords: Machine Learning, GW Approximation, Electronic Structure, Molecular Dynamics