Abstract
Machine learning offers a transformative approach to overcoming the computational bottlenecks of quasiparticle energy calculations in many-body perturbation theory. While the G 0 W 0 approximation provides a rigorous description of excited-state electronic properties, its high computational cost—driven by the evaluation of the inverse dielectric matrix—renders it prohibitive for molecular dynamics (MD) simulations that require sampling thousands of thermal configurations. We present a gradient-boosting framework (LightGBM) that predicts G 0 W 0 quasiparticle energies with high efficiency using only computationally inexpensive mean-field density functional theory eigenvalues and exchange-correlation potentials. Trained on a minimal subset (25%) of MD snapshots for silicon and boron nitride, the model achieves root-mean-square errors below 0.1 eV, accurately reproducing full k -resolved band structures and density of states. Crucially, we demonstrate the model’s transferability and robustness: it successfully generalizes to unseen simulation times, different thermal regimes, and distinct crystal polymorphs excluded from the training data. This data-driven approach effectively bypasses the explicit calculation of the self-energy operator, offering a scalable pathway for high-throughput excited-state simulations and the study of finite-temperature electronic properties in complex materials.
Funding Information
  • Asian Office of Aerospace Research and Development (FA2386-23-1-4104)
  • National Science and Technology Council (113-2112-M-032-013)

This publication has 51 references indexed in Scilit: