Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 17: Data-driven Materials Science: Big Data and Workflows II
MM 17.7: Talk
Tuesday, March 10, 2026, 15:30–15:45, SCH/A251
Learning G0W0 Self-Energies in Real Space with Equivariant Neural Networks — •Elisabeth Keller, Karsten W. Jacobsen, and Kristian S. Thygesen — CAMD, DTU Physics, Kongens Lyngby, Denmark
Many-body G0W0 calculations provide highly accurate quasiparticle energies for semiconductors and insulators beyond standard density-functional theory, but at a much higher computational cost.
To overcome this limitation, we use equivariant neural networks to replace the explicit G0W0 self-energy evaluation. The networks are trained on G0W0 self-energies from GPAW projected onto an atom-centered LCAO basis. Using this representation, we investigate how the real-space localization of the self-energy enables learning from atomic configurations.
Keywords: GW; Equivariant neural networks; Self energy; Many-body electronic structure