AI and the World's Hard Problems · Quantum physics & quantum computing · arXiv · 2026-10-19
Learning the self-energy through equivariant neural networks
The paper trains an E(3)-equivariant message-passing network to predict the dynamical correlation part of the G0W0 self-energy matrix for molecules on the imaginary frequency axis. The predictions reproduce matrix elements to below 1 meV, and after continuation to the real axis, the reconstructed spectral functions follow the G0W0 reference peak by peak for water clusters and QM9 molecules. Frontier quasiparticle levels are recovered to a few tens of meV, and once trained, a new geometry needs no explicit G0W0 calculation.