From Ground State to Electronic and Vibronic Spectra via Memory Kernel Coupling Theory and Neural Quantum States

Abstract

Modelling molecular absorption spectra from scalable ground-state representations is challenging because conventional approaches often require explicit excited states or time propagation. Here, we introduce a method to obtain Absorption Using Recursive Operator Response with vibronic Augmentation (Aurora), which combines neural quantum states (NQSs) with continued-fraction memory-kernel coupling theory (CF-MKCT). Aurora maps the zero-temperature Liouvillian response hierarchy onto repeated wavefunction-space applications of a shifted Hamiltonian to a dipole-prepared state, enabling direct spectral reconstruction from variational ground states. The implementation combines variational Monte Carlo, implicit-restarted-Lanczos-assisted NQS optimization, matrix-free Hamiltonian application, and adaptive recursion control. The resulting electronic spectra closely agree with full-configuration-interaction and conventional MKCT references for tractable molecules, with the restarted-Lanczos strategy improving difficult NQS cases. For electron-phonon Hamiltonians, Aurora combines Huang-Rhys mode selection with local singular-value-decomposition (SVD) phonon encoders. The H2 benchmark indicates that compression retains the dominant vibronic envelope and improves upon bare local-basis truncation, while paired calculations for N2, H2O, H2S, and NH3 yield similar principal features with and without encoding. We further obtain vibronic spectra for C6H6 and a 10-electron, 10-orbital C10H8 active space. Overall, Aurora provides a unified route from variational molecular ground states to electronic and vibronic spectra without constructing excited-state manifolds.