Surface Hopping Nested Instances Training Set for Excited-state Learning.

Abstract

Theoretical studies of molecular photochemistry and photophysics are essential for understanding fundamental natural processes but rely on computationally demanding quantum chemical calculations. This complexity limits both direct simulations and the development of machine learning (ML) models trained on this data. To address this, we introduce SHNITSEL, a data repository containing 418,870 ab-initio data points of nine organic molecules in their ground and electronically excited states. Each data point includes high-accuracy quantum chemical properties such as energies, forces, and dipole moments in the ground state and electronically excited singlet or triplet states as well as properties that arise from the coupling of electronic states, namely nonadiabatic couplings, transition dipoles, or spin-orbit couplings. Generated with state-of-the-art methods, SHNITSEL provides a robust benchmark for ML models and facilitates the development of ML-based approaches for excited state properties.