Neural Networks Accelerate Ab Initio Multiple Spawning Simulations: A Case Study of Using Machine Learning Potentials for Excited State Dynamics

<jats:p>Nonadiabatic simulations are critical for understanding photochemical reactions but are often computationally prohibitive, limiting their application to relatively small molecules and short timescales. Machine learning interatomic potentials (MLIPs) promise to remove most of the cost by replacing the electronic structure calls that dominate these simulations, while retaining chemical accuracy. We evaluate independent MLIPs fit to the ground and first excited state for two prototypical photochemical molecules — ethylene and the deprotonated green fluorescent protein chromophore. Despite sub-chemical-accuracy errors on held-out test data, the MLIPs predict qualitatively wrong potential energy surfaces (PESs) near conical intersections: artificial gaps, spurious near-degeneracies, and even reordering of the adiabatic states. We repair this with a hybrid “holey-ML” approach that switches from the MLIPs to ab initio quantum chemistry (QM) wherever the ML-predicted gap between states is small. This hybrid approach reduces the computational cost by an order of magnitude while reproducing the excited state population decay of reference fully ab initio QM simulations. Therefore, hybrid QM/ML schemes are a practical route to routine nonadiabatic dynamics.</jats:p>