Machine Learning-Based Excited-State Reactive Force Field: A New Approach for Modeling the Photodissociation Dynamics of 2-Fluorothiophenol.

The absence of accurate yet efficient excited-state reactive force fields has emerged as a critical bottleneck hindering advancements in photochemical dynamics. To overcome it, we develop a machine learning-based excited-state reactive force field (ML-ES-RFF) that implements an innovative divide-and-conquer strategy. Specifically, this approach systematically classifies degrees of freedom (DOFs) into chemically active and inactive ones based on their contributions in reactions. For active DOFs, which exhibit a significant anharmonicity, the potential energy surfaces are constructed using high-level quantum mechanics-based neural network potentials (QM-NNPs). In contrast, for inactive DOFs, which satisfy the harmonic approximation near equilibrium, the potential energy surfaces could be described by conventional molecular mechanics (MM) force fields. This multiscale methodology simultaneously achieves quantum chemical accuracy and remarkable computational efficiency. To validate the reliability of this approach, we applied it to study the nonadiabatic dynamics of 2-fluorothiophenol using ML-ES-RFF, successfully obtaining complete atomistic characterization of the photodissociation process. Our framework establishes a new paradigm for studying excited-state processes in complex molecular systems.