Evaluating MS-ANI vs. SS-ANI for Surface Hopping Simulations: A Cyclohexadiene Photochemical Ring-Opening Case Study

Machine learning potentials (MLPs) can offer a powerful, cost-effective alternative to traditional quantum mechanical (QM) methods for simulating excited-state dynamics. In this work, we present a comprehensive comparison of two MLP models, SS-ANI and MS-ANI, by simulating the photochemical ring-opening dynamics of 1,3-cyclohexadiene (CHD). The ring-opening dynamics, which involves three lowest singlet electronic states, were investigated using Landau-Zener-based surface hopping dynamics with MLPs. The training datasets were generated from QM-based surface hopping simulations at the SA-3-CASSCF/cc-pVDZ level of theory. Our results show that MS-ANI significantly outperforms SS-ANI in modelling multiple electronic states, yielding more accurate energy predictions, stable energy gaps, and reliable population dynamics. MS-ANI effectively captures interstate correlations, ensuring consistent energy profiles and correct state crossings, making it robust for non-adiabatic molecular dynamics. In contrast, SS-ANI exhibited unstable energy gap behaviors producing negative gaps between the adiabatic potential energy surfaces and inadequate population behavior during the dynamics. Dataset quality and diversity were also found critical, with MS-ANI benefiting most from larger, well-curated datasets. These findings establish MS-ANI as a powerful, scalable approach for sufficiently accurate excited-state simulations in complex photochemical systems.