Machine-learning-accelerated simulations of vibrational activation for controlled photoisomerization in a molecular motor.

The precise manipulation of photochemical reactions across broad configurational spaces requires sophisticated design of external control fields. Using the photoisomerization of a molecular motor as a prototype, this study integrates enhanced sampling and active learning to construct accurate machine-learned multi-state potential energy surfaces. By combining active-learning trajectories with enhanced sampling, our approach efficiently covers substantial reaction regions, enabling trajectory propagation extending to tens of picoseconds at a low computational cost within the machine learning framework. Furthermore, local control theory (LCT) is employed to selectively activate specific vibrational motions, leading to accelerated access to reactive regions, enhanced nonadiabatic transitions, and significantly improved selectivity toward the dominant photoproduct. This combined strategy of machine-learning potentials and LCT offers an efficient and generalizable framework for controlling excited-state dynamics in complex systems.