Capturing Excited State Proton Transfer Dynamics with Reactive Machine Learning Potentials

Excited state proton transfer is a fundamental process in photochemistry, playing a crucial role in fluorescence sensing, bioimaging, and optoelectronic applications. However, fully resolving its dynamics remains challenging due to the prohibitive computational cost of ab initio simulations and the need for ultrafast experimental techniques with high temporal resolution. Here, we tackle this challenge by using machine learning-driven excited state molecular dynamics simulations. We propose an active learning framework powered by enhanced sampling techniques for constructing high-quality training set for excited state machine learning potentials, which we then use to map the reaction free energy landscape and capture the real-time evolution of photorelaxation. Using 10-hydroxybenzo[h]quinoline as a test case, our simulations reveal a barrierless proton transfer occurring within ∼50 fs, accompanied by a significant red shift in the emission energy (∼1 eV), in agreement with experimental findings. Furthermore, our results highlight a strong coupling between proton transfer and charge redistribution, which facilitates the rapid tautomerization process. These findings showcase the power of machine learning-driven molecular dynamics in accurately capturing photochemical dynamics while enabling large-scale statistical sampling beyond the reach of conventional ab initio methods.