Ab initio-machine learning framework for non-adiabatic dynamics Insights from azomethane photoisomerization

Upon photoexcitation, molecular and condensed systems can undergo major structural changes, requiring accurate modeling of coupled electronic and nuclear dynamics over multiple electronic energy surfaces. To simulate such nonadiabatic dynamics, approximate quantum-classical schemes like surface hopping are typically employed, which require however numerous trajectories and can therefore be computationally very expensive. Here, we study azomethane, a benchmark system for photoisomerization, using high-level electronic structure methods to explore how different electronic descriptions affect relaxation pathways. To reduce the high cost of simulating many trajectories, we train a neural network on quantum data to efficiently predict energies, forces, and couplings. We are now retraining this model using quantum Monte Carlo data and plan to incorporate new, chemically intuitive wave functions developed by collaborators at the University of Groningen to better capture electronic excitations. Our goal is to extend this hybrid approach to larger systems like cyanine dyes relevant to photovoltaic applications.