Machine-Learned Excited-State Dynamics with quantum Monte Carlo: Insights from Azomethane

To simulate photo-induced dynamics, an electronic structure method is required that can accurately describe multiple electronic states and different regions of the potential energy surfaces, while remaining computationally feasible for large numbers of trajectories. We focus on azomethane as a test system, whose excited-state dynamics are particularly sensitive to the choice of many-body wave functions used to describe the states of interest.
We employ quantum Monte Carlo (QMC) to achieve a compact description of the excited-state wave functions, addressing some of the limitations of more conventional approaches. To enable trajectory-based simulations, we train a machine learning model on QMC reference data, with particular attention to reducing the size of the training set, and discuss the quality of the predictions for azomethane’s excited-state dynamics in terms of both the choice of QMC wave function and the data efficiency of the machine learning model.