Exploring active learning strategies for excited state dynamics: application to uracil

In this work, we implement and benchmark the construction of an effective model for the excited state dynamics of uracil using a machine-learned approach based on the polarizable atom interaction neural network (PaiNN) architecture. Using trajectory surface hopping dynamics data at a multireference level of theory, we trained a neural network model that predicts energies, forces, and nonadiabatic couplings. We benchmark the effect of the loss function and weights for the three properties on the training, and find that the inverse hyperbolic sine loss function (Asinh) for training forces is superior to the more standardized mean square error loss function, while the best performing weights favor forces and nonadiabatic couplings. The model is then fine-tuned using an adaptive active learning approach based on the energy gap between electronic states in order to enhance sampling around conical intersections. The accuracy of the population is improved within just one cycle of active learning. The model can be used to extend photochemical dynamics simulations of uracil up to 4-6 ps with a minimum number of additional structures (≈1% of the total data used to train the model). We also showed that we can accurately predict the potential energy surfaces around conical intersections by adding data from the branching plane. So, a combination of active learning using energy gaps plus selected points around conical intersections can provide both accurate surfaces and dynamics.