Neural networks for the prediction of electronic excitation dynamics

We demonstrate a machine-learning-based approach that can learn the time-dependent electronic excitation dynamics of small molecules subjected to ion irradiation. Ensembles of neural networks are trained on data generated by time-dependent density-functional theory to relate atomic positions to occupations of molecular orbitals. New data are incrementally and efficiently added to the training data using an active learning process, thereby improving model accuracy. Predicted changes in orbital occupations made by the neural network ensemble are found to have errors and one-standard-deviation uncertainties which are two orders of magnitude smaller than the typical values of the orbital occupation numbers. The trained neural network ensembles demonstrate a limited ability to generalize to molecules not used to train the models. In such cases, the models are able to identify key qualitative features, but struggle to match the quantitative values. The machine-learning procedure developed here is potentially broadly applicable and has the potential to enable study of broad ranges of both materials and dynamical processes by drastically lowering the computational cost and providing surrogate model for multiscale simulations.