Iterative training set refinement enables reactive molecular dynamics via machine learned forces.

Machine learning approaches have been successfully employed in many fields of computational chemistry and physics. However, atomistic simulations driven by machine-learned forces are still very challenging. Here we show that reactive self-sputtering from a beryllium surface can be simulated using neural network trained forces with an accuracy that rivals or exceeds other approaches. The key in machine learning from density functional theory calculations is a well-balanced and complete training set of energies and forces. We have implemented a refinement protocol that corrects the low extrapolation capabilities of neural networks by iteratively checking and improving the molecular dynamic simulations. The sputtering yield obtained for incident energies below 100 eV agrees perfectly with results from ab initio molecular dynamics simulations and compares well with earlier calculations based on pair potentials and bond-order potentials. This approach enables simulation times, sizes and statistics similar to what is accessible by conventional force fields and reaching beyond what is possible with direct ab initio molecular dynamics. We observed that a potential fitted to one surface, Be(0001), has to be augmented with training data for another surface, Be(011̄0), in order to be used for both.