DIRECT DEEP LEARNING MODELS FOR ATOMIC FORCES: REACHING DFT ACCURACY

The computation of large systems atomistic dynamics is required in a wide range of fields. An approach that was developed in the last few years is to use Machine Learning (ML) and Deep Learning (DL) algorithms to build an "on the fly", computationally cheap, predictors for the energy, forces and other physical properties. Those approaches allow to do calculation with accuracy that is close to a fully quantum MD but with speeds that are more than 100 times faster.

An important challenge is to be able to learn the forces for a large system from quantum simulations of small systems. For this, the input to a ML or DL training and model should have a "local" description, taking into account only some local environment around the atom in question. In this work we describe the construction and use of local environments for a DL based model for a direct estimation of the forces. We show that with these DL models we can reach accuracy that is comparable to state of the art ML and DL models. We also analyze the dependence of such models on parameters such as the number of neighbors, system size and temperature.