LAMMPS implementation of rapid artificial neural network derived interatomic potentials

While machine learning approaches have been successfully used to represent interatomic potentials, their speed has typically lagged behind conventional formalisms. This is often due the complexity of structural fingerprints describe local atomic environment and large cutoff radii neighbor lists in calculation these fingerprints. Even recent learned methods are at least 10 times slower than traditional An implementation a rapid artificial neural network (RANN) style potential LAMMPS molecular dynamics package presented here which utilizes angular screening reduce computational without reducing accuracy. For smallest architectures, this formalism rivals modified embedded atom method (MEAM) for accuracy, while networks approximately one third as fast MEAM were capable reproducing training database with chemical The numerical accuracy assessed by verifying conservation energy agreement between calculated forces pressures observed derivatives well assessing stability dynamic simulation. tested using force field magnesium efficiency variety architectures compared models alternative ANN predictive found rival that methods. Additionally, transferability demonstrated correctly predicting Mg phase diagram include pressure dependence on melting temperature presence high BCC phase.