LAMMPS implementation of rapid artificial neural network derived interatomic potentials

While machine learning has been successful for representing interatomic potentials, their speed has lagged conventional formalisms. This is often due to the complexity of the structural fingerprints used to describe the local atomic environment and the large cutoff radii and neighbor lists used in the calculation of these fingerprints. Even recent machine learned methods are at least 10 times slower than traditional formalisms. An implementation of a rapid artificial neural network (RANN) style potential in the LAMMPS molecular dynamics package is presented here which utilizes angular screening to reduce computational complexity without reducing accuracy. For the smallest neural network architectures, this formalism rivals the modified embedded atom method (MEAM) for speed and accuracy, while the networks approximately one third as fast as MEAM could reproduce the training database with chemical accuracy. The computational efficiency for a variety of architectures is compared to a traditional potential model as well as alternative ANN formalisms. The predictive accuracy is found to rival that of slower methods. Additionally, the transferability of the formalism is demonstrated by correctly predicting the Mg phase diagram include the pressure dependence on melting temperature and the presence of a high pressure BCC phase.