{\ae}net-PyTorch: a GPU-supported implementation for machine learning atomic potentials training

TL;DR

{\

Contribution

{\

Findings

Training on 10-20% of force data yields optimal accuracy.

GPU implementation reduces training time by 10-100x.

Effective for large systems beyond small molecules.

Abstract

In this work, we present {\ae}net-PyTorch, a PyTorch-based implementation for training artificial neural network-based machine learning interatomic potentials. Developed as an extension of the atomic energy network ({\ae}net), {\ae}net-PyTorch provides access to all the tools included in {\ae}net for the application and usage of the potentials. The package has been designed as an alternative to the internal training capabilities of {\ae}net, leveraging the power of graphic processing units to facilitate direct training on forces in addition to energies. This leads to a substantial reduction of the training time by one to two orders of magnitude compared to the CPU implementation, enabling direct training on forces for systems beyond small molecules. Here we demonstrate the main features of {\ae}net-PyTorch and show its performance on open databases. Our results show that training on all the force information within a data set is not necessary, and including between 10% to 20% of the force information is sufficient to achieve optimally accurate interatomic potentials with the least computational resources.