Machine-Learning Interatomic Potentials for Bulk Metallic Glasses

Recently, metallic glass nanowires have been shown to have high catalytic activity, prompting further investigation for applications in fuel cells and electrocatalytic water splitting. However, empirical potentials used in molecular dynamics are conventionally tailored for periodic crystal structures and are inadequate for modeling amorphous interfaces. Consequently, machine learning algorithms have progressively gained popularity due to their low-cost implementation and recent successes. In this study, we generate a performant interatomic potential for NiP binary alloys using an artificial neural-network (ANN) algorithm coupled with an automated framework. An extensive training set of configurations is generated starting from available embedded-atom model (EAM) potentials in the literature. Accurate first principles simulations are performed on the training set in a high-throughput fashion using the AiiDA infrastructure. The total workflow is automated for exploring other multinary alloys, including CoP and alloys in combination with Pt-group elements. As additional data is provided to the neural network, the ANN potential becomes increasingly inclusive and generalizable, making performance improvement effortless and only as expensive as necessary.