Efficient training of machine learning potentials for metallic glasses: CuZrAl validation

Abstract

Interatomic potentials are key to uncovering microscopic structure-property relationships, essential for multiscale simulations and high-throughput experiments. For metallic glasses, their disordered atomic structure makes the development of potentials particularly challenging, resulting in the scarcity of chemistry-specific parametrizations for this important class of materials. We address this gap by introducing an efficient methodology to design machine learning interatomic potentials (MLIPs), benchmarked on the CuZrAl system. Using a Lennard-Jones surrogate model, swap-Monte Carlo sampling, and single-point Density Functional Theory (DFT) corrections, we capture amorphous structures spanning 14 decades of supercooling. These representative configurations, competing with the experimental time scale, enable robust model training across diverse states, while minimizing the need for extensive DFT datasets. The resulting MLIP matches the experimental data and predictions of the classical embedded atom method (EAM) for structural, dynamical, energetic, and mechanical properties. This approach offers a scalable path to develop accurate MLIPs for complex metallic glasses, including emerging multi-component and high-entropy systems.