Assessing the accuracy of machine learning interatomic potentials in predicting the elemental orderings: A case study of Li-Al alloys

Abstract

In atomistic modeling, machine learning interatomic potential (MLIP) has emerged as a powerful technique for studying alloy materials. However, given that MLIPs are often trained on a limited set of materials, a concern remains regarding the MLIP's capability to make accurate predictions for a wide variety of phases, compositions, lattice structures, and elemental orderings across alloy systems. This paper presents a detailed analysis of MLIP's performance in the Li-Al alloy system. Even trained only on a very limited number of phases, the MLIPs exhibit good accuracies in predicting a vast array of known and generated intermediate phases and their elemental orderings across the alloy system. We propose and demonstrate several evaluation metrics to assess and quantify the relative stabilities of complex elemental orderings, which is critical for studying the thermodynamics of alloys. Our testing process combined with the evaluation metrics is valuable for quantifying the performance and the transferability of MLIPs and for future improvements of MLIPs.