Machine learning is funny but physics makes the money: How machine-learning potentials can advance computer-aided materials design in metallurgy

Abstract

Over the past decade, machine learning interatomic potentials (MLIPs) have advanced many areas of computational materials science and chemistry. In metallurgy, however, substantial work remains to be done when computing critical mechanical properties of metals (e.g., strength, ductility, hardness, etc.). This is due to three distinct characteristics of metals that are, in this combined form, not necessarily relevant for predicting other properties (e.g., elastic constants). First, metals are inherently multiscale and, so, simulating realistic microstructures may require millions of atoms. On the other hand, to be predictive, MLIPs must be trained on quantum-mechanical simulations that are limited to a few hundred atoms, which raises the question of transferability to larger systems. Second, there are quantum-mechanical effects, such as magnetism, that are not yet to full extent accounted for in the functional form of state-of-the-art MLIPs. Third, the search space of metallic alloys spans over billions of compositions, and screening for the best possible alloy is like searching for the needle in the haystack without sophisticated high-throughput methods.While the concept of MLIPs is extremely promising regarding quantitative accuracy and automation, these challenges must be addressed in the years to come in order to lift computer-aided materials design to the next level. This review attempts to provide a state-of-the-art survey of ongoing developments towards achieving this goal.