A Review of Atomistic Machine Learning for Grain Boundaries in Advanced Manufacturing: From Metastable Structures to Processing–Structure–Property Design

Abstract Grain boundaries are central to the performance of polycrystalline materials, yet their structures and properties become especially difficult to predict in advanced manufacturing, where rapid thermal cycling, steep gradients, severe deformation, and chemical complexity drive interfaces far from equilibrium. Machine learning offers new opportunities to represent, predict, and ultimately design such complex grain-boundary states. This review surveys recent progress in atomistic machine learning for grain boundaries, with emphasis on manufacturing-relevant challenges. We first discuss digital representations of grain boundaries, including crystallographic descriptors, SOAP-based local atomic environment fingerprints, and physics-informed descriptors for predicting grain-boundary energy and segregation behavior. We then review machine learning interatomic potentials and active-learning strategies that enable larger-scale, higher-fidelity simulations of chemically complex and dynamically evolving interfaces. Particular attention is given to frontier challenges in predicting finite-temperature and kinetic properties, including grain-boundary mobility, shear coupling, defect absorption, disordering, and complexion transitions. Finally, we outline an outlook in which interpretable machine learning, uncertainty-aware models, grain-boundary diagrams, and inverse design workflows connect processing history to interfacial structure and materials performance.