An Overview of Machine Learning Potentials for Structural Materials

Machine-learning interatomic potentials (MLIPs) are emerging as a practical route to bridge the accuracy-cost gap between empirical force fields and first-principles methods for atomistic modeling of structural materials.This article reviews two major families-descriptor-based models (e.g., BPNN, GAP/SOAP, SNAP, DeePMD) and graph neural network approaches (e.g., SchNet, NequIP, MACE, Allegro)-highlighting differences in symmetry handling, scalability, and data requirements.We outline dataset construction workflows, including active-learning loops, uncertainty estimation, and transfer learning, that reduce labeling cost while improving coverage of phases, defects, and compositions.Representative applications to metals and ceramics are summarized, such as screw-dislocation core energetics in bcc Fe, diffusion and chemical short-range order in CrCoNi medium-entropy alloys, and large-scale plasticity in functional ceramics.Practical guidelines are discussed for choosing between descriptor-based and equivariant GNN models, balancing accuracy and throughput, leveraging GPU acceleration, and validating models via thermo-mechanical properties and dynamical simulations.