Machine Learning Interatomic Potentials for Energy Materials: Architectures, Training Strategies, and Applications

ABSTRACT Machine learning interatomic potentials (MLIPs) provide a practical route to bridge first‐principles accuracy with the spatiotemporal scales required for modeling complex energy materials. By learning potential‐energy surfaces from density functional theory (DFT) data, MLIPs enable atomistic simulations up to 10 4 –10 6 atoms over nanosecond timescales, thereby extending access to phenomena inaccessible to ab initio molecular dynamics. This review presents a phenomenon‐oriented synthesis of MLIP development, spanning descriptor‐based models, Gaussian process frameworks with uncertainty quantification, equivariant graph neural networks, and foundation models trained on large‐scale datasets. We assess how architectural choices, training strategies (including active learning and multi‐fidelity approaches), and dataset design influence accuracy, transferability, and computational efficiency across representative energy systems. Applications in solid‐state electrolytes, battery electrodes, electrocatalysts, perovskite photovoltaics, and high‐entropy materials are discussed with both performance metrics and methodological rationale. While MLIPs capture qualitative trends and relative property variations, absolute predictions remain sensitive to DFT functional choices, dataset representativeness, and treatment of long‐range interactions. We address a structured validation framework linking model fidelity to experimentally relevant observables and system complexity. Also, we outline emerging directions including generative inverse design, autonomous discovery workflows, and foundation‐model fine‐tuning, highlighting opportunities and limitations for predictive deployment in energy materials research.