Polyhedral-Cluster Machine Learning Potential for Efficient and Accurate Modeling of Complex Li-Ion Cathodes

Li-ion layered cathodes undergo complex structural evolution during delithiation, where local coordination changes, transition-metal migration, and oxygen-framework distortion are coupled across multiple length scales. Machine learning interatomic potentials can enable efficient atomic modeling of these processes, but conventional local descriptors often struggle to capture medium-range structural correlations without introducing redundant features. Here, a polyhedral-cluster potential (PCP) is proposed for Li-Mn-Ti-O cathode materials to combine short-range local coordination and medium-range oxygen-framework connectivity in a compact representation. SOAP descriptors encode the local environments within oxygen-centered polyhedra, while O-O-mediated cluster features describe interpolyhedral correlations. Compared with a conventional local potential, PCP improves DFT energy and force predictions for both ordered rocksalt-like and delithiated distorted structures, reducing the force MAE of the distorted structure from 101.22 to 41.99 meV Å-1. Descriptor analysis further shows that PCP decreases the feature number from 952 to 250 and lowers redundancy from 45.6% to 18.0%, while increasing the contribution of physically meaningful O-O correlations. Applications to equation-of-state (EOS) and Li-ion migration-path configurations demonstrate that PCP reproduces DFT energy trends under lattice strain and nonequilibrium defect environments. This work may provide a compact and interpretable descriptor framework for machine learning potentials of complex Li-ion cathode materials.