Elucidation of Mg-Ion Diffusion Factors in Mg–O-Based Materials Using Machine-Learning Interatomic Potentials and Causal Analysis

Mg-ion batteries are promising next-generation energy-storage systems because Mg metal offers high volumetric capacity, abundant resources, and intrinsic safety. However, Mg2+ usually diffuses slowly in solids because of its strong interaction with surrounding anion frameworks and cation environments. Here, Mg–O-based candidate materials extracted from The Materials Project were investigated to clarify the compositional and structural factors governing Mg-ion diffusion. Representative materials were selected by clustering composition descriptors and structural descriptors derived from radial distribution functions (RDFs) and angular distribution functions (ADFs). High-throughput molecular dynamics simulations using an M3GNet-based machine-learning interatomic potential were then performed to construct a Mg-ion diffusion-coefficient data set. LightGBM regression, SHAP-based feature importance analysis, and nonlinear NOTEARS causal discovery were integrated to analyze the hierarchical relationships among descriptors and log D. The inferred causal structure indicates that Mg-ion diffusion is not governed by a single compositional factor. Instead, composition affects the short- and medium-range order of the oxygen framework, which regulates the local Mg environment and migration-pathway connectivity. Excessively close-packed and ordered oxygen frameworks restrict Mg-ion migration, whereas moderately disordered oxygen frameworks can promote continuous migration-accessible sites around Mg ions. These results suggest that Mg-ion-conducting oxides should be designed by controlling the oxygen framework and Mg local environment through host-cation selection, particularly using lightweight host materials containing d-block cations with sizes similar to that of Mg.