Domain-specific dataset enable accurate MLIPs for disordered halide-based solid electrolytes

Abstract Efficient exploration of chemical space with machine-learning interatomic potentials (MLIPs) requires a comprehensive, application-focused training database. Solid-state electrolytes for Li-ion batteries have attracted growing attention owing to their ability to improve safety and energy density. Among these, halide-based solid electrolytes have emerged as promising candidates because of their high ionic conductivity and stability. We present an MLIP training database for Li 3 MX 6 -type halide electrolytes, where M can be group-3, group-13, and group-15 elements as well as lanthanides, and X represents halogens. The dataset covers a broad chemical space generated by varying crystallographic space groups, introducing isovalent and aliovalent doping, and including anti-site defects. It contains approximately 400k density functional theory calculated structures with energies, forces, and stresses, together with hierarchical metadata, providing a resource to accelerate the discovery and optimization of halide electrolytes. We demonstrate the utility of the database by fine-tuning the MACE-MP-0b2-large foundation model: the energy root-mean-square error (RMSE) decreased from 154.8 to 10.0 meV atom − 1 and the all-force-component RMSE from 168.3 to 39.0 meV Å − 1 on a held-out test set, confirming that domain-specific training data is necessary for structurally disordered halide configurations. We further validate the fine-tuned model against lattice parameters, room-temperature ionic conductivities, and Li-ion migration activation energies from literature for two representative compositions Li 3 YCl 6 and Li 3 YCl 5 F , demonstrating utility beyond per-configuration error metrics.