High-fidelity Off-equilibrium Dataset and Universal Interatomic Potentials for Li-Halide Solid-State Electrolytes

Universal machine-learning interatomic potentials (MLIPs) enable computational screening of solid-state electrolytes, but elevated-temperature accuracy for halides depends on off-equilibrium coverage of their strongly distorted soft lattices. AQVolt26 contains 322,656 consistent r$^2$SCAN calculations derived from 4,911 lithium-halide candidates and more than 200 million molecular-dynamics frames. A controlled eSEN ablation assessed its inclusion with MatPES and MP-ALOE. On a lithium-halide holdout, AQVolt26 reduced energy, force, and stress errors from 187.3 to 21.3 meV/atom, 351.7 to 70.0 meV/$\mathring{A}$, and 23.71 to 14.75 kbar, respectively. Materials Project relaxation data improved near-equilibrium predictions but increased the failure rate under extreme hydrostatic strain from 0.2\% to 12.9\%. The comparison separates the data requirements of near-equilibrium prediction from those of high-temperature dynamics.