Lithium-ion and sodium-ion batteries underpin portable electronics, electric vehicles, and grid-scale energy storage. Their performance and safety are strongly constrained by cathode behavior, and practical screening repeatedly requires three types of predictions, i.e., voltage profiles during (de)intercalation, alkali-ion migration kinetics, and thermal stability against oxygen release and decomposition. Density functional theory (DFT) can provide accurate baselines, but its cost limits supercell sizes, configuration sampling, and the time/length scales accessible to ab initio molecular dynamics (AIMD). Pretrained machine-learning interatomic potentials (MLIPs) offer near-DFT accuracy at far lower cost, thus enabling high-throughput evaluation and long-time molecular dynamics. However, the fidelity required for voltage thermodynamics, migration kinetics, and high-temperature decomposition is not identical, and a model that performs well on one task may be unreliable on another. In this study, eight representative pretrained MLIPs (i.e., MACE, M3GNet, CHGNet, ALIGNN, SevenNet, AlphaNet, GRACE, and Orb) were systematically benchmarked across these three cathode-relevant tasks under matched protocols.

All crystallographic prototypes were obtained from the Materials Project. All the MLIPs were used as released pretrained models without fine-tuning to reflect a "drop-in" screening scenario. For voltage profiles, four representative cathodes spanning layered oxides and polyanion frameworks were selected, i.e., NaMnO₂, LiCoO₂, NaFePO₄, and Na₃V₂(PO₄)₃. Each parent structure was expanded to a supercell containing 8-12 alkali sites. Partially (de)intercalated configurations were generated by exhaustive enumeration of alkali-vacancy patterns and symmetry-deduplicated with a space-group analysis tool. Each MLIP evaluated energies for all configurations. Formation energies were referenced to alkali-rich/poor endmembers, and lower convex hulls were constructed to obtain piecewise average voltages. Accuracy and curve-shape fidelity were quantified using metrics including mean absolute error (MAE), root mean square error (RMSE), integrated absolute error (IAE), and bias indicators.

For ion migration, Li₃Sc₂(PO₄)₃ was chosen as a representative fast-ion conductor. Molecular dynamics was performed at 600 K for 2 ns with each MLIP, and the Li self-diffusion coefficient was extracted from the mean-square displacement (MSD) via the Einstein relation in three dimensions. A fixed regression window (200-1500 ps) was applied consistently to reduce transient and finite-sampling effects.

For thermal stability, oxygen-release/decomposition was benchmarked on phosphate systems with clear stepwise pathways, i.e., MnPO₄ and FePO₄ in olivine and trigonal polymorphs. Each step was represented by balanced solid-state reactions with explicit O₂ release. Reaction energies from MLIPs were combined with an ideal-gas oxygen chemical potential in an oxygen grand-potential framework to estimate onset temperatures and cumulative O₂-evolution curves.

The benchmark reveals a pronounced task dependence among pretrained MLIPs. For voltage profiles, most models reproduce the qualitative multi-plateau behavior, yet quantitative accuracy differs because voltages are sensitive to small energy differences among near-degenerate vacancy orderings that determine convex-hull vertices and plateau boundaries. Across the four cathodes, M3GNet achieves the minimum overall voltage error with a small systematic bias, followed by AlphaNet and CHGNet. GRACE is generally mid-tier, whereas Orb and ALIGNN exhibit larger offsets and shape errors. A notable system dependence occurs for LiCoO₂, where most MLIPs tend to underestimate voltages relative to the reference, indicating that layered Co-O chemistry remains a challenging regime for the existing universal pretrained models.

For ion migration in Li₃Sc₂(PO₄)₃ at 600 K, the ranking differs from the voltage benchmark. MACE and SevenNet yield diffusivities closest to the reference, with Orb and GRACE moderately higher. CHGNet and M3GNet overestimate a diffusivity by severalfold, while AlphaNet underestimates it more strongly. Importantly, ALIGNN can not reliably complete the long-timescale molecular-dynamics runs under our resource constraints, highlighting a practical limitation (memory footprint) that can matter in screening pipelines. These support the interpretation that diffusion is governed primarily by local force accuracy and barrier-region smoothness along hopping pathways, whereas voltage profiles depend on consistent ranking of many competing configurations across composition.

For thermal stability, all the MLIPs qualitatively reproduce the existence of stepwise oxygen evolution and pyrophosphate-forming decomposition, but predicted onset temperatures can deviate substantially. For MnPO₄, deviations span a wide range. ALIGNN and M3GNet provide the smallest deviations for this system, while AlphaNet and CHGNet delay oxygen release. Across the full phosphate set, M3GNet and MACE deliver the optimum overall agreement with experimental step positions and temperatures, and GRACE often captures intermediate steps with moderate errors. The results indicate that generalization to high-temperature, oxygen-deficient pathways remains sensitive to both training coverage and architectural bias.

This work established a unified, extensible benchmark for comparing pretrained MLIPs on cathode-relevant thermodynamics, kinetics, and high-temperature stability. No single MLIP could be universally optimal. M3GNet was most reliable for voltage prediction on the tested cathodes, MACE and SevenNet were strongest for long-time MSD-based diffusion in Li₃Sc₂(PO₄)₃, and M3GNet together with MACE best captured oxygen-release/decomposition trends in representative phosphate systems. The proposed workflow could provide a practical guidance for task-aware MLIP selection and a foundation for larger, more decoupled benchmarks in future studies.