Reliability of foundation-dataset ML interatomic potentials as substitutes for ab initio in complex materials screening

Determine whether machine-learning interatomic potentials trained on broad r^2SCAN foundational datasets—specifically datasets such as MatPES and MP-ALOE—can reliably substitute for ab initio density functional theory calculations during high-throughput screening of specific, complex material classes, including dynamically soft systems such as halide solid-state electrolytes under elevated-temperature, highly distorted regimes.

Background

Foundational r2SCAN datasets like MatPES and MP-ALOE were recently created to provide consistent, off-equilibrium coverage for training universal ML interatomic potentials. These datasets aim to overcome issues in earlier resources (e.g., PBE-only, inconsistent settings) and enable general-purpose models.

However, dynamically soft materials such as halide solid-state electrolytes exhibit shallow, anharmonic potential energy basins and require elevated-temperature simulations to probe ion transport. It remains uncertain whether models trained only on broad, general-purpose datasets provide sufficient fidelity and stability to replace ab initio DFT in high-throughput workflows for such complex classes.

References

Recently, foundational r$2$SCAN-level datasets containing off-equilibrium states performed with consistent and higher fidelity electronic settings, such as MatPES and MP-ALOE, have been developed to address these issues and enable the development of highly capable universal ML force fields. However, it remains unclear whether models trained on these broad, general-purpose datasets can reliably substitute for ab initio methods during high-throughput screening of specific, complex material classes.

— AQVolt26: High-Temperature r$^2$SCAN Halide Dataset for Universal ML Potentials and Solid-State Batteries  (2604.02524 - Kim et al., 2 Apr 2026) in Introduction