pescert arXiv preprint

Ground-truth-free evaluation of universal MLIPs. Each probe compares a model against a value the exact Born–Oppenheimer surface satisfies by construction. No DFT reference is used.

Other MLIP benchmarks

These measure accuracy against DFT references and downstream tasks, which is the axis this table does not cover.

Matbench Discovery Crystal stability, geometry optimisation, phonons, MD and diatomic curves against DFT. MLIP Arena Physics awareness, chemical reactivity and stability under extreme conditions. FAIR Chemistry Leaderboard Energy and force errors across the OMol25 and OC20 tasks, from ligand strain to spin gaps. MLIPAudit Molecular and biomolecular systems: conformers, torsions, tautomers, reactivity and MD stability. JARVIS Leaderboard NIST leaderboard with a machine-learning force-field category.

Contact