Diatomics

Disclaimer: Take the results on this page with a grain of salt. We checked most data but not as carefully yet as the discovery task.

The PBE/r2SCAN DFT reference curves shown below are documented in the DFT reference provenance note, including the VASP setup, spin-ladder generation, warm-start strategy, postprocessing, and quality checks. The TMI page shows each reference curve’s per-atom magnetic moments vs separation, useful for spotting residual SCF spin-state artifacts and for benchmarking spin-aware MLIPs. Expert advice on how the DFT reference curves might be improved would be appreciated (thanks to Andrew Rosen for guidance on obtaining correct dimer PECs)!

Metrics in the table below aggregate over the same element set for all models: the 87 Materials Project elements between H and U. Po, At, Rn, Fr, Ra are excluded since they’re absent from MP and hence unsupported by models trained on MP-derived data (MPtrj, OMat24, sAlex) — only a minority of models predicts them. Self-consistency metrics (smoothness, flips, jumps) average over all 87 elements. Reference-relative PBE errors skip 8 further lanthanides (Pr, Pm, Sm, Tb, Dy, Ho, Er, Tm) whose reference curves fail a smoothness quality gate (energy jump ≥ 1.5 eV and ≥ 3 energy-difference sign flips inside the scored window — f-electron SCF convergence issues make these references too jumpy to score models against). Curves excluded from metrics are still shown in the plots below. Models missing curves within the scored set (marked *) are scored on their remaining elements pending backfill runs.

With thanks to MLIP Arena by Yuan Chiang, who first benchmarked MLIPs on self-consistency metrics of homonuclear diatomic curves and inspired this task.

We welcome backfill PRs from model authors for models still missing diatomic metrics; see the diatomics runner, calculator registry, and model enum.