Method & caveats

For every (potential × elemental cubic crystal) the campaign relaxes the bulk to the model's own equilibrium a₀ (from its EOS), enumerates the deterministic CSL grain-boundary set at that a₀, and runs a full grain-boundary study on each boundary: grain-length optimisation along the boundary normal, a +20 Å-vacuum relaxation (→ GB-plane surface energy), and a rigid + relaxed cleavage scan across the boundary planes (→ work of separation). One row per boundary carries the GB energy (J/m²), excess volume, surface energy and W_sep; per (potential, structure) they are aggregated over the structure's converged boundaries. Each study is ~30 MLIP relaxations, so cost is dominated by cell size and boundary hardness, not a single forward pass.

There is no DFT grain-boundary reference (unlike the surface-energy track). Accuracy is therefore the cross-model Consensus dev: per (structure, boundary) the reference is the median GB energy over the 15 curated KEEP potentials (only where ≥ 5 KEEP have a converged value), and each of the 51 potentials is scored by the median-over-structures of its per-structure mean |ΔE_gb| from that consensus. It ranks "most typical", not "most correct" — 11 of the 15 KEEP models are GRACE-family, so the consensus leans GRACE, and a physically-correct outlier is penalised. It is also coverage-confounded: a model that only ran a handful of easy low-Σ boundaries scores a deceptively small deviation, which is why the leaderboard defaults to sorting by coverage and why the nequip_OAM-L checkpoint (100 % terminal clamp but 0 usable rows — AOTI-broken on B200) is flagged. The coverage column is the clamp metric over the 36 industrial structures (Σ min(distinct_terminal, n_gb_total) / Σ n_gb_total, bounded ≤ 100 %). The curated deliverable is 15 KEEP potentials × 36 industrial structures; toggle the scope off to see all 51 potentials and all structures, where coverage is uneven and the consensus deviation is no longer comparable across models. GB-plane γ is the free surface exposed by cleaving the boundary slab — not the clean low-index facets of the surface-energy track; do not cross-compare. Full protocol, the CSL enumeration, the clamp coverage metric and the consensus definition: Grain-boundary methodology →.