Method & scope

The overview is precomputed (corpus/summary, one small parquet) so it renders instantly and matches the deployed Elemental PES leaderboard by construction — it reads the unary partition (one clean row per element, pure-element structures only), the same per-element basis the leaderboard uses, with the same keep-filter (status = done, |corrected energy| ≤ 5 eV/atom, max force ≤ 10 eV/Å) and metrics. Force is shown as the per-structure median RMSE (the summary carries force distribution stats, not a per-element force RMSE); energy and stress are RMSE. The energy scheme toggle picks the correction basis: Corrected (systemspin) is the headline per-(system, ISPIN) subset-pooled offset, Raw is uncorrected, Elemental-transfer is the per-element-offset variant. The magnetic state facet splits by the DFT magnetic classification (overview only — the per-structure partition carries no magnetic column, so the drill-in shows all states).

The drill-in is deliberately wider than the overview: it reads every structure tracked to the selected element — its pure-element set and the alloy datasets where it is the tracked (solute) element (Fe–X, Ni–X, carbide, …) — from the selected potential's partition, using the element-clustered parquet so a WHERE element = X predicate prunes to just that element's row groups. It applies the same keep-filter as the overview, projects ~10 columns, downsamples the parity scatter, and SQL-bins the histograms, so the browser never loads the full corpus and the fetch is proportional to the element you pick. Points carry their source dataset, so the scatter distinguishes the pure-element cloud from the alloy datasets. The DFT reference is derived in-table (DFT = MLIP − energy_error); the heavy corpus/structures table is never loaded.

For programmatic / full-corpus access, use the Python client or MCP server over the same lake — see the API and Data pages.