Editorial Policy

What qualifies as a node

Any published machine-learning interatomic potential with a public code repository is eligible. To merit its own node (versus a description inside a parent entry), a model must introduce a distinct architecture, training dataset, or use case. Minor version bumps or re-trainings are folded into their parent.

Distinct model vs. variant

If a model shares architecture and dataset with an existing entry and differs only in hyperparameters or training budget, we list it as a variant in the parent description — not as its own card. Forks with a genuinely new inductive bias or dataset get their own node.

Review cadence

Every entry carries a lastReviewed date. We aim to re-review each model at least once every 12 months, plus on any reported broken link or significant upstream change. Entries older than 12 months without a review surface in the coverage table below as candidates for refresh.

What “maintained” means

• active — commits, releases, or issue responses in the last 6 months.
• maintained — receives bug fixes and compatibility updates but no new features.
• archived — repository frozen or marked archived upstream.
• experimental — research code, may break; no maintenance commitment.

Required capability fields for new entries

Every new model card must declare three capability fields so that users searching the table or compare views by capability get reliable results:

• supportsCharges — boolean or null. Whether the model can be conditioned on total or atomic charge.
• supportsSpins — boolean or null. Whether the model can be conditioned on spin multiplicity or magnetic moments.
• elementsCovered — string or null. An explicit list (“H, C, N, O”), a shorthand (“all elements up to Z=94”), or “—” when no information is available.

Use null (or “—” forelementsCovered) only when the information is not documented in the model’s paper or repository. Existing model cards without these fields are progressively being back-filled.

Corrections & additions

To report a broken link, stale description, or request a new model, open an issue on the GitHub repository or email support@mliphub.com. See Contribute for the full workflow.