Skip to content

Materials MLIP candidate measurement v2

  • Stability: frozen
  • Revision: v2-frozen
  • Scope: atom- and graph-cost-matched OMat24 training environments measured under the frozen Nequix-MP-1 starting potential before post-training

This convention projects the frozen MLIP phonon training panel into Anarres environment profiles. The native final-panel artifact remains authoritative. Profiles contain measurements made before candidate training and exclude every development or prospective phonon outcome.

The measurement contract uses one deterministic starting-model evaluation per retained structure. Every candidate contains exactly 50,000 processed atoms, 2,100,000 Nequix neighbor edges, and at most one configuration from each OMat24 parent. Candidate membership is identified by the SHA-256 digest of its ordered native structure identities.

signal_density_per_atom

Loss-weighted starting-model residual mass divided by processed atoms. For one structure with atom count $n$, mean force-vector residual $r_f$, and mean absolute stress residual $r_s$, the native mass is

$$ A = 20 n r_f + 5 r_s. $$

The candidate coordinate is $\sum A / \sum n$. Its unit is loss_weighted_residual_per_atom, its estimator is the atom-normalized sum over retained structures, and its direction is contextual. The coefficients belong to the frozen measurement contract and do not make force and stress residuals physically interchangeable.

effective_signal_coverage

Normalized effective sample size of the structure-level signal masses. For $m$ retained structures with masses $A_i$,

$$ H = \frac{(\sum_i A_i)^2}{m\sum_i A_i^2}. $$

The value lies between zero and one, with zero defined when total mass is zero. Its unit is fraction, its estimator is the normalized effective sample size, and its direction remains contextual until downstream outcomes validate a decision rule.

chemical_system_entropy

Normalized Shannon entropy over the retained structures' element-set labels. The unweighted label frequency is divided by the log of the number of observed labels. The value is zero for one observed label and lies between zero and one otherwise. Its unit is fraction, and its direction is contextual.

mean_target_affinity

Arithmetic mean of each retained structure's maximum element-set Jaccard similarity to the 16 checkpoint-qualification targets. Only element identities from the qualification split enter this measurement. Development and prospective phonon outcomes did not enter its construction. Its unit is fraction, and its direction is contextual.

neighbor_edges_per_atom

Total Nequix neighbor edges divided by processed atoms. Every frozen candidate equals 42 edges per atom. This is a model- and cutoff-specific compute proxy, not a universal hardware-cost estimate. Its unit is neighbor_edges_per_atom, and its direction is contextual within the matched panel.

processed_atom_budget

Total atom instances in the candidate environment. Every frozen candidate equals 50,000 processed atoms. Its unit is processed_atoms, and its direction is contextual within the matched panel.

Qualification and edge cases

The adapter requires the pinned OMat24 source qualification, Nequix-MP-1 model hash, frozen Alexandria target roles, overlap audit, complete 80,000-record measurement validation, and frozen cost-matched panel. Missing or nonfinite measurements fail export. Zero measurements are preserved.

The panel contains two candidates from each of eight physical source families. All six outcome-blind selectors appear two or three times. Twelve candidates form the development set, while four candidates from distinct source families and selectors remain sealed for prospective evaluation. These roles were assigned from pre-training coordinates without candidate-training outcomes.