{
"access": {
"files": "public",
"record": "public"
},
"files": {
"enabled": true,
"default_preview": "README.md"
},
"custom_fields": {
"dsmd": []
},
"metadata": {
"title": "Goldilocks CSLR quantile forest for k-point mesh index (PSDI 52713-55d86)",
"description": "<p>A quantile random forest that predicts which mesh a periodic material needs from Goldilocks Core's ordered ladder of k-point meshes. It returns a rung on that ladder, not a k-point spacing: rung 1 is the Gamma-only (1, 1, 1) mesh, indices are 1-based, and the ladder was enumerated down to a minimum k-distance of 0.03 inverse angstrom rather than a fixed count per axis.</p><p>Trained on PSDI record 52713-55d86: the same 17757 MC3D structures as the superseded record d5ds2-64f16, relabelled under the corrected floor-based ladder -- not a uniform shift, since the 1-based enumeration also drops a small number of repeated meshes that the old per-axis count did not. Convergence is judged on total energy alone, to within 1 meV/atom across three consecutive rungs; no force criterion is applied. Split 70/10/10/10 and grouped by reduced composition so that polymorphs of one composition cannot straddle the split. Inputs are 174 composition, structure, direct-lattice and reciprocal-lattice descriptors, matching the extractor Goldilocks Core already uses. No SOAP and no learned metallicity representation.</p><p>The model no longer publishes one quantile for every structure. A single shared level means every structure pays the safety margin the hardest ones need, or the hardest ones are denied the accuracy the easy ones keep -- measured on the previous q0.6 policy, under-prediction reached 76.7% for structures that truly need rung 12 or above while the bulk of the dataset (rung < 7) already sat at 12.1%. This version cuts structures into bands by the model's own median (at rung 7 and 12, the same cut used for coverage below) and picks a separate published level per band -- the cheapest one keeping that band's own under-prediction at or below 5%, chosen on the validation split. The bands land on q0.9 below rung 7 and q0.975 at or above it. On the held-out test split, under-prediction falls to 4.2% overall and to 14.7% for structures that truly need rung 12 or above (from 76.7% under q0.6), at the cost of mean excess rising to +2.78 rungs.</p><p>Because the published rung is deliberately conservative, average-error metrics describe the cost of that choice rather than the accuracy of the estimate; the full metric suite behind this trade is recorded in the training run in the source repository (stfc/goldilocks-ml) for anyone auditing it in detail. Bands are cut on the model's own median rather than the true difficulty, so a structure that turns out harder than the model expects can still be served a lower band's level -- on the held-out test split this still affects 14.7% of the structures that genuinely need rung 12 or above, well down from earlier policies but not eliminated. Very small, simple cells can also be recommended a denser mesh than common practice calls for; Goldilocks Core applies an additional ceiling on mesh density based on the structure's metallicity to keep this bounded. For a very small number of simple, high-symmetry cells, the published rung can exceed the number of meshes that exist on that specific structure's own ladder -- Goldilocks Core clamps to the densest mesh available for that structure, and a consumer using this record directly should apply the same check.</p><p>The record contains the fitted forest, its separately recorded calibration, and <code>model.json</code>, which was written by the run that fitted the model and records the feature contract, all 174 column names, the target contract, the decision rule and the rule that chose it, the hyperparameter search, and the digests.</p>",
"creators": [
{
"affiliations": [
{
"name": "UKRI STFC"
}
],
"person_or_org": {
"family_name": "Yin",
"given_name": "Junwen",
"identifiers": [
{
"identifier": "0000-0001-7374-9352",
"scheme": "orcid"
}
],
"name": "Yin, Junwen",
"type": "personal"
}
}
],
"contributors": [
{
"person_or_org": {
"type": "personal",
"name": "Elena, Alin-Marin",
"given_name": "Alin-Marin",
"family_name": "Elena"
},
"role": {
"id": "projectmanager"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
},
{
"person_or_org": {
"type": "personal",
"name": "Yin, Junwen",
"given_name": "Junwen",
"family_name": "Yin"
},
"role": {
"id": "projectmember"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
},
{
"person_or_org": {
"type": "personal",
"name": "Basak, Susmita",
"given_name": "Susmita",
"family_name": "Basak"
},
"role": {
"id": "projectmember"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
},
{
"person_or_org": {
"type": "personal",
"name": "Cha, Jaehoon",
"given_name": "Jaehoon",
"family_name": "Cha"
},
"role": {
"id": "projectmember"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
},
{
"person_or_org": {
"type": "personal",
"name": "Teobaldi, Gilberto",
"given_name": "Gilberto",
"family_name": "Teobaldi"
},
"role": {
"id": "projectmember"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
},
{
"person_or_org": {
"type": "personal",
"name": "Sparks, Willow",
"given_name": "Willow",
"family_name": "Sparks"
},
"role": {
"id": "projectmember"
},
"affiliations": [
{
"name": "UKRI STFC"
}
]
}
],
"rights": [
{
"id": "cc-by-4.0"
}
],
"resource_type": {
"id": "model"
},
"subjects": [
{
"subject": "Materials science"
},
{
"subject": "Density functional theory"
},
{
"subject": "k-point sampling"
},
{
"subject": "Quantile regression"
},
{
"subject": "Random forests"
}
],
"publisher": "PSDI",
"publication_date": "2026-09-13",
"version": "v4.0"
}
}