Skip to content

mMACE + MLP magnetism classifier

Not a default model

External validation on an independent dataset has not been run yet. See "How good it is" below.

Answers one question: is this structure's DFT ground state spin-polarised? Goldilocks needs the answer early, the same way it needs metallicity early — a magnetic structure needs nspin=2 and a starting magnetisation, and several other inputs depend on that. The frozen mMACE backbone the features come from is bundled with this record, licensed CC-BY-4.0 by the training collaborator.
Predicts is_magnetic — magnetic or non_magnetic
Trained on MatPES PBE (materialyze/matpes/pbe-2025.2), not Materials Project
Record 1g8rw-q8128
Source ported from 2-research/2-mace/1-magnetic-mace (outside this ecosystem)

Use it

from goldilocks_ml.inference import load_model

model = load_model("path/to/the/record")
prediction = model.predict(structure)

prediction.value  # True for a spin-polarised ground state

The decision threshold lives in the record and is applied for you. It is 0.324, not 0.5.

How good it is

On an in-distribution MatPES PBE test split (the research checkpoint's own numbers, not yet reproduced by a training run in this repository's own pipeline):
ROC-AUC 0.986
Recall 0.962
Precision 0.926

External validation on an independent dataset has not been run. Treat these numbers as in-distribution only.

The label has a dead zone

A structure's maximum on-site moment magnitude, from MatPES's own DFT calculation, decides the label:

<= 0.05 bohr magnetons  ->  non_magnetic
>= 0.50 bohr magnetons  ->  magnetic
strictly between        ->  excluded from training and test entirely

A prediction for a structure whose true answer would land in that gap is extrapolation, not interpolation — the model was never shown an example like it, in either class.

What is not modelled

ordering (a label: NM/FM/AFM/FiM) and magnetic_moments (one value per site) are not here. The mechanism to declare either as a contract exists (ContractSpec.labels, ContractSpec.index_convention); what is missing is a model. Antiferromagnetic sublattice detection in particular has no implementation anywhere in the project this was ported from to build on — only an FM/ferrimagnetic-only initial-moment guess (goldilocks_ml.models.magnetism.magnetic_moments.fm_fim_relax.seed_moments) and a ported moment relaxation on the same frozen backbone (...fm_fim_relax.relax), collinear only, on a fixed geometry, never a ground-state search.

When to be careful

  • A prediction near the decision threshold, or near the 0.05–0.5 uB dead zone, is not confident. Treat it as unverified and check with a real spin-polarised calculation.
  • The score is not a probability, and 0.5 is not its decision point. Use prediction.value, not your own threshold.
  • External validation is still pending. Do not depend on this as a default model until it is recorded — see the warning at the top of this page.