Predicts whether a periodic material's DFT ground state is spin-polarised —
magnetic or non_magnetic. Goldilocks needs this 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.
A frozen mMACE foundation model (mmace-matpes-pbe-pretrained.model) turns a structure into a 384-number embedding from one non-SCF forward pass. A small MLP (is_magnetic.pt) reads that embedding and predicts magnetic or non-magnetic.
The backbone runs on mace-torch[magnetic]==0.3.17 from PyPI. Version 2.0 of
this record holds the same weights as version 1.0's
mace_matpes_pbe_baseline_run-3.model, re-exported by the training
collaborator for upstream mace and converted for sphericart-torch 2. It gives
the same decisions as version 1.0: embeddings and scores are bit-identical on
13,156 MatPES structures. Upstream mace caps a moment above the element's
learned maximum, so a moment relaxation that strays past it can end
differently than on version 1.0.
MatPES PBE (materialyze/matpes/pbe-2025.2), not Materials Project. Labels
come from the dataset's own DFT magnetic moments: <=0.05 uB is non-magnetic.
from goldilocks_ml.inference import load_model
model = load_model(
"path/to/the/record",
artifacts={"mace_backbone": "path/to/the/record/mmace-matpes-pbe-pretrained.model"},
)
prediction = model.predict(structure)
prediction.value # True for a spin-polarised ground state
The decision threshold -- 0.324, not 0.5 -- is applied for you. The backbone
is a whole pickled torch module, so loading it, through load_model or
directly, executes pickle code. Verify both files against the SHA-256 digests
in manifest.json first.
ROC-AUC 0.986, recall 0.962, precision 0.926, on a held-out MatPES PBE test split. External validation on an independent dataset hasn't been run yet -- treat these numbers as in-distribution only, and treat a prediction near the threshold, or near the 0.05-0.5 uB gap, as unverified.