Goldilocks CGCNN metallicity classifier model(Matbench mp_is_metal)

A crystal graph convolutional neural network that predicts whether a periodic material has a zero DFT band gap. Goldilocks uses metallicity to decide how densely reciprocal space must be sampled and whether smearing applies.

Trained on the Matbench mp_is_metal task: 106113 Materials Project structures, split 70/10/10/10 by reduced composition so that polymorphs of one composition cannot straddle the split, and stratified by label. Training stops on validation ROC-AUC rather than loss. On the 10625-structure test split it reaches 0.748 accuracy, 0.569 Matthews correlation and 0.951 ROC-AUC, against 0.544 accuracy and 0.000 MCC for a majority-class baseline.

The decision threshold is 0.0478, not 0.5. It is chosen on the validation split subject to a floor of 0.97 recall, because missing a metal understates the mesh a calculation needs and produces a wrong answer that does not look wrong, while an unnecessary dense mesh only costs compute. Test recall is 0.972 and test precision 0.649: roughly a third of what this model calls metallic is not, and that is a deliberate trade.

The record contains the weights and model.json, which records the architecture, the feature contract, the target contract, the threshold and the rule that chose it, the digests, and the atomic embedding table it requires from record m742g-g0k14. It was written by the run that fitted the model.