Data Quality
Machine learning models are only as reliable as the labels they learn from, and large high-throughput databases inevitably contain some mislabeled samples, for example from unconverged calculations, metastable solutions or outdated settings. Data quality work covers detecting such anomalies, recomputing or removing them, and the feedback loop in which better data give better models that in turn reveal more bad data.
Embedding-based anomaly detection in Materials Project moments (2025)
Xu and co-workers used atom embeddings from a trained spin-informed GNN to audit magnetic moment labels in the Materials Project trajectory data. Embeddings were clustered with a Gaussian mixture model (group count by BIC) and labels with within-group z-scores beyond ±5 were flagged, on the premise that similar embeddings should carry similar labels. Recomputing 234 Gd/Eu samples with 7 µB starting moments moved many labels from about ±1 µB to about ±7 µB, revealing poorly converged originals; about 0.7% of the Gd/Eu subset and 0.8% of a Cr subset were flagged, suggesting roughly 1% of the data are anomalous. Removing them and retraining markedly improved the fit, and the authors frame this as a closed loop for the “chicken-and-egg” problem of building a dataset while developing a model on it (Xu et al. 2025). Summary: xu2025.