Anomaly-Detection-Driven Screening of Thermodynamic Stability from Composition Descriptors Alone

Materials informatics tends to rely on existing structural database searches that constrain exploration by omitting unregistered compositions. In this study, an autoencoder-based anomaly detector was developed using composition-only descriptors as input features. The model was trained on thermodynamically stable phases─defined as those on the convex hull with an energy above hull (Δ<i>E</i>_<i>hull</i>) of 0 eV/atom─as well as nearly stable phases with Δ<i>E</i>_<i>hull</i> < 0.01 eV/atom, sourced from the Materials Project inorganic database. The reconstruction error (RMSE) was used as the anomaly score. It was shown that the RMSE increased systematically with apparent thermal destabilization─that is, increasing energy above hull. It was also shown that for 50,000 dummy oxides with an intentionally perturbed charge balance, the RMSE increased in proportion to the magnitude of the total charge imbalance, indicating that departures from charge neutrality could be captured even without explicit charge information. Feature-importance analysis suggested that element pairs (two-element combinations) were the principal factors governing the reconstruction RMSE. Accordingly, for 50,000 charge-compensated virtual oxides, we ordinally encoded the valence-shell type as s = 1, p = 2, d = 3, and f = 4 and defined, for each element pair, a coarse indicator given by the product of the two codes (hereafter, the spdf product). For each pair, we evaluated the correspondence between the median RMSE (taken over all compositions containing that pair) and its spdf product and obtained an approximately monotonic relationship─that is, pairs with smaller spdf products tended to have a lower median RMSE, whereas those with larger spdf products tended to have higher values. By contrast, pairs containing Ta consistently deviated downward from this relationship (i.e., exhibited a lower RMSE), suggesting that not only the spdf product but other descriptor information could also influence the assessment of synthesizability.