Fig. 5: Prediction of maximum property values and similarity with synthetically accessible materials. | npj Computational Materials

Fig. 5: Prediction of maximum property values and similarity with synthetically accessible materials.

From: Element selection for functional materials discovery by integrated machine learning of elemental contributions to properties

Fig. 5

Circled materials are reported in ICSD2, and are not found in SuperCon-v20183,28 or MPDS1. a Unexplored ternary phase fields with >70% probability of superconductivity at transition temperature Tc > 10 K and normalised reconstruction error < 0.2. b Unexplored ternary phase fields with >75% probability of energy bandgap > 4.5 eV, and reconstruction error < 0.1 c Unexplored ternary phase fields with >71% probability of Curie temperature TC > 300 K and reconstruction error < 0.1. d Receiver operating characteristics (ROC) of the classification models trained and tested on the random 80-20% train-test splits demonstrate high sensitivity and specificity of classifications for the range of thresholds of probabilities. The corresponding areas under the curves (AUC) indicate excellent performance for magnetic materials and good performance for superconducting transition temperature and energy gap classifications. PhaseSelect considerably increases AUC for all datasets in comparison to default Random Forest classifiers trained on the same data. Inset: distributions of 105995 ternary phase fields with respect to reconstruction errors for all three datasets.

Back to article page