Figure 5.

Predicted phase diagrams from the hold-out test. These charts show the convex hulls predicted for the (a) Ti-O binary and (b) Na-Mn-O from ML models that were trained without any data from each system in their training set. We compare the performance of a Random Forest model trained using only element fractions (RF-Comp), RF trained using physical features (RF-Phys) and a deep learning model (ElemNet). Each vertex on the convex hull corresponds to the composition of a stable compound. The black lines on each chart show the OQMD convex hull. We find that the deep learning model has the fewest predictions outside the regions where compounds are known to form, for both the Ti-O and Na-Mn-O phase diagrams.