Fig. 3: Universal potentials are more reliable classifiers because they exit the red triangle earliest.
From: A framework to evaluate machine learning crystal stability predictions
The lines represent rolling MAE on the WBM test set as a function of distance to the MP training set convex hull. The red ‘triangle of peril’ indicates regions where the mean error exceeds the distance to the stability threshold (0 eV). Within this triangle, models are more likely to misclassify materials as the errors can flip classifications. Earlier exit from the triangle correlates with fewer false positives (right side) or false negatives (left side). The width of the ‘rolling window’ indicates the range over which prediction errors were averaged.
