Fig. 4: Performance of predicted formation energies on stability classification.
From: A critical examination of compound stability predictions from machine-learned formation energies
Classification of materials as stable (ΔHd ≤ 0) or unstable (ΔHd > 0) using each of the six compositional models. “Correct” predictions are those for which the ML models and MP both predict a given material to be either stable or unstable. The histograms are binned every 5 meV/atom with respect to ΔHd,MP to indicate how the correct and incorrect predictions and the number of compounds in our dataset vary as a function of the magnitude above or below the convex hull. Acc is the classification accuracy. F1 is the harmonic mean of precision and recall. FPR is the false positive rate. The moving average of the accuracy (computed within 20 meV/atom intervals) as a function of ΔHd,MP is shown as a blue line (right axis). As expected, the accuracy is lowest near the chosen stability threshold of ΔHd,MP = 0.
