Fig. 7
Atomization energy prediction errors on the 784,875 converged molecules from the VQM24 dataset. All predictions were made by KRR models employing the cMBDF94,95 representation after training on 200k molecules. Errors for all molecules were obtained by training four such models with disjointed training sets and using the smallest prediction error for all out-of-sample molecules. Mean absolute error (MAE) for the predictions is shown in the top right inset along with the standard deviation. Insets show molecules with the largest prediction errors. Atomic colours correspond to: grey-Carbon, white-Hydrogen, blue-Nitrogen, red-Oxygen, dark green-Fluorine, cream-Silicon, orange-Phosphorous, yellow-Sulfur, green-Chlorine, dark red-Bromine.
