Fig. 2. Learning curves for the lattice energy predictions of pentacene, 5A and 5B datasets on a logarithmic scale. All hyper-parameters of our ML model are fixed except for the regularization parameter σ_n in the GPR model which is optimized on the fly at each training. We use 4-fold cross validation on the randomly shuffled dataset and randomly draw N times an increasing number of training samples from 75% of the dataset for each fold. The test MAE and error bars are, respectively, average and standard deviation over the folds. The left-hand panel corresponds to the prediction of W99 energies computed for W99-optimized geometries, the middle panel correspond to the prediction of DFT energies on such structures, and the right-hand panel to the prediction of the difference between DFT and a W99 baseline.