Fig. 6

Toy example of the active learning sampling evaluation approach, illustrating the usage an artificially OOD region of molecular structure space to test the generalization performance of the sampling. The full distribution of PCA component values (Upper Left) is binned and one bin is removed from the model training data (Upper Right). The remaining data (Middle Left) is sampled to augment the initial training set, either randomly (Middle Right) or based on UQ estimates (Lower)