Figure 27.

Using Gaussian process regression (GPR) to learn the Lennard-Jones potential (same as in Figure 25). Here, we trained two different GPR models: First, on the same 80 points we used for Figure 25, and then one for a bad training set with “holes”, i.e., areas from which we did not sample training points. Again, we tuned the hyperparameters of the kernel and then predicted for all points. We can observe that, similar to our KRR results, our model cannot predict the strong repulsion due to the lack of training points. But, in contrast to the KRR, the GPR gives us an estimate for the uncertainty that is larger when we lack examples in a particular region.