Fig. 4: GAP and MTP predictions for the transition pathway.

Since there were no configurations in the training set along the middle of the path, the MTP prediction is purely extrapolative and presents a false, deep, local minimum. Because of its tight regularization, the GAP prediction for this transition is reasonable. It underestimates the barrier significantly due to the lack of training data, but does not introduce a false minimum. Note the symmetric log scale (linear between −1 and +1) on the y-axis in the upper plot. Above the plots are configurations along the pathway. The upper Pd atom (blue) at index 1 moves further into the page as the unit cell distorts. The furthest Ag atom (red) simultaneously moves out of the page. The lowest panel plots the expected variance of the GAP model and the extrapolation grade of the MTP model. Predictive variance is essentially the expected error per atom (in eV) along the transition pathway and the extrapolation grade is a dimensionless quantity indicating that if it exceeds ~10, then a configuration must be added to the training set because the prediction of its energy is not reliable.