Fig. 2

Schematic illustrating the ability (or inability) to extrapolate predictions far from the training set of a traditional blind machine learning (trained based on arbitrary descriptors α) and b topology-informed machine learning (trained based on topological descriptors β). In both panels, the dashed red curve represents the true function relating the inputs to the targeted output. The squares indicate the known points from the training set. The solid green curve represents the “guessed” function interpolated by the ML model. The gray window indicates a range of systems (i.e., specific values of descriptors α) that is not represented within the training set and for the predictions from the ML models are tested. Note that this window is outside the training set in a, but not in b—since several systems with different descriptors β may present the same topology