Machine Learning that predicts well may not learn the correct physical descriptions of glassy systems.*

Machine Learning (ML) has drawn attention as a possible way for gaining new insights into complex physical systems. However, it is not understood whether or not ML methods learn the correct physical description. We answer this question in the context of the application of Support Vector Machines (SVMs) to modeling glassy dynamics. SVM was one of the first methods to show the correlation between local structure and rearrangement dynamics. It was previously found that an SVM model output, the distance from the separating hyperplane (softness), predicted rearrangements and, surprisingly, was linearly related to the activation energy for the rearrangement process. We introduce a simple model for the rearrangement process and study the ability of SVM to learn the known activation energy in this model. We show that the distance from the separating hyperplane is linearly related to this energy, but the SVM inference of energy is strongly biased for finite data, despite high cross-validation accuracy. We also show the linear relationship between softness and energy is dependent on the representation of the structural features. Thus, SVM does not necessarily learn the correct energy even when it seems that it does or should.