The role of feature space in atomistic learning

Abstract Efficient, physically-inspired descriptors of the structure and composition molecules materials play a key role in application machine-learning techniques to atomistic simulations. The proliferation approaches, as well fact that each choice features can lead very different behavior depending on how they are used, e.g. by introducing non-linear kernels non-Euclidean metrics manipulate them, makes it difficult objectively compare methods, address fundamental questions one feature space is related another. In this work we introduce framework sets descriptors, ways transforming them means kernels, terms induce. We define diagnostic tools determine whether alternative spaces contain equivalent amounts information, common information substantially distorted when going from compare, particular, representations built n -body correlations atom density, quantitatively assessing loss associated with use low-order features. also investigate impact choices basis functions hyperparameters widely used SOAP Behler–Parrinello features, Wasserstein-type metric, change comparison simpler linear space.