Interpretably learning the critical temperature of superconductors: Electron concentration and feature dimensionality reduction

Abstract

A machine learning model for predicting the critical temperature of novel superconductors is proposed. The novelty of this approach is based solely on the choice of ab initio features, that is, descriptors directly and solely related to the electronic and atomic information of the single elements that chemically bond to form superconductors. We could show that selecting features, such as the electron concentration in the materials and the electronegativity from the available superconductor data, allows for a significant reduction in the learning dimensionality. At the same time, this choice provides a prediction accuracy in critical temperature up to 93% (relevant to a mean absolute error of 4.2 K) similar to more complex models using a significantly higher feature space. In total, the features could be reduced down to 11 in addition to the multidimensional electronic concentration (including 17 features for the atomic orbitals), emphasizing the significantly higher importance of electronegativity and electron concentration. The latter is mainly influenced by the 3s orbital followed by the 3p orbital. This choice is physically intuitive as it directly links to the electronic orbitals in the superconductors that mostly influence the learning, thus defining the accuracy of the predicted critical temperature. At the same time, the learning process is interpretable, providing a deep insight that could be proven invaluable in developing new theories on novel superconductor materials.