Material informatics for layered high-TC superconductors

Abstract

Superconductors were typically discovered by trial-and-error aided by the knowledge and intuition of individual researchers. In this work, using materials informatics aided by machine learning (ML), we build an ML model of superconductors, which is based on several material descriptors with apparent physical meanings to efficiently predict critical superconducting temperature TC. The descriptors include the average atomic mass of a compound, the average number of electrons in an unfilled shell, the average ground state atomic magnetic moments, the maximum difference of electronegativity, etc. To fully optimize the ML model, we develop a multi-step learning and multi-algorithm cross-verification approach. For known high TC superconductors, our ML model predicts excellent TC values with over 92% confidence. When the ML model is applied to about 2500 layered materials in the inorganic crystal structure database, 25 of them are predicted to be superconductors not known before, including 12 cuprates, 7 iron-based crystals, and 6 others, with TC ranging from ∼32 K to ∼138 K. The findings shed considerable light on the mapping between the material descriptors and TC for layered superconductors. The ML calculates that in our descriptors, the maximum difference of electronegativity is the most important one.