Discovery of interpretable descriptors in conventional superconductors guided by symbolic regression
Phys. Rev. Materials 10, 054803 – Published 22 May, 2026
DOI: https://doi.org/10.1103/4pxk-z8ww
Abstract
Advancement of superconductivity is hindered by complexity of underlying mechanisms. Machine learning models have potential to accelerate materials discovery but face data limitations and interpretability challenges. This study identifies material genes governing conventional superconductivity through systematic two-stage analysis, where atomic descriptors are directly related to their phonon-mediated mechanism, unlike unconventional superconductors where strong electronic correlations that go beyond the Bardeen-Cooper-Schrieffer paradigm dominate in the formation of pairing glue. We first apply the random forest classifier and regressor to superconductors from NIMS database, screening over 114 MAGPIE and MEREDIG atomic descriptors to identify the 30 most important features. Combining these atomic descriptors with DFT-calculated bulk properties, we then perform SISSO (sure independence screening and sparsifying operator) symbolic regression on 60 carefully selected conventional superconductors. The resulting three-dimensional model [, AFD (average factor difference) ] reveals optimal requires an average of half filled or near half filled electrons with moderate unfilled orbital heterogeneity. The discovered descriptors integrate -valence electrons, unfilled orbitals, and electronegativity variations, providing actionable guidelines for materials design.
Physics Subject Headings (PhySH)
Collections
This article appears in the following collection:
Machine Learning for Materials Discovery and Understanding
The Editors of Physical Review Materials are pleased to present the Collection on Machine Learning for Materials Discovery and Understanding, highlighting cutting-edge advances in machine learning method development and applications for materials discovery and fundamental understanding of the structure-property-function relationship. The Collection is being guest-edited by Deyu Lu of Brookhaven National Laboratory (USA) and Jinlan Wang of Southeast University (China). Every article published in this collection underwent a rigorous peer review process, adhering to the same high standards applied to all papers. The Physical Review Materials editorial team managed the peer review and made all editorial decisions.
