Machine Learning Prediction of BCS Superconductors without BCS Theory

We show that machine learning (ML) approaches can expedite the future discoveries of BCS superconductors, providing a faster computationally inexpensive alternative to density functional theory for estimating whether a material is a BCS superconductor. The key to the technological applications of superconductors is the superconducting critical temperature. Although the electron pairing mechanism for unconventional superconductors is not known, the pairing mechanism for BCS superconductors is well understood by the Eliashberg function and the McMillan equation. Nevertheless, the first-principles calculations needed are exceedingly expensive. For this reason, a rapid screening of candidate BCS superconductors and their T _c using first-principles calculations is prohibitive. We leverage machine learning to eliminate the need for first-principles calculations when predicting whether or not a material is a BCS superconductor. A database of experimentally relevant crystal structure data is carefully curated and materials descriptors suitable for describing the experimental data are chosen. Machine learning models are trained to classify if a material is a BCS superconductor. This framework provides a roadmap for accelerating the discovery of novel BCS superconductors without BCS theory.