Skip to content

Machine-accelerated extraction of experimental superconducting parameters for physical insight and prediction across thousands of materials

Mihir Manium, Benjamin Geisler, Peter J. Hirschfeld, Amalia I. Coldea

cond-mat.supr-conarXiv:2611.15627

Abstract

New superconductors could unlock next-generation technologies from nuclear fusion to lossless electronics. Machine learning can predict candidate superconductors but lacks validated experimental training data beyond the critical temperature. Here we present AutoSuperCon, a database of eight superconducting parameters, containing 20354 entries for 6793 superconductors. The data were extracted from over 47000 arXiv preprints using a large language model and validated against reference datasets (R2 = 0.96 with NIMS SuperCon). Using AutoSuperCon, we show that median gap ratios rise from 3.6 in conventional superconductors to 6.4 in cuprates, 28% of upper critical fields exceed the Pauli limit, and critical temperatures lie below a ceiling proportional to the superfluid density. Further, we train composition-based predictors for the critical temperature (R2=0.83) and the upper critical field (R2=0.69), boosted by transfer learning. Within known superconducting families, composition alone is sufficient to screen candidates for both parameters. AutoSuperCon thus enables physical insight and data-driven discovery across thousands of superconductors.

Create a lesson