Machine learning-accelerated search of superconductors in B-C-N based compounds and R3Ni2O7-type nickelates

Abstract

Superconductor research has traditionally depended on experiments and theoretical approaches. However, the rapid advancement of data-driven methods and machine learning (ML) has opened avenues for accelerating superconductor discovery. Here, we integrated ML with density functional theory (DFT) calculations to efficiently screen conventional B-C-N based superconductors and identify potential high-TC candidates among R3Ni2O7-type bilayer nickelates. We identified 12 new binary and ternary B-C-N based superconductors with TC >= 10 K, including 3 with TC >= 25 K, such as two structural forms of B2CN (TC = 44.8 K and 41.5 K) and TiNbN2 (TC = 26.2 K). These materials share a common feature of strong {\sigma}-bonds, which is key to achieving relatively high TC. Moreover, we proposed Tb3Ni2O7 (TC = 61.6 K) and Ac3Ni2O7 (TC = 70.3 K) as potential high-TC nickelate superconductors under high pressure. Their electronic structures closely resemble those of La3Ni2O7, especially in the hole-type band dominated by Ni-3dz2 orbital character. We also analyzed feature importance in the ML results for both conventional and high-TC superconductors. These results advance the search for new superconductors and enhance the fundamental understanding of superconducting mechanisms.