Machine-learning accelerated prediction of two-dimensional conventional superconductors

Abstract

We perform a large scale search for two-dimensional (2D) superconductors, by using electron–phonon calculations with density-functional perturbation theory combined with machine learning models. In total, we screened over 140 000 2D compounds from the Alexandria database. Our high-throughput approach revealed a multitude of 2D superconductors with diverse chemistries and crystal structures. Moreover, we find that 2D materials generally exhibit stronger electron–phonon coupling than their 3D counterparts, although their average phonon frequencies are lower, leading to an overall lower T_c. In spite of this, we discovered several out-of-distribution materials with relatively high-T_c. In total, 105 2D systems were found with T_c > 5 K. Some interesting compounds, such as CuH₂, NbN, and V₂NS₂, demonstrate high T_c values and good thermodynamic stability, making them strong candidates for experimental synthesis and practical applications. Our findings highlight the critical role of computational databases and machine learning in accelerating the discovery of novel superconductors.