High-Tc superconductor candidates proposed by machine learning

We cast the relation between chemical compositions of solid-state materials and their superconducting critical temperature (Tc) in terms of a statistical learning problem with reduced complexity. Training of query-aware similarity-based ridge regression models on experimental SuperCon data with (implicit) and without (ambient) high pressure entries achieves average Tc prediction errors of  10 K for unseen out-of-sample materials. Subsequent utilization of the approach to scan  153k materials in the Materials Project enables the ranking of candidates by Tc while taking into account thermodynamic stability and small band gap. Stable top three high-Tc candidate materials with large band gaps for implicit and ambient pressures are predicted to be Cs2Sn(H2N)6 (324 K), CsH5N2 (315K), Rb2Sn(H2N)6 (305 K), and H15IrBr3N5 (189 K), H12OsN5Cl3O (161 K), B10H13I (151 K), respectively. Stable top three high-Tc candidate materials with small band gaps for implicit and ambient pressures are predicted to be RbLiH12Se3N4 (255 K), CeH14Cl3O7 (246 K), Li(H3N)4 (234 K), and ReH30Ru2(NCl)10 (127 K), AlH18Ru(NF)6 (120 K), Sr(Li2P)2 (117 K), respectively.