Deep learning model for finding new superconductors

Exploration of new superconductors still relies on the experience and intuition experts, is largely a process experimental trial error. In one study, only 3% candidate materials showed superconductivity [Hosono et al., Sci. Technol. Adv. Mater. 16, (2015)]. Here, we report deep learning model for finding superconductors. We introduced method named ``reading periodic table'' that represented table in way allows to learn read law elements purpose discovering novel which are outside training data. It recognized it difficult predict something Although used chemical composition as information, obtained an ${R}^{2}$ value 0.92 predicting ${T}_{\text{c}}$ database also ``garbage-in'' create synthetic data nonsuperconductors do not exist. Nonsuperconductors reported, but must be required distinguish between nonsuperconductors. three remarkable results. The can material with precision 62%, shows usefulness model; found recently discovered superconductor ${\mathrm{CaBi}}_{2}$ another ${\mathrm{Hf}}_{0.5}{\mathrm{Nb}}_{0.2}{\mathrm{V}}_{2}{\mathrm{Zr}}_{0.3}$, neither database; Fe-based high-temperature (discovered 2008) from before 2008. These results open discovery families. list, data, openly available Internet.