Neural network prediction of Tc for conventional and unconventional superconductors

<!--StartFragment-->We demonstrate the use of artificial neural networks to predicting the experimental superconductor transition temperature for conventional and unconventional superconductors. The training sets consist of 580 BCS superconductors, 6,489 uncategorized superconductors, 1,375 iron based superconductors, and 4,226 copper and oxygen containing superconductors. Descriptors are limited to quantities which can be obtained from the chemical formula and standard tables (e.g. atomic masses, electronegativities). Despite not explicitly accounting for crystal structure, neural networks are shown to predict Tc with mean absolute errors for BCS superconductors of 2 K, iron based superconductors of 5 K, and cuprate superconductors of 12 K. The approach fails to produce a usable single network model if multiple classes are combined in a training set. Several potential new superconductors are predicted by the neural network, and their Tc are compared to values computed using Migdal-Eliashberg theory for BCS-type systems.<!--EndFragment-->