Discovery of Novel Superconducting Materials with Deep Learning

Superconducting devices are crucial to emerging quantum technologies, contributing to innovations in the areas of computing, metrology, and communication systems. However, many of these devices are based on conventional BCS (Bardeen-Cooper-Schrieffer) superconductors, which have operating critical temperatures
$(T_{c})$
that require expensive helium-based cryogenic cooling. This has motivated the exploration of unconventional non-BCS superconductors as an alternative platform for these technologies, since some of these superconductors exhibit desirable properties such as high
$T_{c}$
and topologically robust states. Although unconventional superconductors have been known to exist for several decades, a comprehensive theory of superconductivity in these materials has not yet been developed, making the discovery of new superconductors a challenging endeavor. In this paper, we present ongoing work in the autonomous discovery of conventional, unconventional, and exotic superconductors with deep learning methods. Using experimental data from over 36,000 superconducting materials, we demonstrate that graph neural network models trained on raw crystalline structure can be used to discover superconductors with desirable properties such as high
$T_{c}$
.