Data-Driven Discovery of Superconductivity within the Inorganic Crystals Graphlet Bank

Machine learning offers powerful opportunities for materials discovery, although realizing this potential depends sensitively on how materials are represented. Approaches that incorporate experimentally derived chemical and structural information provide a more physically meaningful basis for connecting atomic-scale organization with emergent quantum phenomena [1].

The Inorganic Crystals Graphlet Bank (ICGB) provides a large, experimentally grounded dataset that encodes local chemistry, geometry, and symmetry for nearly 200,000 materials. Building on previous work [2] where Gaussian Process (GP) models were trained to capture superconducting behavior, we applied these models in a high-throughput search across ICGB to predict new materials with potential superconductivity. This large-scale screening identifies new candidate superconductors and demonstrates how data-driven approaches can leverage ICGB to explore vast materials spaces. The superconductivity example highlights the broader potential of ICGB as a phenomenological, data-driven platform for predicting target properties, leveraging existing measurements and reported properties.

[1] Sommer et al., Sci. Data 10, 816 (2023).

[2] Lesser et al., arXiv:2510.07373 (2025).