Machine learning using structural representations for discovery of high temperature superconductors

The compositional phase space in condensed matter systems is vast, making exhaustive searches with current theoretical tools impractical. Machine learning can uncover correlations in high-dimensional parameter spaces and serve as predictive tools for materials discovery. The authors develop a structural representation that differentiates polymorphs across pressure landscapes—critical for understanding high-temperature superconductivity—and integrate it into a machine learning model to predict superconducting transition temperatures (Tc). By encoding periodic mass and charge densities of crystal structures and combining them with numerical descriptors in a complex-valued convolutional neural network, the approach enables fast predictions of Tc with r^2 above 0.94.