A structure-informed machine learning approach for understanding superconductivity

Superconductivity remains one of the most remarkable quantum phenomena in materials. However, when it comes to how to increase the transition temperature, our understanding remains limited. The past decade’s effort in curating databases of superconductors have recently become available[1-2], presenting fresh opportunities to attempt to learn what trends are informative from the curated database. However, machine learning efforts to date have not considered structural information or space group symmetry of materials. Moreover, most existing approaches used the machine learning algorithm as a black box to output predictions of transition temperatures without reasoning. In this work, we introduce theoretically motivated feature representations of materials that systematically reflect structural information, space group symmetry, and atomistic information associated with all the elements in a given material. We then use the features as input into an interpretable model, in the spirit of the “Materials Expert-AI (ME-AI)” approach[3]. We resulting new insights and plans to extend the features to include more experimental results.

[1] SuperCon, https://doi.org/10.48505/nims.3837 (2022).

[2] Sommer, T., Willa, R., Schmalian, J. et al. 3DSC - a dataset of superconductors including crystal structures. Sci Data 10, 816 (2023).

[3] Y. Liu, M. Jovanovic, K. Mallayya, W. J. Maddox, A. G. Wilson, S. Klemenz, L. M. Schoop, and E.-A. Kim. Materials expert-artificial intelligence for materials discovery. Preprint at https://arxiv.org/abs/2312.02796 (2023).