From individual elements to macroscopic materials: in search of new superconductors via machine learning

An approach to supervised classification and regression of superconductive materials is proposed which builds on the DeepSet technology. This enables us to provide the chemical constituents of the examined compounds as an input to the algorithm, while avoiding artefacts that could originate from the chosen ordering in the list. The performance of the method are successfully challenged for both classification (tag a given material as superconducting) and regression (quantifying the associated critical temperature). We then searched through the International Mineralogical Association list with the trained neural network. Among the obtained superconducting candidates, three materials were selected to undergo a thorough experimental characterization. Superconductivity has been indeed confirmed for the synthetic analogue of michenerite, PdBiTe, and observed for the first time in monchetundraite, Pd2NiTe2, at critical temperatures in good agreement with the theory predictions. This latter is the first certified superconducting material to be identified by artificial intelligence methodologies.

Automated summary

We propose an approach based on DeepSet technology for supervised classification and regression of superconductive materials. The method takes the chemical constituents as input, avoiding artefacts from ordering in the list. Successful performance is achieved in classifying superconducting materials and quantifying their critical temperature. Using the trained neural network, we searched the International Mineralogical Association list and identified three superconducting candidates, confirming superconductivity in the synthetic analogue of michenerite and observing it for the first time in monchetundraite with critical temperatures in good agreement with theory predictions. This marks the first certified superconducting material identified through artificial intelligence methodologies.