Interpretable Machine Learning Surrogate Model for Critical Temperature Prediction of Superconductors

A general theory of superconductivity has been the focus of research over the last decades. Machine learning (ML) approaches based on chemical and structural features have been developed in order to predict both the critical temperature Tc and potential novel candidates. Nevertheless, these models lack interpretability. Either the feature matrix is reduced via mathematical transformations such as SVD/PCA or augmented with statistical feature generation. In this work, we introduce a ML model based only on electronic and structural descriptors derived from the composition formula and the individual elements. We reach an R2 >91% a reduced number of descriptors while keeping the physical meaning of the feature space. In the end, we test our model predicting the Tc of new superconductors.