Prediction of Tc in superconducting materials: A machine learning approach

Knowing whether a material becomes a superconductor or not is of utmost importance, given the interesting physics behind it and also from a practical point of view. Similarly, predicting the critical temperature of a superconductor is also important. In this work, we carried out a machine learning approach to predict the critical temperature of a superconductor. Different machine learning methods, namely elastic net regressor, linear regressor, decision tree regressor, random forest regressor, multilayer perceptron regressor, and extreme gradient boosting (XGB) regressor, are applied to two different datasets containing physical properties and chemical composition of 21 263 superconductors. The primary objective was to predict the critical temperature of these superconductors by leveraging the aforementioned algorithms and their underlying statistical methodologies. Among the models, the random forest regressor demonstrated superior performance, achieving a Root Mean Square Error (RMSE) of 9.35 K on the dataset containing chemical properties. This was closely followed by the XGB regressor, which attained an RMSE of 9.97 K on the physical dataset, with other algorithms exhibiting relatively higher RMSE values. The critical temperatures are estimated based on the chemical composition and physical properties, such as thermal conductivity, atomic radius, valence, electron affinity, and atomic mass, of the materials considered independently.