The application of machine learning models based on feature selection for the prediction of the superconducting critical temperature

Superconducting critical temperature has great significance in superconducting applications. Traditional critical temperature prediction using experimental and computational physics methods, there are problems such as high experimental cost and long calculation time. Machine learning model has the advantages of fast prediction speed and low threshold of use. At present, traditional machine learning algorithms use more superconducting property characteristic parameters, resulting in large calculations, over-fitting, and other problems. In this paper, aiming at the situation of excessive features in the prediction algorithm, the method based on superconducting microscopic theory combined with the analysis of physical and chemical properties of materials is adopted to select the feature parameters and use them for machine learning model training and prediction, which can improve the prediction accuracy, effectively prevent over-fitting. For example, reduce RMSE from 8.4K to 2.7K through mutual information screening features, improving the prediction accuracy by 68%.