The impact of the feature space and the optimal approach for selecting significant features in predicting the Tc of superconducting materials using machine learning methods.

Gashmard, Hassan; Shakeripour, Hamideh; Alaei, Mojtaba

The impact of the feature space and the optimal approach for selecting significant features in predicting the transition temperature of superconducting materials using machine learning methods

(2024) 31 Jan., The 6^th Iranian Conference on Computational Physics, Iran. IUT.

Recently, machine learning methods have become one of the areas of interest in materials science, especially in the field of condensed matter, and are increasingly expanding. The existence of large data sets as well as improving the performance of algorithms are two key points in this new field of study. In this study, the SuperCon dataset, which is the largest dataset of superconducting materials, was used, and after performing various steps of data pre-processing, the DataG dataset was finally presented. CatBoost algorithm was used to predict the transition temperature of superconducting materials. Finally, by designing the Jabir package that produces 322 atomic descriptors and implementing an innovative hybrid method in the form of the Soraya package to select the most important features, we managed to achieve values of 0.952 and 6.45 K, respectively, for R2 and RMSE evaluation criteria.

 Keywords: Machine learning, CatBoost algorithm, superconducting materials, Jabir package, Soraya package