Superconducting critical temperatures of metals from machine learning techniques

Abstract

Superconductivity is one the most sophisticated and interesting phenomena of nature in which the electrical resistance of a material completely disappears at a certain temperature Tc. Since the superconductivity discovery by the Dutch physicist Heike Kamerlingh Onnes in 1911, enormous research e↵orts have been focusing on predicting the critical temperature Tc, which control the range and type of industrial applications that can benet from di↵erent superconductors. It is a very dicult task to predict Tc for new superconductors, even for electronphonon coupling superconductors. Recently, because of the advances in computing power and di↵erent online repositories containing huge amount of data of measured and calculated materials properties have been created over the years, researchers started using Machine Learning (ML) to nd the connection between the superconducting state and properties of materials and also design materials with specify values of Tc. In this research we developed ML models to determine the superconducting transition temperature Tc arising from electron-phonon interaction for any metallic material. Several machine learning models based only on the chemical compositions are developed to estimate the value Debye temperature ✓D for ⇠ 5200 materials obtained from an online database [”AFLOW”]. Also, several models are trained to predict the density of states near the Fermi level N(0) for ⇠ 13000 metallic materials existing in the same AFLOW online repository. These models are employed to predict ✓D and N(0) for ⇠ 4000 compounds of metallic materials obtained from the Supercond databases. We used these characteristics/descriptors and elemental properties to develop ML models for predicting the value of the transition temperature TC for those (⇠ 4000) metallic materials. The best model shows strong predictive power, with learned predictors o↵ering potential insight into the mechanism behind superconductivity in this family of materials. Eventually, the best model in predicting ✓D, N(0) and Tc was employed to search for potential new superconductors in the Inorganic Crystallographic Structure Database (ICSD). The compounds from ICSD having the top twenty predicted values of Tc are listed for further studies.