Prediction of the critical temperature of superconducting materials using image regression and ensemble deep learning

Abstract

Due to the exponential growth and availability of data in recent years, machine learning and deep learning technologies have become highly effective in a wide variety of fields, particularly those that are data-driven. Since the discovery of superconducting materials, thousands of variations have been found and synthesized, and their data is now available. As a result of the unique properties of superconducting materials and the large number of effective features that contribute to determining the critical temperature ( T c ), predicting and estimating their critical temperature remains a challenge. Since data on superconducting materials is considered tabular data, deep learning models are still not well suited for working with this type of data despite their success in machine vision. This paper proposes a method for predicting the critical temperature of superconductors by converting their tabular data to images and using a deep learning model optimized for image regression in conjunction with a shallow learning model. The proposed method outperforms existing shallow machine learning models for tabular data, including XGBoost.