Predictive modeling of critical temperatures in magnesium compounds using transfer learning

This study presents a transfer learning approach for discovering potential Mg-based superconductors utilizing a comprehensive target dataset.Initially,a large source dataset(Bandgap dataset)comprising approximately∼75k compounds is utilized for pretraining,followed by fine-tuning with a smaller Critical Temperature(T_(c))dataset containing∼300 compounds.Comparatively,there is a significant improvement in the performance of the transfer learning model over the traditional deep learning(DL)model in predicting Tc.Subsequently,the transfer learning model is applied to predict the properties of approximately 150k compounds.Predictions are validated computationally using density functional theory(DFT)calculations based on lattice dynamics-related theory.Moreover,to demonstrate the extended predictive capability of the transfer learning model for new materials,a pool of virtual compounds derived from prototype crystal structures from the Materials Project(MP)database is generated.T_(c) predictions are obtained for∼3600 virtual compounds,which underwent screening for electroneutrality and thermodynamic stability.An Extra Trees-based model is trained to utilize E_(hull)values to obtain thermodynamically stable materials,employing a dataset containing Ehull values for approximately 150k materials for training.Materials with Ehull values exceeding 5 meV/atom were filtered out,resulting in a refined list of potential Mg-based superconductors.This study showcases the effectiveness of transfer learning in predicting superconducting properties and highlights its potential for accelerating the discovery of Mg-based materials in the field of superconductivity.

Crystal structure graph neural networks for high-performance superconducting critical temperature prediction

The utilization of machine learning methods to predict the superconducting critical temperature(T_(c))traditionally necessitates manually constructing elemental features,which challenges both the provision of meaningful chemical insights and the accuracy of predictions.In this work,we introduced crystal structure graph neural networks to extract structure-based features for T_(c)prediction.Our results indicated that these structure-based models outperformed all previously reported models,achieving an impressive coefficient of determination(R^(2))of 0.962 and a root mean square error(RMSE)of 6.192 K.From the existing Inorganic Crystal Structure Database(ICSD),our model successfully identified 76 potential high-temperature superconducting compounds with T_(c)≥77 K.Among these,Tl_(5)Ba_(6)Ca_(6)Cu_(9)O_(29)and TlYBa_(2)Cu_(2)O_(7)exhibit remarkably high T_(c)values of 108.4 and 101.8 K,respectively.This work provides a perspective on the structure-property relationship for reliable T_(c)prediction.