Automatic hydrogen clathrate recognition neural networks for high-pressure superconductor prediction

Abstract

Recently, significant progress has been made in ultra-high-pressure hydride superconductors. Using machine learning to predict and design materials is an important means of accelerating materials development. Here, we propose an algorithm that automatically recognizes hydrogen clathrate structures and incorporates the interactions of hydrogen clathrates, significantly improving the accuracy of predictions for critical temperatures compared to traditional graph neural networks and elemental convolutions. Simultaneously, we demonstrate that multi-task learning and a non-superconducting dataset with distributions similar to a superconducting dataset can enhance the predictive performance. Our model provides extensive predictions for hydrogen-rich compounds across the 0–500 GPa range. Validated by density functional theory, our model predicts multiple superconductors with critical temperatures exceeding 200 K. This study provides an effective method for predicting the critical temperature of high-pressure superconductors, facilitating efficient discovery and development of superconductors.