Deep Learning Based Superconductivity: Prediction and Experimental Tests

TL;DR

The paper addresses the challenge of discovering new superconductors by developing a deep-learning framework that classifies materials as superconductors and predicts their critical temperature $T_c$ using only chemical composition. It contrasts composition-only DL models against traditional descriptor-based tree ensembles, and demonstrates two network architectures (fully-connected and convolutional) with a dual-branch setup for classification and regression, trained on the SuperCon dataset. Experimentally, a DL-guided suggestion, Mo20Re6Si4, is synthesized and shown to be superconducting with $T_c = 5.4$ K, validating the approach and illustrating the potential for AI-assisted materials discovery. The study finds that the DL models achieve performance on par with random forests while requiring less detailed atomic information, and it outlines limitations and future directions to enhance predictive power, such as incorporating crystal structure, neighborhood information in the periodic table, and larger training datasets.

Abstract

The discovery of novel superconducting materials is a longstanding challenge in materials science, with a wealth of potential for applications in energy, transportation, and computing. Recent advances in artificial intelligence (AI) have enabled expediting the search for new materials by efficiently utilizing vast materials databases. In this study, we developed an approach based on deep learning (DL) to predict new superconducting materials. We have synthesized a compound derived from our DL network and confirmed its superconducting properties in agreement with our prediction. Our approach is also compared to previous work based on random forests (RFs). In particular, RFs require knowledge of the chem-ical properties of the compound, while our neural net inputs depend solely on the chemical composition. With the help of hints from our network, we discover a new ternary compound $\textrm{Mo}_{20}\textrm{Re}_{6}\textrm{Si}_{4}$, which becomes superconducting below 5.4 K. We further discuss the existing limitations and challenges associated with using AI to predict and, along with potential future research directions.

Figures (9)

• Figure 1: Histogram representing the data and $T_c$ distribution. Dashed line is the mean $T_c$ value across the dataset. $T_c$ values are in $K$.
• Figure 2: DNN structure with a shared backbone and two prediction branches: one for the $T_c$ value and the other for classification.
• Figure 3: Fully-connected Model: Accuracy in classification during training with epoch.
• Figure 4: Fully-connected Model: MSE loss in classification during training.
• Figure 5: Convolutional Model: Accuracy in classification during training with epoch.