Developing a Complete AI-Accelerated Workflow for Superconductor Discovery

Abstract: The evolution of materials discovery has continually transformed, progressing from empirical experimentation to virtual high-throughput screening, which leverages computational techniques to fully characterize a material before synthesis. Despite the successes of these approaches, significant bottlenecks remain due to the high computational cost of density functional theory (DFT) calculations required to determine the thermodynamic and dynamic stability of a material and its functional properties. In particular, discovering new superconductors is massively impeded by the cost of computing the electron-phonon spectral functions, which limits the feasible materials space. Recent advances in machine learning offer an opportunity to accelerate the superconductor discovery workflow by developing machine-learning-based surrogates for DFT. Here we present a Bootstrapped Ensemble of Equivariant Graph Neural Networks (BEE-NET), a ML model that predicts the Eliashberg spectral function achieves a test mean absolute error of 0.87 K for the superconducting critical temperature ($T_c$), relative to predictions of the Allen-Dynes equation using DFT-derived spectral functions. BEE-NET simultaneously identifies candidate structures with $T_c > 5$ K at a precision of 86% and a true negative rate of 99.4%. Combined with elemental substitution and ML interatomic potentials, this models puts us in position to develop a complete AI-accelerated workflow to identify novel superconductors. The workflow achieved 87% precision, narrowing 1.3 million candidates to 741 stable compounds with DFT-confirmed $T_c > 5$ K. We report the prediction and successful experimental synthesis, characterization, and verification of two novel superconductors. This work exemplifies the potential of integrating machine learning, computational methods, and experimental techniques to revolutionize the field of materials discovery.