AI-Assisted Superconductor Discovery—Reducing Computational Bottlenecks with Machine Learning

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.
While high-throughput screening has been successful, there are significant bottlenecks in the screening process due to the high computational cost of the Density Functional Theory (DFT) calculations required to determine a material's thermodynamic and dynamic stability and its functional properties.
In the search for superconducting materials, the cost of computing the electron-phonon spectral functions significantly reduces the material space that can be feasibly searched.
Recent advancements in machine learning present an opportunity to accelerate the superconductor discovery workflow by enabling machine learning to act as a DFT surrogate.
As such, we have vastly improved our previously published machine learning model, Bootstrapped Ensemble of Tempered Equivariant Graph Neural Networks (BETE-NET), obtaining a test MAE of 0.7K for the superconducting critical temperature.
We will then leverage this improved model in conjunction with elemental substitution and machine-learned interatomic potentials to develop an AI-accelerated workflow to identify novel superconductors.
This workflow obtained a final screening precision of 87\% and reduced \~1.3 million candidate structures to ~1000 dynamic and thermodynamically stable candidate structures with a DFT computed T_c greater than 5K, exemplifying the potential of integrating machine learning, computational methods, and experimental techniques to revolutionize the field of materials discovery.