From Data to Discovery: Accelerating Superconductor Discovery with Machine Learning

Historically, superconductor discovery relied on empirical methods, but computational advancements and ab initio methods like Density Functional Theory (DFT) have ushered in a new era of in-silico materials design. Our predictions and the experimental confirmation of superconductivity in WB2 and Nb-substituted MoB2 exemplify this progress. However, DFT's high computational cost, especially for electron-phonon interactions, limits its applicability for large-scale screening. While machine learning (ML) has accelerated discoveries in other materials, its application to superconductors is constrained by the limited accuracy of existing models.

In this talk, I will present the development of robust ML models that redefine superconductor discovery through enhanced predictive accuracy and AI-driven workflows. Central to our efforts are equivariant graph neural networks designed to predict the Eliashberg spectral function, from which critical temperature (Tc) is derived. By leveraging bootstrapping and tempered overfitting, we trained BETE-NET (Bootstrapped Ensemble of Tempered Equivariant Graph Neural Networks) on 600 crystal structures, achieving remarkable accuracy. I will discuss two strategies to improve BETE-NET: i) embedding site-projected phonon density of states to introduce inductive bias, achieving substantial gains without expanding the dataset, and ii) increasing the dataset size while engineering the loss function. These enhancements enabled Tc predictions with a mean absolute error of 0.9 K relative to DFT.

Beyond Tc prediction, designing experimentally practical superconductors also requires assessing material stability. Addressing this, I will introduce an AI-accelerated workflow integrating our models with elemental substitution and machine-learned interatomic potentials to explore vast material spaces efficiently. This workflow reduced ≈1.3 million candidates to 700 stable superconductors with DFT-calculated Tc > 5 K, achieving 87% screening precision. Finally, the transformative potential of this approach is exemplified by its prediction and subsequent experimental discovery of new Hf-Be-Nb superconductors. This work marks a pivotal shift, establishing AI and ML as transformative tools in superconductor discovery, moving from promise to reality.