| Abstract Scope |
Discovering superconductors with higher critical temperatures, Tc, and fields, Hc, is limited by the cost of electron-phonon calculations and the vastness of chemical and structural space. We present a closed-loop discovery effort coupling artificial intelligence, first-principles simulations, materials synthesis, and characterization. Guided diffusion models and our Open Materials Generation framework, based on stochastic interpolants, generate candidate superconductors across chemical and structural space. BEE-NET, an equivariant graph neural networks, then predicts the Eliashberg spectral function α2F(ω) and Tc directly from crystal structures. Trained on a substantially enlarged dataset and using machine-learned interatomic potentials for phonons, BEE-NET screens millions of candidates and guides synthesis, already yielding experimentally realized superconductors. Synthesis revealed that many predicted compounds form with substitutional disorder, which our generative models, trained on ordered structures, could not anticipate. We therefore developed a disorder model that screens for materials that remain ordered, aligning predictions with BEE-NET and enabling synthesizable designs. |