Diffusion-Based Discovery of Ordered Superconductors
Abstract: AI-driven materials discovery pipelines have consistently generated theoretically stable structures which turn out to be disordered solid solutions upon synthesis. This disconnect between computational predictions and experimental realization has been especially prominent in recent attempts in the generation of novel superconductors. To address this, we fine-tune a DiffCSP foundation model on 7,183 superconductors labeled with crystallographic disorder probability, and employ classifier-free guidance to steer generation towards ordered superconductors. We show that generated structures have an 88% chance of having a lower disorder probability than randomly selected structures from an unguided baseline. Candidates are then screened through a multi-stage pipeline which employs machine learning and density functional theory calculations to assess disorder probability, thermodynamic stability, and superconducting properties, with selected structures advancing to experimental synthesis.
