CatEdit: A Generative Editing Framework for Heterogeneous Catalyst Discovery

Generative models for materials and catalysts have largely focused on de novo structure generation. In practice, materials are often improved by doping or alloying an existing structure. However, the associated combinatorial space is too large to enumerate, even with machine learning interatomic potentials (MLIPs). We introduce CatEdit, a generative editing framework inspired by image-editing processes with generative models. It edits structures by replacing several atoms to reach a target adsorption energy while preserving a host structure. Across unary and binary metal adslabs, CatEdit finds more target-matching structures than random substitution, including configurations with elements absent from the agentic framework's substitutions. Applied to non-precious unary/binary oxygen reduction reaction (ORR) catalyst screening, CatEdit expands the two candidates to 319 distinct doped candidates, a $160\times$ larger candidate set. CatEdit's model-agnostic formulation offers a general route to redesigning existing materials toward target properties.