Designing Ligated Nanoclusters with Flow Matching and Natural Language

Generative models have transformed molecular and materials design, yet systems such as ligated metal nanoclusters that lie at the interface of molecules and periodic crystals remain largely unaddressed. Here, we introduce a framework for property-driven inverse design of ligated nanoclusters. First, we construct a dataset of 3.3M ligated nanocluster conformations, expanding available data for this materials class by two orders of magnitude. We then introduce a chemistry-informed flow matching model that uses a domain-adapted prior to generate accurate all-atom geometries of ligated nanoclusters. Our model shows strong performance even for highly ligated clusters with over 20 ligands and systems exceeding 500 atoms. Finally, we embed our generative model in an inverse-design pipeline in which we train a large language model in a closed loop to propose compositions from user-defined natural-language requests. This allows the generative design of ligated nanoclusters with targeted catalytic descriptors at high accuracy.