CatalyticMLLM: A Graph-Text Multimodal Large Language Model for Catalytic Materials

ArXi:2605.17254v2 Announce Type: replace Property prediction and inverse structural design of catalytic materials are typically modeled as two independent tasks: the former predicts target properties from given structures, whereas the latter generates candidate structures according to desired properties. Although the decoupled paradigm facilitates the implementation of a ``generation--evaluation--screening'' workflow, the inconsistency between the generative model and the property prediction model in terms of representation spaces and.