A multimodal semantic learning framework for data-driven discovery of nitrate-to-ammonia electrocatalysts

Data-centric approaches are increasingly driving the discovery of functional materials; however, the heterogeneity of experimental conditions and the fragmented nature of data, especially when embedded in unstructured scientific literature, hinder the development of generalizable predictive models. Here, we present a multimodal artificial intelligence (AI) framework, the Material Semantics Network (MatSemNet), which learns materials semantics by integrating textual, numerical, and reaction pathway information extracted from the literature. By constructing a unified representation of structure-property relationships, MatSemNet enables transferable prediction of catalytic performance across material systems, outperforming conventional single-modal or descriptor-based approaches. As a proof of concept, we apply MatSemNet to the nitrate reduction reaction (NO 3 RR), achieving high predictive accuracy for Faradaic efficiency (FE) and identifying promising catalyst candidates. Guided by the model, we design and experimentally validate rare-earth (RE)-doped Co 3 O 4 catalysts, which deliver an FE of 85.5% and an ammonia yield rate of 28.3 mg h −1 cm −2 under alkaline conditions. Mechanistic studies show that RE incorporation modulates the electronic structure and enhances reaction selectivity. Beyond this specific application, MatSemNet offers a scalable and transferable paradigm for data-driven materials discovery by bridging heterogeneous knowledge with predictive modeling, and it is broadly applicable to diverse catalytic and functional material systems.