Multimodal Machine Learning for Intelligent MOF Design: Adsorption Mechanism Analysis and Visual Prediction

Abstract Metal–organic frameworks (MOFs) show great potential for pollutant adsorption due to tunable pore structures and modifiable surface chemistry, but predicting their adsorption performance and screening optimal MOFs for specific scenarios remain challenging. Traditional experimental and empirical methods are inefficient, driving machine learning (ML) to accelerate MOF research; however, most existing ML models adopt single-modal approaches that fail to integrate heterogeneous data (e.g., MOF topological features, pollutant simplified molecular input line entry system SMILES data, and environmental parameters), limiting prediction accuracy. Herein, we propose a multimodal fusion framework, graph self-attention convolutional neural networks–fully connected network (GACNN-NN), which uses GACNN to decode MOF crystal spatial electronic distributions and a fully connected network to process pollutant physicochemical properties and environmental variables. We also developed a web-based prediction platform for real-time adsorption capacity estimation via inputting MOF synthesis and pollutant molecular parameters. Interpretability analysis further provides intuitive visualizations of key MOF adsorption sites, advancing MOF screening for environmental applications.