Atomic-level interpretable multimodal graph neural network for predicting carbon dioxide adsorption in metal-organic frameworks

Metal–organic frameworks (MOFs) are porous crystalline materials with strong potential for CO₂ capture because their structures and chemistries are highly tunable. However, their vast compositional and structural search space hampers efficient prediction of adsorption performance. Here we introduce MMAGNN, a multimodal MOF adsorption graph neural network that integrates graph representations and three-dimensional atomic coordinates. MMAGNN employs graph neural networks (GNNs) with attention mechanisms to extract and fuse molecular and spatial features, and incorporates a Transformer-based interpreter to quantify the contribution of individual atoms to adsorption. We find that the multimodal design reduces redundant information and feature interference, improving prediction accuracy and stability. Validation against adsorption density distributions confirms that MMAGNN learns interatomic interactions and pore-structure information and accurately predicts CO₂ adsorption sites. This interpretable framework provides high-throughput screening and rational design of high-performance MOFs for carbon-capture applications. Metal–organic frameworks (MOFs) are porous crystalline materials with tunable chemistry and pores, making them promising for CO₂ capture. This work introduces the Multimodal MOF Adsorption Graph Neural Network, fusing atomic graphs and 3D point clouds to predict CO₂ adsorption and produce atom-level attention maps highlighting adsorption sites.