Multimodal deep learning-based prediction of activated carbon adsorption capacities for volatile organic compound removal

The adsorption of volatile organic compounds (VOCs) onto activated carbons (ACs) is primarily governed by its pore structure. However, previous studies have often reduced AC surface structural properties to a single numerical variable for predicting adsorption performance. This simplification limits the representation of adsorption mechanisms and affects model generalization, resulting in reduced predictive accuracy for ACs with untrained compositions. To address these limitations, this study developed a multimodal model that integrates raw brunauer-emmett-teller (BET) data to enhance adsorption capacity prediction. Generalization performance was assessed by testing on ACs that were not included in training and validation datasets. The proposed joint fusion-based multimodal model outperformed single-modal models and alternative fusion strategies, achieving high predictive accuracy, with R² values of 0.9465 for validation and 0.9318 for testing. Furthermore, model interpretation identified BET data, initial surface area, and ion type as key factors influencing equilibrium adsorption capacity. The relative importance of these factors varied based on VOC type, revealing distinct adsorption mechanisms. Additionally, the analysis of BET data underscored the critical role of micropore size in adsorption performance, aligning with the pore-filling mechanism. This study highlights the potential of multimodal learning in adsorption modeling by improving both predictive accuracy and mechanistic interpretation. The proposed approach provides a scalable framework for analyzing adsorption processes across diverse adsorbents and environmental conditions.