Data-Mechanism Fusion for Deep Learning-Enabled Membrane Performance Prediction and Material Discovery

Deep learning (DL) and related data-driven models are increasingly being used to predict membrane performance, guide material discovery, and optimize membrane processes. However, membrane separation is governed by coupled effects of material structure, solute properties, interfacial interactions, transport mechanisms, and operating conditions, making purely data-driven prediction difficult to generalize across membrane systems and water matrices. This review examines data-mechanism fusion as a membrane-science-oriented framework that integrates experimental data, molecular simulations, material databases, imaging information, process time-series data, and physicochemical knowledge throughout the modeling workflow. We first summarize membrane data types and assess how feedforward, convolutional, graph-based, recurrent, Transformer, and hybrid architectures align with data modalities and tasks. We then review recent progress in ion-selective separation, micropollutant and emerging contaminant removal, membrane material screening and inverse design, fouling diagnosis, and operational optimization. Particular attention is given to how mechanistic descriptors, physical constraints, hybrid coupling, simulation-informed surrogates, and closed-loop learning improve model reliability and interpretability. Finally, we discuss persistent challenges, including data scarcity, data set heterogeneity, limited transferability, uncertainty quantification, weak physical interpretability, and the gap between computational prediction and experimental or pilot-scale validation. This review highlights data-mechanism fusion as a practical route toward more reliable, interpretable, and adaptive membrane research workflows.