Flux Decline Prediction in Dead-End Ultrafiltration Combining Fluorescence Spectroscopy and Mechanism-Informed Machine Learning

Membrane fouling, which causes a water flux decline over time, is the greatest challenge in membrane filtration. The inherent complexity of fouling phenomena combined with the heterogeneous and fluctuating characteristics of dissolved organic matter makes predicting fouling a difficult task. This work proposes, develops, and validates a novel approach for predicting fouling and flux decline under varying organic loads. A semiempirical mechanistic model (Hermia) is enhanced with machine learning to address the complexity of fouling phenomena. The machine learning inputs include initial flux, organic load, and excitation–emission fluorescence spectra. The model was trained by using different initial fluxes, organic carbon loads, and compositions. Validation on new experiments (i.e., experiments not used during training) demonstrated good predictive accuracy, with R2 values ranging from 0.87 to 0.99 and an average of 0.95. The proposed approach is expected to have significant potential in the field of membrane processes and water technologies in general. Mechanism-informed machine learning enhances the prediction of flux decline during ultrafiltration, even with complex and variable feed compositions.