Flux decline prediction in dead-end ultrafiltration combining fluorescence spectroscopy and mechanism-informed machine learning

Membrane fouling, which leads to water flux decline with time, is the greatest challenge in membrane filtration. The complexity of phenomena itself, as well as the heterogeneous and fluctuating characteristics of dissolved organic matter, make fouling prediction an arduous task. In this work, a novel approach to predict fouling and flux decline under fluctuating organic load is proposed, developed and validated. A semi-empirical mechanistic model (Hermia) is empowered with support vector machine to address the mechanistic complexity of fouling phenomena. The inputs of the machine learning steps are the initial flux, the organic load, and three pre-selected combinations of excitation-emission fluorescence spectra. The model was trained varying initial flux, organic carbon load and composition. Model validation on unknown experiments (e.g., experiments the model was not trained on) revealed a good predictive accuracy, with R2 ranging from 0.89 to 0.99, with an average of 0.96. The proposed approach is expected to have great potential in the field of membrane processes and water technologies in general.