Physics-informed neural networks for modeling and diagnosing degradation in reverse osmosis membranes

Abstract

A physics-informed neural network (PINN) framework is proposed for modeling the performance of reverse osmosis (RO) membranes and diagnosing membrane degradation. Using governing mass and momentum transport equations as part of the loss function, the model simultaneously learns physically meaningful membrane parameters while identifying outliers from regular behavior. A loss weighting of λ=10 enables the identifiability of interpretable parameters, whereas λ=0 represents a baseline purely data-driven model. Both configurations identify anomalies in RO performance by displaying rising residuals between predicted and observed values. Such findings attest to the potential of PINN-based models as early-warning mechanisms for membrane degradation in real-time RO monitoring. In addition to detecting anomalies, the framework provides interpretable transport parameters and reveals trade-offs between intrinsic salt permeability (Bs) and defect-driven transport (β). This underscores its diagnostic value and highlights the need for future work on parameter identifiability.