Physics-informed neural network-based prediction of permeation performance in reverse osmosis membrane elements

Abstract

Reverse osmosis (RO) technology, as a key support for modern water treatment systems, is significantly affected by membrane fouling in its long-term operational performance. Accurately predicting the permeability of membrane elements is of great significance for determining the fouling status of membrane elements. Although traditional data-driven methods have achieved modeling of membrane fouling trends to some extent, they generally suffer from problems such as poor model monotonicity and insufficient ability to model physical laws. To overcome the above limitations, this paper proposes a Physics-Informed Neural Network (PINN) framework that integrates physical knowledge. It innovatively introduces the physical monotonicity reflected by the variation of reverse osmosis membrane permeability with operating conditions as a constraint, and constructs a predictive model with physical consistency and data-driven capabilities. The model is developed based on the experimentally measured data obtained from the test bench of the pure water special station. It selects operating time, inlet salt content, concentrated water salt content, inlet pressure, concentrated water pressure and temperature as inputs, and membrane permeability coefficients as outputs. The results indicate that the constructed PINN model outperforms traditional data-driven methods in both error evaluation metrics and coefficient of determination evaluation metrics, and partial dependency analysis shows that its prediction results have high consistency at the physical trend level. This study provides an effective paradigm for embedding physical constraints into reverse osmosis performance prediction models, and offers a more universal and interpretable modeling approach for state monitoring and performance optimization of reverse osmosis systems.