Machine Learning-Driven Dynamic Prediction and Optimization for Ammonia Recovery in Membrane Distillation System

Abstract

Ammonia recovery is crucial for environmental protection and resource sustainability, as it helps mitigate nitrogen pollution and enables recycling of valuable resources. Membrane distillation (MD) is a promising approach for high-purity ammonia recovery, but optimizing its performance relies on understanding dynamic ammonia transport under varying conditions. Conventional theoretical models often lack the flexibility to capture complex system dynamics, while empirical models fail to generalize beyond specific experimental scenarios. Here, we integrate machine learning (ML) with dynamic simulation techniques to model complex nonlinear interactions, embedding empirical inference into theoretical frameworks. Key input variables influencing the instantaneous rate of ammonia concentration change—such as temperature, pH, and ammonia partial pressure gradient—were identified through theoretical analysis. An artificial neural network (ANN) was developed to simulate the instantaneous rate of change of ammonia concentration on the feed side, achieving a high accuracy on the test set (R² = 0.8537). The ANN was further combined with the fourth-order Runge-Kutta (RK4) algorithm to predict real-time ammonia concentrations and the cumulative ammonia recovery rate. To optimize the instantaneous MD performance for ammonia, a wide range of ML-based operational condition grid predictions was performed to identify optimal parameters. This study enhances the comprehension of ammonia transport mechanisms in MD systems while simultaneously advancing real-time process control and adaptive optimization, promoting both theoretical development and practical implementation.