Experimental and chemometric insights into sustainable magnesium hydroxide recovery from RO reject brine

The study presents an integrated experimental modeling strategy to selectively recover magnesium from real desalination brine (RO reject). To bridge the gap between bench-scale recovery and stand-alone prediction, controlled precipitation is coupled with machine learning models in a single predictive workflow. Magnesium hydroxide (Mg(OH)₂) was produced via alkaline treatment, with pH, temperature, and brine concentration systematically tuned to map the operating window. Across the tested conditions, pH emerged as the primary control variable, with temperature exerting a secondary influence. To predict magnesium recovery (Mg-R), Gaussian process regression (GPR), bagged trees (BT), and artificial neural network (ANN) models were trained and validated using Bayesian optimization (BO), and performance was evaluated on a held-out test set. ANN was the best-performing model (Coefficient of Determination (R²) = 0.9967, Nash–Sutcliffe Efficiency (NSE) = 0.9968, Root Mean Square Error (RMSE) = 0.1184), and model predictions closely tracked experimental results near the optimum (e.g., GPR: 97.96 %; ANN: 96.91 %; BT-BO: 97.19 %). At the best operating point (pH 11, temperature 35 °C, magnesium concentration (Mg conc.) 3148.5 ppm), the process achieved 96.91 % recovery and 97.05 % purity of Mg(OH)₂, indicating an efficient, practically deployable approach with ANN-guided set-point selection and GPR-BO as a robust alternative. The findings highlight the potential of coupling brine valorization processes with data-driven modeling to enhance recovery efficiency and support sustainable resource utilization in desalination systems.