From lab to pilot: Predicting and validating caffeine and DEET removal from treated greywater using response surface methodology and artificial neural networks

• ANN and RSM models predict removal efficiency based on lab-scale experiments. • ANN outperformed RSM for caffeine; both performed well for DEET removal predictions. • Caffeine concentration had the highest influence on caffeine removal. • DOC concentration had the highest influence of DEET removal. • ANN performed better in predicting caffeine and DEET removal for unseen pilot tests. Greywater, a valuable resource in water-scarce regions, contains organic micropollutants (OMPs) from household chemicals, pharmaceuticals and personal care products. Ozonation is a promising technology for removing these OMPs, enabling the safe reuse of treated greywater (TGW) for irrigation. In this study, TGW was analyzed and found to contain high levels of caffeine and N, N‑diethyl‑meta-toluamide (DEET). Laboratory-scale ozonation experiments were conducted using the Box-Behnken design, varying ozone dosage, dissolved organic carbon (DOC), and pollutant concentrations. Response surface methodology (RSM) and artificial neural network (ANN) models were developed to examine the effects of these variables and to compare the prediction of removal efficiency (RE) for these compounds. The models were validated with pilot-scale ozonation experiments using treated greywater. The ANN model was more effective than RSM in predicting caffeine RE, while both models showed comparable performance for DEET, with RSM yielding lower relative errors. Sensitivity analysis revealed that caffeine RE was most influenced by its initial concentration, whereas DEET RE was primarily affected by DOC levels, followed by ozone dosage. The pilot study results indicated an effective removal efficiency of caffeine (80 %) and DEET (91 %) at an ozone dosage of 2 mg/L. The ANN model demonstrated better prediction and trend capturing than the RSM model when subjected to pilot-scale data. The study demonstrated that models developed from lab-scale data, particularly ANN, can be effectively applied to design, optimize, and control pilot-scale ozonation experiments.