Title : Classical optimization vs. artificial intelligence models for color removal in wastewater from denim washing
Abstract:
The classic optimization of a wastewater treatment process such as coagulation-flocculation (C-F) aims to establish operating conditions that maximize efficiency through the dosage of reagents, reaction times, and mixing rates. In recent years, artificial intelligence (AI) has developed various optimization models of interest for application in diverse fields, with the aim of validating their application and evaluating their potential use. The objective of this study was to determine and compare the optimal conditions for a C-F process for color removal in textile wastewater from denim washing, using a multilevel factorial (MF) experimental design with 24 experiments, as a classical optimization approach. The factors evaluated were reagent dose (g/L), agitation speed (RPM), and agitation time (min). Gradient Boosting (GB) and Evolutionary Differential (DE) algorithms were used as AI optimization models, and various metrics were used to compare the models’ performance. By operating the C-F process under the optimal conditions estimated by the models, parameters such as color, turbidity, electrical conductivity (EC), final pH, total dissolved solids (TDS), and redox potential (RP) were monitored. The results indicated that all models estimate a theoretical efficiency > 90% for color removal. Validation tests of the operating conditions obtained based on the % color removal were as follows: for FM, 80.9% at 199.8 RPM, 6.3 min, 0.99 g/L; GB from 38.9% at 20 RPM, 5 min, 0.75 g/L, and DE from 83.9% at 20 RPM, 30 min, 1 g/L. The DE algorithm demonstrated the best overall performance, exhibiting the lowest errors (MAE = 12.9 and SES = 22.1), a low PRESS value (2404), and the highest generalization capacity (R_(adj–predicted )^2= 0.925). These results indicate that this approach offers more stable and accurate predictions under the experimental conditions analyzed. The DE AI model performed best in the validation tests, with an average removal of 82% for color, 95.9% for turbidity, 43% for total dissolved solids, and a final pH of 9.8. It is possible that AI models can yield operating conditions that better align with real-world situations. Methodological constraints regarding the availability of training data can be overcome by prioritizing the applicability of these advanced models in various water treatment optimization processes. The DE model achieved the best performance even though it was trained using an experimental dataset consisting of only 24 runs, demonstrating its ability to extract reliable patterns from small datasets.
