Evaluation of Machine-Learning Models for Reliable Prediction of Gabion Weir Aeration Efficiency for Water Quality Enhancement

Abstract Aeration efficiency is a key indicator of the environmental performance of weirs, particularly for enhancing dissolved oxygen levels in natural water bodies. This study focuses on predicting gabion weir aeration efficiency ( E 20 ) and presents a comparative evaluation of empirical equations alongside advanced machine-learning (ML) models. Gabion weirs, due to their cost-effectiveness, ease of construction, and ecological compatibility, provide a sustainable means of improving oxygen transfer through air entrainment. The modeling framework incorporates the adaptive neuro-fuzzy inference system with varying membership functions, artificial neural network with logsig and purelin transfer functions (ANN_LP), gradient boosting machine, stacked ensemble (SE), generalized linear model, and extreme random trees models. Statistical metrics and visual analyses including scatter plots, Taylor diagrams, box plots, and performance diagrams identified ANN_LP as the best-performing model, followed by SE, for both dimensional and nondimensional datasets. Uncertainty analysis using Monte Carlo simulations confirmed the robustness of ANN_LP, while SE exhibited narrow uncertainty bounds. ANOVA indicated minimal deviation between predicted and experimental values for all ML models, in contrast to empirical relations. Sensitivity and variable importance analyses revealed that the discharge per unit width ( q ) and Reynolds number ( R e ) were the most influential parameters for dimensional and nondimensional datasets, respectively. These findings show that combining ML models with sustainable hydraulic designs such as gabion weirs can help improve aeration and support the ecological health of river systems.