Machine learning enhanced aeration systems for optimizing oxygen transfer efficiency for sustainable and safe wastewater management.

This study models oxygen-transfer efficiency (OTE) in circular solid-jet aerators using a laboratory dataset of 320 observations collected under controlled conditions. Experiments varied jet count (1-8), opening area (49.24-124.03 mm²), jet length (170-470 mm), and discharge (1.05-3.04 l s⁻¹); dissolved oxygen was measured, and OTE was computed and standardized to 20 °C. Five regressors-Linear Regression (LR), M5P, Random Tree (RT), Reduced Error Pruning (REP) Tree, and Random Forest (RF)-were trained with a 70/30 train-test split and evaluated using CC, RMSE, MAE, NSE, and SI. Residual histograms with kernel-density overlays and an uncertainty summary (U95, bounds) indicated compact, slightly negative-centered errors for the tree-based models and broader, heavy-tailed errors for LR; a Taylor diagram and a Spearman heatmap supported these patterns. Among all models, RF achieved the highest test performance and the lowest errors, with results statistically superior to alternatives by paired t-tests on residuals (α = 0.05); the Spearman heatmap also showed the strongest concordance between RF predictions and observations, while a leave-one-input-out sensitivity analysis identified discharge (Q) as the dominant driver. Taken together, the results identify RF as the most accurate and generalizable predictor across the tested operating envelope, providing a practical basis for the design and optimization of aeration systems in water and wastewater treatment.