ABSTRACT This study evaluates the predictive performance of five well-established machine learning models, Gaussian process (GP), random forest (RF), random tree (RT), M5 tree, and linear regression (LR), for estimating trapping efficiency (TE) of silt ejectors. A dataset of 252 hydraulics laboratory samples was used, with 189 randomly selected for model training and the remaining 63 reserved for testing. The input dataset comprises concentration, silt size, and extraction ratio, while TE is indicated as the output. Model performance was assessed with correlation coefficient (CC), root mean square error (RMSE), mean absolute error (MAE), and Nash–Sutcliffe efficiency (NSE). Results reveal that the M5 tree model, despite achieving lower training accuracy (CC = 0.885, RMSE = 7.20), outperformed all models on unseen data with the lowest prediction errors (MAE = 5.74, RMSE = 6.94, CC = 0.884), highlighting superior generalization and resistance to overfitting. Sensitivity analysis confirmed silt size as the most influential input variable. Findings establish the M5 tree and RF as the most reliable and generalizable models, offering robust, interpretable, and practical predictive frameworks for TE under uncertain and variable conditions.
Dahiya et al. (Thu,) studied this question.