Enhanced prediction of scour dimensions: Temporal variations induced by turbulent plane wall jets using FFNN, CatBoost, and XGBoost models

Turbulent wall jet-induced scours provide a considerable challenge in hydraulic engineering, impacting the stability of adjacent structures. Precisely forecasting variations in scour dimension over time is crucial for the design of safe and effective hydraulic systems. This work tackles this problem by using advanced machine learning (ML) approaches to enhance predictive accuracy. Three ML models—CatBoost, XGBoost, and a feedforward neural network (FFNN)—were assessed using several metrics, including Mean Squared Error (MSE), coefficient of determination (R²), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Willmott Index (WI). The findings indicated that CatBoost had the best predictive accuracy, closely followed by XGBoost. Both tree-based ensemble models surpassed the examined FFNN architecture, but every model had R² values over 0.8. The examination of the model by 5-fold cross-validation and learning curve analysis validated its generalization capabilities and prediction consistency. The little performance disparity between training and validation sets suggested robust generalization without signs of overfitting, while robustness testing under input noise demonstrated that model predictions retained accuracy despite simulated measurement uncertainty. Sensitivity analysis indicated that the most significant factors for forecasting $h_m/D$, $x_d/D$, and $y_s/D$ had relevance ratings of 0.34, 0.45, and 0.33, respectively. The parameter Y_t/D was shown to be the most significant for predicting $x_s/D$, with a score of 0.33. These results validate that CatBoost and XGBoost surpass traditional methods and highlight the importance of reliable input parameter selection in precisely estimating scour hole diameters.