Machine learning and interactive GUI for estimating roller length of hydraulic jumps

Hydraulic jumps reduce kinetic energy after ogee spillways, improve wastewater chlorination, and serve many hydraulic applications. This study utilized seven Machine Learning (ML) models: Artificial Neural Network (ANN), Gene Expression Programming (GEP), Random Forest (RF), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Gradient Boosting (CatBoost) to predict the roller length of a hydraulic jump (RLHJ) over a rough bed channel. Firstly, 367 experimental datasets were gathered from a previous research database, allocating 70% and 30% for the training and testing stages. Then, these datasets were assessed for descriptive statistics and correlation analysis. Three dimensionless inputs were considered: upstream Froude number (F_r1), relative HJ depths (h₂/h₁), and relative bed roughness to the HJ initial depth (k_s/h₁). Relative roller length to the HJ initial depth (L_r/h₁) was the output from the models. The hyperparameters of the models were optimized using the Bayesian optimization technique, with fivefold cross-validation ensuring high performance. Comprehensive analyses were employed to validate model effectiveness, including visual and quantitative methods. Results found that the XGBoost and CatBoost models surpassed the other models, achieving the highest determination coefficients (R²) of 0.938 and 0.936, along with the lowest Mean-Absolute-Relative-Error of 7.3% and 7.1%, respectively, thereby establishing them as trustworthy techniques for RLHJ prediction. RF and LightGBM follow closely behind, exhibit strong generalization, whereas GEP and ANN demonstrate commendable performance, although with a marginal rise in error metrics. The study also incorporated Shapley-Additive-exPlanations (SHAP) and Partial-Dependence-Plot analyses, revealing that the relative HJ depths had the most significant impact on the RLHJ predictions. Finally, an interactive Graphical User Interface was later built to allow designers to predict RLHJ quickly and more economically than costly computational or experimental tests.