Predicting Max Scour Depths near Two-Pier Groups Using Ensemble Machine-Learning Models and Visualizing Feature Importance with Partial Dependence Plots and SHAP
Buddhadev Nandi, Subhasish Das
Abstract
Assessing scour depth (Sd) near side-by-side, tandem, and eccentric bridge piers is crucial for designing resilient structures. Researchers employed soft computing techniques to enhance Sd prediction models, focusing on ensemble machine learning (ML) methods for isolated piers. However, this research is limited on such two-pier groups, which necessitates a detailed study of how pier spacing and positioning collectively affect Sd predictions. A thorough examination is needed to analyze scouring patterns and the collective two-pier impact on estimating Sd using ensemble ML models. This study employs two ensemble ML models, random forest (RF) and extreme gradient boosting (XGBoost), to collectively predict circular two-pier Sd. Input parameters such as flow characteristics, sediment properties, time, pier gaps, and flow skew angle are rigorously evaluated to assess their combined impact on Sd. Partial dependence plots (PDP) and SHapley Additive exPlanations (SHAP) are used to visualize feature importance and effects on predicting Sd after training ML models, providing insights into individual features’ influence on predictions. Performance indicators [coefficient of determination (R2), mean absolute error, and root mean squared error] assess the ML models’ performance. Results demonstrate that XGBoost outperformed RF in training and testing phases with random search cross validation (CV) optimization for both piers. However, RF excelled over XGBoost in training using grid search CV and random search CV for both piers. Flow intensity was identified as the most influential variable, making the phenomenon highly vulnerable during model training with SFS and SHAP. Using ensemble ML models with detailed parameter evaluations and visualizations, engineers can predict Sd more effectively, thus enhancing scouring pattern understanding to design resilient structures.