Abstract
Bridge piers in marine and estuarine environments experience scour driven by the combined action of waves and tidal or river currents — a loading condition significantly more complex than the steady-current case addressed in most design codes. The interaction between oscillatory wave-induced flow and unidirectional current creates highly nonlinear vortex dynamics around the pier, producing scour depths that conventional empirical equations frequently underestimate or fail to predict consistently.
This study, published in Acta Geophysica (Vol. 72, Issue 3, pp. 1895–1911, 2024), by Lalit Kumar and Mohammad Saud Afzal, applies ensemble boosting machine-learning models to estimate equilibrium pier scour depth under combined wave-current conditions.
The methodological core of the study is a systematic comparison of four boosting algorithms: AdaBoost, XGBoost, CatBoost, and LightGBM. Boosting is an ensemble learning strategy that builds a strong predictor by sequentially training weak learners (typically decision trees), with each subsequent model focusing on the errors of its predecessors. This iterative error-correction mechanism makes boosting particularly effective for capturing complex, nonlinear relationships in tabular data — precisely the challenge presented by wave-current scour prediction.
The input parameters include pier geometry (pier width), bed sediment properties (median particle diameter), water depth, and the hydrodynamic characteristics of the combined wave-current field. The target variable is the equilibrium scour depth normalised by pier diameter. Model performance was evaluated using the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE).
For a consulting engineer, the practical takeaway is significant. Traditional scour design relies on empirical formulae — HEC-18 (FHWA), Sumer & Fredsøe, Melville & Coleman — that were largely calibrated for steady-current conditions or simplified wave-only scenarios. When combined wave-current loading governs (as it does for bridges in tidal inlets, estuaries, and coastal zones), these formulae require correction factors that introduce substantial uncertainty. The boosting ML models demonstrated in this study can be trained on project-specific or regional datasets to provide more accurate, site-specific scour predictions — reducing the conservative safety factors that inflate foundation costs.
The ensemble approach also offers interpretability advantages: feature importance rankings from gradient-boosted trees reveal which physical parameters most strongly control scour depth, guiding engineers toward the most critical design variables.
Key Findings
- Boosting ensemble methods (AdaBoost, XGBoost, CatBoost, LightGBM) systematically compared for pier scour prediction under combined wave-current loading.
- Models trained on pier geometry, sediment properties, water depth, and wave-current hydrodynamic parameters to predict equilibrium scour depth.
- Feature importance analysis from gradient-boosted trees identifies which physical parameters most strongly control scour depth.
Methodological Approach
Systematic comparison of four boosting ML algorithms (AdaBoost, XGBoost, CatBoost, LightGBM) trained on combined wave-current scour datasets. Performance evaluated via R², MAE, and RMSE metrics.
Implications for Hydraulic Practice
Provides consulting engineers with data-driven scour prediction tools for combined wave-current conditions, enabling site-specific estimates that reduce reliance on conservative empirical correction factors and potentially lower foundation costs.
Related Insights & Publications
Computational fluid dynamics modeling of abutment scour under steady current using the level set method
Scour — the erosion of riverbed material around bridge foundations — is a leading cause of bridge failure worldwide. This study presents a 3D CFD investigation using the level-set method in REEF3D to predict abutment scour evolution under steady current, with results validated against laboratory flume experiments.
Discharge performance of plan view of multi-cycle W-form and circular arc labyrinth weir using machine learning
Labyrinth weirs increase spillway discharge capacity by folding the crest into a zigzag pattern, but predicting their performance is notoriously complex. This study demonstrates that machine learning models — particularly SVM regression — outperform traditional empirical formulae in predicting discharge coefficients for W-form and circular arc labyrinth weirs.