Data-driven modeling of local scour in cohesive soils: a random forest approach with sensitivity to clay content and moisture

Local scour at bridge piers in cohesive soils remains a critical challenge due to the complex, nonlinear interactions among hydraulic, geometric, and geotechnical factors, particularly clay content C p and bed moisture W p . Traditional empirical models, often developed for non-cohesive beds, fail to capture the conditional behavior of cohesive sediments, which exhibit peak resistance at optimum moisture and brittleness at higher saturation. This study employs the Random Forest (RF) algorithm to predict relative scour depth (d s /D) using over 300 high-quality laboratory datasets encompassing clay percentages (20–100%), moisture contents (20–46%), pier shapes (circular, square, rectangular, rounded-square), and both contracted and uncontracted flow conditions. Dimensionless parameters including Froude number, dimensionless shear stress, and a dynamic pier shape factor were used to enhance physical consistency and model generalizability. Data were intelligently split (75%–25%) while preserving statistical distributions across scenarios. Model performance was rigorously evaluated using R 2 , RMSE, and MAE, and benchmarked against nonlinear regression. RF achieved excellent accuracy, especially for circular piers under contraction (R 2 = 0.96, RMSE = 0.04), while reduced performance for sharp-edged piers (e.g., square piers: R 2 = 0.64) was attributed to highly turbulent, three-dimensional flow separation rather than algorithmic limitations. Sensitivity analysis confirmed that clay content (C p ) and moisture (W p ) are the dominant predictors, omitting them increased prediction error by over 40%. Crucially, modeling the shape factor as a function of C p , W p , and contraction ratio significantly improved accuracy. The study demonstrates that integrating dimensionless hydraulic theory with machine learning yields a robust, assumption-free framework for scour prediction in cohesive environments, substantially outperforming conventional methods and providing actionable insights for resilient bridge foundation design.