Prediction of Local Scour Depth Around Cylindrical Bridge Piers in Cohesive Soils Using Advanced Machine Learning Algorithms

Reliable prediction of local scour around cylindrical bridge piers in cohesive beds remains challenging because hydraulic forcing and soil resistance interact nonlinearly. This study evaluates five data-driven approaches—Artificial Neural Network (ANN), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Gene Expression Programming (GEP), and QNET—for predicting the dimensionless maximum scour depth, \(\:{d}ₘₐₓ/D\), using 122 laboratory records selected from the cohesive-bed database compiled by Choi and Choi (2022). Four predictors were examined: water content (\(\:{W}c\)), clay content (CP), pier Froude number Fr = V/\(\:√gD\), and normalized critical shear strength \(\:{τ\:}c/(ρ\:\:{V}²)\). The 15 input patterns are the complete non-empty subsets of these four variables. On the reported 70:30 holdout split, QNET achieved the highest test performance (R² = 0.9958, RMSE = 0.0111, DC = 0.9836), followed by ANN (R² = 0.9206, RMSE = 0.0303, DC = 0.9285). SVM and KNN showed intermediate performance, whereas GEP showed weaker holdout generalization (R² = 0.6791) despite yielding an explicit formula. Parameter-deletion results for SVM, ANN, and GEP identified Fr as a highly influential hydraulic predictor, while QNET showed more distributed dependence among inputs. Because the sample is modest and the available analysis is based mainly on a single holdout split, the near-unity QNET performance is interpreted cautiously. The results demonstrate the potential of machine learning for cohesive-bed scour while emphasizing physically consistent predictors, leakage-free validation, and explicit assessment of model applicability.