Machine learning prediction of strain dependent soil damping improves seismic analysis of bridge foundations

Abstract

Conventional empirical soil damping models often neglect strain-dependent, site-specific variability in heterogeneous alluvial deposits (sands, silts, gravels), leading to uncertainty in seismic bridge foundation design. This study develops a Random Forest (RF) model to predict strain-dependent damping from routine geotechnical parameters (SPT-N₆₀, D₅₀, uniformity coefficient, curvature coefficient, fines content, and plasticity index) and integrates it into nonlinear soil-pile-structure interaction (SPSI) simulations. A physics-informed dataset of 2,340 soil-layer records from ten bridge foundation sites was compiled and used for model training and validation. The RF model achieves cross-validation R² = 0.88, RMSE = 0.52%, and MAE = 0.35%, outperforming XGBoost (R² = 0.85) and ANN (R² = 0.81). Feature importance analysis identifies shear strain amplitude as the dominant predictor (42%), followed by SPT-N₆₀ (18%) and fines content (12%), consistent with established soil dynamics theory. Compared to conventional empirical damping assumptions (Seed & Idriss, 1970; Vucetic & Dobry, 1991), ML-based damping reduces prediction error in pile bending moment from 26% to 6% (20% points) and lateral displacement from 29% to 8% (21% points). The proposed framework enables site-specific seismic performance assessment of bridge foundations in heterogeneous alluvial deposits using only routine geotechnical investigation data, without requiring specialized dynamic laboratory testing.