Quantifying Easy-to-Repair Displacement Ductility and Lateral Strength of Scoured Bridge Pile Group Foundations in Cohesionless Soils: A Classification–Regression Combination Surrogate Model

Scoured pile-group foundations in bridges are likely to undergo inelastic deformation during earthquakes, which can be utilized to dissipate seismic energy and withstand seismic loads by the post-yield hardening strength of the foundation. However, limit states and associated ductility indices and post-yield strength indices are yet to be well documented. This study develops a surrogate model, namely, the classification–regression combination model (CRCM), for the efficient, interpretable, and high-confidence quantification of displacement ductility factor $({\mu }_{\mathrm{\Delta }}^{\mathrm{ER}})$ and associated strength hardening factor $(R_F^{\mathrm{ER}})$ of scoured bridge pile-group foundations at the easy-to-repair limit state, where the damage of piles is limited to the aboveground region (thereby being relatively easy to repair). To this end, a proper pushover method from those with different load patterns is first identified for efficient nonlinear analyses of scoured bridge pile groups. A large number of bridge samples are then analyzed to prepare a comprehensive database for the development of CRCM, which first classifies the failure process of scoured bridge pile-group foundations and then regresses ${\mu }_{\mathrm{\Delta }}^{\mathrm{ER}}$ and $R_F^{\mathrm{ER}}$ with variables characterizing the soil–bridge systems. It is found that the pushover method with a two-node load pattern (i.e., load at the superstructure and pile-cap centroids) can very well capture ${\mu }_{\mathrm{\Delta }}^{\mathrm{ER}}$ and $R_F^{\mathrm{ER}}$ and the failure process of bridge pile groups. The data-driven CRCM can efficiently provide reasonable predictions of ${\mu }_{\mathrm{\Delta }}^{\mathrm{ER}}$ and $R_F^{\mathrm{ER}}$ with errors mostly within 20%; it is specifically compared with a regression-only model to demonstrate the necessity of incorporating a classifier in advance of the regression model.