Seismic response prediction and variable importance analysis of extended pile-shaft-supported bridges against lateral spreading: Exploring optimized machine learning models

Reliable prediction of seismic responses of bridges against lateral spreading is critical for fragility and resilience assessment of transportation infrastructure. This problem, however, has remained a significant challenge due to the high complexity of the liquefaction phenomenon and its significance for the seismic performance of structures. The present study explores machine learning (ML) approaches, particularly for bridges supported by extended pile-shafts, for reliable estimation of bearing deformation and column drift ratio responses of bridges. The study considers a large set of covariates across soil, structural, and ground motion features. Five ML algorithms are examined including the traditional multiple linear regression (MLR) as the baseline reference, as well as Lasso regression, neural network (NN), random forests (RF), and gradient tree boosting (GTB). A large number of nonlinear soil-bridge finite element models considering the soil and structural uncertainties are dynamically analyzed under 720 ground motions, and the optimal parameters of the ML models are determined via five-fold cross validation. The results indicate that NN and GTB can well predict the seismic responses of the studied soilbridge systems, followed by RF, while Lasso and MLR are generally not able to yield reliable estimates. Furthermore, a variable importance analysis is conducted using the produced regression models. Results indicate that intensity measures (IMs) (particularly the spectral acceleration at 2.0 s) are generally more significant than the soil- and structure-related variables. In addition to the IMs, variables associated with the nonliquefiable crust layer (i.e., thickness, strength, and sloping angle) and the concrete strength and longitudinal reinforcement ratio are generally significant variables, whereas the column diameter and rebar yielding strength are relatively insignificant.