A Surrogate Model for Predicting Hysteresis Model Parameters for Earthquake Simulation of Elevated Transportation Infrastructure

Quick estimation of hysteresis model parameters is a necessary step when simulating the earthquake response of large inventories of structures. To this end, this paper introduces a Gaussian process (GP) based workflow aimed at predicting nonlinear model parameters to simulate the hysteresis behavior of reinforced concrete column plastic hinges. The workflow is presented through an application using a modified Bouc-Wen model. The GP is trained using a dataset of 290 experimental cyclic tests, ranging from columns with stable flexural failure modes to tests failing in shear with significant pinching and strength decay during cycling. Each test is characterized by a set of physical dimensionless predictors and a set of best-fit BW model parameters, obtained from calibrations performed with the Transitional Markov Chain Monte Carlo (TMCMC) algorithm. The BW model was enhanced with a new pinching initiation rule to better cover the multiple hysteresis shapes present in the database. Two GP surrogate models were then trained to map the non-dimensional parameters to the BW model parameters, one for flexure-dominated and another for shear-influenced failure mode. The model performance is evaluated through comparison of the hysteresis curves obtained using the GP-predicted BW model parameters, the curves resulting from the direct calibration procedure, and the real data observed in the experiments with a cross-validation procedure, holding out points in the dataset for testing. Overall, the GP surrogate model exhibits similar performance, in terms of mean average error in the hysteresis curves, as the direct calibration results from the TMCMC algorithm. Better performance is achieved for flexural failure modes than shear failure modes, due to the high variability of the pinching parameters across the dataset, and sparser data when it comes to shear-influenced tests. A sample application is developed to show how the GP surrogate model can be used to rapidly generate medium-fidelity non-linear structural models to simulate the response of cantilever concrete columns from a few easy-to-compute geometric parameters as inputs, and how the responses compare to a higher-fidelity FEM model. The introduced workflow and the presented model were developed with the specific aim of enabling medium-fidelity simulation of structural responses for a large, elevated transportation infrastructure inventory through time-history analyses, for which quick model creation and efficient simulations are needed.