Conditioned Simulation of Ground Motion Time Series using Gaussian Process Regression

Abstract

Ground motion time series are critical elements of earthquake engineering for performance analysis of seismic regions’ built environment. At present, the number of available instruments to record the free-field ground motions in the US is generally sparse. Therefore, ground motion estimation methods are used to obtain input motion estimates at locations where there is no available instrumentation. In this study, the ground motion time series are constructed using a Gaussian Process regression, which models the Fourier spectrum’s real and imaginary parts as random Gaussian variables. The proposed model’s training and validation are carried out using the physics-based simulated ground motions of the 1906 San Francisco Earthquake. The evaluation of the model’s performance is also carried out using the simulated magnitude 7.0 Hayward fault earthquake and the ground motions recorded in the 2019 magnitude 7.1 Ridgecrest Earthquake sequence within the Los Angeles area. All evaluations imply that the trained Gaussian Process regression model can estimate the ground motion time series properly. It is also observed that the trained Gaussian Process regression model has decent performance on the long-period ground motion estimation due to the ground motion directivity pulses. The results also illustrate that the stations’ prediction either at the boundary edges or outside of the network might not be as accurate as other stations’ estimations.