Construction of Site-Specific Ground Motion Model Based on Multi-Output Gaussian Process

Abstract

Site-specific ground motion prediction using insufficient number of observed data at each site is a challenge issue. This paper proposes a framework to construct a site-specific ground motion model using the multi-output Gaussian process technique that makes it possible to predict site-specific ground motion in case of insufficient number of observed records at each site by utilizing the information obtained from records at other sites. In this paper, first, a proposed framework is validated using simulated records of hypothetical crustal earthquakes. The estimation accuracy is estimated depending on the number of sites and the number of observed records at each site. When earthquake locations are uniformly distributed and their magnitudes follow the Gutenberg-Richter relation, it is confirmed that the characteristics of site-specific ground motions can be accurately predicted over any distance range, provided there are at least nine stations and each station has at least 50 observed records. Then, the framework is applied to observed records to discuss the characteristics of site-specific ground motion. The results suggest that site-specific seismic motion predictions could be made using multi-output Gaussian process regression. In the ranges with ample records, the estimated uncertainty varied from site to site, with a logarithmic standard deviation between 0.2 and 0.4. The results also indicate an increase in uncertainty at close distances, which corresponds to statistical errors due to insufficient records.