Application of Machine Learning for Identification of Hidden Rock Sites Using Earthquake Records

Earthquake ground motions recorded at rock sites can be used to reduce seismic hazard uncertainty for dams and other critical infrastructure on rock; however, few ground motion records exist from known rock sites because most seismic stations are on soil or lack measured shear wave velocities. In this study, a dataset of California ground motions is examined to identify additional instrumented “Very Dense Soil” and “Soft Rock” sites which may actually be “Rock.” This approach offers a cost- and time-efficient alternative to installing new seismic stations. To identify possible rock sites, partially-crossed linear mixed-effects regression is used to compute site residuals from empirical ground-motion model predictions and recorded ground motions. Based on these site residuals and other site characteristics, machine learning is then applied to estimate shear wave velocities at sites without measured velocities. Using these estimated shear wave velocities, a ranked list of candidate rock sites is developed. Geophysical testing is proposed at a selection of these sites to verify the high estimated shear wave velocities, with a focus on sites with many recorded ground motions, as an efficient means to expand the catalogue of rock ground motions.