Realistic Synthetic Broadband Ground Motions by Machine Learning

Date: 4/24/2019

Time: 11:00 AM

Room: Elliott Bay

High frequency (f > 1 Hz) ground motions play an important role in seismic hazard for most structures; however, accurately simulation of high frequency ground motion requires details of crustal structure that we do not have and the computation expense of such simulations grow rapidly with frequency. At present, seismologists can only deterministically simulate ground motions with acceptable accuracy up to 1 Hz; above that, ground motion is typically simulated using stochastic or semi-stochastic methods. Here, we attempt to simulate high-frequency ground motions using a machine- learning-based approach. Because the low frequency component (f < 1 Hz) can be physically modeled, we train a deep neural network that takes input of the low frequency component and generates the broadband ground motions from it. We train the network using low frequency waveforms as input, so that the source, path and site effects are all included. The training and test data consist of broadband waveforms of 4,700 M>3 local earthquakes recorded by the Caltech/USGS Southern California Seismic Network (SCSN) from 2000 to 2018. Specifically, the 1-Hz lowpass filtered waveforms serve as input to the network, while the 20-Hz lowpass filtered waveforms are the prediction targets. We evaluate the accuracy of the synthetic broadband waveforms by comparing their key characteristics to a validation subset of the field recordings that were not used to train the network. This provides an alternative approach to generate realistic synthetic broadband ground motions across the frequency range of primary earthquake engineering interest.

Presenting Author: Zefeng Li

Authors