Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

A Machine Learning Approach to Developing Ground Motion Models from Simulated Ground Motions

Kyle B. Withers, Morgan P. Moschetti, & Eric M. Thompson

Published August 6, 2019, SCEC Contribution #9347, 2019 SCEC Annual Meeting Poster #016

The USGS is working towards incorporating regionally specific seismic analyses into the U.S. National Seismic Hazard Model. The large dataset of ground motions generated from simulations can serve to supplement empirical data in areas where observed ground motion data is lacking and help to understand trends in intensity where geologic structures complicate seismic wave propagation. Machine learning provides an avenue to incorporate the information from synthetic datasets, without enforcing specific mathematical formulations that ground motion trends must follow, potentially biasing ground motion models (GMMs). One approach to implementing synthetic information into GMMs is to use simulated ground motions to build a GMM for certain distance/period/magnitude ranges. This technique may also constrain GMMs obscured by the large variability in recorded data. To ensure that any conclusions drawn from the synthetic data are accurate, however, it must first be validated with empirical data.

Here, we combine the ground motions from the SCEC’s CyberShake study to build a database of ground motion metrics. We also calculate additional parameters that may help to better describe ground motions trends, i.e. the basin-edge distance and directivity parameters based on fault finiteness. We use an artificial neural network to estimate the weights and coefficients that fit the data, with several formulations of input parameters. We find that a machine learnt GMM fits the synthetic data with much (0.1-0.3 log units) lower variability than empirical relations. This is expected, as the synthetic simulations don’t include the full complexity of site and path effects ground motions experience. When applying our model (simply a table of coefficients as a function of period) to the global PEER dataset across the same magnitude, distance, and period range used to train our model, we find similar values of total variability to GMMs developed from the PEER database. Furthermore, we include source and site location in the neural network to develop a regional model, specific to Southern California, that implicitly includes path and site effects. Although there are few records in this region, we find reduced uncertainty, particularly at longer periods, where directivity and path effects dominate. We highlight this as motivation to continue to simulate ground motions across a wider range of magnitudes and site effects, in order to better resolve trends in the data.

Key Words
machine learning, deterministic ground motion, GMM

Withers, K. B., Moschetti, M. P., & Thompson, E. M. (2019, 08). A Machine Learning Approach to Developing Ground Motion Models from Simulated Ground Motions. Poster Presentation at 2019 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions