Data-Driven Ground Motion Synthesis Using Deep Generative Models

Session: Near-Surface Effects: Advances in Site Response Estimation and Its Applications [Poster]
Type: Poster
Date: 4/28/2020
Time: 08:00 AM
Room: Ballroom
Description: 

The number and complexity of ground motion prediction equations (GMPEs) has grown exponentially in recent years due to the increased availability of data. GMPEs are crucial for building codes and engineering design; they allow for the prediction of a set discrete intensity measures that approximate the ground motions generated by earthquakes at a range of magnitudes and distances from the site. In spite of their wide adoption, GMPEs do not capture the full acceleration spectrum and, in many cases, lead to poor approximations of the strong motion time series. Also, given their empirical nature, it has become increasingly challenging to calibrate them, typical regression models involve a set of linear and nonlinear terms, whose particular functional forms can be subjective. Here we present a new framework for ground motion synthesis; instead of attempting to predict a limited set of intensity measures, we generate the full broadband 3-component strong-motion time series for a wide range of magnitudes, distances and site characteristics. In particular, we train Generative Adversarial Networks (GANs), conditioned on magnitude, distance and site response, to learn the inherent probability distribution of a massive set of strong-motion recordings from Japan. We are able to synthesize 20 second-long waveforms that capture most of the details of the acceleration spectrum and are therefore suitable for ground motion prediction tasks.

Presenting Author: Manuel A. Florez

Authors