← Back to Sessions

Forecasting Ground Motion Intensity Time Series with a Generative Pre-trained Transformer

Description: 

Typically, earthquake forecasts project the expected rate or number of earthquakes in the hours to weeks following a large earthquake. While grounded in the statistics and physics of seismogenesis, earthquake forecasts inherently do not include the complexities of ground motion physics. Here, we develop a ground motion-based approach to earthquake forecasting, which we call ground motion forecasting. Rather than forecast the rate or number of earthquakes and apply a ground motion model, we directly forecast future ground motion at a particular location from recorded ground motion. To do so, we introduce the Generate Unsupervised Aftershock Velocity Amplitudes (GUAVA) model. GUAVA is a generative pre-trained transformer with ~100 million model parameters that generates time series of ground motion intensity autoregressively at 10 Hz sampling rate with recorded ground motion time series as input. We train GUAVA on all continuous ground motion intensity time series (~225 stations) recorded by the Southern California Seismic Network within 2 degrees of the epicenter of 2019 M7.1 Ridgecrest, CA earthquake sequence from July 1-10, 2019. The GUAVA training is designed to maximize the log-likelihood of the next ground motion amplitude A(t+1) given the previous 8,192 (2^13) ground motion samples (A(t), A(t-1), .., A(t-8192) – effectively the last 819.2 seconds of data at 10 Hz) using stochastic gradient descent. When given a new ground motion time series as input, GUAVA generates suites of aftershock ground motion time series – complete with body, surface, and coda waves. We evaluate GUAVA’s performance using the Kullback–Leibler divergence between the predicted and true distribution of future ground motion intensities on the December 5, 2024, M 7.0 Offshore Cape Mendocino earthquake. We suggest that GUAVA could be run in real-time using streaming seismic data to forecast ground motion from seconds to minutes after intense shaking occurs, as a type of Earthquake Early Warning for aftershock ground motion.

Session: Improving the State of the Art of Earthquake Forecasting Through Models, Testing and Communication [Poster]

Type: Poster

Date: 4/15/2025

Presentation Time: 08:00 AM (local time)

Presenting Author: Tim

Student Presenter: No

Invited Presentation: 

Poster Number: 132

Authors