Benchmarking the GRAPES EEW Model on a Suite of California Earthquakes

Vishal Gondi, & Tim Clements

Published September 8, 2024, SCEC Contribution #13961, 2024 SCEC Annual Meeting Poster #190

Current Earthquake Early Warning (EEW) systems rely on calculations of earthquake source parameters, such as magnitude and location, to predict ground shaking, which can take seconds and thus delay warning. However, the Graph Prediction of Earthquake Shaking (GRAPES) EEW algorithm is a deep learning model that utilizes acceleration waveforms processed through a graph neural network to enhance prediction capabilities. This model is specifically designed to improve early warning accuracy and timeliness by incorporating spatial relationships between seismic stations and their waveform data. The GRAPES model processes three key inputs: a graph of station nodes linked by neighboring connections, acceleration waveforms segmented into 4-second windows and formatted into a 4-dimensional tensor, and current maximum acceleration (PGA) values for each station. The GRAPES model was trained and tested on a suite of Japanese earthquakes (Clements et al., 2024). Here we will benchmark the model, unmodified, on its ground motion accuracy and warning time performance using a test suite of 40 M4+ California earthquakes. In our initial testing of the 2019 M7.1 Ridgecrest, CA earthquake, which had a maximum Modified Mercalli Intensity (MMI) of 9, the GRAPES model warned all stations requiring an alert at a MMI threshold of 4.0, with warning times increasing linearly from zero seconds at the station closest to the epicenter to 30+ seconds at 100 km epicentral distance. While the Ridgecrest test demonstrates the model’s potential to improve the timeliness of earthquake early warnings for large earthquakes (M7+) in California, we will show performance results on additional, moderate magnitude earthquakes (4<M<6) in our dataset, which are the most frequent triggers of EEW systems.

Citation
Gondi, V., & Clements, T. (2024, 09). Benchmarking the GRAPES EEW Model on a Suite of California Earthquakes. Poster Presentation at 2024 SCEC Annual Meeting.


Related Projects & Working Groups
Ground Motions (GM)