## Group A, Poster #021, Seismology

### Earthquake Early Warning with Graph Learning

Poster Image:

#### Poster Presentation

2022 SCEC Annual Meeting, Poster #021, SCEC Contribution #12131 VIEW PDF

We introduce a graph learning-based algorithm for earthquake early warning (EEW). The goal of EEW is to characterize when and where strong ground motion is expected as soon as possible after an earthquake starts. EEW systems currently in use across the world either rapidly estimate earthquake source properties (location and magnitude) then predict shaking using ground motion attenuation relations, or forward predict ground motion as it travels across a seismic network. Ground motion-based EEW algorithms are robust during aftershock sequences but provide shorter warning times whereas source-based EEW algorithms offer longer warning times at further epicentral distance but often fail during in...tense aftershock sequences.

We take a different approach – formulating EEW as a graph learning problem, in which we learn how ground motion spreads across a seismic network in space and time. Here, the nodes of the graph are seismometers, the edges of the graph are a seismometer’s k-nearest neighbors in the seismic network, the features at each node in the graph are real-time seismic waveforms, and the prediction target for the graph is the peak ground acceleration (PGA) at each station (node) in the network over the next 15 seconds. We train a model composed of a set of convolutional, fully connected, and graph neural network (GNN) layers in an end-to-end fashion to predict ground motion across a seismic network in real-time. We train and test our model on earthquake and ambient noise data from the KiK-net and K-NET strong motion networks in Japan. Once trained, our graph network model can predict future ground motion for seismic networks of any size, ranging from tens to thousands of seismic stations. We examine whether a graph learning-based approach to EEW can readily be applied in real-time and predict high PGA values sooner than source or ground-motion-based EEW algorithms.

SHOW MORE

We take a different approach – formulating EEW as a graph learning problem, in which we learn how ground motion spreads across a seismic network in space and time. Here, the nodes of the graph are seismometers, the edges of the graph are a seismometer’s k-nearest neighbors in the seismic network, the features at each node in the graph are real-time seismic waveforms, and the prediction target for the graph is the peak ground acceleration (PGA) at each station (node) in the network over the next 15 seconds. We train a model composed of a set of convolutional, fully connected, and graph neural network (GNN) layers in an end-to-end fashion to predict ground motion across a seismic network in real-time. We train and test our model on earthquake and ambient noise data from the KiK-net and K-NET strong motion networks in Japan. Once trained, our graph network model can predict future ground motion for seismic networks of any size, ranging from tens to thousands of seismic stations. We examine whether a graph learning-based approach to EEW can readily be applied in real-time and predict high PGA values sooner than source or ground-motion-based EEW algorithms.

SHOW MORE