← Back to Sessions

The Transformer Earthquake Alerting Model: A Data Driven Approach to Early Warning

Session: Earthquake Early Warning Live in California! Current Status and Challenges I

Type: Oral

Date: 4/23/2021

Presentation Time: 09:45 AM Pacific

Description: 

Traditionally, earthquake early warning relies on relatively simple empirical relationships between observables, source parameters and ground shaking. Examples are relationships between magnitude and early waveform parameters, but also ground motion prediction equations. These simplified assumptions result in apparent aleatoric uncertainty, that materialize as false or missed warnings.

Given the huge amount of waveform data that became available over the last two decades, we suggest that new models do not need to rely on approximate relationships, but can be constructed in a purely data driven manner. Here, we introduce the transformer earthquake alerting model (TEAM), a deep learning based early warning model that is trained to directly make probabilistic predictions of ground shaking based on a few seconds of recorded earthquake waveforms.

We apply TEAM to two large scale strong motion datasets from Japan (13,512 events, 372,661 traces) and Italy (7,055 events, 494,183 traces), regions with complementary seismic hazard. TEAM consistently outperforms two prototypical early warning baselines in terms of precision and recall, as well as warning times, across PGA thresholds from 1%g to 20%g. For example, for 10%g in Japan, TEAM achieves precision/recall of 0.50/0.60, while a point source approach achieves 0.27/0.36 and a PLUM-like approach achieves 0.18/0.39. Average warning times for TEAM are 0.31 s longer than for the point source and 5.01 s longer than for the PLUM-like approach.

While purely data driven models generally perform well when presented with sufficient training data, they often struggle with scenarios with few training examples. Therefore, we also analyze TEAM on scenarios where the largest test event is considerably larger than the largest training event, using the Norcia and Tohoku sequences. For the Norcia sequence, we show that using domain adaptation, TEAM can accurately assess even the largest events. For the Tohoku sequence, TEAM underestimates the main shock, but considerably outperforms the baselines in the aftershock sequence.

Presenting Author: Jannes Münchmeyer

Student Presenter: Yes

Authors