The Transformer Earthquake Alerting Model: Improving Earthquake Early Warning with Deep Learning

Abstract:

Earthquakes are major hazards to humans, buildings and infrastructure. A key method for reducing earthquake damage is early warning. Here we present the transformer earthquake alerting model (TEAM), a real time deep learning model achieving state of the art performance for early warning. Traditional early warning approaches rely on one of two approaches: estimating the source properties or modelling the propagation of shaking. In constrast, the deep learning approach of TEAM allows the model to dynamically combine a global view of the event with local measurements.

We evaluate TEAM on two large scale strong motion datasets from Japan (13,512 events, 372,661 traces) and Italy (7,055 events, 494,183 traces). We assembled these datasets for this study and make them publicly available for future research and comparable analysis. Due to the different hazard scenarios between the datasets, the combination allows for a comprehensive evaluation of early warning algorithms.

We evaluate the relation between true alerts, false alerts and missed alerts. For PGA thresholds from 1%g to 20%g, TEAM considerably outperforms two state of the art early warning methods, one using source estimation, one using propagation modeling. Furthermore, TEAM offers excellent warning times, performing better or at least on par with an idealistic point source baseline. On the Italy dataset we conduct a particularly challenging experiment, by evaluating TEAM on events considerably larger than any training event, thereby testing its extrapolation capabilities. We show that using tailored domain adaptation, TEAM can assess even those events accurately.

To achieve this early warning performance, TEAM overcomes multiple challenges of previous deep learning models. First, by using a transformer network, it incorporates data from a variable set of stations and explicitly encodes their locations. This enables the application to heterogeneous datasets and stations not present during training. Second, TEAM gives real time predictions, by using a combination of zero-padding and training data augmentation. Third, TEAM provides accurate predictions even for large events, achieved through a combination of training data augmentation and domain adaptation strategies.