Testing the US West Coast ShakeAlert Earthquake Early Warning System in Complex Earthquake Sequences

We test the performance of the ShakeAlert Earthquake Early Warning System using synthetic, composite complex earthquake sequences to understand the current limitations of the algorithms used. ShakeAlert is currently based on two seismic algorithms, EPIC and FinDer. Other algorithms, such as PLUM, are being considered for future integration. Source parameter estimates from the algorithms are combined by a Solution Aggregator (SA) that passes on information to the Decision Module for alert distribution if certain criteria are met. Here, we develop composite event data sets by combining signals from two known events (building blocks) to test the behavior of ShakeAlert during complex earthquake sequences. The goal is to better understand where the current system does not perform well and how some aspects of the algorithms can be improved. We examine four event types: (1) foreshock/aftershock sequences, (2) near-simultaneous events in different source regions, (3) simulated off-shore earthquakes, and (4) seismic swarms. We synthesize the first three sets by summing up earthquake signals recorded at the same station with different time shifts (-60 to 180 s) and scaling up select signals with a frequency-dependent transfer function to larger magnitudes. We generate the third data set from the records of the 2019 M7.1 Ridgecrest earthquake and iteratively remove stations near the epicenter to mimic offshore scenarios where epicenters move further and further from the coast (ranges 25 - 175 km). We find that the ShakeAlert algorithms perform largely as expected, but we identify challenging scenarios where close foreshocks lead to missed or mis-associated alerts for the largest events. A few critical improvements to the seismic algorithms are identified. The current SA does not optimally leverage the algorithms tested here, most notably the contributions from PLUM. Additionally, since each algorithm behaves differently in handling complex sequences, we can explore ways to refine the association configurations. We find moving to ground-motion-based alert association and aggregation will likely obviate some of the identified challenges and produce more robust ShakeAlert warnings.