Donnerstag, 10.03.2022 (14:00 - 15:00 Uhr)

Greg Beroza

(Stanford University, Stanford, United States)

Machine learning is having an impact in nearly all aspects of seismology ranging from forward simulations to seismic imaging to interpretation.  To date it is most well-established, and is having the greatest impact, in earthquake monitoring.  While some of the trends in using machine learning for earthquake monitoring may play out in other applications of machine learning to seismology, there are good reasons that explain with it has been so successful for this purpose.
There is a well-established sequence of tasks: phase detection, phase association, event location, and event characterization, used to develop seismicity catalogs around the world and across scales. Because the number of earthquakes is universally observed to increase rapidly as magnitude decreases, cataloging somewhat smaller earthquakes will dramatically increase the information available.  Appropriate architectures and data augmentation play important roles in developing effective models that generalize well, but access to large, accurately labelled data sets is also critical.  The large existing earthquake catalogs and their corresponding waveforms make seismic monitoring an ideal use case for supervised machine learning.
The simplest approach to earthquake monitoring is modular in which individual earthquake monitoring tasks are replaced one-by-one with neural network models that are applied in serial.  There are advantages, however, in combining steps in multi-task models, with a complete end-to-end model as an extreme end member, to take advantage of contextual information.  AI-based earthquake monitoring is now being deployed for real-time applications, and there is no reason for it not to be applied comprehensively to available archived data.  It is allowing seismologists to develop catalogs for both natural and induced seismicity that are far more comprehensive and information-rich.  The next challenge will be to use this more complete view of seismicity to understand better the mechanics of earthquake processes.  The methods of AI should also be useful for this effort.