Two talks: Automatic Analysis of the Maule, Chile, Aftershock Sequence using AI Techniques & Advancing earthquake event detection pipelines through machine learning approaches.
Geb. 06.42 - Raum 001 (Seminarraum) / Online
Ankitha Pezhery & Jack Woollam
GPI - KIT
Ankitha Pezhery received her Bachelors degree in Physics from Calicut University, India. Since 2019, she is a Masters student in the Geophysics department, KIT.
Jack Woollam is a research geophysicist, currently undertaking his PhD at Karlsruhe Institute of Technology. His research focuses on the statistical analysis of large-scale datasets to improve understanding of earthquake processes. In particular, he seeks to apply machine learning and deep learning routines to extensive seismic datasets to better image faults and their rupture mechanisms.
Seismic phase picking and subsequent event association are inevitable steps of a seismic processing workflow for unravelling the characteristics and structure of Earth’s subsurface with the help of earthquake data. With the increasing deployment of denser seismic networks rendering enormous volumes of frequent seismic data, manual phase picking and automated approaches are falling behind, but the field of seismology is taken by storm at the possibility of using AI techniques to develop efficient phase picking and event association algorithms. In addition to working with data from latest large N array seismic networks, such novel techniques help us to reinvestigate notable highly active seismic sequences already recorded. This study is an attempt to better understand the aftershock sequence of 2010 Maule, Chile earthquake of Mw = 8.8, one of the largest subduction zone earthquakes that was monitored by a dense seismic network, by adopting latest machine learning developments in seismic processing piplelines. We applied a deep neural network based phase picker, PhaseNet and subsequently Hyperbolic Event Extractor (HEX) as an event association algorithm on Maule dataset. To demonstrate if the event associator can keep pace with the vast amounts of picks detected by the phase picker and to tweak associator’s configuration parameters, we developed a synthetic generator routine which was tested for homogeneous and 1D velocity models of this region. We hope that this synthetic data generator along with the results of this study will prove fruitful to better understanding the relative benefits of such novel seismic event detection routines over traditional approaches.
High-resolution earthquake locations are one of the most fundamental components in imaging subduction zone interfaces. The accurate association of seismic events in these regions is vital to understanding the physical processes associated with the largest earthquakes on record. The last decade has seen the continued increases in the scale of both onshore and offshore seismic deployments, aiming to better capture the variety of seismic processes occurring throughout these dynamic environments. Typical approaches for earthquake detection typically involve combinations of traditional automated methods (e.g STA/LTA picking) and manual refinement by a human expert. The trade-off in performance and accuracy afforded by this hybrid approach can result in under-exploited catalogs, when compared to a human expert exhaustively identifying and associating all arrivals manually. In recent years, Machine Learning (ML) methods have shown significant improvement in the task of automatically detecting seismic events. ML methods now potentially operate at a similar standard to a human expert, whilst running orders of magnitude more efficiently than comparable techniques. Here, we show the advances made when applying these latest machine learning-based algorithms to process continuous seismic data across seismic networks in practice.