Automated Seismic Source Characterisation Using Deep Graph Neural Networks

Abstract

Most seismological analysis methods require knowledge of the geographic location of the stations comprising a seismic network. However, common machine learning tools used in seismology do not account for this spatial information, and so there is an underutilised potential for improving the performance of machine learning models. In this work, we propose a Graph Neural Network (GNN) approach that explicitly incorporates and leverages spatial information for the task of seismic source characterisation (specifically, location and magnitude estimation), based on multi-station waveform recordings. Even using a modestly-sized GNN, we achieve model prediction accuracy that outperforms methods that are agnostic to station locations. Moreover, the proposed method is flexible to the number of seismic stations included in the analysis, and is invariant to the order in which the stations are arranged, which opens up new applications in the automation of seismological tasks and in earthquake early warning systems.

Instantaneous tracking of earthquake growth with elastogravity signals

“…Before being fed to PEGSNet, we first sort the input waveform for each example based on station longitude. We found this approach to be effective, but we note that the problem of concatenating station waveforms in a meaningful way in deep learning is an active area of research 35 . Then, on the basis of the theoretical P-wave arrival time ( T P ) at each station for a given event, we set the amplitude of the seismograms to zero for t ≥ T P .…”

“…In this context, we show that a convolutional neural network (CNN) 30 approach can leverage the information carried by PEGS at the speed of light to overcome these limitations for large earthquakes. Successful applications of deep learning in seismology have provided new tools for pushing the detection limit of small seismic signals 31 , 32 and for the characterization of earthquake source parameters (magnitude and location) 33 – 35 with EEWS applications 29 , 36 , 37 . Here we present a deep learning model, PEGSNet, trained to estimate earthquake location and track the time-dependent magnitude, M w ( t ), from PEGS data before P-wave arrivals.…”

A Little Data Goes a Long Way: Automating Seismic Phase Arrival Picking at Nabro Volcano With Transfer Learning

“…Typically, multiple convolutional layers are applied in succession and in combination with other operations, such as nonlinear "activation," down-sampling and normalization, to extract complex patterns from the data using a hierarchy of simpler filter kernels. These extracted features can then be fed into a standard fully-connected neural network or other machine learning architecture for classification, segmentation, regression, clustering, or inference (e.g., Mousavi et al, 2019;Ross, Meier, Hauksson, & Heaton, 2018;van den Ende & Ampuero, 2020). As such, the "convolutional" part of CNNs act as the model's feature extraction system.…”