Application of machine learning in seismic array processing

Abstract:

Back-projections of large earthquakes using local seismic arrays are useful for rapid estimation of the source parameters and are thus suitable for earthquake and tsunami early warning. The source slowness vector determined by individual arrays can be combined to resolve the timing and location of the rupture fronts and hence the rupture history. One of the challenges of local back-projections is the slowness error due to complicated seismic phases and horizontal velocity heterogeneities. To reduce such errors, we design a convolutional neural network (CNN) to detect the wavefronts and predict the slowness vector using local array recordings. The neural network consists of ten convolutional layers and one fully-connected layer. The input dataset is a movie of the seismic wavefield recorded by the local array in map view. The output images predict 2-D probability distribution of source locations at each time step. We consider two sets of training labels based on event location probability distribution and the slowness vector predicted by event locations. We test the performances of our model using the aftershocks of the 2019 Ridgecrest earthquakes. We use the recordings of a seismic array consisting of 48 strong motion stations (epicentral distance ～2.5°) in southern California as empirical Green’s function (EGFs). The training data set is composed of 26,400 synthetic events evenly distributed in the source domain. Each synthetic event is generated by adjusting the travel-times of 11 EGFs according to the event locations. The results show that the azimuthal bias is approximately 4° for an event located 200 km away, which is comparable to the uncertainty of traditional parametric back-projections. Our preliminary results demonstrate the potential advantage of deep-learning-based array processing. We will also explore applying the Recurrent Neural Network (RNN), which is more suitable for temporal behavior. We envision incorporating more real earthquake data in the training data set may improve the accuracy and reduce the processing time of automatic back-projection for earthquake early warning purposes.