A multi-task deep learning scheme using receiver functions to study crustal tectonics and its application to the middle-southern segment of Tanlu Fault Zone and adjacencies

Abstract

We propose a novel scheme that applies a multitasking convolutional neural network to learn the back azimuthal behavior from receiver function seismograms, which can effectively predict the depth and occurrence of the Moho beneath a single seismic station. Our scheme consists of three main steps: 1. Based on the style transfer technique, we generate 9000 synthetic receiver function seismograms blended by realistic noise as training data sets. 2. A multitasking convolutional neural network is trained to predict the depth and occurrence of the Moho. 3. All real receiver function seismograms are reconstructed by the accelerated joint iterative method before prediction. We apply the scheme to study the middle-southern of the Tanlu fault zone and adjacent regions and successfully achieve the depth and occurrence of the Moho beneath 10 permanent seismic stations. The predicted depths are in agreement with the results computed by conventional methods, and the predicted strikes and dip angles present an undulating Moho with near NE-striking. Moreover, the predicted strikes are nearly consistent with the strikes of the normal faults in the upper crust, which implies that intense continental extension in the Cretaceous play a prominent role in the tectonic deformation of the brittle upper crust and the ductile lower crust simultaneously. Besides, it helps to illustrate that the stress field orientation of the major geological event can be recorded and preserved in the lower crust.