Inverting Rayleigh surface wave velocities for crustal thickness in eastern Tibet and the western Yangtze craton based on deep learning neural networks

Abstract

Abstract. Crustal thickness is an important factor affecting lithospheric structure and deep geodynamics. In this paper, a deep learning neural network based on a stacked sparse auto-encoder is proposed for the inversion of crustal thickness in eastern Tibet and the western Yangtze craton. First, with the phase velocity of the Rayleigh surface wave as input and the theoretical crustal thickness as output, 12 deep-sSAE neural networks are constructed, which are trained by 380 000 and tested by 120 000 theoretical models. We then invert the observed phase velocities through these 12 neural networks. According to the test error and misfit of other crustal thickness models, the optimal crustal thickness model is selected as the crustal thickness of the study area. Compared with other ways to detect crustal thickness such as seismic wave reflection and receiver function, we adopt a new way for inversion of earth model parameters, and realize that a deep learning neural network based on data driven with the highly non-linear mapping ability can be widely used by geophysicists, and our result has good agreement with high-resolution crustal thickness models. Compared with other methods, our experimental results based on a deep learning neural network and a new Rayleigh wave phase velocity model reveal some details: there is a northward-dipping Moho gradient zone in the Qiangtang block and a relatively shallow north-west–south-east oriented crust at the Songpan–Ganzi block. Crustal thickness around Xi'an and the Ordos basin is shallow, about 35 km. The change in crustal thickness in the Sichuan–Yunnan block is sharp, where crustal thickness is 60 km north-west and 35 km south-east. We conclude that the deep learning neural network is a promising, efficient, and believable geophysical inversion tool.

Surface Wave Inversion Method Based on Convolutional Neural Network and Support Vector Regression Machine

“…Generally, the use of CNN for surface wave inversion of dispersion curves is for the regression tasks, and the commonly used loss function for regression tasks is the mean square error (MSE) (Cheng et al, 2019;Meier et al, 2007). The loss function is generally de ned as:…”

“…Through the CNN-SVR inversion algorithm, 2×10 5 input dispersion data and output shear wave velocity data samples are used as the training set, and 2×10 4 data samples are used as the validation set, and all data sets are not repeated. In this CNN-SVR model, when we trained the CNN model to improve the training speed and relative accuracy of the network, we chose a custom adam function as the network optimizer.…”

Semi‐Supervised Surface Wave Tomography With Wasserstein Cycle‐Consistent GAN: Method and Application to Southern California Plate Boundary Region

“…The supervised learning such as convolutional neural networks (CNN) based methods have been widely utilized in geophysical studies. The neural networks approach has been proven to be promising in surface wave studies, for instance, extraction of crustal thickness (Cheng et al., 2019; Devilee et al., 1999; Meier et al., 2007) from surface wave data, and automatic surface wave travel time dispersion picking (e.g., Zhang et al., 2020).…”

Semi-supervised Surface Wave Tomography with Wasserstein Cycle-consistent GAN: Method and Application on Southern California Plate Boundary Region

“…The supervised learning such as convolutional neural networks (CNN) based methods have been widely utilized in geophysical studies. The neural networks approach has been proven to be promising in surface wave studies, for instance, extraction of crustal thickness (Cheng et al, 2019;Devilee et al, 1999;Meier et al, 2007) from surface wave data, and automatic surface wave travel time dispersion picking (e.g., Zhang et al, 2020).…”