3‐D S Wave Imaging via Robust Neural Network Interpolation of 2‐D Profiles From Wave‐Equation Dispersion Inversion of Seismic Ambient Noise

Abstract

AbstractAmbient noise seismic data are widely used by geophysicists to explore subsurface properties at crustal and exploration scales. Two‐step dispersion inversion schema is the dominant method used to invert the surface wave data generated by the cross‐correlation of ambient noise signals. However, the two‐step methods have a 1‐D layered model assumption, which does not account for the complex wave propagation. To overcome this limitation, we employ a 2‐D wave‐equation dispersion (WD) inversion method which reconstructs the subsurface shear (S) velocity model in one step, and elastic wave‐equation modeling is used to simulate the subsurface wave propagation. In the WD method, the optimal S velocity model is obtained by minimizing the dispersion curve differences between the observed and predicted surface wave data, which makes WD method less prone to getting stuck to local minima. In our study, the observed Scholte waves are generated by cross‐correlating continuous ambient noise signals recorded by each ocean bottom node (OBN) in the 3‐D Gorgon OBN survey, Western Australia. For every two OBN lines, the WD method is used to retrieve the 2‐D S velocity structure beneath the first line. We then use a robust neural network (NN)‐based method to interpolate the inverted 2‐D velocity slices to a continuous 3‐D velocity model and also generate a corresponding uncertainty model. Moreover, we compared the predicted dispersion curves and waveforms to the observed data, and a robust waveform and dispersion match are observed across all of the Gorgon OBN lines.