Machine Learning-Based Seafloor Seismic Prestack Inversion

Abstract

Machine learning (ML) is emerging as a new approach for developing algorithms in geophysical research. In this study, a six-layer neural network is designed to estimate seafloor properties using seismic prestack data. In conjunction with the designed neural network architecture, a straightforward network training scheme is developed, which is efficient with two matrix inversions. The trained network can be stored for further use and be applied to multiple data sets, which offers a reduction in the corresponding computational cost. The trained neural network performs the inversion by directly mapping the seismic prestack data to seafloor elastic parameters, and therefore, it does not suffer from the typical problems, such as initial model, convergence, and local minima problems encountered in conventional iteration optimization-based inversion methods. The inversion accuracy of the network is comparable to that of the conventional inversion method, which means that the network architecture and corresponding training scheme proposed in this research are reliable for applications involving seafloor seismic prestack inversion. Numerical analysis verifies the efficacy of the ML-based method.