Simultaneous inversion of velocity and vector reflectivity based on the reccurent neural network

Xiaohui Cai, Tianze Zhang, Kristopher A. Innanen, Daniel O. Trad

This study presents an advanced simultaneous inversion method that integrates velocity and vector reflectivity models using a Recurrent Neural Network (RNN) framework. By combining Full Waveform Inversion (FWI) and Least Squares Reverse Time Migration (LSRTM), this approach addresses key limitations of each method. While traditional LSRTM retrieves reflection coefficients at a 0° reflection angle, the introduction of vector reflectivity models enables simultaneous inversion at 0° and 90°, essential for constructing common-angle gathers and improving data utilization efficiency.

FWI excels in low-wavenumber velocity reconstruction but struggles with high-wavenumber details, whereas LSRTM resolves high-wavenumber reflectivity but requires an accurate migration velocity model. The proposed method leverages RNN's sequential learning and wave propagation physics to achieve joint optimization of velocity and vector reflectivity through a three-parameter inversion process. Synthetic tests using the SEG/EAGE overthrust model demonstrate the method's ability to recover detailed velocity structures and high-resolution reflectivity efficiently. These results highlight the potential of RNN-based techniques to enhance simultaneous inversion, offering precise subsurface characterization in complex geological settings.