AD-GRT Deep Learning Waveform Inversion

This paper proposes a novel deep learning-based waveform inversion method that enables rapid and high-precision extraction of nonlinear waveform information from seismic data. To address critical challenges in existing deep learning full waveform inversion (FWI), such as poor network generalization and huge computational cost, we introduce angle-domain generalized Radon transform (AD-GRT). As intermediate data, AD-GRT generated angle-domain gathers are equivalent to the original seismic data. Instead of seismic data, the angle-domain gathers serve as input for the new waveform inversion deep network. To ensure the waveform nonlinearity of the angle-domain gathers is local, this method requires a basically accurate background velocity model to guarantee that the contour lines in angle-domain gathers have no significant curvature, and avoids inversion in areas covered by strong reflectors like salt. Relying on the original data equivalence and the local waveform nonlinearity of the angle-domain gathers, we developed a horizontally layered model-based CNN network for local nonlinear waveform mapping between angle-domain gathers and true models, achieving a highly accurate and strongly generalized waveform inversion network with minimal training data. The generalization and accuracy of the CNN network are validated using random layered models. Furthermore, the CNN network is directly extended to 2D media waveform inversion. Three complex 2D models confirm the high accuracy, stability and robustness of the CNN network. Additionally, the computational cost of the CNN network waveform inversion is negligible, compared with a single finite-difference simulation.