Physics Prior Constrained ResUNet-BiMamba Network for Prestack AVO Inversion
Prestack seismic data contain rich angle-dependent amplitude variation information, which provides an important basis for the simultaneous inversion of P-wave velocity, SV-wave velocity, and density. However, prestack inversion based on the Aki-Richards approximation is applicable only to relatively small-to-moderate incidence angles (<30°). Although the exact Zoeppritz equation is valid for gathers with relatively large incidence angles, it couples the reflection and transmission of PP- and PS-waves and establishes a strongly nonlinear relationship between the PP-wave reflection coefficient and the three parameters, which is difficult to solve using conventional methods. Building on the exact Zoeppritz equation, this article proposes a nonlinear prestack three-parameter inversion method based on a Residual U-Net (ResUNet)-BiMamba network. A bidirectional Mamba structure is introduced into prestack seismic inversion and combined with the multiscale feature extraction capability of ResUNet, effectively suppressing stripe artifacts in the inverted profiles and improving interface continuity in structurally complex areas. The forward modeling process based on the Zoeppritz equations is embedded into the hybrid network, and a masking function and Huber loss function are introduced to reduce the influence of strong-amplitude anomalies on the inversion results. A semi-supervised joint loss function is constructed by combining a low-frequency consistency term with a well log supervised loss, thereby reducing the dependence on labeled data. Tests on the Marmousi2 model and field data demonstrate that, compared with the Mamba and BiMamba network models, the ResUNet-BiMamba network model produces superior inversion results in terms of stratigraphic boundary characterization, lateral continuity, and identification accuracy.