Integrating U-net into full-waveform inversion for salt body building: A challenging case

Abstract

Full-waveform inversion (FWI) applied to regions with large salt bodies often fails without a good initial model, long offsets, and low frequencies. To address this limitation, a previous study embedded U-Nets within a multi-scale FWI to assist in building salt bodies by flooding and unflooding. This approach showed successful results on salt bodies with relatively smooth structures. In this abstract, we test this framework on a more challenging salt body inversion case. We extract seismic velocity slices from the Tiber field in the Gulf of Mexic (GOM) region, where the top of the salt body exhibits a distinctive ”U” shape, like a canyon, with significant variations in depth ranging from 2 km to 6 km. To enable the generalization of neural network, we design training samples that include 1D velocity models with variations in the salt top ranging from 2 km to 6 km. Meanwhile, we use stronger total-variation regularization in FWI to reduce false flooding. Preliminary results from synthetic data indicate that this approach has reasonable potential for complex salt inversion.