High-Wavenumber Extrapolation for Band-Limited FWI Results Using Deep Learning

Abstract

A reliable estimation of high-resolution subsurface properties is a very challenging task for seismic inversion. Seismic data and well logs can jointly contribute to this task by combining their advantages. We have developed a deep learning method to extrapolate high-wavenumber model components for band-limited full waveform inversion (FWI) results by incorporating high-resolution well-log velocity information. A U-Net architecture, that is adapted to 1D data, is used to build the mapping relationship between the FWI result and the well velocities. We apply a window warping approach to the available well logs to augment the training data, which helps reduce potential overfitting. These one dimensional well models, which serve as labels for the neural network (NN) model training, will be used to generate synthetic data for FWI to obtain the corresponding input FWI velocities to the NN model. After the training process, we can employ the trained NN model to infer high-resolution velocities for each profile of the band-limited FWI result. The test and analysis on the Otway model demonstrate that the deep learning can help inject high wavenumbers into the band-limited FWI result.