Applications of deep neural networks for velocity model building

In recent years, machine learning (ML) has become a crucial and integrated component of the seismic processing sequence. Most applications, however, have focused on preprocessing steps such as de-ghosting, de-noising, demultiple, interpolation, regularization, and interpretation (e.g., horizons and well logs). In contrast, there have been fewer applications of ML for velocity model building (VMB) on field data examples. This is likely due to the significant variability in the Earth’s geology, the high dimensionality of seismic acquisition, and the challenges of mapping long-range velocity dependencies across different domains, both in image and data domain gathers. In this work, we utilize a deep neural network comprising a series of Fourier neural operators (FNOs), convolutional neural networks (CNNs), and fully connected neural networks (FCNs) to map velocity errors within the image domain. The network is trained to perform tomographic updates based on synthetic data, effectively replacing the conventional reflection tomographic engine. We demonstrate the performance on several field data examples from different basins worldwide, showcasing how this methodology can improve or assist conventional tomography, and help condition the model for full waveform inversion (FWI).