Direct estimation of local wavefront attributes using deep learning Kirill Gadylshin*, Institute of Petroleum Geology and Geophysics SB RAS, Novosibirsk State University; Ilya Silvestrov and Andrey Bakulin, EXPEC Advanced Research Center, Saudi Aramco Summary Wavefront attributes, such as local dips and curvatures of seismic events, are used in different seismic data processing methods, from prestack data enhancement to migration to tomography. The attributes' estimation for prestack data is a time-consuming and computationally expensive process. We propose a new approach based on U-Net convolutional neural network that directly map prestack seismic data to the local wavefront attributes. Using a 3D real data example, we demonstrate that this deep-learning-based approach can reduce the computational time by two orders of magnitude compared to a classical coherency-based optimization technique while preserving a reasonable quality of results. Introduction Kinematic wavefront attributes form the basis of many seismic data processing methods. Notably, they are used in stacking (Mann et al., 1999), reflection and diffraction imaging (Fomel, 2007; Berkovitch et al., 2009), tomography (Lambaré, 2008), data interpolation (Hoecht et al., 2009), and data enhancement (Baykulov and Gajewski, 2009; Buzlukov et al., 2013, Bakulin et al., 2018b). Different algorithms exist to estimate the wavefront attributes, including radon transform techniques, plane-wave destruction filters, and structure tensor approaches. When the data quality is low, such as in land seismic applications, the classical coherency-based search that uses semblance as a target function for optimization often shows the most robust results. Recent examples of applying such a kinematic-based approach for land prestack data enhancement can be found in Bakulin et al. (2020). For modern high-density seismic volumes, where the data size can reach hundreds and thousands of terabytes, the biggest challenge becomes a lengthy computational time required to estimate the attributes. In this work, we present a deep- learning-based approach to speed up the calculations tremendously. Theory and Method Our work's main idea consists of detecting the local geometrical attributes of the wavefront directly from the 3D prestack seismic data using a deep learning approach. Usually, wavefront attributes are computed on a dense regular spatial and temporal grid. The conventional estimation approaches are based on the semblance optimization procedure. In other words, at each point of the dense 3D X-Y-T grid, one needs to solve an optimization problem, which is very time-consuming. We propose using the specially trained deep neural network (DNN) to overcome this situation. This DNN directly links the prestack seismic data with estimated attributes on a regular grid. Preprocessing The irregular acquisition geometry is a common situation in seismic processing. It is preferable to use a regular input in deep learning, especially if we speak about estimated local wavefront attributes on a regular grid (the output of the DNN). Therefore, the substantial step in our workflow is a regularization, which makes the input seismic data grid consistent with the regular output attributes grid. To do so in an efficient way, we choose a regular grid and collect the signal to this grid using super-grouping within a small grouping aperture (Bakulin et al., 2018a). In this way, we solve two problems at once:  We enhance the signal-to-noise ratio (SNR) of the data and simplify the problem of attributes detection; and  we provide the regular seismic input data to be processed by a DNN. The example of this preprocessing step is presented in Figure 1. On the left (Fig.1a), we plot a single inline section of the 3D common-receiver OBN data using wiggles. As one may observe, the original acquisition is irregular. It has a different trace density with respect to a shot X-coordinate. After supergrouping, we enhanced the SNR and obtained the inline section on a regular grid with a constant X-coordinate density (see Fig. 1b). Figure 2. The acquisition geometry of the marine OBN dataset. Receiver positions are shown in black, and shots are in blue. Red points indicate common-receiver gathers used for generating the training dataset. 10.1190/segam2021-3583265.1 Page 1596 © 2021 Society of Exploration Geophysicists First International Meeting for Applied Geoscience & Energy Downloaded 09/03/21 to 3.88.72.34. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/segam2021-3583265.1