Deep Learning to maximize the value of fast-track 4D seismic processing Arnab Dhara* 1 , Haron Abdel-Raziq* 1 , Denis Kiyashchenko 2 , Asiya Kudarova 2 , Janaki Vamaraju 1 , Albena Mateeva 2 , Pandu Devarakota 1 , Kanglin Wang 2 , Jorge Lopez 3 1 Shell Global Solutions (US) Inc., 2 Shell International Exploration and Production Inc., 3 Shell Brasil Petróleo Ltda Summary Time lapse seismic monitoring plays a vital role in reservoir management. However, extracting such value requires dedicated acquisition, co-processing, and interpretation of multiple seismic vintages to highlight possibly small 4D changes, such as those expected in stiff reservoir rocks. While the standard full production processing is laborious and time consuming, the fast-track processing can be available in few weeks and the results are valuable for early interpretation insights and to inform possible actions related to well operations and field development. However, the signal fidelity of fast-track processing is usually low and may be incapable of identifying accurately subtle 4D signals. This may lead to misinterpretations, missed opportunities, and negatively impact decisions. In this paper we discuss novel methods using deep learning to generate denoised 4D attributes from fast-track processing products. We formulate the problem as an image-to-image translation and introduce a custom loss function to generate meaningful attributes for the reservoir zone. In addition, we show how real data can be augmented with data from a reservoir simulation volume and used as input for training. We demonstrate the efficacy of methodologies on Brazilian pre-salt 4D seismic dataset. Introduction Time-lapse seismic is an important tool for reservoir monitoring (Calvert, 2005), with multiple applications to better understand production-related changes (Tura and Lumley, 1999), enhanced oil recovery processes (Ditkof et al., 2013), field development, and carbon sequestration (Egorov et al., 2017). Recently, time-lapse seismic achieved success in monitoring the deep and stiff reservoirs of the Brazilian pre-salt (Kiyashchenko et al. 2020, Cruz et al. 2021). This became possible by ensuring high survey repeatability using OBN acquisition and a new 4D compliant processing workflow that ensures minimal time-lapse noise. The turnaround time of conventional OBN processing may be many months, while recovery of challenging 4D signals requires beating the noise, which may increase the timeline further. This paper leverages deep convolutional neural networks (CNN) to reduce the turnaround time by denoising time-lapse attributes derived from intermediate (fast-track) processing results. Early 4D signal insights can be valuable for timely decisions. CNNs have recently achieved success in many computer vision tasks including natural image super-resolution and denoising (Lim et al., 2017; Dai et al., 2019). Although numerous applications of deep learning have been proposed for seismic image denoising and super- resolution (Li et al., 2021), the application of such techniques to 4D processing has been limited (Duan et al., 2020; Alali et al., 2020). Unlike 3D processing, the objective of 4D processing is to highlight subtle 4D differences between two vintages acquired at different times over a producing field. Moreover, unlike natural images, the number of training samples in the field of 4D seismic is small. In this work, we are limited to a single time-lapse vintage with one baseline and one monitor survey. This restricts us to training on data from outside the reservoir, to show that our approach provides key denoising and signal recognition in the most important area of interest, the reservoir. To maximize the value of fast-track 4D seismic processing, we propose to train a neural network using full track processing results from previous vintage. To improve CNN learning capabilities, we devise a custom loss function which is a combination of L1 loss and L1 loss with focus on maximum signal values. To address the problem of limited training samples, we augment the training data with realistically simulated data that provide the deep neural network with information on signal estimates in the reservoir, as this 4D signal is not present in training data from outside the reservoir. Dataset used in this study The dataset used in this study is from the Brazilian Santos basin, where the water depth ranges from 2000 to 2200 m. The pre-salt carbonate reservoir is covered by a massive salt body having locally a rugose top, with internally stratified salt sequences (see Fig. 1). The 4D OBN seismic survey was acquired in 2015/2017. The porosity of the reservoir ranges from 5 to 15% which, in addition to a stiff rock framework, is one of the main challenges for the visibility of 4D signal. Figure 1: Brazilian Pre-salt field setting. 10.1190/image2022-3751640.1 Page 1664 Second International Meeting for Applied Geoscience & Energy © 2022 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Downloaded 08/17/22 to 18.234.237.68. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/image2022-3751640.1