Saudi Aramco: Public Semi-Automated Prestack Seismic Inversion Workflow using Temporal Convolutional Networks Hussain Alfayez, Robert Smith, Ayman Suleiman and Nasher AlBinHasan, Saudi Aramco Summary Seismic inversion is the process of transforming seismic data into subsurface rock properties. Recent applications of deep learning in the seismic inversion domain have demonstrated great potential in revealing subsurface properties. In this study we test a temporal convolutional network (TCN) for the inversion of acoustic impedance, Vp/Vs ratio and density from a field dataset. The deep neural network learns to map sequences of angle-stack data to elastic properties using only data acquired at well locations. We show that the trained model produces promising volumetric property predictions despite the limited training data,. The method has the potential to semi-automate the prestack inversion workflow for relatively simple geological scenarios, since common steps such as the construction of an initial model and wavelet extraction are not explicitly required. Introduction Seismic inversion plays an important role in the exploration and development phases in the upstream oil and gas life cycle. The ultimate goal of inversion is to predict the 3D distribution of subsurface rock properties from seismic data. A high level of confidence in the inverted elastic properties can help in the estimation of petrophysical subsurface properties, which ultimately reduces uncertainty around the identification and evaluation of prospects. Most conventional methods are model-based, where physics- based forward modelling drives the optimization process. Recent advances in machine learning provide an opportunity to address the seismic inversion challenge using new techniques. A number of machine learning algorithms have been tested. For instance, Liu et al. (2018) used Bayesian- based support vector machines (BSVM) to estimate velocity and density. Das et al. (2018) utilized convolutional neural networks (CNN) to estimate acoustic impedance using post- stack seismic data. Later, Das et Al. (2020) extended his work to estimate petrophysical properties using (CNN). However, these networks typically require large training datasets, which can be a challenge for seismic inversion where only a small number of wells are usually accessible. Alfarraj and AlRegib (2018) proposed the use of a recurrent neural networks (RNN) as a solution, which are networks designed to handle sequential data. They developed a model that could take a series of seismic amplitudes and convert them into a series of acoustic impedance. RNNs suffer from a number of issues however, including problems modeling long sequences and the computational time required to train models. As an alternative, temporal convolutional networks (TCN) were introduced which is an updated CNN to handle sequential problems (Lea et al. 2016). Mustafa et al. (2019) demonstrated that a TCN may be utilized to estimate acoustic impedance using post-stack synthetic seismic data. Later, Smith et al. (2021) demonstrated that a TCN model could be trained to predict acoustic impedance from noisy field data, obtaining better results than conventional model- based inversion. A key part of this work was the generation of a synthetic dataset (with realistic complex noise) to provide more samples for training. While this approach can enable us to incorporate additional knowledge from the geoscientist in the synthetic training set, it can be time consuming to produce and relies on having a method for mimicking the noise encountered in the field seismic. The same authors later showed that reasonable results could be produced by a model trained only on real data acquired at the well locations (Smith et al., 2022). Using this approach, it may be possible to semi-automate the seismic inversion workflow. In this paper we extend the work of Smith et al. (2022) to test the implementation of the TCN workflow for prestack inversion on a new field dataset with more wells (16) available. Here the TCN was modified to accept multiple angle stack traces as input and trained using only the field data at the wells. In this case, we tested the model’s ability to predict acoustic impedance, Vp/Vs ratio and density. The ultimate aim is for a semi-automated inversion workflow that can produce reasonable predictions with minimal manual tasks for the user. Figure 1 (a) Receptive field for a conventional CNN with four layers and filter width of two, and (b) the much larger receptive field for a TCN with four layers and filter width of two. A. Convolutional Neural Network B. Temporal Convolutional Network 10.1190/image2022-3746095.1 Page 1780 Second International Meeting for Applied Geoscience & Energy © 2022 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Downloaded 08/17/22 to 3.81.173.119. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/image2022-3746095.1