SEP-184 (2021)

M. Bader, G. Barnier, E. Biondi, J. Claerbout, S. Farris, J. Frigerio, P. Given, F. Huot, J. Jennings, A. Lellouch, B. Luo, M. Park, R. Sarkar, J. Voyles, S. Yuan

DOWNLOAD

SEP-184 (June, 2021)

Moment tensor inversion of perforation shots using DAS data
Milad Bader, Robert G. Clapp, and Biondo Biondi
Elastic full-waveform inversion of DAS data in unconventional wells requires accurate modeling of the seismic source representing the perforating gun. Unlike surface seismic sources, a complete representation of the perforating gun in a fluid-filled borehole with casing is unavailable. Theoretical considerations provide possible structures for the moment tensor describing such seismic sources. However, they remain uncertain in representing perforation shots and do not provide the time evolution function. We propose the alternating direction method to reconstruct the full moment tensor with the time evolution function when the elastic model is known, and the data quality is high. Alternatively, we propose a two-step strategy to recover the moment tensor components and the macro parameters of a predefined time function. We illustrate and test both methods using 2D synthetic DAS data. We test the second method to estimate the seismic source parameters of fifty-six perforation shots from a field DAS data set usingvisotropic and VTI elastic models. The VTI results lead to significantly smaller data residuals than the isotropic ones, giving additional evidence of anisotropic reservoir layer. However, the moment tensor variability across the examined shots is larger in the VTI case. In all cases, the estimated moment tensors indicate the presence of a tensile crack component in addition to a radially symmetric explosive component in the source mechanism.

Skewed Pulses
Jon Claerbout and Shuki Ronen
Analogous to bandpass filters, Jon provides a two-parameter definition of a compact pulse with a sharp onset and a slow decay. Its decay/risetime ratio defines the pulse skewness. The immediate application is for modeling and image estimation. Since pulse skewness is occasionally observable on field data we should design a nonstationary process to highlight it in hopes of identifying an interpretive value. Marine sources always have an inherent low cut. This low cut is in addition to the low cut that is coming from a source ghost, possibly also a receiver ghost, and inherent low cut in the transfer functions of hydrophones and geophones. One way to produce such inherent low cut is to take the difference between two skewed wavelets. The two skewed wavelets are normalized so they would have the same area, so their difference has no zero zero frequency (\DC") content.

Automatic velocity model building in the presence of salt
Stuart Farris, Robert Clapp, and Biondo Biondi
We train a convolutional neural network (CNN) to map from streamer seismic data to two-dimensional velocity profiles which have been projected onto a Bspline grid representation that allows for coherent up sampling to fine Cartesian grids. The CNN is trained with synthetically generated feature-label pairs with a seismic streamer acquisition that is general enough to be projected onto modern field streamer data footprints. We show the ability of the network to predict velocity models with complex salt models from the validation set.

Tomographic waveform inversion (TWI) with dipping reflectors
Julio Oliva Frigerio, Biondo Biondi, and Robert Clapp
Tomographic waveform inversion is a velocity model building inversion algorithm to recover the low (including zero) to medium (beyond conventional reflection seismic tomography) wavenumbers in the final model from a very simple and inaccurate initial model. By design it is immune to cycle-skipping and robust with respect to amplitude variations and AVA e↵ects of the input seismic data. In this work we generalize TWI to handle reflectors with arbitrary dips. The proposed algorithm is computationally efficient and its inversion scheme does not require complex preconditioning. In the end of the paper we show the result of the application of the new algorithm to the Marmousi 2 model where a very simple and inaccurate initial model. The inversion process was successful even though only reflection events - no diving waves - were present in the input data.

Quantitative Analysis of Training Size for Microseismic Detection Convolutional Neural Network
Paige Given, Fantine Huot, Ariel Lellouch, Bin Luo, Robert G. Clapp, Tamas Nemeth, Kurt Nihei, and Biondo L. Biondi
The ability to efficiently and accurately implement Convolutional Neural Networks (CNN) is pivotal to their success. We designed a CNN that is able to predict microseismic events with an accuracy of 98% given an input size of approximately 6,000 manually labeled microseismic events. The manual labeling process for CNN training can be quite time consuming, taking days to weeks to complete. Here, we analyze the results for our CNN using various training input amounts, to determine the minimal sample training size necessary to achieve accurate CNN results. This allows us to maintain efficiency when implementing our model. We found that given our CNN architecture, we can drastically reduce our training input size, using only 15% of our initial input for model training while still maintaining an accuracy of 94%. We also found that when using the reduced training data input, it is best to also use this decimated data size to optimize the model hyperparameters. Tailoring our model hyperparameters to the 15% of our original training data, we achieved results of up to 99% accuracy when training using the full 100% input dataset and 97% accuracy training the model on only 25% of the original data.

Microseismic event detection with super-human accuracy
Fantine Huot, Ariel Lellouch, Paige Given, Bin Luo, Robert G. Clapp, Tamas Nemeth, Kurt Nihei, and Biondo L. Biondi
Microseismic analysis is the primary tool available for fracture characterization in unconventional reservoirs. As DAS fibers are installed down wells and are thus close to the microseismic events, they hold vast potential for their high-resolution analysis. However, accurately detecting microseismic signals in continuous data is challenging and time-consuming. DAS acquisitions generate substantial data volumes, and microseismic events have a low signal-to-noise ratio in individual DAS channels. Herein we design, train, and deploy a machine learning model to automatically detect microseismic events in DAS data acquired inside an unconventional reservoir. We create a curated dataset of 6,786 manually-picked microseismic events. The machine learning model achieves an accuracy of 98.6% on the benchmark dataset and even detects low-amplitude events missed during manual picking. Our methodology detects over 100,000 events allowing us to reconstruct the spatio-temporal fracture development accurately.

Focusing unfocused faults with deep learning and residual migration: 2D field data application
Joseph Jennings, Bob Clapp, Mauricio Araya-Polo, and Biondo Biondi
Seismic images are often difficult to interpret when they are poorly focused due to velocity error. While velocity model building approaches can be used to correct the velocity error and remigrate the image, they can require significant effort and computational resources. We present a novel and computationally efficient approach to seismic image refocusing that relies on prestack Stolt residual depth migration and a deep convolutional neural network (CNN). Through the application of prestack Stolt residual migration to a poorly focused image, we generate many residually focused images which are then provided as input to a CNN that predicts the optimally focused regions of these images. With these optimally selected regions, a refocused image can be reconstructed that will be better suited for tasks such as automatic fault detection. We apply our refocusing approach to an unfocused image from the Gulf of Mexico and use an uncertainty-aware intersection over union metric (IOU) to quantify the improvement in the fault segmentation. We observe that after refocusing, the IOU of the segmented faults improves from 0.86 to 0.89.

Microseismic analysis in unconventional plays with a single horizontal DAS well
Ariel Lellouch, Bin Luo, Fantine Huot, Paige Given, Robert G. Clapp, Tamas Nemeth, Kurt Nihei, and Biondo L. Biondi
A single horizontal DAS fiber is notoriously challenging for microseismic analysis despite its proximity to recorded events. Due to its uni-axial measurement, locations suffer from cylindrical ambiguity. Nonetheless, in unconventional plays, recorded guided waves can partially resolve such ambiguity. They are generated only by microseismic events occurring inside, or close to, the low-velocity reservoir, and their propagation is con ned to the reservoir. We first train and apply a machine-learning approach for microseismic event detection yielding more than 100,000 event detections over ten stimulation stages from two offset wells. Detection capabilities surpass manual labeling. Detected events undergo a location procedure that is based on known kinematic dispersion properties of guided waves and allows us to accurately reconstruct the spatio-temporal fracture development. Fracture growth is validated by perforation shot analysis.

Guided wave analysis for lateral variation of unconventional reservoir structure
Bin Luo, Ariel Lellouch, Milad Bader, and Biondo Biondi
We present a guided wave analysis for structural characterization of unconventional reservoirs by extending the propagator-matrix-based, 1D dispersion inversion approach to a pseudo-2D scheme. The pseudo-2D scheme aims to extract lateral variations of the unconventional reservoir layer and provide more accurate structural information to help build precise velocity models and aid operational decision making. We verify the scheme by conducting 2D numerical tests, which show substantial sensitivity of the guided wave dispersion curves in response to lateral inhomogeneity of the multilayered media. We then apply the scheme to leaky-mode guided waves in field DAS data generated by same-well perforation shots. Our result reveal a generally at reservoir structure with a local anomaly of slightly high velocity confirmed by the horizontal sonic log.

Toward a real-time CO2 monitoring system using Deep Learning
Min Jun Park and Stuart Farris
We present a workflow for carbon dioxide (CO2) leakage detection that uses a supervised learning approach trained with realistic synthetic seismic data. We generate a sophisticated geologic model with realistic textures using simplex noise and simulate the injection of CO2 into a target reservoir. Further, we generate CO2 fault-based leakage scenarios at the top of the CO2 plume. Finally, we perform elastic wave modeling to produce the data for both baseline and monitor models. We define the training input of the deep neural networks (DNNs) as a time-lapse shot gathers. The trained DNNs can determine if there is a CO2 leakage for a given input. Validation results show that the trained model's performance is reliable.

A numerical scheme to solve the Lippmann-Schwinger equation for a linearly varying background
Rahul Sarkar and Biondo Biondi
We consider the problem of solving the Helmholtz equation in 2D and 3D, where the velocity can be decomposed into a sum of two components: (i) the background velocity component that varies linearly with depth, and (ii) a smooth, compactly supported velocity perturbation. The integral equation equivalent of this problem yields the Lippmann-Schwinger equation (LSE), and in this paper we present a new numerical scheme to solve it. Iterative methods to solve the resulting linear system involves repeated computation of the integral appearing in the LSE. Using the truncated kernel approach introduced by Vico et al. (2016), we are able to exploit the translation invariance of the background in the horizontal direction and the compact support of the perturbation, to design a scheme that handles the singularity of the integrand and efficiently computes this integral. We present 3D numerical examples where we compute the solution to the LSE using our method for different velocity models.

Seismic Imaging for Carbon Capture and Sequestration using Distributed Acoustic Sensing
Jonathan Voyles and Stuart Farris
We build an elastic model reflective of the geology of the Red Trail Energy carbon capture and sequestration (CCS) project and simulate elastic wave propagation recorded by strain rate sensitive distributed acoustic sensing (DAS) sensors. We perform reverse time migration (RTM) imaging to establish a baseline in order to monitor changes resulting from CO2 diffusion after injection.

Urban system monitoring using combined Vehicle Onboard Sensing and roadside Distributed Acoustic Sensing
Siyuan Yuan, Jingxiao Liu, Hae Young Noh, and Biondo Biondi
We explore the value of combined Vehicle Onboard Sensing (VOS) and roadside Distributed Acoustic Sensing (DAS) to enable a cost-effective urban infrastructure monitoring system. We conducted experiments to analyze simultaneous recordings of the VOS system and DAS to better estimate the fiber location using GPS recordings of the car. We utilize DAS to measure car speed with speed meter measurements as a benchmark. We characterize the sources of the traffic-induced surface waves leveraging vertical accelerations recorded next to car wheels. Lastly, we introduce an automated approach for traffic recording extraction using Neural Networks, which can be used as a first step for car detection to fully automate the aforementioned tasks about locating DAS fibers using controlled drives, monitoring traffic and imaging near-surface.