ABSTRACT Distributed Acoustic Sensing (DAS) is a rapidly growing novel sensing method for seismic data acquisition. DAS arrays are particularly well-suited for dense recording of low-frequency ambient surface waves on long (>5 km) linear sections of deployed optical fiber cable. Applying multi-channel analysis of surface waves (MASW) to ambient wavefield DAS recordings characterized by a large number of sensing points and long recording times may enable 1D characterization of the S-wave velocity profile to depths of 750 m or greater. We present a low-frequency ambient wavefield investigation using a DAS dataset acquired on an array deployed in suburban Perth, Australia. We extract storm-induced swell noise from the nearby Indian Ocean in a low-frequency band (0.1–1.8 Hz) and generate virtual shot gathers by applying cross-correlation and deconvolution seismic interferometric analyses. The resulting gathers are transformed into dispersion images through two different methods: phase shift and high-resolution linear Radon transform. To recover the near-surface S-wave velocity model, we first pick and then invert the recovered 1D Rayleigh-wave dispersion curves using a particle-swarm optimization algorithm. Inversion results show that low-frequency ambient-wavefield DAS data can constrain the Vs model to 750 m depth, which helps validate the potential of DAS technology as a tool for large-scale surface-wave investigation. Presentation Date: Monday, September 16, 2019 Session Start Time: 1:50 PM Presentation Time: 3:30 PM Location: 221C Presentation Type: Oral

ABSTRACT As full-waveform inversion (FWI) is a nonunique and typically ill-posed inversion problem, simultaneous-source noisy data, which suffers from crosstalk noise caused by shots overlap and environmental noise, indeed requires a proper regularization to avoid cycle skipping and accelerate convergence. We propose normalized low-rank regularization to achieve simultaneous-source noisy data waveform inversion, in which the crosstalk noise caused by shots overlap is suppressed by nonlocal low-rank regularization on the updated velocity while normalized shaping operator is to shape the noisy misfit between recorded and simulated data into the least-squares sense. We test the performance of proposed method using synthetic simultaneous-source noisy data, and use the results of conventional full-waveform inversion with Tikhonov regularization as comparisons. Results show that the proposed normalized lowrank regularization can obtain clearer and smoother waveform inversion results. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 3:05 PM Location: Poster Station 8 Presentation Type: Poster

ABSTRACT We present the modeling of blended acquisition for a node survey acquiring long offsets (50 km) with a focus on the low-frequency diving waves. The synthetic data simulates six sources, with shots fired both simultaneously and in flip-flop formation. The diving waves are recovered using an iterative inversion-based deblending methodology, which would subsequently be used for refraction FWI. Acquisition options are studied using different shot firing orders, source vessel spacing, dither time, and background noise. We show that this deblending methodology successfully recovers useful signal across a wide range of acquisition geometries and firing patterns. The success of deblending in recovering the diving waves is quantitatively established. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 4:45 PM Location: 221A Presentation Type: Oral

ABSTRACT Recently the Reflection Full waveform Inversion method has been proposed as an alternative to classical ray-based velocity model building methods. This method considers a scale separation between the reflectivity and the velocity models. In the original RFWI approach, the reflectivity and the velocity models are inverted by sequentially solving two optimization sub-problems. During the velocity model estimation subproblem, the reflectivity model is keep fixed therefore ignoring reflectivity-velocity coupling effects. This may lead to a slow convergence toward the final velocity model and also generate some in-consistencies between the reflectivity and the velocity model. We propose, here, to reformulate the conventional RFWI approach in order to better account for the coupling between the velocity and reflectivity models. Compared to the conventional RFWI this formulation induces several changes. First, the gradient expression contains the gradient of the conventional RFWI plus an additional term. We show that under particular circumstances, the variable projection method applies allowing to ignore this supplementary gradient term. Another consequence of this formulation is that the reflectivity model has to be re-migrated for every perturbation of the velocity models. Finally, we verify with a synthetic experiment that the reflectivity-velocity consistent RFWI approach has a faster convergence than the conventional approach. Presentation Date: Monday, September 16, 2019 Session Start Time: 1:50 PM Presentation Time: 3:55 PM Location: 302B Presentation Type: Oral

ABSTRACT It is widely known that anisotropic inversion is a nonlinear ill-posed problem, which are usually simplified as a linear problem. However, the linear problem is still unstable which is usually avoided by adding a specific mathematical regularization item (suppose the parameters follow Gaussian distribution) on anisotropic inversion. The regularization item can only works for a simple model which limits the applications in subsurface complicated structure. We proposed a data-driven regularization method to improve the stability of anisotropic inversion. The method uses dictionary learning algorithm to obtain dictionary of anisotropic characteristics and elastic parameters from logs. The obtained dictionary is used to sparsely represent prior information which is eventually used to constrain the anisotropic inversion procedure. Requirement that the model conforms to a particular mathematical distribution is no longer needed in this sparse representation anisotropic inversion (SRAI) method. The only requirement is to provide sufficient logs in the same work area to learn the parameters’ characteristic. The stratigraphic characteristics in logs are contained in the inversion method which makes the inversed result approaching to the subsurface structures. The EAGE model and the field data are used to test the reliability of the algorithm. Presentation Date: Monday, September 16, 2019 Session Start Time: 1:50 PM Presentation Time: 3:05 PM Location: 217D Presentation Type: Oral

ABSTRACT Anelastic attenuation causes energy dissipation of seismic waves propagating inside the earth, which leads to amplitude decay and phase distortion of migrated images. To correct for this dissipation effect, we use Gaussian beams as wavefield propagator to formulate an efficient and accurate linearized inversion scheme, denoted as least-squares Q Gaussian beam migration (LSQGBM), to iteratively solve for optimal images in viscoacoustic media. With this method, we can not only compensate for the amplitude loss and phase shift caused by attenuation loss during migration, but also produce images with better balanced amplitudes and higher resolution. Numerical tests on synthetic and field datasets demonstrate the validity and effectiveness of our proposed method. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Start Time: 2:15 PM Location: Poster Station 10 Presentation Type: Poster

ABSTRACT Reservoir characterization based on amplitude variation with offsets/angles (AVO/AVA) is the core of geophysics, which focuses on the quantitative prediction of lithology and pore geofluid properties. In this study, we have proposed one simultaneous stochastic inversion approach for discrete fluid-facies and continuous geofluid parameters based on differential evolution Markov Chain Monte Carlo (DE-MCMC) algorithm and Gaussian mixture model. The Gaussian mixture model influenced by reservoir pore fluid-facies is assumed to be the prior probability density distribution of elastic parameters in Bayesian inference. The likelihood probability function is first established with time-frequency joint domain seismic data. Then, the standard posterior probability distribution expressed with a multi-dimensional Gaussian mixture model is derived. Due to the superiority of multi-solution global optimization in DE-MCMC model, the DE-MCMC model is first introduced into the optimization of mixture posterior PDF. The fluid-facies is predicted effectively through the posterior probabilities of discrete facies. The main advantage of DE-MCMC AVO inversion is that the probability distributions of estimated parameters can be easily built based on the multiple-solutions optimization. Thus, the uncertainty and reliability of geofluid parameters and fluid-facies are assessed effectively. Model and field application examples demonstrate the feasibility and stability of the proposed methodology evidently. Presentation Date: Monday, September 16, 2019 Session Start Time: 1:50 PM Presentation Time: 3:55 PM Location: Poster Station 8 Presentation Type: Poster

ABSTRACT Full waveform inversion (FWI) is a challenging data-fitting procedure based on full wave-field modeling to extract quantitative information from seismograms, and it has been widely knows as the most prospective quantitative seismic imaging technology. However, as an ill-posed and highly nonlinear inverse problem, there are still challenges for the application of FWI in land seismic data. In this paper, we briefly introduce a strategy of the FWI application to land seismic data: the geology structural information as the prior model to decrease the non-linear problem and the compensation attenuation coefficient to scale the gradient. Then we show an application of FWI to the land seismic data from Southwest Oil Filed of CNPC in Sichuan basin. The final results show the feasibility and validity in the land data application, and have the potential to the 3D land data. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 302B Presentation Type: Oral

ABSTRACT Reflection full-waveform inversion (RFWI) is a wave equation based model building approach that utilizes reflection data and can update deeper parts of a velocity model not accessible by diving waves of standard FWI. In this paper, we propose an efficient RFWI workflow that uses full-waveform information contained in reflection data to update kinematics of reflections. It is based on the linearized MBTT (Migration Based Travel Time) approach and employs (Least-Square Reverse-Time Migration) LSRTM type iterations to construct true-amplitude data-space reflectivity that can be re-used in several iterations of RFWI velocity updates. The approach uses a regular L2-norm data misfit objective function and provides a fully automatic workflow. We illustrate the approach with synthetic and real-data examples. Presentation Date: Monday, September 16, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 302B Presentation Type: Oral

ABSTRACT Geophysical inversion and machine learning both provide solutions of inverse problems in which we estimate model parameters from observations. In this study, we compare geophysical inversion and machine learning approaches in an aspect of solving inverse problems and show similarities and differences of these approaches in a mathematical form and numerical tests. We take reflectivity inversion as an example of the inverse problem. We apply geophysical inversion based on the least-squares method and artificial neural network (ANN) as a machine learning approach to solve reflectivity inversion using 2D synthetic data sets and 3D field data sets. Neural network with multiple hidden layers successfully generates the non-linear mapping function to predict reflectivity. The different levels of noise were applied to data to examine noise sensitivity of each method. Also we test different L1 regularizers to investigate how regularizers affect on inverting clean or noisy data for both approaches. L1 regularization alleviate effect of noise in seismic traces and enhance sparsity, especially in leastsquares method. The 2D synthetic model and field data examples show neural network yield high spatial resolution. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 2:15 PM Location: Poster Station 2 Presentation Type: Poster

ABSTRACT We present a novel workflow based on a convolutional neural network (CNN) for automatic salt interpretation from a seismic image. By using a smoothed P-wave migration velocity as a second input channel, along with an RTM image for training the network, artifacts in the predictions are reduced significantly. Additionally, a second CNN built on top of the first one is trained along a different image plane, mitigating oscillations generated by 2-D predictions, leading to overall improved prediction results. Examples based on field data demonstrate the effectiveness of our automatic salt interpretation workflow with machine learning. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 4:20 PM Location: 221D Presentation Type: Oral

ABSTRACT We present an algorithm to address the solution to large-scale full waveform inversion (FWI) in the framework of Bayesian inference. This is a sampling method guided by the local gradient information of the L2 norm misfit function. We employ the local gradient and the diagonal of the approximate Hessian information to construct a proposal distribution, which is a local Gaussian distribution. New samples/models can be easily drawn from the proposal distribution. The new samples/models are then subjected to the accept/reject Metropolis-Hasting (M-H) criterion. The solutions to FWI are then described by point-wise posterior PDFs. Uncertainties associated with FWI can be easily estimated by the PDFs. The proposed method can achieve the neighborhood of the target distribution with poor initial models. We term the proposed method a gradient based Markov chain Monte Carlo (GMCMC) method. We apply the proposed method to 2D Marmousi model with acoustic FWI. Even though the proposed method requires more computational effort than traditional FWI based on local optimization methods, numerical results suggest that the proposed method can provide a more reliable inversion results and statistical assessments to FWI Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 225B Presentation Type: Oral

ABSTRACT We studied the feasibility of deblending of ocean bottom node (OBN) data with the purpose of using low frequency long offset signals for full waveform inversion (FWI) velocity-model building. We used unblended field records to simulate heavily blended data up to 8 Hz with a maximum offset of 30 km. Our inversion-based deblending method involves many iterations of separating the most coherent hyperboloidal events in a cube of data formed by all traces recorded in a single node. Coherent hyperboloidal components are modeled by using an amplitude threshold in the 3D domain, while various temporal shifts are applied. The leakage of blending noise into the estimated model is then reduced by using further coherency enhancement algorithms during the inversion process. This study shows that it is possible to reveal the signals required for an FWI workflow. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Start Time: 2:40 PM Location: 214D Presentation Type: Oral

ABSTRACT Unconventional reservoirs cannot be characterized by relying on the juxtaposition of single-attribute analysis. To understand their complexity, many correlated static and dynamic reservoir properties need to be carefully and quantitatively analyzed as part of integrated multidisciplinary studies. The goal of quantitative interpretation is to predict reservoir properties away from control points in order to improve vertical and lateral well placement to maximize stimulated reservoir volume (SRV) and productivity. The dense lateral spacing of three-dimensional (3D) seismic lends itself to estimating reservoir properties in 3D. Therefore, prediction of elastic attributes from seismic amplitude away from wells is common practice. However, transforming these elastic properties into geomechanical and other reservoir properties requires integration of multiple data sets (e.g., wireline logs, core, and cuttings) at various scales. Integration of all of this information is essential for de-risking seismic amplitude-supported interpretations. This study demonstrates this approach and the methods used to help predict geomechanical properties in 3D space, away from well control. The computed geomechanical volumes (pore pressure (Pp), fracture pressure (Fp), and minimum horizontal stress (SHmin)) were used to pick intermediate casing points (IPCs) and to predict recommended maximum and minimum mud weights, both of which provided significant capital savings in casing design and/or lost bottom hole assemblies (BHAs). Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Start Time: 3:30 PM Location: 303B Presentation Type: Oral

ABSTRACT PP-wave and PS-wave event registration is an essential element in multicomponent data analysis. In seismic exploration, multi-wave data usually shows big difference in quality, which makes it difficult to obtain stable Vp/Vs ratios, and leads to inaccuracy registration between PP-wave and PS-wave image volumes. In this paper, we developed a practical approach for PP-wave and PS-wave image volumes registration based on the 3D non-rigid matching algorithm. First, with a relatively accurate model of Vp/Vs ratios derived from prior information such as VSP, a rough registration of PS-wave and PP-wave image volumes in the PP time scale can be realized. After that, for the residual time difference between PS-wave and PP-wave, using the PP-wave image volume as the reference model, the 3D non-rigid automatic iterative matching algorithm is performed to eliminate the time difference. The real data processing in Block SNL from Sichuan Basin shows that the proposed approach realizes a pretty registration between PP-wave and PS-wave image volumes, and effectively improves the precision of joint multi-wave inversion. Therefore, the feasibility of multi-wave non-rigid matching is demonstrated. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 9:20 AM Presentation Time: 11:00 AM Location: Poster Station 1 Presentation Type: Poster

ABSTRACT An accurate earth-model is critical in correct depth imaging of a borehole seismic survey. Borehole seismic survey generally provides excellent control on the vertical velocities at well location and can be used to estimate anisotropy and to a lesser extent lateral heterogeneity, all essential components of an earth model for depth imaging. In this study, direct- and reflection-tomography is used to construct a precise anisotropic earth-model. Existing anisotropic earth-model building workflows are adapted to use P and S direct- and reflected-energy to resolve the complete VTI model for imaging. S direct- and reflected-waves from the horizontal-vibrator source are used to further constrain Thomsen’s parameter. The earth-model thus generated, provides well focused, high resolution walkaway VSP images with good Pp-Ss event correlation, and flat VSP common-image-point gathers. The tomography based anisotropic earth-model parameters are also compared to slowness-phase analysis within the receiver depth interval. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 8:30 AM Presentation Start Time: 11:25 AM Location: 303B Presentation Type: Oral

ABSTRACT From the academic to the oil industry, it is known that wellto-seismic-tie is an important step in seismic processing and interpretation. Deconvolutional statistical methods to estimate the proper wavelet, in general, are based on the assumptions of the classical convolutional model, which implies a random process reflectivity and a minimum-phase wavelet. The estimative of the wavelet for well-to-seismic-tie purposes through least squares minimization with zero order quadratic regularization is compared with the results obtained from the deterministic homomorphic deconvolution. Both methods do not make any assumptions regarding the wavelet’s phase or the reflectivity. The best-estimated wavelet was used as input to the sparse-spike deconvolution to recover the reflectivity near to the well location. The results show that the wavelets estimated from both deconvolutions are similar, which suggests their feasibility. The reflectivity of the seismic section was recovered according to known stratigraphic markers (from gamma-ray logs) present in the real dataset from the Viking Graben field, Norway. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 1:50 PM Presentation Time: 3:05 PM Location: Poster Station 4 Presentation Type: Poster

ABSTRACT In 2018, BP acquired an ocean-bottom-node (OBN) dataset at Mad Dog field in the Gulf of Mexico. We used these data to rebuild the salt and sediment velocity model, employing a recently developed full-waveform inversion (FWI) algorithm. This resulted in a step-change improvement in image quality, as well as significant time savings compared to traditional salt model building workflows. We conclude that FWI applied to OBN data is the preferred method for salt model building. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 10:10 AM Location: 302B Presentation Type: Oral

ABSTRACT 3D modeling of marine controlled source electromagnetic (CSEM) survey can be computational expensive because of the distributed sources and receivers that cover large areas. We propose a novel approach to achieve second/minute response time for large-scale marine CSEM modeling in a massively parallel computing environment. Different from conventional parallel methods that mostly rely on spatial domain decomposition, our approach decomposes the original data set using relations in survey configurations. We consider a large marine CSEM survey as a simple assembly of many small "atomic" problems, each of which only models one datum from a single source (S) and single receiver (R) at a single frequency (F). Each S-R-F problem is then modelled using a local mesh customized to the specific frequency and S-R geometry. Such a mesh is usually small, so the solving time is at the order of seconds or minutes if efficient and accurate direct solvers are used. The decomposition is also fine-grained and embarrassingly parallelizable, therefore inter-communication between subproblems is minimal and a variety of parallel strategies can be used in a scalable manner. Our approach is demonstrated with a 10 × 10 km 3D synthetic reservoir model and a deeptowed EM system. Upon decomposition, 6912 independent sub-problems are generated; the number of cells in a local mesh ranges from 11200 to 103680 and the computing time from 5 s to 158 s. Because sub-problems do not communicate with each other, the entire problem is solved when individual problems are solved. A total computing time of less than 10 minutes for such a large survey was achieved by distributing the sub-problems over 448 cores on multiple nodes of a computer cluster. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 1:50 PM Presentation Time: 1:50 PM Location: 225C Presentation Type: Oral

ABSTRACT Full-waveform inversion is an iterative data-fitting procedure used to resolve complex subsurface elastic property models. It has the potential to become a tool for quantitative interpretation at stratigraphic and reservoir scales. To illustrate, we make use of a geologically modeled data set, and conventional marine seismic from the Northern Carnarvon Basin, Western Australia. We utilize a data-driven workflow that mitigates bias and exemplifies interpretational constraints on inversion results. Presentation Date: Tuesday, September 17, 2019 Session Start Time: 8:30 AM Presentation Time: 10:35 AM Location: 221B Presentation Type: Oral