- Home
- Conferences
- Conference Proceedings
- Conferences
83rd EAGE Annual Conference & Exhibition Workshop Programme
- Conference date: June 6-9, 2022
- Location: Madrid, Spain / Online
- Published: 06 June 2022
-
-
Searching the Parameter Space for Resolution and Uniqueness in Elastic Anisotropic Waveform Inversion
Authors T. Alkhalifah and Y. LiSummaryFull waveform inversion (FWI) can retrieve high-resolution subsurface medium parameters from the observed data. However, the inverse problem is typically ill-posed and non-unique, especially for the multi-parameter elastic FWI (EFWI) in complex media. Besides, high-resolution EFWI is computationally expensive because it requires fine discretization for the whole computational domain. The redatuming approach allows us retrieve the virtual data at the target level using mainly a kinematically accurate overburden, thus, focusing the high-resolution inversion on the target zone to reduce the computational cost. In multi-parameter inversion, even at the target zone, we will need to utilize a prior information and we do that through deep learning to find the connection between well information and the a prior needed by FWI. In such a framework, we take into consideration the proper parameter makeup for reducing the ill posedness of the problem. Numerical tests on the synthetic SEAM model are used to demonstrate the performance of the proposed inversion scheme, and its robustness in the multi parameter inversion case.
-
-
-
Machine-Learning Seismic Processing Tasks by Fine Tuning a Pre-Trained Attention-Based Neural Network: Storseismic
Authors T. Alkhalifah and R. HarsukoSummaryMulti-dimensional seismic data often contain signatures and geometrical features that can help processing of such data provide valuable subsurface information. We propose a framework to first learn (and store) those features in a pre-trained neural network model, we refer to as StorSeismic. We then use this network for specific seismic processing tasks in an efficient fine-tuning stage. We use the Bidirectional Encoder Representation from Transformers (BERT), utilized in natural language processing (NLP), for the pre-training and fine-tuning stages. The traces, as opposed to words in NLP, are randomly masked to allow the network to learn the structure of the shot gathers. We demonstrate this approach on synthetic data resembling a marine setting, and we will share real data applications in the workshop.
-
-
-
Physics-Coupled Deep Learning Inversion: Geophysical Data Applications
Authors D. Colombo, E. Turkoglu, E. Sandoval-Curiel and D. RovettaSummaryWe develop a heuristic method for enhancing geophysical inversion based on the coupling of a standard optimization problem with a machine learning training and prediction scheme. The coupling mechanism is implemented for the model term by imposing one or multiple penalty functions. The procedure is complemented by an iterative self-feeding approach for the deep learning (DL) network that implicitly learns the requirements of the physics-based (Phy) optimization while the latter is constrained by the DL predictions. The result is a coupled physics-deep learning inversion (PhyDLI) scheme. The PhyDLI procedure is applied iteratively with automated stopping criteria evaluated on the convergence of the model term between Phy inversion and DL prediction, as well as on the data misfit reduction across several cycles. Network training is performed through a probabilistic/statistical sampling of the model space. This enables DL predictions to occur within the statistical distribution of the testing data and ensure adequate generalization properties for the derived neural network. The developed scheme is tested on challenging synthetic examples of transient electromagnetic data (TEM) where training and testing datasets have distinct distributions in the parameter space.
-
-
-
Multi-Parameter Viscoelastic Full Waveform Inversion for Subsurface Property Estimation.
Authors A. Baumstein, K. Basler-Reeder and P. RouthSummaryWe describe a practical workflow for applying viscoelastic full waveform inversion for the purpose of obtaining estimates of elastic properties from seismic amplitudes. We provide intuitive explanations for choices behind the proposed multi-stage hierarchical algorithm, outline its key steps, and briefly describe the field dataset which will be shown in the oral presentation
-
-
-
Constant Versus Variable-Density Asymptotic Linearized Direct Waveform Inversion: a Case Study from an Australian Seismic Dataset
Authors M. Farshad, H. Chauris and L. LopesSummaryFor quantitative seismic imaging, least-squares reverse time migration can be formulated as a linearized full-waveform inversion problem. A successful application of such a method requires many migration/demigration cycles. An alternative is asymptotic linearized direct waveform inversion, estimating quantitative results within a single iteration while having approximately the same computational burden as reverse time migration. The direct inverse was originally proposed for constant-density acoustics, and recently has been extended to variable-density acoustics. In this paper, to examine the importance of accounting for density variations, we compare constant- and variable-density linearized direct waveform inversion techniques applied to a 2D towed-streamer real dataset acquired in the northwest of Australia, in the Carnarvon Basin. The variable-density linearized direct waveform inversion not only delivers higher resolution subsurface images but also better reconstructs the field data than does the constant-density approach.
-
-
-
Attenuation Sensitivity Kernel Calculation in Anelastic Wave Equation Tomography
More LessSummaryImaging the subsurface attenuation models using wave equation based tomography methods is an important and challenging task. The forward modeling of wave propagation in anelastic media can be considered as solved based on superposition of rheological bodies with the generalized standard linear solid model. However, in previous literature, different approaches are employed to calculate the attenuation sensitivity kernels, which are essentially important for anelastic wave equation tomography. This study examines different theoretical frameworks for constructing the attenuation sensitivity kernels. In the numerical examples, these methods are used to calculate the anelastic sensitivity kernels for comparison. We have found that in the presence of velocity errors, the attenuation sensitivity kernels calculated with memory strain variables can resolve the attenuation anomalies better suffering from fewer trade-off artifacts.
-
-
-
Value of Multi-Parameter FWI
Authors P. Trinh, B. Duquet, C. Rivera and S. WhiteheadSummaryFull waveform inversion (FWI) has become an indispensable step in our internal VMB workflow, especially in complex geological contexts. Each main component of FWI is well documented in the literature but the optimal combination of different components to maximize the values of FWI in different imaging contexts remains unclear. Our internal experiences show that the multi-parameter kernel is one of the key factors for successful inversion applications and also to push FWI beyond its standard limit. To enhance the quality of the constructed velocity model, we have built an eco-system around multi-parameterization with adapted cost-functions, data, and model weightings. The proposed strategy appears to be robust in different acquisition settings and geological contexts. Besides, the multi-parameterization kernel does not prevent us to go to higher frequencies. We show that it has the capacity to build reliable high-resolution velocity and Acoustic Impedance models, which fit well data and have genuine interpretative benefits.
-
-
-
Elastic Full-Waveform Inversion for Improved Salt Model Building in the Central North Sea
Authors N. Masmoudi, W. Stone, A. Ratcliffe, R. Refaat and O. LeblancSummaryConventional acoustic full-waveform inversion (FWI) workflows exhibit limitations when the assumption of slowly varying elasticity is violated. Mode conversions between P- and S-waves occur at sharp interfaces and lead to differences in the wave-equation modelling. This indicates velocity model building should benefit from using the more accurate elastic FWI solution in areas of strong elasticity. We demonstrate this with results from a typical salt diapir in the Central North Sea, where the acoustic FWI result has already generated a good velocity model and subsequent image. Even with a basic assumption related to the P- and S-wave velocity (Vp/Vs) ratio, the elastic FWI result can further improve both the velocity model and resulting reverse time migration image, especially in areas of structural and velocity complexity that are associated with strong velocity contrasts.
-
-
-
Viscoelastic Full Waveform Inversion by Alternating Direction Method of Multipliers
Authors K. Aghazade, H. Aghamiry, A. Gholami and S. OpertoSummarySimultaneous reconstruction of velocity and attenuation models by FWI is challenging due to inter-parameter cross-talks. Moreover, classical FWI suffers from slow convergence and cycle skipping due to insufficient low-frequency content or inaccurate starting models. In this abstract, we develop the recently proposed extended FWI formulation based on the augmented Lagrangian (AL) method for the viscoelastic media where density is considered as a passive parameter. Linearization of the velocity and attenuation subproblems recast the FWI as tri-linear optimization problem where data augmented system (DA), seismic phase velocities and attenuation factors are updated in an alternating mode. We investigate three main features that lead to more accurate estimation of these variables and mitigate inter-parameter cross-talk: (i) the properties of the DA wavefields for viscoelastic media, (ii) the implementation of the Hessian matrix and (iii) robust regularization implementation to handle crosstalk issue.
-
-
-
Elastic Fwi to Directly Estimate Subsurface Elastic Properties in Sub-Salt Environment – Field Data Example
Authors P. Routh, Y.H. Cha, V. Brytik, O. Burtz, J. Sitgreaves, J. Barr, D. Chu, M. Gutierrez and R. SainSummaryPre-Salt seismic data typically suffers from significant challenges due to illumination, non-primary mode noises (converted waves, multiples) arising from hard contrast. The cumulative effect leads to loss of signal quality and loss of frequency in the data. Elastic FWI provides a path forward in addressing many of these challenges by directly inverting the raw seismic shots to produce elastic properties such as impedance, Vp/Vs and P-wave velocity. Using a field example we will demonstrate the efficacy of the eFWI approach.
-
-
-
First Steps in the Application of Physics-Informed Neural Networks to Full Waveform Inversion
Authors L. De Souza, N. Desassis, H. Chauris and E. HachemSummaryPhysics-informed neural networks are a method used for solving differential equations by approaching their solution with a neural network. This method has been used successfully for solving inverse problems based on many types of partial derivative equations. We discuss the first steps in the application of physics-informed neural networks to full waveform inversion. We leverage an automatic differentiation framework available in Julia and explore different optimization techniques, as well as some architecture ideas to determine the optimal velocity and the corresponding pressure wave field. In a simple one-dimensional case, our approach can retrieve the correct velocity perturbation as well as the up-going part of the wave field.
-
-
-
Image-to-image Seismic Interpolation
Authors M. Fernandez, R. Durall, N. Ettrich, M. Delescluse, A. Rabaute and J. KeuperSummaryIn this work, we explore three deep learning algorithms apply to seismic interpolation: deep prior image (DPI), standard, and generative adversarial networks (GAN). The standard and GAN approaches rely on a dataset of complete and decimated seismic images for the training process, while the DPI method learns from a decimated image itself, without training images. We carry out two main experiments, considering 10%, 30%, and 50% of regular and irregular decimation. The first tests the optimal situation for the GAN and the standard approaches, where training and testing images are from the same dataset. The second tests the ability of GAN and standard methods to learn simultaneously from three datasets, and generalize to a fourth dataset not used during training. The standard method provides the best results in the first experiment, when the training distribution is similar to the testing one. In this situation, the DPI approach reports the second best results. In the second experiment, the standard method shows the ability to learn simultaneously and effectively three data distributions for the regular case. In the irregular case, the DPI approach is more effective. The GAN approach is the less effective of the three deep learning methods in both experiments.
-
-
-
Physics Assisted Deep Learning for Full Waveform Inversion
More LessSummaryWe develop a deep learning framework for full waveform inversion in which the network learns the inverse operator together with regularization. Seismic shot gathers are used as input to the network and its output is treated as the desired model parameters. We use the output of the network to compute synthetic seismograms using finite differences that are compared against the input gathers. Realistic acoustic (salt model) and elastic synthetic data examples are used to demonstrate the effectiveness of our algorithm.
-
-
-
Geophysical and Core Data Integration to Characterise and offshore Freshened Groundwater System in the Canterbury Bight
Authors Z. Faghih, M. Jegen, A. Haroon, C. Berndt, R. Gehrman, A. Micallef and K. SchwalenbergSummaryOffshore groundwater systems have been suggested as alternative sources of potable water in islands and coastal regions. In this study, we integrate offshore controlled-source electromagnetic (CSEM) with borehole data to identify an offshore groundwater system in the Canterbury Bight, New Zealand. CSEM data were acquired with a seafloor-towed system along four profiles and 2-D inversion was carried out using MARE2DEM to derive resistivity models along each profile. Moreover, a trans-dimensional Bayesian inversion was conducted to assess the distribution of plausible resistivity-depth models. The study area was previously investigated during IODP Expedition 317 in which a pore-fluid salinity anomaly (24 psu at 40 mbsf) was recorded in borehole U1353. A comparison between the CSEM resistivity model and the resistivity-depth profile converted from pore-water salinity within the borehole shows a strong correlation between CSEM and borehole data at the closest waypoint to site U1353. We show through a Markov-Chain Monte Carlo approach that our estimates of seafloor resistivity agree with the measured borehole data. The computed resistivity distributions at the borehole provide significant evidence that the CSEM inversion models can be used to extrapolate groundwater inferences from the borehole onto a basin scale providing improved geophysical imaging capabilities for offshore freshened groundwater.
-
-
-
Towards Integrated Ground Models - an Example from TNW Offshore Windfarm
Authors G. Sauvin, M. Vanneste, R. Klinkvort, J. Dujardin and C. ForsbergSummaryGround modelling for large offshore wind project is challenging. Site investigation data should be integrated as much as possible up front and/or at early stages, allowing improved decision-making ahead of finalizing the ultimate design layout at the site. Addressing the uncertainty chain in the prediction of soil properties in a reliable, systematic and geostatistically robust way is also a key aspect of ground modelling which may have implications on the way one will conduct next generation site investigations. A generic workflow to build a 3D model of the subsurface populated with soil properties is proposed and applied to the Ten Noorden van de Waddeneilanden Wind Farm Zone, offshore the Dutch coast.
-
-
-
Rapid Replication of Best Practices through Machine Learning. Never mind the network, look at the data!
Authors H. Rynja, F. Bazargani, A. Chandran, Z. Liu, C. Sutton, J. Vamaraju, J. Vila, S. Ye and W. ZhangSummaryWe aim to demonstrate that machine learning (ML) can reduce turn-around substantially without sacrificing quality for seismic processing. We show this in the context of velocity model building based on Residual Moveout and focus on two specific elements: gather clean-up and RMO picking. We find that it is critical not only to curate well labelled data, but also to better condition the data as per learning problem requirements. While we optimized the supervised-learning ML networks for training, our results got a substantial boost through the label and data conditioning workflows developed for the problems. When applying on seismic data it was not trained on, the ML networks generalized well not only for other seismic data from the Gulf of Mexico and Brazil but also for deep seismic data from the Black Sea, Nigeria, and Oman. Applying the machine-learning networks reduced turn-around both for gather clean-up and for RMO picking from 1–2 weeks to less than a day. Less time is needed for tuning the workflow, testing parameters, laborious QC and climbing the learning curve. Execution on the compute is also faster. Turn-around for velocity model building was reduced both for a ray-based workflow and for a wave-equation-based workflow.
-
-
-
Deep Learning in Seismic Inverse Problems with Recurrent Inference Machines
Authors I. Vasconcelos, H. Peng, M. Ravasi and D. KuijpersSummaryMachine learning approaches are rapidly finding their way into many applications in processing and imaging seismic data. More specifically, various convolutional deep-learning architectures are currently being explored for seismic data processing tasks from denoising to imaging. Here, we present Recurrent Inference Machines (RIMs): a recurrent network architecture designed specifically for inverse problems, where a known forward operator is known and used as a constraint. We describe how both the original RIM and its invertible counterpart (iRIM) are designed to mimic gradient-based optimisation methods, and thus learn to perform data-driven regularisation and implicit model shaping due to their deep learning nature. We show examples of using RIMs to perform seismic data interpolation and image-domain inversion by deblurring, benchmarking them against UNets as a more widely-used deep learning architecture. Our examples show that RIMs outperform UNets particularly in dealing with features not necessarily present in the training data, due to the role of the forward operator as an additional constraint in training.
-