Seismic Velocity Model Building Research Articles

Self-supervised, active learning seismic full-waveform inversion

A novel recursive, self-supervised machine-learning (ML) inversion scheme is developed. It is applied for fast and accurate full-waveform inversion of land seismic data. ML generalization is enhanced by using virtual super gathers (VSGs) of field data for training. These are obtained from midpoint-offset sorting and stacking after applying surface-consistent corrections from the decomposition of the transmitted wavefield. The procedure implements reinforcement learning concepts by adopting an inversion agent to interact with the environment and explore the model space under a data misfit optimization policy. The generated parameter distributions and related forward responses are used as new training samples for supervised learning. The active learning (AL) paradigm is further embedded in the procedure, for which queries on data diversity and uncertainty are used to generate fully informative reduced sets for training. The procedure is recursive. At each cycle, the physics-based inversion is coupled to the ML predictions via penalty terms that promote a long-term data misfit reduction. The resulting self-supervised, AL, physics-driven deep-learning inversion generalizes well with field data. The method is applied to perform full-waveform inversion (FWI) of a complex land seismic data set characterized by transcurrent faulting and related structures. High signal-to-noise VSGs are inverted with a 1.5D Laplace-Fourier FWI scheme. The AL inversion procedure uses a small fraction of data for training while achieving sharper velocity reconstructions and a lower data misfit when compared with previous results. AL FWI is highly generalizable and effective for land seismic velocity model building and for other inversion scenarios.

Investigating seismic mode conversions from an ultra-high-velocity caprock by physical modelling, numerical simulations and a Gulf of Mexico salt proximity VSP survey

SUMMARY Salt structures are widely distributed in many basins worldwide and play an important role in understanding tectonic movements, offering underground storage and sealing hydrocarbon traps. In addition to the acknowledged difficulties in mapping complex salt structures through seismic methods, when an evaporitic layer, such as anhydrite, forms over salt, it can introduce strong multimode conversions that can couple with the primary compressional wavefields and generate artefacts in resulting acoustic images. From two well-log suites from the Gulf of Mexico, we identify thin evaporitic caprocks on top of their salt bodies and analyse their elastic properties. Through controlled experiments, including physical and numerical modelling with a vertical seismic profiling survey geometry, we observe significant shear-mode conversions at the top of the ultra-high-velocity caprock, which further result in a family of prominent S and P (converted from S wave) modes across the top of the salt region. Similarly, in a field survey, we identify evident converted S waves and a multimode P wave (converted S wave in the anhydrite layer, and P wave elsewhere) following the primary P transmission inside the salt body. While separating the converted Smodes at the receiver end is more unambiguous, the multimode P waves could behave very similarly to the primary Pmodes and are more difficult to suppress. Under the common acoustic assumption of seismic velocity model building and imaging, complex mode conversions on top of the salt are generally ignored. Through controlled experiments and a field survey, we analyse the pitfalls associated with this omission. We emphasize the importance of understanding the physics of wave partitioning in the presence of a thin ultra-high-velocity layer on the top of the salt.

Automated salt model building: From compaction trend to final velocity model using waveform inversion

Conventional seismic velocity model building in complicated salt-affected areas requires the explicit identification of salt boundaries in migrated images and typically involves testing of possible subsurface scenarios through multiple generations. The resulting velocity models are slow to generate and may contain interpreter-driven features that are difficult to verify. We show that it is possible to build a full final velocity model using advanced forms of full-waveform inversion applied directly to raw field data, starting from a model that contains only a simple 1D compaction trend. This approach rapidly generates the final velocity model and migrates processed reflection data at least as accurately as conventionally generated models. We demonstrate this methodology using an ocean-bottom-node data set acquired in deep water over Walker Ridge in the Gulf of Mexico. Our approach does not require exceptionally long offsets or the deployment of special low-frequency sources. We restrict the inversion so it does not use significant energy below 3 Hz or offsets longer than 14 km. We use three advanced forms of waveform inversion to recover the final model. The first is adaptive waveform inversion to proceed from models that begin far from the true model. The second is nonlinear reflection waveform inversion to recover subsalt velocity structure from reflections and their long-period multiples. The third is constrained waveform inversion to produce salt- and sediment-like velocity floods without explicitly identifying salt boundaries or velocities. In combination, these three algorithms successively improve the velocity model so it fully predicts the raw field data and accurately migrates primary reflections, though explicit migration forms no part of the workflow. Thus, model building via waveform inversion is able to proceed from field data to the final model in just a few weeks. It entirely avoids the many cycles of model rebuilding that may otherwise be required.

Multiscale Fusion Network With SR-Attention for Seismic Velocity Model Building

Multiscale Fusion Network With SR-Attention for Seismic Velocity Model Building

Spyro: a Firedrake-based wave propagation and full-waveform-inversion finite-element solver

Abstract. In this article, we introduce spyro, a software stack to solve wave propagation in heterogeneous domains and perform full waveform inversion (FWI) employing the finite-element framework from Firedrake, a high-level Python package for the automated solution of partial differential equations using the finite-element method. The capability of the software is demonstrated by using a continuous Galerkin approach to perform FWI for seismic velocity model building, considering realistic geophysics examples. A time domain FWI approach that uses meshes composed of variably sized triangular elements to discretize the domain is detailed. To resolve both the forward and adjoint-state equations and to calculate a mesh-independent gradient associated with the FWI process, a fully explicit, variable higher-order (up to degree k=5 in 2D and k=3 in 3D) mass-lumping method is used. We show that, by adapting the triangular elements to the expected peak source frequency and properties of the wave field (e.g., local P-wave speed) and by leveraging higher-order basis functions, the number of degrees of freedom necessary to discretize the domain can be reduced. Results from wave simulations and FWIs in both 2D and 3D highlight our developments and demonstrate the benefits and challenges with using triangular meshes adapted to the material properties.

Seismic velocity model building using neural networks: Training data design and learning generalization

Data-driven artificial neural networks (ANNs) offer several advantages over conventional deterministic methods in a wide range of geophysical problems. For seismic velocity model building, judiciously trained ANNs offer the possibility of estimating high-resolution subsurface velocity models. However, a significant challenge of ANNs is training generalization, which is the ability of an ANN to apply the learning from the training process to test data not previously encountered. In the context of velocity model building, this means learning the relationship between velocity models and the corresponding seismic data from a set of training data, and then using acquired seismic data to accurately estimate unknown velocity models. We have asked the following question: What types of velocity model structures need to be included in the training process so that the trained ANN can invert seismic data from a different (hypothetical) geologic setting? To address this question, we create four sets of training models: geologically inspired and purely geometric, with and without background velocity gradients. We find that using geologically inspired training data produces models with well-delineated layer interfaces and fewer intralayer velocity variations. The absence of a certain geologic structure in training models, however, hinders the ANN’s ability to recover it in the testing data. We use purely geometric training models consisting of square blocks of varying size to demonstrate the ability of ANNs to recover reasonable approximations of flat, dipping, and curved interfaces. However, the predicted models suffer from intralayer velocity variations and nonphysical artifacts. Overall, the results successfully determine the use of ANNs in recovering accurate velocity model estimates and highlight the possibility of using such an approach for the generalized seismic velocity inversion problem.

Seismic Velocity Model Building Using Recurrent Neural Networks: A Frequency-Stepping Approach

Data-driven artificial neural networks (ANNs) demonstrably offer numerous advantages over conventional deterministic methods in a wide range of geophysical problems. For seismic velocity model building, judiciously trained ANNs offer the possibility of estimating subsurface velocity models; however, there are substantial challenges with effective and efficient network training. Motivated by the multi-scale approach commonly used to address FWI non-linearity challenges, we develop a frequency-stepping velocity model building approach that uses a sequence-to-sequence recurrent neural network (RNN) with built-in long short-term memory (LSTM). The input sequences to the LSTM-RNN consist of the frequency-domain seismic data ordered by frequency from lowest available to highest usable or chosen, while the corresponding output sequences are frequency-dependent smoothed velocity models. We compare models recovered using the trained RNN to the true models qualitatively and quantitatively. The normalized root-mean-square (NRMS) misfit between the true and predicted models has a mean of 6%, which confirms that the network recovers highly accurate models.

Acquisition/Processing

In seismic data processing, static corrections for near-surface velocities are derived from first-break picking. The quality of the static corrections is paramount to developing an accurate shallow velocity model, a model that in turn greatly impacts the subsequent seismic processing steps. Because even small errors in first-break picking can greatly impact the seismic velocity model building, it is necessary to pick high-quality traveltimes. Whereas various artificial intelligence-based methods have been proposed to automate the process for data with medium to high signal-to-noise ratio (S/N), these methods are not applicable to low-S/N data, which still require intensive labor from skilled operators. We successfully replace 160 hours of skilled human work with 10 hours of processing by a single NVIDIA Quadro P6000 graphical processing unit by reducing the number of human picks from the usual 5%–10% to 0.19% of available gathers. High-quality inferred picks are generated by convolutional neural network-based machine learning trained from the human picks.

Integrated VTI model building with seismic data, geologic information, and rock-physics modeling — Part 2: Field data test

Velocity model building is the first step of seismic inversion and the foundation of the subsequent processing and interpretation workflow. Velocity model building from surface seismic data only becomes severely underdetermined and nonunique when more than one parameter is needed to characterize the velocity anisotropy. The traditional seismic processing workflow sequentially performs seismic velocity model building, structural imaging/interpretation, and lithologic inversion, modifying the subsurface model in each step without verifications against the previously used data. We have developed an integrated model building scheme that uses all available information: seismic data, geologic structural information, well logs, and rock-physics knowledge. We have evaluated the accuracy of the anisotropic model in the image space, in which structural information is estimated. The lithologic inversion results from well logs and the dynamic seismic information (amplitude versus angle) are also fed back to the kinematic seismic inversion via a cross-parameter covariance matrix, which is a multivariate Gaussian approximation to the numerical distribution modeled from stochastic rock-physics modeling. The procedure of building the rock-physics prior information and the improvements using these extra constraints were tested on a Gulf of Mexico data set. The inverted vertical transverse isotropic model not only better focused the seismic image, but it also satisfied the geologic and rock-physics principles.

Salt delineation via interpreter-guided 3D seismic image segmentation

Although it is a crucial component of seismic velocity model building, salt delineation is often a major bottleneck in the interpretation workflow. Automatic methods like image segmentation can help to alleviate this bottleneck, but issues with accuracy and efficiency can hinder their effectiveness. However, a new graph-based segmentation algorithm can, after modifications to account for the unique nature of seismic data, quickly and accurately delineate salt bodies on 3D seismic images. In areas where salt boundaries are poorly imaged, limited manual interpretations can be used to guide the automatic segmentation, allowing for interpreter insight to be combined with modern computational capabilities. A successful 3D field data example demonstrates that this method could become an important tool for interactive interpretation tasks.

3D interpolation and extrapolation of sparse well data using rock-physics principles: Applications to anisotropic velocity model building

Mapping sparse well data into 3D volumes is a challenging problem. Because seismic-velocity model building requires 3D velocity and anisotropy volumes and most anisotropy estimates are calibrated to borehole data, the generation of 3D anisotropic models relies heavily on extrapolation of sparse well data. This spatial-extrapolation problem is ill posed, and many subjective decisions are made during model building. New advances in basin modeling and analysis facilitate the modeling of temperature and compaction history using simple physical principles. These fields can be employed as auxiliary fields to solve spatial-interpolation problems of sparse well-log data using the concept of interpolation in the rock-physics domain. The principle of interpolation in the rock-physics domain suggests that sparse well-log measurements in the physical (x, y, z) 3D space might still adequately sample the rock-physics space of temperature, porosity, and effective stress to enable proper reconstruction of the anisotropic velocity field in a manner consistent with diagenetic processes. Synthetic and field examples support this direct link between basin modeling and anisotropic velocity model building.

Seismic velocity model building: CE in Dallas, 2 November

Depth conversion and depth imaging (depth migration) of 3-D seismic data require a geologically and geophysically competent velocity model. The constantly improving quality of seismic data and increasing survey accuracy demand a correspondingly high quality and accuracy standard for velocity models. In depth conversion of time‐based maps, for example, using a single regional velocity function is rapidly being replaced by depth conversion of both the interpreted surfaces and the seismic data using a single unified high quality 3-D velocity model.