Construction and Application of a Foundation Model Based on Prestack Seismic Data

Seismic reflection pattern analysis or seismic facies analysis is crucial for subsurface reservoir prediction. Supervised pre-stack reflection pattern analysis utilizing well-logging data can fully utilize abundant reservoir information in pre-stack data, and provide clearer physical interpretations than unsupervised methods. However, pre-stack seismic data is high-dimensional, and the well-logging data labels are limited. Traditional convolutional neural network-based approaches face challenges in capturing long-range dependencies across different angle-gathers in pre-stack seismic data due to the limitations of their receptive fields. Additionally, existing data augmentation methods lack constraints and physical guidance. To tackle these problems, we introduce a supervised pre-stack seismic reflection pattern analysis method based on the ConvNext network and incorporating physics-guided and active learning for label augmentation. The ConvNext model incorporates the large kernel attention (LKA) mechanism, enhancing the model's sensitivity to stratigraphic and spatial features in pre-stack seismic data. To reduce model ambiguity, we develop a label augmentation algorithm that combines active learning and physical attributes. The experiments on synthetic data and real data demonstrate that our method has better performance than the traditional approaches in pre-stack seismic reflection analysis.

Semantically consistent multi-view representation learning

Seismic derivable elastic attributes, e.g., elastic impedance, lambda-rho, mu-rho, and Poisson impedance (PI), are routinely being used for reservoir characterization practice. These attributes could be derived from inverted VP, VS, and density, and usually indicate high sensitivity to reservoir lithology and fluid. Due to the high sensitivity of such elastic attributes, errors or measurement noise associated with the acquisition, processing, and inversion of prestack seismic data will propagate through the inversion products, and will lead to even larger errors in the computed attributes. To solve this problem, we have developed a two-step cascade workflow that combines linear inversion and nonlinear optimization techniques for the improved estimation of elastic attributes and better prediction and delineation of reservoir lithology and fluids. The linear inversion in the first step is an inversion scheme with a sparseness assumption, based on L1-norm regularization. This step is used to select the major reflective layer locations, followed in the second step by a nonlinear optimization process with the predefined layer structure. The combination of these two procedures produces a reasonable blocky earth model with consistent elastic properties, including the ones that are sensitive to reservoir lithology and fluid change, and thus provides an accurate approach for seismic reservoir characterization. Using PI, as one of the target elastic attributes, as an example, this workflow has been successfully applied to synthetic and field data examples. The results indicate that our workflow improves the estimation of elastic attributes from the noisy prestack seismic data and may be used for the identification of the reservoir lithology and fluid.

We present an integrated study of fracture detection by analyzing the seismic anisotropy in the 3D P-wave seismic data from Caidamu Basin in Northwest China. The study area lies on a big anticline which runs from southeast to northwest with the area around 400 km2, and each flank of the anticline develops a reverse fault. The target layer is a fractured gas reservoir at a depth of around 3 km just under the anticline. We first study the fracture characteristics from outcrop analogues, core samples and well logs. We then analyze the coherence and curvature attributes from post-stack seismic data, and azimuthal variation of amplitude and traveltime from pre-stack seismic data. After a careful and integrated analysis, we find that fractures are widespread and regularly distributed in the middle of study area along the anticline strike, and form good pathways and store space for fluid. The final results estimated using azimuthal analysis of pre-stack seismic amplitude data are consistent with results of core and log analysis at the well locations, which extended our knowledge of the fracture distribution in the reservoir.

Different from the stacked seismic data, pre-stack data includes abundant information about shear wave and density. Through inversing the shear wave and density information from the pre-stack data, we can determine oil-bearing properties from different incident angles. The state-of-the-art inversion methods obtain either low vertical resolution or lateral discontinuities. However, the practical reservoir generally has sharp discontinuities between different layers in vertically direction and is horizontally smooth. Towards obtaining the practical model, we present an inversion method based on the regularized amplitude-versus-incidence angle (AVA) data to estimate the piecewise-smooth model from pre-stack seismic data. This method considers subsurface stratum as a combination of two parts: a piecewise smooth part and a constant part. To fix the ill-posedness in the inversion, we adopt four terms to define the AVA inversion misfit function: the data misfit itself, a total variation regularization term acting as a sparsing operator for the piecewise constant part, a Tikhonov regularization term acting as a smoothing operator for the smooth part, and the last term to smoothly incorporate a priori information for constraining the magnitude of the estimated model. The proposed method not only can incorporate structure information and a priori model constraint, but also is able to derive into a convex objective function that can be easily minimized using iterative approach. Compared with inversion results of TV and Tikhonov regularization methods, the inverted P-wave velocity, S-wave velocity and density of the proposed method can better delineate the piecewise-smooth characteristic of strata.

Summary We develop a nonlinear iterative inversion algorithm for estimating three-dimensional (3D) reservoir parameters and initial fluid saturations jointly from pre-stack seismic data and fluid production history. The production measurements and the seismic data are synthetically generated from the output of a multi-phase fluid flow simulator. A Biot-Gassmann rock physics/fluid substitution model is used to enforce a deterministic link between the multi-phase flow parameters and the elastic properties. Iterative nonlinear optimization is used to solve the leastsquares minimization problem associated with the inversion. Fluid production measurements are sensitive to initial fluid saturations, fluid properties, porosity, and permeability. Pre-stack seismic data, on the other hand, provide good lateral and vertical control on lithology and fluid distributions. The proposed joint inversion approach, therefore, allows one to effectively integrate the best of the two measurement sets into a consistent 3D distribution of petrophysical variables and fluid saturations. Such distribution can be used for the accurate assessment of infill drilling and enhanced-oil-recovery operations. Synthetic test examples are presented with the sole objective to advance a proof of concept for the proposed joint inversion technique. The inversions require large computer resources and hence efficient numerical algorithms. We analyze the relative value of both sets of data and propose an extension of our algorithm to assimilate time-lapse 3D seismic data.

Facial expression recognition (FER) in uncontrolled, in-the-wild settings is a challenging task due to variations in pose, illumination, occlusion, and background clutter. The state-of-the-art methods on FER employ fixed size pooling in attention networks, resulting in suboptimal robustness in unconstrained environments. In this work, a novel Multi-Scale Attention Network (MSAN) is integrated with the YOLOv8 backbone architecture for expression classification. The YOLOv8 backbone provides efficient feature extraction with a lightweight, detection-oriented architecture, enabling real-time processing and strong spatial feature representation. The Convolutional Block Attention Mechanism (CBAM) is employed in the proposed work. The standard convolutional blocks in YOLOv8 backbone are replaced with MSAN, wherein the fixed-size pooling in the spatial attention mechanism of CBAM is replaced with a Pyramid Pooling Mechanism (PPM). The PPM employs adaptive average pooling at multiple resolutions to capture multi-scale contextual information, which is subsequently fused into a unified feature representation. This enables the spatial attention mechanism in MSAN to capture both global facial context and local region-level cues simultaneously, leading to superior robustness to scale variations, partial occlusions, and diverse expression variations. The work is evaluated on benchmark wild datasets including RAF-DB, FER2013, and AffectNet, with all experiments performed using 5-fold cross-validation. Experimental results demonstrate that the multi-scale attention network with PPM in YOLOv8 backbone achieved improvement in classification accuracy and robustness when compared with the state-of-the-art methods for expression classification in the wild. These findings highlight the effectiveness of integrating MSAN into lightweight detection-based backbones for real-world FER applications.

Delineating the extend of a producing reservoir is one of the main challenge of exploration and development interpreters. Our approach to address this challenge is to use ?spectral decomposition? of pre-stack data couple with an AVA/AVO studies (Amplitude Versus Angle/Offset). Seismic spectral decomposition is a known method to determine more accurately the lithology of an area of interest. AVO effect are used to identify and quantify the extend of a fluid anomaly in a potential reservoir. Combine together, the aim of this method is to provide a better delineation of the producing area. A shared earth approach will be used in order to perform cross validation of the results at each step. The first step is a 1D study well bore centred: Using well logs information: Vp (compressional velocity), Vs (Shear velocity) and the density to compute a first set of synthetic gather that will constitute our reference. The chosen well has an AVO effect and should be visible on this reference set. A real set of pre-stack gather at the well bore will then go through a series of tests to identify the best frequency bandwidth that identify more accurately the AVO effect. In a second phase, this frequency bandwidth will be used to produce two results: A 3d set of gather with this optimum bandwidth. Producing Gradient and the Intercept of this set, a cross plot of those attributes should confirm the result of the 1D study and will help to delineate the AVO effect spatial extent. An inversion based on this spectral decomposition should help to delineate more accurately the reservoir lithology Those results unified well bore information, seismic pre and post stack data, frequency analysis and inversion results. The analysis will tend to demonstrate that lithology and gas anomaly can be better mapped using combined methodology, compare to a juxtaposition of methods.

Recognizing seismic phase as a primary attribute in seismic processing workflows, we apply circular statistics, a robust data-driven approach for correcting phase distortions in prestack seismic data. Unlike traditional linear methods that struggle with wrapped phase and often defer phase diagnostics to the final processing stages, the proposed approach treats phase as a circular variable. We compute the circular mean, variance, and von Mises concentration parameter directly from phase ensembles in the frequency domain. These parameters provide insights into phase stability and coherence without needing phase unwrapping or wavelet assumptions. Synthetic tests using additive and multiplicative noise models confirm that phase distributions follow the von Mises distribution, an analog of the normal distribution for circular variables, with circular statistics reliably tracking the true phase even in low-signal-quality scenarios. Field examples demonstrate how this framework can map phase behavior across frequency and offset, enabling the detection of coherence bands and assessing the impact of each processing step on phase fidelity. The proposed approach can be particularly valuable in land acquisition, where prestack data often exhibit a low signal-to-noise ratio. Circular statistics allow us to evaluate phase integrity at each frequency, facilitating novel data conditioning and acquisition design strategies.

The bioclastic limestone reservoirs of Cretaceous period occupy an important position in the petroleum industry of Middle East. It is the carbonate heterogeneity that is challenging the accuracy of the reservoir prediction, which brings forward higher requirements for the seismic data quality. Besides, some seismic data are processed more than 10 years ago, the signal to noise ratio (SNR) is relative low due to the random noise and coherent noise like acquisition footprint anomalies. The acquisition footprint artifacts caused by acquisition and processing seriously suppress the true stratigraphic features, which can result in pitfalls in seismic interpretation, seismic attribute analysis as well as seismic inversion. While the pre-stack seismic data is usually unavailable, which means that the noise can hardly be subtracted by conventional pre-stack seismic processing workflows, such as statics, high-resolution velocity analysis and ground roll attenuation. Consequently, a comprehensive post-stack seismic data conditioning workflow is necessary to solve the above problems. In order to improve the post-stack seismic data quality, a comprehensive data conditioning workflow are applied for noise suppression. Firstly, structural-oriented filtering is utilized to attenuate random noise and partial acquisition footprint artifacts. Then 2D waveform transform of seismic amplitude and filtered seismic attribute in x-y domain are calculated, to separate acquisition footprint anomalies (large wave number in kx-ky domain) from true structural signal (small wave number in kx-ky domain) by interactive analysis. The application of Laplace-Gaussian (LoG) filter deserves an obvious improvement in acquisition footprint suppression workflow. The comprehensive noise attenuation workflow in this paper can effectively remove both periodic and non-periodic noise to obtain higher signal to noise ratio (SNR) for post-stack seismic volume. In this way, the stratigraphic features (tidal-channel, reef-beach complex) can be more clearly depicted and some artifacts caused by noise will disappear in seismic attribute calculation, seismic inversion and reservoir prediction.

Reservoir characterization has a leading role in the reservoir geophysics and reservoir management. Since the interests of the reservoir geophysics and reservoir managements are the elastic properties and reservoir properties of the subsurface rock for their purposes, a robust method is required for converting seismic data into elastic properties. Accordingly, by employing a rock physics model and using the inverted seismic data, one can describe the reservoir for purposes such as improvement in the production of the reservoir. In the present study, we employ one of the methods for converting the seismic data into the elastic properties. This method of inversion is known as simultaneous inversion, which is grouped in amplitude-variation-with-offset (AVO) inversion category. In this method, unlike the other methods of AVO inversion, the pre-stack seismic data are directly inverted into the elastic properties of the rock and an excellent lithology and fluid indicator (VP/VS) are provided. Then, this indicator is tested on one of the oilfields of the Persian Gulf. Moreover, by means of this method, one can locate the fluids contact and the lithological interlayers; also, by the inversion results, which are the cubes of the seismic properties of the rock, one can generate sections of the elastic properties of the rock such as Poisson’s ratio and Young modulus which are useful for geomechanical analysis. Therefore, this kind of method is a quick way for the prior analysis of the studied area.

Seismic inversion requires two main operations relative to changes in the frequency spectrum. The first operation is deconvolution, used to increase the high frequency component of the observed seismic data and the second operation is integration of a reflectivity function to decrease the high frequencies and increase low frequencies of the seismic signal. The first operation is very unstable and non-unique for noisy seismic data. The second operation is very sable in high frequencies but has problems in low frequencies due to undefined low frequency data in seismic traces. By performing both of these operations simultaneously the operation will be stable in high frequency area and can be effectively stabilized in low frequency area based on an a priori acoustic impedance power spectrum and use Tikhonov and Arsenin's (1979) regularization technique. This approach can be applied to post-stack and pre-stack seismic data.

Surface reflection tomography is an inversion method that attempts to determine simultaneously the subsurface elastic parameters and reflector depths from pre-stack seismic data. To date, this technique has been applied to acoustic and isotropic elastic media. In this paper, we study surface reflection tomography in the anisotropic elastic case. Anisotropic media are described by many more elastic parameters than are needed to describe isotropic media; these additional parameters are useful for obtaining rock properties such as crack density, pore shape and fracture strike that cannot be found with isotropic methods. Similar to isotropic surface reflection tomography, many features of an anisotropic model are poorly resolved due to limited ray path coverage; the vertical smearing of isotropic tomography occurs in each elastic parameter in the anisotropic problem. Unlike the isotropic problem, however, there is additional indeterminacy in the solution of the anisotropic tomography problem because of ambiguity amongst the several elastic parameters needed to describe anisotropic media. Also unlike the isotropic problem, the reflector depths participate very strongly in this ambiguity. We investigate the nature of this indeterminacy by studying the null space for linearized tomography, that is, the class of model perturbations of a background medium which, to first order, cause no perturbation at all in the surface reflection traveltimes. Such model perturbations cannot be determined from the traveltime perturbations; describing these null space model perturbations gives insight into the indeterminacy in the anisotropic problem. Complementary to computational approaches towards identifying the null space for discrete formulations of tomography, we study a continuum formulation. Our first set of results concerns the ‘elastic’ null space: perturbations of the elastic parameters (with zero depth perturbation) which cause no perturbation in the traveltimes. The results here are very similar to previous results for cross-well transmission tomography. As expected, the ‘elastic’ null space is larger than the isotropic null space due to the ambiguity amongst the elastic parameters. We identify three categories of model perturbations in the ‘elastic’ null space. The first category consists of model perturbations for which the perturbation in each of the individual elastic parameters is itself either in the isotropic null space, or in a closely related set of model perturbations which we call the odd isotropic null space. Elements in the second category are anisotropic versions of the most well-known isotropic null space elements: perturbations which are polynomials in the horizontal variable with coefficients which are functions of the depth variable satisfying certain linear integral constraints; unlike the isotropic problem, the integral constraints in the anisotropic problem couple together the several elastic parameters. The third category consists of model perturbations satisfying zero boundary conditions at the surface and at the reflector for which a specific linear combination of integrals and derivatives of the several elastic parameters is in the isotropic null space. In particular, there are model perturbations in this third category which represent anomalies that are completely contained in the interior of the model and yet are in the ‘elastic’ null space; this behaviour is different from the isotropic problem. These categories are sufficient to describe the ‘elastic’ null space completely. We demonstrate that every model perturbation in the ‘elastic’ null space is the sum of an element in the first category (indicating an indeterminacy of the same nature as in the isotropic problem in each of the elastic parameters separately) and an element in the third category (indicating an ambiguity amongst the parameters). The second category gives a rich family of examples of sums of null space elements in the first and third categories, and thereby gives a sense of just how large the ‘elastic’ null space is. Moreover, we show that the traveltime perturbations caused by an elastic perturbation determine only a small number of features of the elastic perturbation which distinguish between the several elastic parameters. We identify these features pre-cisely: they are functions of the horizontal variable representing vertical averages of combinations of the elastic parameters and their derivatives. Elastic parameters influencing the vertical velocity appear more prominently in these features than those influencing the horizontal velocity, and these features are closely related to the zero-offset traveltimes and the normal moveout velocities. Our second set of results concerns the so-called velocity–depth ambiguity, more precisely here the elastic parameter versus the depth parameter ambiguity. We demonstrate that for any smooth depth perturbation, there is a perturbation in the elastic parameters for which the combined elastic and depth perturbations cause zero traveltime perturbations; moreover, this can be accomplished by an elliptically anisotropic elastic perturbation with a vertical axis of symmetry, perturbing the vertical velocity alone. This implies that the velocity–depth ambiguity cannot be resolved here as it can be in the isotropic case. We show an example of an anticlinal structure embedded in a homogeneous background whose reflection traveltimes exactly match those from a model with elliptically anisotropic elastic perturbations overlying a flat reflector. We also draw a number of conclusions from these results for the special case of transversely isotropic media with a vertical symmetry axis.

SUMMARY In this paper, we present a method for the direct stratigraphic inversion of lithology from pre-stack seismic data. This approach is a generalization to discrete variables of a methodology for petrophysical seismic inversion of continuous attributes.

AVO/AVA (amplitude variation with an offset or angle) inversion is one of the most practical and useful approaches to estimating model parameters. So far, publications on AVO inversion in the Fourier domain have been quite limited in view of its poor stability and sensitivity to noise compared with time-domain inversion. For the resolution and stability of AVO inversion in the Fourier domain, a novel robust Bayesian pre-stack AVO inversion based on the mixed domain formulation of stationary convolution is proposed which could solve the instability and achieve superior resolution. The Fourier operator will be integrated into the objective equation and it avoids the Fourier inverse transform in our inversion process. Furthermore, the background constraints of model parameters are taken into consideration to improve the stability and reliability of inversion which could compensate for the low-frequency components of seismic signals. Besides, the different frequency components of seismic signals can realize decoupling automatically. This will help us to solve the inverse problem by means of multi-component successive iterations and the convergence precision of the inverse problem could be improved. So, superior resolution compared with the conventional time-domain pre-stack inversion could be achieved easily. Synthetic tests illustrate that the proposed method could achieve high-resolution results with a high degree of agreement with the theoretical model and verify the quality of anti-noise. Finally, applications on a field data case demonstrate that the proposed method could obtain stable inversion results of elastic parameters from pre-stack seismic data in conformity with the real logging data.