A Physics‐Guided Deep Learning Workflow for Partial‐Stack Seismic Inversion

Seismic reflection amplitude data are increasingly used in reservoir modeling to provide information on changes in earth properties away from well locations. In geostatistical reservoir modeling, the most common application is to use seismic data as background data in some form of comodeling. Seismic data image reflectors and not earth layer properties. Therefore, prior to use in comodeling, seismic data must first be transformed into an earth layer property. Typically, the transform is to acoustic impedance using an appropriate seismic inversion method. Seismic inversion methods generate results that are generally band limited in nature, resulting in limits to vertical resolution. The vertical resolution achieved can be an order of magnitude below the vertical model resolution required from geostatistical reservoir modeling, which is in the order of well-log resolution. Hence, in using seismic data, geostatistical modelers encounter a problem of downscaling, not the more commonly encountered upscaling problem. This difference in scale introduces scatter between the primary data with well-log order resolution and the secondary seismically derived rock property data used in the comodeling. As a result, to preserve vertical heterogeneity, only limited use of the secondary data can be made in comodeling procedures. This results in models that only partially fit the seismic data, i.e., only limited use is made of the seismic information. If the secondary data are more strongly imposed, the fit to the seismic data improves, but the required vertical heterogeneity is not preserved. The inability to overcome this difference in scale issue, therefore, limits the value of the application of comodeling methods to integrate seismic data into reservoir models. One class of geostatistical methods that overcomes this limitation relies on iterative geostatistical modeling. In these methods, referred to as geostatistical seismic inversion, the iterative modeling process is conditioned such that the final models generated closely match the seismic data while maintaining the required vertical heterogeneity. The application of these methods is computationally expensive relative to comodeling methods but is now practical for large models on today's desktop hardware. Relative to comodeling, geostatistical seismic inversion methods make full use of the information carried in the seismic data, resulting in a significant reduction in model uncertainty away from well control.

High Resolution, Simultaneous VSP and Land Seismic Acquisition

Accurate characterization of subsurface reservoirs is a core objective in hydrocarbon exploration and development. The quality of seismic reservoir characterization directly influences well placement, field development strategies, and reserve estimation workflows. When seismic inversion is constrained by limited bandwidth and tuning effects, thin beds are merged into composite responses, subtle facies transitions are obscured, and small-scale structural or stratigraphic features are either poorly resolved or completely masked. These limitations have direct operational consequences: missed pay intervals, incorrect mapping of reservoir connectivity, suboptimal placement of production and injection wells, and development plans burdened by unnecessary uncertainty. Challenges are especially acute in geologically heterogeneous settings where vertical and lateral variations occur at scales close to or below seismic tuning thickness. In these cases, the blending of reflectors caused by limited frequency content in the seismic data can obscure stratigraphic terminations, make thin impermeable baffles invisible, and reduce confidence in both qualitative and quantitative interpretations. The Jurassic Plover Formation in the Poseidon area of the Browse Basin, offshore North-West Australia, is a compelling example. Its interbedded sandstones, siltstones, and shales, deposited in variable-energy environments, form a complex internal architecture that strongly influences reservoir performance. Thin, laterally variable sand bodies and shale drapes exert significant control on fluid flow and sweep efficiency. In such settings, conventional post-stack inversion often reproduces regional impedance trends but fails to resolve the finer-scale architecture that governs production behavior. Over the past four decades, inversion methods have evolved considerably. Early deterministic acoustic impedance inversions, grounded in the convolutional model, provided a first step in transforming seismic amplitudes into impedance volumes for structural mapping and gross lithologic interpretation. Pre-stack inversion expanded capabilities by integrating amplitude variation with offset (AVO) information to estimate elastic parameters and improve lithology and fluid discrimination. Stochastic and geostatistical inversions sought to inject high-frequency detail through simulation, aiming to honor both seismic data and geologic plausibility. Spectral inversion and thin-bed reflectivity methods attempted to extend usable bandwidth and de-tune seismic responses, offering improvements in vertical resolution. Despite these advances, the vertical resolution achievable with most conventional methods remains fundamentally tied to the seismic bandwidth and physics of wave propagation (Russell and Hampson, 1991; Partyka et al., 1999). Features thinner than approximately one-quarter of the dominant wavelength are difficult to recover reliably, and any attempt to artificially increase resolution risks introducing noise or geological inconsistency without tight constraints.

Accurate mapping of structural faults and stratigraphic sequences is essential to the success of subsurface interpretation, geologic modeling, reservoir characterization, stress history analysis, and resource recovery estimation. In the past decades, manual interpretation assisted by computational tools — i.e., seismic attribute analysis — has been commonly used to deliver the most reliable seismic interpretation. Because of the dramatic increase in seismic data size, the efficiency of this process is challenged. The process has also become overly time-intensive and subject to bias from seismic interpreters. In this study, we implement deep convolutional neural networks (CNNs) for automating the interpretation of faults and stratigraphies on the Opunake-3D seismic data set over the Taranaki Basin of New Zealand. In general, both the fault and stratigraphy interpretation are formulated as problems of image segmentation, and each workflow integrates two deep CNNs. Their specific implementation varies in the following three aspects. First, the fault detection is binary, whereas the stratigraphy interpretation targets multiple classes depending on the sequences of interest to seismic interpreters. Second, while the fault CNN utilizes only the seismic amplitude for its learning, the stratigraphy CNN additionally utilizes the fault probability to serve as a structural constraint on the near-fault zones. Third and more innovatively, for enhancing the lateral consistency and reducing artifacts of machine prediction, the fault workflow incorporates a component of horizontal fault grouping, while the stratigraphy workflow incorporates a component of feature self-learning of a seismic data set. With seven of 765 inlines and 23 of 2233 crosslines manually annotated, which is only about 1% of the available seismic data, the fault and four sequences are well interpreted throughout the entire seismic survey. The results not only match the seismic images, but more importantly they support the graben structure as documented in the Taranaki Basin.

Self-adaptive denoising net: Self-supervised learning for seismic migration artifacts and random noise attenuation

Recent experience in applying recurrent neural networks (RNNs) to interpreting Permannt Downhole Gauge (PDG) records has highlighted the potential utility of machine learning algorithms to learn reservoir behavior from data. The power of the RNN resides in its ability to retain information in a form of memory of previous patterns and information contained in the previous behavior of phenomena being modeled. This memory plays a role of informing the decision at the present time by using what happened in the past. This property suggests RNNs as a suitable choice to model sequences of reservoir information, even when the reservoir modeler is faced with incomplete knowledge of the underlying physical system. Convolutional Neural Networks (CNNs) are another variant of machine learning algorithms that have shown promise in sequence modeling domains, such as audio synthesis and machine translation. In this study, RNN and CNN were applied to tasks that traditionally would be modeled by a reservoir simulator. This was achieved by formulating the relationship between physical quantities of interest from subsurface reservoirs as a sequence mapping problem. In addition, the performance of a CNN layer as compared to RNN was evaluated systematically to investigate their capabilities in a variety of tasks of interest to the reservoir engineer. Preliminary results suggest that CNNs, with specific design modifications, are as capable as RNNs in modeling sequences of information and as reliable when making inferences to cases that has not been seen by the algorithm during training. Design details and reasons pertaining to the way these two seemingly different architectures process information, and handle memory are also discussed.

Seismic inversion in geophysics is a method that uses certain prior information, such as known geological laws and well logging and drilling data, to infer the physical parameters of underground media, such as wave impedance, velocity, and density, from seismic observation data, and thereby obtain the spatial structure and physical properties of underground strata. Seismic inversion is a highly complex problem with multiple solutions, and with the advancement of collection equipment, the volume of geophysical observation data is increasing at an astonishing rate. This presents new challenges for the accuracy and speed of seismic data inversion methods. There is an urgent need to develop intelligent and efficient inversion technologies for seismic inversion.Deep learning networks have powerful nonlinear fitting capabilities and can be used to solve complex nonlinear problems, such as seismic inversion. However, the predictive ability of deep learning networks largely depends on the quantity of training data. In the early stages of oil and gas exploration and development, the amount of well logging label data available for training is very limited, which poses a challenge for the application of deep learning in seismic inversion. Semi-supervised learning seismic inversion methods consider both data mismatch issues and well logging data mismatch issues, and can better adapt to inversion problems in real-world scenarios. Unlike supervised learning approaches, semi-supervised learning does not require a large amount of labeled data, thus it can better handle situations of data scarcity or mismatch.This paper utilizes a semi-supervised learning workflow to perform inversion on post-stack seismic data and has conducted experimental validation on the Marmousi 2 model. The experimental results show that, compared to supervised learning networks, the semi-supervised learning network still exhibits good predictive performance with a limited amount of data, demonstrating better stability in the presence of noise and geological variations, and effectively learns the mapping relationship between seismic data and artificial intelligence. Furthermore, as the amount of training data increases, the performance of the network also improves, confirming the importance of data quantity for training deep learning networks. The application results of the network on actual data indicate that the network has broad application prospects and feasibility. However, since the network is based on a channel-by-channel inversion method, there is still a lack of representation in terms of lateral continuity, which requires further exploration and improvement in subsequent research.

Seismic data is an essential source of information often contaminated with disturbing, coherent and random noise. Seismic random noise has degenerative impacts on subsequent seismic processing and data interpretation. Thus, seismic noise attenuation is a key step in seismic processing. Convolutional Neural Networks (CNNs) have proven successful for various image processing tasks in multidisciplinary fields and this paper aims to study the impact of three CNN architectures (autoencoders, denoising CNNs (DnCNN) and residual dense networks (RDN)) on improving the signal to noise ratio of seismic data. The work consists of three steps: Data preparation, model training and model testing. In this study we have used real seismic data to prepare the training dataset we have manually added noise. Most studies on seismic noise attenuation, study only a single kind of noise. However this paper suggests making our approach by exposing the model to many kinds of noises and noise levels such as Guassian noise, Poisson noise, Salt and Pepper and Speckle noise. In this paper we have analysed the performance of three models. Autoencoders: This architecture consists of two parts, the encoders and the decoders. The encoder consists of convolutions on the input image to extract all key information and map it to a latent space with loss of unnecessary data(noise) while the decoder reconstructs the image from the latent space to a seismic image while high signal to noise ratio. DnCNNs: This architecture is a combination of residual learning and batch normalization and mainly consists of three kinds of blocks. The model is trained to predict the residual image, that is the difference between the noisy observation and the latent clean image. RDNs: This architecture comprises of shallow feature extraction net, residual dense blocks (RDBs), dense feature fusion, and lastly up-sampling net. The data prepared as mentioned above is trained on all three CNN models across different noise levels and the performance of these models was compared. The model is finally tested on a batch of unseen noisy seismic sections and the performance is measured by an l2 loss namely mean squared error and the improvement in signal to noise ratio. The resultant images from all three architectures across different noise levels have drastically improved signal to noise ratio and thus the application of CNN as a denoiser for seismic images proves to be successful. It is important to note that when comparing the difference plots(Noisy image minus the denoised image) we found minimal signal leakage. While the application of CNN for image pre-processing has seen great success in other fields, mathematical denoising techniques such as F-K filter, tao-p filter are still used in oil and gas industry particularly in seismic denoising. After thorough review, this paper studies some of the most successful denoising CNN architectures and its success in seismic denoising.

Deep learning (DL) applications of seismic reservoir characterization often require the generation of synthetic data to augment available sparse labeled data. An approach for generating synthetic training data consists of specifying probability distributions modeling prior geologic uncertainty on reservoir properties and forward modeling the seismic data. A prior falsification approach is critical to establish the consistency of the synthetic training data distribution with real seismic data. With the help of a real case study of facies classification with convolutional neural networks (CNNs) from an offshore deltaic reservoir, we have highlighted several practical nuances associated with training DL models on synthetic seismic data. We highlight the issue of overfitting of CNNs to the synthetic training data distribution and propose regularization strategies to address it. We demonstrate the efficacy of our proposed strategies by training the CNN on synthetic data and making robust predictions with real 3D partial stack seismic data.

Model-constrained and data-driven double-supervision acoustic impedance inversion

Seismic interpretation is an essential factor in the success of field development projects because it allows geologists and engineers to better understand subsurface formations and identify potential resources; however, noisy seismic data, especially in regions with complex geological structures, frequently hinder this process. Such noise emerging from equipment limitations, environmental interference, and inherent geological complexities can obscure essential geological features, making accurate subsurface imaging challenging and reducing the effectiveness of traditional interpretation methods. This paper presents a machine learning (ML) method to enhance the conditioning of 2D and 3D post-stack seismic data by targeting noise and migration artifacts. These issues are often responsible for compromising the quality of seismic data sets. The method this paper presents leverages convolutional neural networks (CNNs) to identify and correct spatial inconsistencies within the data. With minimal preprocessing, requiring only the conversion of data into a 3D NumPy array format, this approach simplifies the workflow while improving data clarity and accuracy. Early statistical analysis of outliers and data normalization provides the ML model with stable inputs, allowing it to iteratively train on raw data and infer the original signal patterns in an unsupervised manner without the need for labeled data sets. The model outputs a conditioned version of the seismic data, significantly reducing noise while preserving critical geological features. The results demonstrate that ML seismic conditioning improves fault imaging, horizon continuity in areas with low signal-to-noise ratios (SNR), and localized smoothing without compromising resolution. Furthermore, the method presented in this paper retains signal fidelity and relative amplitude strengths, ensuring that geological interpretations remain accurate and reliable. By automating the traditional manual task of data conditioning, the proposed solution optimizes interpreter workflows, allowing them to focus on higher-level geological analysis and decision-making. This approach ultimately increases the accuracy of subsurface interpretations and resource exploration, improving the overall quality and efficiency of seismic interpretation projects. In addition, for 2D seismic, traditional methods of denoising rely on expert judgement using tools such as f-x deconvolution. 2D CNNs offer an efficient alternative, especially when training data can be synthesized and leveraged in supervised learning. This method can help the CNN identify and remove noise while preserving pertinent subsurface structures. This paper presents a 2D CNN architecture containing a model trained with synthetic seismic data to improve SNR of field seismic data; hence, interpretation.

Hydrocarbon exploration and production depend on the classification of seismic facies, a basic component of subsurface reservoir characterization. Conventional interpretation techniques mostly rely on subjective, time-consuming, and inconsistent manual examination of seismic data. Automated methods have become effective instruments for increasing seismic interpretation efficiency and accuracy as deep learning has advanced. In this work, a deep learning framework for classifying seismic facies using a Convolutional Neural Network (CNN) based on U-Net is presented. To assess the model's prediction power, it was tested on an unlabeled dataset after being trained on labeled seismic data. To improve model performance, data preprocessing methods such as one-hot encoding, normalization, and image scaling were used. Because of the magnitude of the dataset and the computational complexity, training was done on Google Collab. Techniques for data augmentation were also used to enhance generalization. Numerous performance measures were used to assess the algorithm, proving its efficacy in classifying facies. The findings show that by automating classification, decreasing manual labor, and enhancing the dependability of subsurface analysis, deep learning considerably improves seismic interpretation. The trained model's potential for scalable deployment in oil and gas exploration was demonstrated by the fact that it was preserved for later use.

Natural fractures occur over several orders of size magnitude. Accurately predicting the three-dimensional and multi-scale distribution of fractures in subsurface reservoirs is very difficult. Direct observations are limited to large-scale faults visible on seismic data and high-resolution, mostly small-scale, measurements of fractures intersecting the well, valid only in the direct vicinity of the wellbore. Analogue outcrop studies can be used to help fill in the gaps. The multi-scale properties of the natural fracture network exposed in the Cambro-Ordovician siliciclastic sequence in Petra, Jordan, are captured and it is investigated how these properties can be applied for the characterization of subsurface fractured reservoirs. The thesis consists of a series of chapters describing the geology of the study area, the results of the multi-scale fracture study in Jordan, followed by an analysis of the network connectivity using percolation theory and discrete fracture network models. The thesis is concluded by the results of numerical experiments where the influence of stress rotation on the re-activation of pre-existing fractures is investigated to help understand the evolution of complex networks consisting of multiple fracture sets.

SummaryWell data reveal reservoir layering with high vertical resolution but are areally sparse, whereas seismic data have low vertical resolution but are areally dense. Improved reservoir models can be constructed by integrating these data. The proposed method combines stochastic seismic inversion results, finer-scale well data, and geologic continuity models to build ensembles of flow models.Stochastic seismic inversions operating at the mesoscale generate rock property estimates, such as porosity, that are consistent with regional rock physics and true-amplitude imaged seismic data. These can be used in a cascading workflow to generate ensembles of fine-scale reservoir models wherein each realization from the Bayesian seismic inversion is treated as an exact constraint for a subensemble of fine-scale models. Exact constraints ensure that relevant interproperty and interzone correlations implied by rock physics and seismic data are preserved in the downscaled models. Uncertainty in the rock physics and seismic response is included by using multiple stochastic inversions in a cascading workflow. In contrast, inexact constraints generally do not preserve these correlations. We use two-point covariance at the fine scale to provide prior model thickness and porosity distributions of multiple facies. A Bayesian formulation uses the kriged data as the prior with the coarse constraints as the likelihood, and this posterior is sampled using a Markov Chain Monte Carlo (MCMC) method in a sequential simulation framework.These methods generate rich pinchout behavior and flexible spatial connectivities in the fine-scale model. These flow models are easily represented on a cornerpoint grid. 2D examples illustrate the interactions of prior and constraint data, and 3D examples demonstrate algorithm performance and the effects of stratigraphic variability on flow behavior.

The transformation of seismic reflection data into quantitative rock properties can be very valuable in all phases of oil and gas exploration and production. In this study, public domain data from Australian offshore have been used to explain the logical workflow, which consists of systematic steps from seismic petrophysics to rock physics modeling and seismic inversion for quantitative rock property estimation. The application of seismic inversion technology offers many rewards, such as: better reservoir definition and management; better resource estimation; and reductions in uncertainty. The main benefit is that it improves direct and intangible interpretation of seismic data to provide meaningful geological boundaries in the subsurface. In order to provide the reliable estimate of rock properties for reservoir modeling the workflow is tested rigorously and divided into three categories which are based on the basic inversion types. The results reveal that relative impedance inversion should be performed first, which then followed by revision of seismic interpretation on impedance data. Full bandwidth rock properties are estimated by deterministic inversion. The results can be improved by several iterations of well to seismic tie, wavelet estimation, and low frequency models. As resolution with accuracy has always been challenging, the stochastic inversion approach is trailered for reservoir characterization and high resolution rock properties have been predicted. This produces a number of possible rock property models, as well as litho facies models and the results can be optionally constrained by the well data. It can be concluded that, to some degree, stochastic inversion is able to overcome the limited seismic bandwidth by integrating the rock physics properties of different lithologies. The examples of seismic inversion are discussed to demonstrate the significant benefits can be obtained by following an optimized workflow that is tailored to deal with the uncertainties that impact the end product. Introduction Seismic surveys are routinely performed to gather information about the geology of hydrocarbon fields. Seismic data, even when it has relatively low resolutions, can still deliver a structural model for the reservoir. The availability and coverage area of the data provides valuable information to constrain the reservoir model. Geophysicists are now able to extract greater quantitative information about reservoir rocks and fluid parameters from seismic data by working backwards using seismic inversion methods. Described as the flip-side of forward modeling, seismic inversion is the process of characterizing rocks and fluids that have produced the seismic data. In a sense, the process involves trying to determine the input (rock properties) by looking at the output (seismic data). It starts with actual measured data, removes the wavelet and derives acoustic impedance from reflection coefficient (Figure-1) with a low frequency model to infer rock properties and deliver earth model.