Seismic volumetric dip estimation via a supervised deep learning model by integrating realistic synthetic data sets

Deep learning has been applied to tackle the seismic inversion problem, bringing more efficiency and accuracy. However, bad spatial continuity and poor generalizability limit the practical application. To solve these problems, we propose a 2D end-to-end seismic inversion method based on domain adaption. Firstly, the proposed 2D network learns the inversion mapping of seismic data under the constraint of domain adaption layer, which can reduce the difference between the features of real seismic data and synthetic seismic data, improving the generalization ability on real seismic data. Then, the trained model is finetuned with well logging data. In the first process, the spatial continuity of the inversion result is guaranteed by the 2D training scheme. Meanwhile, due to the constraint of the domain adaption layer, our model not only performs well on the synthetic data but also has good generalization ability on the real seismic data. And we carefully discuss the mechanism of domain adaption layer. In the second process, finetuning introduces well logging information, which can further improve the ability to invert details. Moreover, in order to improve the inversion accuracy on real seismic data, we develop a new training data generation method that can generate the synthetic samples close to the real samples, and a 2.5D training strategy is adopted to improve the continuity of the 3D data. The experiments on both synthetic and real seismic data show that our method performs better than both the recursive inversion method and the 1D closed-loop CNN methods.

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.

Historically, machine learning techniques and particularly neural networks have been applied in the domain of Geophysics to discern rock properties through a quantitative approach from seismic data and associated seismic attributes. However, the success or failure of neural network is dependent on amount of available data. In general, more labelled data during training of the neural network results in accurate predictions. Therefore, seismic based rock property prediction models and their calibration needs to be fed with labelled data coming from well logs in adequate amount. The fact that the number of wells are often limited can be a huge challenge in adopting neural networks predictions. Downton and Hampson (2019) developed a Hybrid Theory-guided Data Science (TGDS) technique to improve the training of the network by augmenting the amount of data used for training. This technique aims at overcoming the issue of scarcity of the well log data. By using the outputs of theory based components like Rock Physics theory as input to the data science component of augmenting the amount of the data. In this method, Rock Physics is applied to model the elastic response due to the changes in rock and fluid properties of the available well control to create a huge dataset of pseudo wells. Then, synthetic seismic gathers are modeled at the pseudo well locations to train a Deep Feed-Forwards Neural Network (DFNN). The parameters estimated during DFNN training with synthetic seismic gathers are then applied on a real seismic dataset. The final predicted P-impedance from this formerly described method was compared to conventional inverted Pre-Stack Elastic Inversion. This paper shows the application of the method of TGDS based synthetic seismic catalog generation followed by DFNN applied on real seismic dataset for a commercially oil producing field in North Sea. The reservoirs are injectite sands which are difficult to image but can be key pay zones with quite high porosity and permeability. Well log curves e.g. P-Impedance, Porosity and Water Saturation serve as guidelines to create the synthetic seismic data. Additionally, with the added advantage of increased synthetic data, pseudo wells were broken into training and validation datasets. So, the DFNN training is carried out on a portion of the training data subset. During the process of training the DFNN, the seismic data remains blind. The estimated DFNN non-linear operator is applied to the seismic data, however, it needs to be scaled to the real data as the training was performed on the synthetic seismic data. The final result of the data science and machine learning approach driven derived P-impedance showed higher frequency and better lateral continuity than the conventional theory based approach of Pre-Stack Elastic Inversion. Augmenting the amount of training data by generating synthetic data based on the statistics of from nearby well control and rock physics relationships is prime differentiator of this workflow. By using rock physics theory a large number of pseudo wells can be generated that gives a range of expected geologic scenarios. Key wells from nearby fields, that do not necessarily tie the seismic under consideration, can be incorporated into the analysis and also, if required additional wells could be incorporated into the analysis as drilled.

Reducing the uncertainty of reservoir characterization requires to better identify the small‐scale structures of the subsurface from the available data. Studying the seismic response of meter‐scale, stratigraphic heterogeneities typically relies on the generation of reservoir models based on outcrop examples and their forward seismic modelling. To bridge geological information and seismic modelling, these methods allocate values of acoustic properties, such as mass‐density and P‐wave velocity, according to discretized properties like layer‐type lithology or facies units. This strategy matches the current workflow in seismic data inversion in industry, where modelling workflows are based on lithofacies distributions. However, from stratigraphic modelling, we know that meter‐scale heterogeneities occur within certain facies and lithologies. Here, we evaluate the difference on the seismic response between allocating acoustic properties in a grain size–based, semi‐continuous manner versus discretized manners based on lithology and facies classifications. To do so, we generate a reference geological simulation that we populate with acoustic properties, mass‐density and P‐wave velocity, using three different strategies: (1) based on grain size distribution; (2) based on facies distribution; and (3) based on lithology. The method we propose includes the generation of realistic geological simulations based on stratigraphic modelling and the transformation of its output into acoustic properties, honouring the intra‐lithology and intra‐facies, small‐scale structures. We, then, generate seismic data by applying a forward seismic modelling workflow. The synthetic data show that the grain size–based simulation allows the identification of small‐scale, stratigraphic heterogeneities, such as beds with strong density and velocity contrasts. These stratigraphic structures are smoothened or may completely disappear in the facies and lithology discretized simulations and, therefore, are not (well) represented in the synthetic seismic data. Recognizing meter‐scale, stratigraphic heterogeneities is relevant for the characterization of the fluid flow in the reservoir. However, current discrete and lithology‐based strategies in seismic inversion are not able to resolve such heterogeneities because real subsurface properties are not discrete properties but continuous, unless there are stratigraphic discontinuities such as erosional surfaces or faults. This research works towards a better understanding of the relationship between changes in these continuous properties and the observed seismic data by introducing greater complexity into the discretized geological simulations. Here, we use synthetic seismic images with the goal of eventually aiding in fine‐tuning seismic inversion methodologies applied to real seismic data. One pathway is to foster the development of inversion approaches that can leverage stratigraphic modelling to get stronger geological priors and replace the standard but inadequate multi‐Gaussian prior.

The aim of this study is to calibrate synthetic seismic data to the real reservoir for a proposed seismic modelling and inversion studies. For this purpose, well tie was done by generating reflectivity series from log data in TWT, convolving it with Ricker wavelet. Optimal match was then explored by adapting the phase of the wavelet and correlating the resulting synthetic trace with neighbouring traces of the real data. Thus, synthetic post-stack seismic data was simulated using a convolution model in an iterative approach, extracting a wavelet from the seismic data. After a satisfactory calibration, forward prestack seismic modelling was applied to investigate the AVO response of different interfaces. Both the simulated post-stack and prestack models were then verified against real seismic data. Results show a perfect fit of synthetic-to-seismic calibration at WTR well. Good match between the petrophysical and the seismic attributes then implies that well data has been tied to the seismic data (lithologic units can be assigned to logs and seismic parameters at these wells). Therefore, further seismic modelling and inversion can be performed to study the acoustic response of the reservoir in-fills and extract the seismic attributes in this reservoir using these data sets.

ABSTRACTThe training and updating of the supervised deep learning model are both dependent on labelled data. Unfortunately, the data‐labelling process is usually expensive and time consuming, which makes obtaining labelled data a huge challenge, especially in the field of geophysics. Because of the above‐mentioned factors, some seismic data denoising methods based on supervised deep learning use synthetic seismic data for network training. Since synthetic seismic data are more or less different from real seismic data, the denoising results may have some amount of residual noise and some false signals when dealing with real seismic data. In response to the above problems, this paper introduces an unpaired domain‐to‐domain translation method based on the framework of two‐way generative adversarial networks, which does not need to use labelled data for training, and effectively solves the problem of lacking labelled data in the seismic data denoising tasks. We use an unpaired mixed training set containing synthetic seismic data and real seismic data to train the network, which effectively improves the denoising ability of the network for the real seismic data. The experimental results show that compared with the state‐of‐the‐art denoising methods such as denoising convolutional neural network, the proposed method can suppress the random noise and ground roll more thoroughly, and the denoising results basically would have no false signals.

Summary Detecting fault from seismic images is really useful for interpretation of geologic structures and stratigraphic features. With the recent developments in deep learning, this study makes it possible for efficient seismic fault detection based on convolutional neural network. In this work, we propose a general segmentation network that works on real 3D seismic data. We propose a novel network (GCS-Net) which includes a global context block (GC) and channel attention module together with spatial attention module (CS) between the encoder and decoder, instead of the U-Net based convolutional neural network. The GC block is able to capture the long-range dependencies and the CS block is utilized to further integrate local features with their global dependencies adaptively. The proposed approach was trained on a synthetic seismic data and tested by the real data. Experimental results show that our method has better performance to some extent.

Most deconvolution algorithms try to transform the seismic wavelet into spikes by designing inverse filters that remove an estimated seismic wavelet from seismic data. We assume that seismic trace subtle discontinuities are associated with acoustic impedance contrasts and can be characterized by wavelet transform spectral ridges, also called modulus maxima lines (WTMML), allowing us to improve seismic resolution by using the wavelet transform. Specifically, we apply the complex Morlet continuous wavelet transform (CWT) to each seismic trace and compute the WTMMLs. Then, we reconstruct the seismic trace with the inverse continuous wavelet transform from the computed WTMMLs with a broader band complex Morlet wavelet than that used in the forward CWT. Because the reconstruction process preserves amplitude and phase along different scales, or frequencies, the result looks like a deconvolution method. Considering this high-resolution seismic representation as a reflectivity approximation, we estimate the relative acoustic impedance (RAI) by filtering and trace integrating it. Conventional deconvolution algorithms assume the seismic wavelet to be stochastic, while the CWT is implicitly time varying such that it can be applied to both depth and time-domain data. Using synthetic and real seismic data, we evaluated the effectiveness of the methodology on detecting seismic events associated with acoustic impedance changes. In the real data examples, time and in-depth RAI results, show good correlation with real P-impedance band-pass data computed using more rigorous commercial inversion software packages that require well logs and low-frequency velocity model information.

With the development of the information age, the awareness of archives informatization of governments, organizations and groups at all levels is constantly increasing. The key content extraction of scanned archive images has been widely concerned by researchers. Based on the deep learning method, we extract text and key elements (ID photo, seal, date of birth) by analyzing the archive images. Based on the real archive images data set and synthetic archive images data set, we use the PP-OCR [1] model to finish the text structured extraction and YOLOv5 [2] to train a target detection model of ID photos, seals, and dates of birth to achieve the extraction of key elements. Since the real data set cannot be made public, we use ordinary archive data for explanation.KeywordsArchive imagesSynthetic archive imagesPP-OCRYOLOv5Text structured extractionKey elements extraction

Conventional sparse spike deconvolution algorithms that are based on the iterative shrinkage-thresholding algorithm (ISTA) are widely used. The aim of this type of algorithm is to obtain accurate seismic wavelets. When this is not fulfilled, the processing stops being optimum. Using a recurrent neural network (RNN) as deep learning method and applying backpropagation to ISTA, we have developed an RNN-like ISTA as an alternative sparse spike deconvolution algorithm. The algorithm is tested with both synthetic and real seismic data. The algorithm first builds a training dataset from existing well-logs seismic data and then extracts wavelets from those seismic data for further processing. Based on the extracted wavelets, the new method uses ISTA to calculate the reflection coefficients. Next, inspired by the backpropagation through time (BPTT) algorithm, backward error correction is performed on the wavelets while using the errors between the calculated reflection coefficients and the reflection coefficients corresponding to the training dataset. Finally, after performing backward correction over multiple iterations, a set of acceptable seismic wavelets is obtained, which is then used to deduce the sequence of reflection coefficients of the real data. The new algorithm improves the accuracy of the deconvolution results by reducing the effect of wrong seismic wavelets that are given by conventional ISTA. In this study, we account for the mechanism and the derivation of the proposed algorithm, and verify its effectiveness through experimentation using theoretical and real data.

The ultimate goal of this effort is to quantify lithology, fluid, and porosity from seismic data. We approach this goal by generating synthetic seismograms and comparing them to real seismograms, with the expectation that a similarity in the seismic response indicates a similarity in the underlying rock properties, specifically, lithology, fluid, and porosity (LFP). The synthetic seismic approach also can be used to understand the sensitivity of real data to LFP. The starting point for synthetic seismic generation is rock physics analysis and geologic interpretation of well data. Then a geological model of the subsurface with geobodies and expected depositional and diagenetic patterns outlined. Next, geobodies are populated with clay content and the corresponding porosity and hydrocarbon saturation at the well log scale. These rock properties are translated into the P- and S-wave velocity and bulk density by using a rock physics model. Then synthetic seismograms are produced from the elastic properties. We illustrate this methodology using well data from a fluvial environment in which the clay content is related directly to the total porosity by the dispersed clay model and to oil content via irreducible water saturation. These properties are related to the velocity and density by the Raymer-Greenberg-Castagna model, which is appropriate for mature sand/shale sequences. We use standard schemes of depositional models to construct a vertical geological section of grain size distributions, with a fluvial channel that includes the well. We examine two - the massive and gradational - depositional sequences within the channel. In the case under examination, synthetic seismic appears to be sensitive to lithology, fluid, and porosity. This result indicates that LFP information can be extracted from real seismic. Introduction The depositional environment, diagenesis, burial, and compaction can vary substantially within the same basin or within the same stratigraphic facies. These factors may directly affect lithology, fluid, and porosity (LFP). LFP, in turn, affects the elastic properties of the rock and, subsequently, the seismic response. The main question of exploration is how to translate seismic data into LFP. Our approach to this problem is synthetic seismogram generation and comparison thereof to real seismic data. If the real and synthetic data are similar, we assume a similarity in the underlying LFP. Of course, this approach carries nonuniqueness because multiple combinations of reservoir properties and geometries can produce the same seismic signature. However, the power of the synthetic seismic approach is that responses can be generated easily, rapidly, and massively for various scenarios of lithology, porosity, and fluid distributions in the subsurface. By so doing, the geophysicist can create a site-specific catalogue of seismic signatures to assess various possibilities and probabilities contained in the field data. Also, synthetic seismic data helps to understand the sensitivity of real data to LFP. Well data represent the rock properties along a one dimensional spatial trajectory. Therefore, information contained in the well data lacks lateral continuity and cannot be used alone for reliable predictions of rock properties where lateral geologic variation is significant.

Infrared (IR) images are essential to improve the visibility of dark or camouflaged objects. Object recognition and segmentation based on a neural network using IR images provide more accuracy and insight than color visible images. But the bottleneck is the amount of relevant IR images for training. It is difficult to collect real-world IR images for special purposes, including space exploration, military and fire-fighting applications. To solve this problem, we created color visible and IR images using a Unity-based 3D game editor. These synthetically generated color visible and IR images were used to train cycle consistent adversarial networks (CycleGAN) to convert visible images to IR images. CycleGAN has the advantage that it does not require precisely matching visible and IR pairs for transformation training. In this study, we discovered that additional synthetic data can help improve CycleGAN performance. Neural network training using real data (N = 20) performed more accurate transformations than training using real (N = 10) and synthetic (N = 10) data combinations. The result indicates that the synthetic data cannot exceed the quality of the real data. Neural network training using real (N = 10) and synthetic (N = 100) data combinations showed almost the same performance as training using real data (N = 20). At least 10 times more synthetic data than real data is required to achieve the same performance. In summary, CycleGAN is used with synthetic data to improve the IR image conversion performance of visible images.

Enhancing the detection of small coal structures by seismic diffraction imaging

ABSTRACTAutomatic feature detection from seismic data is a demanding task in today's interpretation workstations. Channels are among important stratigraphic features in seismic data both due to their reservoir capability or drilling hazard potential. Shearlet transform as a multi‐scale and multi‐directional transformation is capable of detecting anisotropic singularities in two and higher dimensional data. Channels occur as edges in seismic data, which can be detected based on maximizing the shearlet coefficients through all sub‐volumes at the finest scale of decomposition. The detected edges may require further refinement through the application of a thinning methodology. In this study, a three‐dimensional, pyramid‐adapted, compactly supported shearlet transform was applied to synthetic and real channelised, three‐dimensional post‐stack seismic data in order to decompose the data into different scales and directions for the purpose of channel boundary detection. In order to be able to compare the edge detection results based on three‐dimensional shearlet transform with some famous gradient‐based edge detectors, such as Sobel and Canny, a thresholding scheme is necessary. In both synthetic and real data examples, the three‐dimensional shearlet edge detection algorithm outperformed Sobel and Canny operators even in the presence of Gaussian random noise.

SUMMARY One of the aims of time-lapse seismic feasibility studies is to determine whether a desired time-lapse signal is distinguishable from incoherent noise. Acquisition parameters play a key role in repeatability, with sourcereceiver positioning errors commonly being regarded as the most important issue. A normalised root-mean-square (NRMS) variogram analysis measures the effect of source-receiver geometry differences on observed nonrepeatability of traces. The overall trend of the variogram is strongly controlled by the heterogeneous characteristics of the overburden. We investigate the influence of heterogeneity on seismic repeatability using a NRMS variogram on synthetic data. We generate synthetic seismic data (with no incoherent noise) from velocity models with a variety of overburden characteristics and run finite-difference simulations over them. Variograms are generated from the synthetic data and show similar trends to those observed in real seismic data. We demonstrate that the length of the coherent signal of a target reflector (derived from the variogram) is directly related to the size and position in depth of the heterogeneity.