The research object of oil and gas exploration is gradually shifting toward complex oil and gas reservoirs. The periphery of the Banghu Syncline in the Qianjiang Sag is a typical inland salt lake deposit. The complex thin sand and mudstone interbedded reservoir structure in the Qian-3 section of the Banghu Syncline area requires high-precision and high-resolution exploration technology to support actual production. In view of this, the study conducts forward modeling analysis and complex lithology classification based on well logging data. First, it analyzes and calculates the lithological data based on the original logging data, analyzes the lithological characteristics of sandstones containing different fluids (water-bearing sandstones, oil-bearing sandstones, and dry-layer sandstones), establishes different wedge forward models based on the convolution theory, and investigates the seismic response characteristics of different lithological combinations. Then, lithology curves are reconstructed using the K-means algorithm with known logging lithology data, and density attributes are used to correct natural gamma values for lithology classification. Finally, a geological model of the $ \mathrm{E}{\mathrm{q}}_{3}^{4} $ oil formation is designed to study the effects of changes in reservoir thickness and fluid content on amplitude.Model analysis and the example demonstrate K-means algorithm can effectively divide salt rock, sandstone, and gypsum mudstone with a prediction accuracy of 90.4%, and the forward model based on high-resolution logging information is consistent with actual geological characteristics. Therefore, it is feasible to analyze the reflection characteristics of salt rock and mudstone interbedded layers and thin sandstone layers by the forward model established through lithology curve reconstruction with logging data.

With the development of offshore oil and gas exploration and development, the application of dual-sensor ocean bottom receivers is becoming more and more extensive. As a key technology of dual-sensor seismic data processing, the separation of the up-going and down-going wavefield determines data processing qua- lity. In response to the failure of conventional methods to complete separation of up-going and down-going wavefields (up-going wavefield separated include some of the down-going wavefield, and the down-going wavefields separated also include some of the up-going wavefield), this paper proposes a separation method of up-going and down-going wavefields for dual-sensor seismic data based on direct wave calibration. First, the spatial weighted function of the dual-sensor seismic data is calculated by using the dual-sensor direct waves in the frequency-space domain. Then the calibration filter operator of the dual-sensor seismic data is directly computed by using the dual-sensor seismic data with and without direct waves in the time-space domain. Finally, the dual-sensor seismic data is calibrated, and up-going and down-going wavefield are separated. The data examples demonstrate the effectiveness and practicality of the proposed method. The separated up-going wavefield data not only eliminates the interference of the ghost multiples but also improves the signal-to-noise ratio and resolution of seismic data. The method provides high-fidelity up-going and down-going wavefield data for subsequent joint deconvolution and migration imaging processing.

With increasingly complex oil and gas exploration targets, seismic exploration faces challenges such as low signal-to-noise ratio (SNR), low resolution, and difficulties in velocity modeling and imaging of seismic data. Conventional seismic data processing and interpretation methods have certain limitations in accuracy or efficiency when applied to massive seismic data. The artificial intelligence (AI)-based seismic data processing and interpretation methods can effectively improve accuracy and efficiency. To this end, this paper provides an overview of supervised, semi-supervised, and unsupervised deep learning techniques, and summarizes the applications of deep learning in data processing such as first break picking, SNR improvement, data reconstruction, velocity spectrum interpretation, migration, and resolution enhancement. Meanwhile, it discusses the applications of deep learning in identifying geological bodies such as faults, seismic facies, river channels, and salt domes, as well as in wave impedance inversion, AVO inversion, full waveform inversion, lithology identification, reservoir parameter prediction, and fluid identification. The production of training sets, optimization of neural networks, training strategies, and large models are discussed, with an outlook on the development trend of AI-based processing and interpretation methods for seismic data provided. It is pointed out that the generalization of networks should be continuously increased and large models suitable for seismic exploration should be studied.

The total organic carbon (TOC) content is an important evaluation index for shale gas exploration and development. Logging data can efficiently assess TOC, but it cannot be used for TOC prediction in inter-well areas. The TOC-sensitive factors extracted from seismic data can achieve three-dimensional (3D) prediction. However, due to the thin thickness and strong heterogeneity of shale reservoirs, it is difficult to achieve the required resolution by relying solely on seismic data. Therefore, it is necessary to comprehensively use multiple data sources to improve the accuracy of TOC assessment. For this purpose, a high-precision quantitative prediction method for shale TOC based on a convolutional neural network (CNN) is proposed. Firstly, the correlation analysis between the measured TOC data of the core from drilling and multiple logging characteristic curves is conducted on the Longmaxi Formation shale in southern Sichuan, and the most representative and sensitive features are selected. Secondly, based on the identified sensitive parameters, a CNN prediction model is constructed. The measured TOC samples and the training samples constructed by sensitive logging parameters are divided into datasets at a ratio of 7:3 for model training and validation. Finally, the high-resolution sensitive parameter inversion results obtained by simulation of seismic waveform indication are used as the feature input for 3D TOC content prediction. The sensitive parameters are rearranged, reorganized, and then input into the CNN model to achieve 3D TOC content prediction. The research results show that CNN has more advantages than multiple regression and back propagation (BP) neural networks in fitting the nonlinear relationship between TOC content and sensitive parameters. The average absolute error and root mean square error are both less than 0.6% between the predicted TOC data and the measured values from drilling. The prediction results are consistent with the actual situation. This method has high accuracy and obvious advantages in 3D TOC content prediction of thin shale reservoirs.

In the field of oil and gas exploration, the occurrence state of oil and gas resources can be determined from rock microstructure, and the accuracy and efficiency of this approach can be improved by an effective lithology identification method. This paper proposes a rock thin section lithology identification method based on improved ConvNeXt V2. First, ConvNeXt V2-T is used as the core feature extraction network, and the global attention mechanism is embedded to improve the perception of global features. Then, a multi-scale feature fusion module is designed, which can effectively fuse feature maps at different scales. Finally, the model optimizer is improved by using Lion optimizer instead of the original Adamw optimizer, which is faster, achieves better generalization performance, and saves more memory. The experimental results show that the average values of accuracy, precision, recall, specificity, and F₁ score of the algorithm proposed in this paper are 96.1%, 95.5%, 96.2%, 99.1%, and 95.8%, respectively. The improved algorithm has a faster convergence rate and higher accuracy, which can realize accurate classification and identification of rock thin section images.

Intelligent seismic velocity inversion is currently a hot and challenging topic in seismic exploration research. Nevertheless, the complex structure of deep learning networks demands significant computing power from hardware devices, which restricts the application of the model in scenarios with large data volumes and high timeliness requirements. To address these practical issues, in this paper, the U-Net is improved based on the concepts of feature engineering and model lightweighting, and the inversion networks U-Net vG for GPU and U-Net vC for CPU are proposed. Firstly, the characteristics of the velocity inversion network are analyzed to deduce the lightweighting principles of convolutional neural networks. Subsequently, lightweight processing is conducted on the multi-scale module, attention gate module, and feature extraction module to obtain a lightweight convolutional neural network for velocity modeling, which reduces the network volume while maintaining prediction accuracy. Data test results demonstrate that the training process of the proposed network has lower requirements for high-performance hardware resources, and that the network enables efficient velocity inversion, possesses higher seismic velocity inversion accuracy, and exhibits superior noise resistance. It provides a new idea for solving the computing power bottleneck problem in seismic data inversion.

Seismic facies identification is a crucial link in seismic data interpretation. Deep learning technology can enhance the efficiency and accuracy of automatic seismic facies identification. However, deep learning methods typically rely on large amounts of labeled data, and in practical applications, the labeling cost of seismic data is high, with great difficulty. Additionally, basic logging data cannot be directly utilized. To this end, this paper proposes a semi-supervised automatic seismic facies identification method based on ultra-sparse logging labels. First, based on the HRNet, a seismic facies identification model that uses one-dimensional logging labels is built for for supervision. Second, to preserve the vertical characteristics of seismic data, this paper develops a sparse label sampling module (SLSM) that conducts samples around the logging labels without slicing the seismic data vertically, thus retaining its vertical depth features and laying a solid foundation for subsequent semi-supervised learning tasks. Third, in terms of the lateral correlation of seismic data, the region growing training strategy (RGTS) is proposed, which expands the information from logging labels to the entire seismic volume through an iterative growing process. Experiments on real-world data show that the proposed model achieves a mean intersection over union (MIoU) of 79.64% by using only 32 one-dimensional logging labels, which account for less than 0.5% of the total data volume. This approach provides references for conducting seismic facies identification in areas with sparse and locally distributed logging data, demonstrating promising application potential.

Karst caves exhibit distinctive "string-of-beads" reflection configurations on seismic profiles, with their spatial distribution governed by intricate fracture networks and thus forming complex fracture-cave systems. Conventional methods, constrained by ambiguities in reservoir architecture and limited sample availability, face challenges in achieving accurate delineation. This study proposes a knowledge graph-guided intelligent identification technique based on coupled fracture-cave modeling, which innovatively integrates geological prior knowledge with deep learning through encoding geological topological relationships into adjacency matrix constraints. The methodology establishes a multi-task learning framework by synergistically combining forward modeling-derived label data volumes with expert-annotated data volumes. The approach employs knowledge graphs to characterize connectivity relationships between fractures and karst caves and designs geologically interpretable loss functions to dynamically adjust model optimization trajectories. Application in the Ordovician Lianglitage Formation of the Tarim Basin demonstrates substantial reduction in manual interpretation workload and significant enhancement in boundary delineation precision for fracture-cave systems. This methodology presents an innovative solution integrating knowledge-driven and data-driven approaches for prediction of strongly heterogeneous carbonate reservoirs.

The shear wave velocity and thickness of near-surface strata can be obtained through the inversion of Rayleigh wave dispersion curves. However, traditional nonlinear inversion algorithms often have disadvantages such as poor convergence effect and being prone to fall into local extremity. A new improved beluga whale optimization (DWBWO) algorithm is proposed in this paper and applied to the inversion of the Rayleigh surface wave dispersion curve. Based on the beluga whale optimization (BWO) algorithm, this algorithm introduces the Cubic chaotic initialization strategy to improve the uniformity of the initial population. Meanwhile, the dimensional reverse learning strategy is used to improve the convergence efficiency of the algorithm, and the whirlwind foraging strategy (WFS)is adopted to improve the local optimization ability of the algorithm. The DWBWO algorithm is tested by applying the multi-extremum functions, simulated data and measured data, and compared with the grey wolf optimization (GWO) algorithm, sparrow optimization (SSA) algorithm, whale optimization (WOA) algorithm and BWO algorithm. It was proved that the improved algorithm in this paper has higher stability and accuracy.

The "stamp-like" acquisition method in complex areas results in fragmented three-dimensional seismic data of a zone, with differences in energy, frequency, phase, and time between datasets. This poses challenges to regional seismic interpretation and comprehensive studies. Conventional pre-stack joint processing helps address boundary splicing problems between 3D blocks but the significant disparities in the quality of original data across different blocks prevent full utilization of high-quality data and ultimately reduce imaging precision. In response to the greatly different seismic data in complex areas, this paper proposes a technical scheme of block and joint seismic data processing based on key parameters of the zone. This method preserves the advantages of native bin processing for individual blocks while possessing the advantages of joint data for facilitating comprehensive studies. Meanwhile, it resolves the problems in conventional joint processing of a large number of existing 3D data, such as "difficult blending of single-block processing" and "insufficient fineness in joint processing". Case studies from the complex mountainous structural belt in Kuche and the Fuman East area of Tarim Basin demonstrate the effectiveness of this technical method. This method constructs joint data volumes by adopting building blocks, which supports the trap well location and basic geological research and provides useful references for the industrial sector.

Multiples reduce the signal-to-noise ratio of seismic data, affecting the identification of primaries, thereby increasing the difficulty of seismic processing, reducing the authenticity and reliability of seismic imaging, and even forming geological illusions, which affect subsequent seismic exploration and development. The multiple suppression method based on wave theory can better adapt to complex media, which mainly consists of two steps, multiple prediction and adaptive matching subtraction. Both steps impact the final accuracy of multiple suppression. The paper compares three adaptive matching subtraction methods, respectively based on minimum energy principle, pattern recognition, and deep learning. The advantages, disadvantages, and adaptability conditions of each method are also analyzed. The test results of model data containing surface-related multiples and field data with internal multiples show that the adaptive subtraction algorithm based on the principle of minimum energy assumes wavelet consistency, while the pattern recognition based adaptive subtraction technique requires high lateral consistency of seismic data. Compared with the two traditional methods, adaptive matching subtraction based on deep learning can avoid assumed conditions and effectively protect primaries, achieving higher computational accuracy.

Continental shale oil has become the major unconventional oil resource in China. The reservoir features include the development of lamellation crack, low porosity, large difference between horizontal and vertical permeability, and complex pore structure. Therefore, clarifying the relationship between the elastic and physical parameters of shale oil reservoirs is important.In addition, there are few studies on the characteristics of shale lamellation crack and its influencing factors. Based on anisotropic background media, a rock physical modeling method for fracture-porosity shale rich in kerogen is proposed in this paper.Through rock physical modeling, the effects of matrix porosity, aspect ratio, connectivity coefficient and the lamellation crack length, width and numbers of the shale are compared and analyzed. The results show that the P wave velocity is greatly affected by the connectivity coefficient, while the S wave velocity is not affected by the connectivity coefficient. The effect of lamellation crack over 100 microns should not be ignored and its induced anisotropy could not be summarized as the effect of intrinsic anisotropy.Vp/Vs positively correlates with the length of lamellation crack, while the relationship with the number of lamellation crack decreases first and then increases. The method in this paper is applied to the logging data of a practical working area. The P‐wave velocity and S‐wave velocity prediction in shale reservoirs through this method are highly consistent with actual data, which verifies the effectiveness and applicability of the proposed method, and the method can be used as a bridge in the characterization of unconventional oil and gas reservoirs.

The typical 3D numerical model and seismic simulation are of great significance for the study of seismic processing and interpretation methods, as they can test and validate the effectiveness of various new technologies and approaches. In the double-complex exploration areas in western China, where both neasurface and subsurface structures present significant complex, there have long been challenges of difficult seismic migration imaging and delineation of structures, making it urgent to build a typical 3D model and forward modeling data that represent the double-complex characteristics of China's foreland basins. By taking the representative double-complex area of the Keshen block in the Kelasu area in western China as an example, this paper combines the typical surface and subsurface geological and geophysical features of foothill areas in the western foreland basins and fully utilizes data of seismic depth migration velocity, acoustic logging, VSP logging, and uphole survey to construct typical numerical models for double-complex exploration zones in western China(BGP-DC²GModels). The BGP-DC²GModels include an acoustic velocity model, an isotropic elastic medium model, TTI/TORT models, and a viscoelastic medium model, typically featuring complex near-surface conditions, complex subsurface structures, a high-speed conglomerate fan body, salt-gypsum formations, detachment layers (coal seams), a Q anomaly, and anisotropy. Based on the BGP-DC²GModels, five sets of high-quality forward modeling data for surface and borehole 3D observation are generated by adopting the spectral-element method, including the isotropic acoustic wave, isotropic elastic wave, isotropic viscoelastic wave, TTI viscoelastic wave, and TORT viscoelastic wave. Additionally, the acoustic reverse time migration and anisotropic reverse time migration results of the data are compared and analyzed, further demonstrating that typical models and numerical simulation data can test and validate the effectiveness of new acquisition, processing, and interpretation methods.

Microseismic monitoring is an essential technology in the field of unconventional oil and gas reservoirs exploration. It has been widely used in hydraulic fracturing fracture monitoring, CO₂ storage, and so on. However, the microseismic signal is weak in energy and easy to be polluted by noise. Its low signal-to-noise ratio makes it difficult to obtain good results in subsequent processing. Therefore, microseismic data denoising is a highly important processing step. The denoising effect has a key impact on the accuracy of subsequent source location and the reliability of focal mechanism inversion results. In this paper, a Monte Carlo non-negative dictionary learning (MCNDL) method is proposed for microseismic data denoising. The Monte Carlo block can obtain the initial dictionary containing relatively many effective signal features in a small amount of time. In the process of dictionary updating, non-negativity constraints are used to ensure the sparsity of data transformation and reduce the solution space, thus reducing the computational cost and improving denoising accuracy. This study evaluates the performance of the proposed method by using both synthetic and real-world microseismic datasets, comparing it with band-pass (BP) filtering, frequency-wavenumber (F-K) filtering, and K-singular value decomposition (KSVD) techniques. The findings highlight the superior denoising effect and efficiency of the proposed approach.

In the field of oil exploration, multiple attenuation is an important means to improve the migration imaging quality of seismic data, especially for marine streamer data that is strongly disturbed by free-surface-related multiples. To better improve the multiple attenuation effect and data processing efficiency, this paper proposes a free-surface-related multiple attenuation method using an image translation technology based on deep learning. Firstly, the seismic data processing task is regarded as an image translation task in deep learning, and the Pix2Pix network is used to process the seismic data converted into the image form. Secondly, by improving the form of the target data set from the conventional single form to the combining form, this study carries out multi-task training to improve the output effect of the Pix2Pix network. According to the correlation before and after data processing, an additional loss function is designed to further constrain and improve the output effect of the network. Finally, a layered model and a complex model are established for numerical testing, and additional interference items are added to the input data for quantitative testing. The numerical test results show that the proposed method can achieve multiple identification and attenuation by exploring the common features of data, improve the clarity of migrated images, and more accurately identify horizon information with high computational efficiency in data processing.

Conventional post-stack seismic data are usually affected by absorption attenuation, resulting ina lower peak frequency, narrower frequency band, low resolution, and poor inversion prediction effect. Therefore, an improved Q estimation and inverse Q-filtering method is proposed to improve the resolution of seismic data and obtain the fidelity and amplitude-preserving seismic data. Firstly, in Q estimation, given the problem that the extracted wavelet amplitude spectrum deviates from the actual situation due to the thin-layer tuning effect that affects the Q estimation accuracy, the complete ensemble empirical model decomposition with adaptive noise (CEEMDAN) method is introduced, which eliminates the interference of reflection coefficients by decomposing and reconstructing the log amplitude spectrum and thus obtains the wavelet amplitude spectrum that removes the tuning effect. Combined with the centroid frequency shift of the energy spectrum, higher-accuracy Q values are obtained. Then, in terms of inverse Q filtering, the amplitude compensation function of the time-varying gain stabilization factor method is optimized to overcome the density dependence and obtain more stable inverse Q-filtering results. The actual data processing results show that the proposed method can obtain fidelity and amplitude-preserving high-resolution post-stack seismic data, which lays a reliable data foundation for subsequent exploration and development of the study area.

The accurate prediction of tight sandstone reservoir parameters is a key scientific issue and technical challenge in unconventional oil and gas exploration. Traditional prediction methods based on linear or nonlinear regression have limitations in characterizing the complex nonlinear relationship between logging curves and reservoir parameters, leading to insufficient prediction accuracy. This study takes the Chang 6 reservoir in the Tang 157 well area of the Ganguyi Production Plant in Yanchang Oilfield as an example. Based on logging data and core analysis porosity data, multi-source data fusion preprocessing is conducted, and a novel neural network architecture (CNN-Transformer Network) that integrates the core advantages of CNN and Transformer is innovatively proposed. The prediction performance of the CNN-Transformer model is comprehensively compared with that of traditional linear regression (LR), TCN-LSTM, GRU, and ResNet models using RMSE, MAE, and R² metrics. Experimental results show that the prediction accuracy of the CNN-Transformer model reaches 96.7%, significantly outperforming the other comparative models. This model effectively captures the unique complex nonlinear mapping relationship between logging curves and porosity in tight sandstone reservoirs, significantly improving the accuracy of reservoir parameter prediction and providing reliable technical support for the efficient exploration and development decision-making of tight sandstone reservoirs.

The sedimentary water body of Enping Formation in Huibei Area of the Pearl River Mouth Basin is relatively shallow. It is mainly in a sedimentary environment of shallow water delta-lacustrine swamp-shore and shallow lake facies, among which the floodplain swamps of the delta plain and lacustrine swamps are the main sedimentary facies belts where coal measures source rocks develop. Peripheral drilling has confirmed that a large number of coal seams have developed in Enping Formation. The maximum thickness of a single coal seam is 4.5 m, and the minimum thickness is 0.5 m, which is far less than the minimum thickness that can be identified by conventional seismic inversion. Therefore, how to predict the development scale of thin coal seams is the key to the evaluation of the coal measures source rocks in this area. Thus, based on a comprehensive analysis of the geological and sedimentary backgrounds of coal measures source rocks, a "three-step method" is explored to achieve the prediction of thin coal seams in coal measures source rocks. Firstly, through forward modeling and analysis, it is confirmed that thin coal seams in different sedimentary environments all have the characteristics of strong amplitude seismic reflection, while sandstone and mudstone show weak amplitude characteristics. The root-mean-square amplitude attribute is extracted as a sensitive attribute to distinguish the development range of coal seams from that of sandstone and mudstone. Secondly, the frequency band of seismic data is broadened through high-resolution seismic processing techniques to obtain broadband seismic data, which provides a high-precision data foundation for seismic inversion. Finally, by using the root-mean-square amplitude attribute, a stepwise fusion inversion method of low-frequency, medium-frequency, high-frequency dominant frequency bands is adopted to gradually characterize coal seams of different thicknesses. Finally, the prediction of the development range of thin coal seams in coal measures source rocks is completed. The actual application results in Block L of Huibei Area show that the predicted development range of coal seams have a relatively high degree of consistency with the drilling results, and this method can effectively improve inversion accuracy.

As the rotational components of seismic waves capture both the rotational motion of particles and spatial gradient field information, they offer certain advantages in seismic source localization, P-and S-wave field separation, and subsurface velocity inversion imaging. However, the high cost and limited availability of rotational seismometers have resulted in scarce observational data on rotational components, making it difficult to conduct in-depth research on their application value using limited data. Consequently, many traditional methods and techniques developed for translational seismic components are restricted in the study of rotational components.This paper, based on simulated and measured data, conducts experimental comparisons of the conversion of rotational components in dense seismic arrays using the differential and wavefield gradient methods. The results reveal that the wavefield gradient method exhibits superior conversion accuracy compared to the differential method. Furthermore, this study places particular emphasis on analyzing the impact of inter-array spacing and subsurface velocity on the conversion accuracy and discusses the optimal frequency range for array-based rotational component calculations.Additionally, this paper focuses on analyzing the influence of inter-array spacing and site phase velocity on conversion accuracy and derives a relationship for calculating the frequency range of translational-to-rotational component conversion. The accuracy of converting rotational components in dense arrays is affected by array spacing, site phase velocity, and target frequency. Given the conditions of the observational system, the derived relationship can be used to calculate the target frequency range, within which the computed rotational components show a high degree of fit with the observed rotational components, with a maximum correlation coefficient reaching 0.98.

The absorption and attenuation of seismic wave energy in complex loess plateaus has always been a focal point for geophysicists. This problem makes it difficult to improve the resolution and signal-to-noise ratio of seismic data and satisfy the requirements for describing reservoirs and exploring residual oil in mature oilfields. The distributed acoustic sensing (DAS) multi-well borehole-surface joint exploration (BSJE) technology, characterized by identical-source excitation, consistent reception, small downhole trace distance, and high coverage, can reduce absorption and attenuation of near surface by receiving seismic waves in the well, which is one of the main approaches to solve the aforementioned problem. However, the most critical issue facing BSJE technology is the difficulty in vertical seismic profile (VSP) imaging due to uneven seismic excitation points on complex surfaces. At the same time, there are problems such as the lack of BSJE scheme demonstration method and immature techniques for synchronous reception and quality control of DAS multi-well BSJE. Therefore, based on wide-azimuth, wide-bandwidth, and high-density 3D seismic, a technical study on the DAS BSJE, processing, and interpretation of 3D-VSP is conducted with seven wellsin a block in the eastern part of the Ordos Basin.This paper, as the first part, "Acquisition", of the technical study systematically studies the key issues of DAS BSJE technology in complex loess plateau areas. A sector method-based BSJE excitation scheme demonstration technique aimed at VSP imaging is proposed, and a scheme with uniform excitation, DAS multi-well continuous reception, and quality control for BSJE is developed. High-quality DAS 3D-VSP data are obtained by applying the proposed method, which lays a solid foundation of high-quality data for multi-well BSJE-based 3D-VSP migration imaging and reservoir prediction.

Shale gas has become an important strategic alternative field for China's oil and gas resources. Shale is characterized by low porosity and low permeability, and only after going through large-scale volume fracturing can industrial production capacity be obtained. The fine characterization and quantitative characterization of fracture parameters after shale fracturing are the key to fracturing effect evaluation and development parameter optimization. By taking the three-dimensional CT images of shale cores after fracturing as the research object, this paper conducts intelligent fracture extraction based on the deep learning semantic segmentation model. Firstly, a U-Net deep learning model integrating the pyramid convolution and attention mechanism is built to alleviate the influence of image category imbalance and improve fracture extraction accuracy. Secondly, a digital core model is built based on the semantic segmentation results, and quantitative characterization of the spatial distribution of fractures is realized by combining parameters such as the porosity and tilt index. Finally, the complexity of the fracture network is characterized by the peak and width of the multi-fractal spectrum. The research results show that compared with the traditional image segmentation model, the sensitivity of the improved model is increased by 6.69%, and the intersection over union grows by 0.48%. This study systematically characterizes the three-dimensional fracture features by image segmentation algorithm optimization, digital core modeling, and multi-fractal analysis, which is applicable to the characterization of fracture networks in unconventional reservoirs such as shale and can provide a reference for the evaluation of reservoir stimulation effects after hydraulic fracturing.

Accurately segmenting pores in scanning electron microscope (SEM) images can provide a scientific basis for oil and gas exploration and development, and more. At present, pore segmentation methods mainly rely on data-driven approaches, requiring a large amount of manual annotation of data, which is time-consuming and costly. To this end, this paper proposes the semi-supervised pore segmentation network PoreSeg for SEM images. Firstly, a semi-supervised framework is constructed based on consistency regularization and pseudo labeling. Secondly, a high-intensity combined perturbation strategy is introduced to enhance data diversity. Finally, combined with the pore aware fusion (Pore-CutMix) method, the sparse pore information is fully utilized to improve the segmentation ability of the model for pores. Experimental results show that under the condition of equal labeled samples, PoreSeg improves the pore intersection over union (IoU) by 15.10% compared with the fully supervised network. At the same time, compared with existing semi-supervised methods, PoreSeg is more sensitive to pores and has higher segmentation accuracy. PoreSeg significantly reduces dependence on annotated data while maintaining high accuracy, and has huge application potential.

CO₂ flooding plays an important role in improving oil recovery and reducing greenhouse gas emissions, and is an effective means to achieve the goals of carbon peak and carbon neutrality, for which the seismic monitoring technology is the key. This paper analyzes and summarizes the research status and progress of seismic monitoring technology for CO₂ flooding at home and abroad, including time-lapse seismic feasibility analysis, consistency processing technology and comprehensive interpretation. The application of seismic monitoring technology for CO₂ flooding in Gao89 block is also discussed. Feasibility analysis is an important prerequisite for time-lapse seismic monitoring in the study area. Only when reservoir geological conditions, petrophysical conditions and seismic conditions are met, time-lapse seismic monitoring can be carried out effectively. In order to realize reservoir dynamic monitoring, it is particularly important to deal with the consistency between basic seismic and monitoring seismic (time-lapse seismic), and it is necessary to carry out the matching filtering of time difference, amplitude, frequency, phase and other factors. Time-lapse seismic comprehensive interpretation is helpful to predict accurately the CO₂ plume. The pre-stack method is mainly AVO attribute analysis method. Post-stack difference analysis based on basic seismic data and monitoring seismic data is still the main method, and frequency domain information such as frequency division processing, velocity dispersion, low-frequency shadows and so on is worthy of use. The prediction method of the CO₂ plume based on deep learning is in the ascendancy, but its operational efficiency and generalization ability need to be further improved. Finally, the paper provides an outlook on the development potential of time-lapse seismic technology in improving monitoring accuracy, developing monitoring methods and expanding application market.

Dense drilling platforms, coral reefs, docks, pipelines, and other obstacles lead to the necessity of obstacle avoidance for source and receiver points in the design of marine seismic data acquisition by the ocean bottom node (OBN). Planning a path with obstacle avoidance is a key to the success of marine seismic exploration. Reasonable obstacle avoidance method and path planning scheme can minimize the influence of obstacles on bin attributes in the acquisition area, avoid the invalid work, and reduce the acquisition risk. Traditional obstacle avoidance methods generally move the source and receiver points in the obstacle area laterally to the outside without considering the influences of the turning radius of the source vessel and the width of the expander. The manual obstacle avoidance method leads to technical problems such as low obstacle avoidance efficiency, a large error, serious loss in the coverage area, and potential safety hazards caused by inaccurate turning radius design of vessel. This paper proposes an intelligent obstacle avoidance method for complex sea area observation system. The shot line of the observation system is fitted by the trajectory theory, which is then combined with algorithms of multi-obstacle tangent circle against trajectory, obstacle merging, and dynamic feedback correction of trajectory. In this way, the paper forms an intelligent obstacle avoidance technique based on a backtracking algorithm and real-time calculation and adjustment of various parameters, greatly improving the efficiency and accuracy of obstacle avoidance. The production practice shows that the method can improve the efficiency and accuracy of exploration in complex sea areas and provide technical support for safe and efficient exploration in complex marine obstacle areas.

Current data transmission technologies face challenges such as high coding complexity, low encryption efficiency, and limited data compression, thus restricting the development of efficient and safe communications. To resolve these problems, this paper proposes an innovative ultra-low-traffic data transmission scheme, the Temporal-base data pulse transmission method. The core of this method is the combination of high-accuracy time synchronization and subdivision technology, and the establishment of the "Temporal-base" theoretical framework. By adopting the accurate time datum as the data coding basis, the information is mapped to the specific pulse time sequence or phase via super large binary representation to achieve pulse transmission of data. This mechanism notably simplifies the traditional coding process and improves the confidentiality (due to accurate synchronization requirements of time pulses) and compression potential (due to efficient pulse coding of information in the time dimension) of data. By carrying out case analysis, this study verifies the advantages of the proposed data transmission method in improving the transmission rate, significantly saving communication bandwidth, and strengthening data confidentiality, thus fully proving the huge application potential and value of this method in future efficient and safe data transmission scenarios.

In seismic surveys of complex surfaces such as loess plateau areas, coherent noise greatly reduces the signal-to-noise ratio (SNR) of seismic data, which seriously affects the accuracy of subsequent seismic imaging and physical inversion. To this end, a new coherent noise suppression strategy is proposed in the paper. The method mainly consists of the following core steps. Firstly, anisotropic diffusion filtering is adopted to effectively suppress the incoherent noise components in the data, and initially improve the overall quality of the low SNR data. Next, the dictionary learning method is employed to sparsely characterize the seismic data, and statistical indicators are applied to precisely locate and eliminate the dictionary atoms with a large variance of gradients. These atoms tend to be the main carriers of linear coherent noise and random noise. Then, the dictionary atoms and their corresponding sparse coefficients that can effectively characterize the effective signals are filtered and retained to reconstruct the seismic data. Finally, the effective signals are further extracted from the removed noise through the principle of signal-to-noise local orthogonalization. The simulated data and typical real data tests show that the method ensures the intact preservation of the effective signals while suppressing the coherent and random noises, which further improves the SNR of the data. The method can provide a reference for the treatment of linear coherent noise.

In ocean bottom node (OBN) seismic exploration, significant discrepancies often exist between the actual node positions and the initial surveyed locations due to factors such as ocean currents, tides, topographic variations, and fishing vessel dragging, necessitating secondary positioning. To address this, a high-precision secondary positioning method based on four-quadrant stacking is proposed. Specifically, within a relative coordinate system, shot points are divided into four quadrants according to their azimuth. Linear moveout correction and common receiver point stacking are performed separately for shot points in each quadrant. Subsequently, the displacement of the node from the surveyed position to the actual position is decomposed into two mutually perpendicular components. These two components are determined by analyzing the first arrival time differences from the common receiver points in each quadrant, which thereby enables precise estimation of the node's actual coordinates. Application to field data demonstrates that this method achieves high-precision secondary positioning of nodes. Moreover, the four-quadrant stacking strategy significantly improves the reliability of first arrival picking. Compared with traditional methods, this approach offers higher positioning accuracy and efficiency, showing excellent engineering applicability and promising potential for broader adoption.

Ensuring strong correlation among samples of the same category in seismic facies images and determining the number of seismic facies categories are the core of unsupervised pre-stack seismic facies analysis. This paper proposes an unsupervised pre-stack seismic facies analysis algorithm with correlation threshold constraints for category determination. First, the pre-stack seismic trace set data is transformed into two-dimensional images, and the high-level nonlinear, discriminative, and invariant features of the images are extracted using unsupervised deep learning networks, which can highlight strongly concealed information. Subsequently, the threshold for the number of seismic facies categories is determined based on the cross-correlation values of deep features of pre-stack seismic images corresponding to different categories of seismic facies within the study area, ensuring that samples within the same class in the obtained pre-stack seismic facies images exhibit highly strong correlation, and the number of seismic facies categories is determined based on discriminant thresholds. Finally, the obtained pre-stack seismic facies images are calibrated using existing drilling information to provide a basis for geological experts to infer sedimentary environments and reservoir distributions. Theoretical model testing confirms that this method not only determines the number of pre-stack seismic facies depending on discriminant thresholds but also ensures strong correlation among samples within the same class in the seismic facies images, demonstrating greater robustness compared to other methods. Application of actual data shows that this method improves the accuracy of predicting seismic facies of fracture-cavity reservoirs in the Permian Maokou Formation and provides a reliable scientific basis for well deployment and the discovery of undrilled fracture-cavity reservoirs.

Clastic rock reservoirs are widely distributed in the Barbados accretionary wedge basin with the primary reservoir type of deepwater turbidite sandstone and the hydrocarbon type of biogas reservoir. Since gas-bearing sandstones and water-bearing sandstones in the basin are both presented as "bright spot" anomalies in seismic data, it is difficult for conventional post-stack hydrocarbon detection methods to accurately predict the distribution of gas reservoirs, which leads to a low success rate of early exploration. Thus, after analyzing logging response characteristics of drilled reservoirs, this paper proposes a new fluid detection method based on post-stack relative amplitude decoupling for "bright spot" sandstone reservoirs. Firstly, the main controlling factors of the seismic response of reservoirs are quantitatively analyzed by means of fluid substitution and forward modeling. Secondly, the post-stack relative amplitude decoupling factor is designed for "bright spot" sandstone reservoirs, and the relative amplitude relationship template is established for different lithologies and fluids. Finally, gas-bearing sandstones and water-bearing sandstones are distinguished utilizing the relative amplitude relationship template to realize reservoir prediction and fluid detection with multi-age post-stack seismic data. The application results show that the new methodology has a stronger ability to distinguish gas-bearing sandstones and water-bearing sandstones in the study area, which overcomes the blind area of traditional post-stack hydrocarbon detection methods, significantly improves the prediction accuracy of biogas reservoirs, and provides an important basis for exploration decision.

Seismic-well tie is an important step in seismic data interpretation. The traditional seismic-well tie method synthesizes seismic records by using well logging data and extracted seismic wavelets and matches them with the seismic traces beside the well by dragging. This method has significant human factors, is highly time-consuming, and can easily cause overstretching. Therefore, a deep learning method based on convolutional neural network (CNN) and gated recurrent unit (GRU) network is proposed to achieve automatic seismic-well tie. Firstly, seismic records are synthesized using typical models, and time correcting values are introduced to correct the records of seismic traces beside the well. Secondly, the relationship between two seismic traces and the time correcting values is established through a trained CNN-GRU network, and the correlation coefficients of the two seismic traces are used as constraint conditions to directly predict the time correcting values by using the synthetic seismic records and seismic traces beside the well. Finally, the neural network is tested using actual data from 30 wells, and the obtained results are compared with manual calibration results. The correlation coefficients between the calibrated synthetic seismic records and the seismic traces beside the wells are calculated. The following findings are obtained. ① The correlation coefficients of automatic calibration with the network are greater than or equal to those of manual calibration for 25 wells and are basically consistent for the other wells. ② Manually calibrating 30 wells takes about 30 min, while calibrating them with the network only takes 5 s. Therefore, compared with the traditional method, the proposed method has higher accuracy and better efficiency in seismic-well tie, which verifies the feasibility and progressiveness of the method.

In seismic exploration, multiples pose a significant issue, particularly in marine exploration. This study developed a method for suppressing interbed multiples based on the inverse scattering series method in the curvelet domain. The curvelet transform (CT), as a multi-scale and multi-directional transformation method, can sparsely represent seismic data, facilitating the capture of the main features of seismic data. By combining CT with the inverse scattering series method, the seismic data was first converted to the curvelet domain using CT. Then, the interbed multiples were predicted in the curvelet domain through the inverse scattering series method. This approach did not require building a model of the subsurface medium, offering high applicability and practicality. Furthermore, CT can sparsely represent seismic data, reducing the computational load of the inverse scattering series method. By performing an inverse CT on the predicted multiples, they were reconstructed in the time-space domain and subtracted from the original seismic data, ultimately obtaining the effective signal with suppressed multiples. The numerical experiments demonstrate that this method not only significantly improves computational efficiency but also reduces memory consumption while ensuring accuracy. This study provides an innovative approach for suppressing multiples in seismic data processing and is expected to play an important role in practical seismic exploration.