To address irregular missing seismic data caused by environmental interference during field acquisition，this paper proposes a local learning-based reconstruction method that integrates particle swarm optimization and grey wolf optimizer （HPSOGWO） algorithm with the XGBoost model. The proposed method establishes a nonlinear mapping relationship between seismic trace spatial coordinates （trace number and sampling number） and amplitude values. By adaptively optimizing feature windows using the HPSOGWO algorithm，this method achieves intelligent selection of adjacent trace data and high-precision prediction of missing values. Compared with the traditional convex-set projection method based on the Curvelet transform （Curvelet-POCS），the proposed approach significantly improves reconstruction accuracy in complex structural areas. In contrast to deep learning methods such as U-Net，it reduces the reliance on large training datasets and lowers computational costs. Tests on a three-layer horizontal layered model with 20% random missing traces show that the proposed method achieves a peak signal-to-noise ratio （PSNR） improvement of 11 dB over Curvelet-POCS and 7 dB over U-Net. F-K spectrum analysis further confirms its effectiveness in preserving seismic wavefield characteristics in the frequency domain. Tests on real onshore 2D seismic data show that the reconstructed profile with 20% missing traces achieves a relative amplitude error of 5.72%，demonstrating high amplitude fidelity and phase consistency. The method thus provides an effective and practical solution for seismic data reconstruction under complex geological conditions.

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 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.

To address the problem of limited generalization ability in traditional models for logging prediction of total organic carbon （TOC） content in the Longmaxi Formation shale reservoirs of the Sichuan Basin， this study systematically evaluates the performance of various deep learning models in both single-well and cross-well prediction tasks and identifies practical models suitable for different scenarios. Based on the acoustic time difference， density， and natural gamma-ray logging data from three wells （well1， well2， and well3） in the Longmaxi Formation， multiple models are constructed， including multiple linear regression （MLR）， support vector regression （SVR）， convolutional neural network （CNN）， bidirectional long short-term memory network （BiLSTM）， and hybrid CNN-BiLSTM and CNN-BiLSTM-Attention models. In the single-well prediction experiment， a random split strategy is applied to well1 for modeling and validation. The results show that the CNN model achieves the best performance， with the coefficient of determination （R²） reaching 0.9519 on the prediction set， demonstrating excellent local feature extraction capability and strong resistance to overfitting. To further evaluate model generalization ability， a leave-one-well-out （LOWO） cross-validation strategy is designed for cross-well prediction. The results indicate that the CNN-BiLSTM-Attention model exhibits the strongest generalization performance， achieving the highest R² of 0.9653 on the prediction set， with mean absolute error （MAE） and root mean square error （RMSE） as low as 0.131% and 0.170%， respectively， which significantly outperforms other models. The attention mechanism effectively integrates the local features extracted by CNN with the long-term sequential dependencies captured by BiLSTM， enhancing the model’s ability to focus on key information and adapt to inter-well variations. This study verifies the effectiveness and robustness of deep learning models integrated with an attention mechanism for TOC prediction under complex geological conditions， emphasizes the importance of cross-well validation in practical applications， and provides a reliable methodological foundation for shale gas sweet-spot prediction.

Deep-learning-based seismic impedance inversion methods have received wide attention due to their ability to handle nonlinear mapping problems. The conventional deep-learning-based seismic impedance inversion methods have the problem of an overwhelming dependence on labeled data, which results in a decrease in the model's ability to extract local features and poor precision of inversion results when training data is insufficient. To address these issues, a new atrous spatial pyramid pooling and U-Net (ASPP-UNet) based seismic impedance inversion method is proposed. The multi-scale feature extraction ability of U-Net is enhanced by the atrous spatial pyramid pooling operation. Based on this, the training datasets were constructed using seismic data and a small amount of logging data. To verify the effectiveness of the proposed method, we conducted two simulation experiments on the Marmousi2 and SEAM public datasets and compared the results with those of CNN, U-Net, and Attention-UNet under the same experimental conditions. The experimental results show that, under the same experimental conditions, the single-trace impedance inversion produced by the proposed method contains richer high-frequency details, and the inverted impedance profile displays smooth vertical continuity between layers and at fault locations. The inversion results also depend less on labeled data and exhibit the least information loss at positions far from the training wells, which is reflected in the strong lateral continuity between traces in the inverted impedance profile. Compared with the comparison methods, the ASPP-UNet inversion results show the best statistical indicators. To further validate the applicability of the ASPP-UNet method, it was applied to real seismic impedance inversion data from East Sichuan Province. The impedance profile obtained by ASPP-UNet is consistent with the actual geological structure. Compared with the three deep-learning methods mentioned above, the inversion results have the highest accuracy, and the impedance profile error is the smallest.

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.

Accurate prediction of geological sweet spots is a core component for the efficient exploration and development of shale oil and gas. Intelligent prediction of sweet spots of shale oil and gas based on logging data holds significant importance for hydrocarbon development. Within the field of sweet spot prediction， traditional model-based prediction methods suffer from limitations such as weak generalization， insufficient feature representation， and cumulative error propagation. This study integrates novel deep learning frameworks with the aim of achieving accurate prediction of geological sweet spots using logging data. Porosity （POR） and total organic carbon （TOC） are key parameters defining geological sweet spots. This study establishes deep learning models between logging curves and these two critical parameters to achieve intelligent prediction of geological sweet spot parameters. Firstly， a convolutional neural network （CNN） model is constructed. By analyzing and comparing different combinations of input logging curves and varying CNN layer architectures， the model achieves high-precision predictions for both POR and TOC on the test set， with correlation coefficients exceeding 0.96. Secondly， innovatively fusing CNN with an attention mechanism， a composite Transformer architecture termed the CNN-Transformer compound （CTNet） model is proposed. This model possesses the dual capability of capturing local lithological features and global inter-layer dependencies. After validation and testing on the dataset， the CTNet model achieves a correlation coefficient above 0.91 on the test set. Experimental results indicate that the CNN model demonstrates higher overall prediction accuracy than the CTNet model. However， the CTNet model exhibits significantly superior prediction accuracy within limestone intervals compared to the CNN model （e.g.， in the limestone section of well GY6-1， the correlation coefficient of the CTNet model and the CNN model is 0.848 and 0.300， respectively）， showcasing its unique potential for addressing prediction challenges in complex heterogeneous reservoirs. The findings demonstrate that both the proposed CNN and CTNet deep learning mo-dels can effectively achieve high-precision prediction of key geological sweet spot parameters in shale oil and gas， offering new methodologies for intelligent exploration.

Traditional manual cutting logging faces on-site technical challenges, including inability to quantitatively identify cutting components and accurately name the lithology, thereby hindering the comprehensive and precise acquisition of cutting lithology information. To this end, this paper develops an intelligent lithology identification system based on cutting images collected during the logging-while-drilling process. Firstly, by establishing unified standards for image acquisition and annotation, systematic sample annotation is conducted manually to generate standard samples. Secondly, the deep convolutional neural network algorithm YOLOv5 is adopted for sample training, inference, and post-processing, and an attention mechanism for small target identification is added, with the focus on the influence of Fitness function adjustment on target identification accuracy. Finally, the ONNX(Open Neural Network Exchange) model is adopted for cross-platform support, developing an intelligent lithology identification system based on cutting images. Practical applications show that the system can identify six major lithologies (mudstone, sandstone, limestone, dolomite, coal, and carbonaceous mudstone), with the overall accuracy exceeding 85%. Meanwhile, the system can analyze the component contents of each lithology and characteristics of sandstone cuttings including the roundness, grain size, and sorting, thus enabling the accurate description of cutting component characteristics and developing a new lithology identification technology.

Full waveform inversion (FWI) is currently the most accurate velocity field inversion technique. FWI employs gradient-based local optimization algorithms to minimize the error between the forward data and the original data, thereby obtaining an accurate subsurface velocity field. The success of FWI on real data relies on the careful processing of original data and the prudent selection of inversion strategies. This paper discusses and experiments with the parameter control and implementation strategies of FWI, considering the characteristics of marine data, and proposes practical techniques for full waveform inversion of marine data. Specifically, the technology is as follows. ① The fraction frequency noise suppression technology is used for turning waves, which primarily improves the signal-to-noise ratio of low-frequency turning waves while avoiding damage to the effective signal. ② The wavelet adjustment techniques to enhance the accuracy of the initial wavelet are applied, which allows for more precise calculation of the error between forward data and actual data and leads to a more accurate calculation of the velocity update gradient. ③ The first-arrival tomography inversion velocity models for near-seafloor velocity modeling improve the accuracy of the initial velocity field, better avoid cycle skipping, and ensure that the objective function converges towards a more accurate direction and at a faster rate. ④ The offset increment strategy is adopted during FWI iterations, progressively updating the velocity field from shallow to deep, reducing the solution multiplicity of FWI, and ensuring more accurate convergence of the FWI velocity field. This technology has achieved relatively good results in practical marine data projects, obtaining more accurate velocity fields and better depth migration imaging effects.

In recent years, oil and gas has been repeatedly discovered in the western margin of the Ordos Basin, which reveals its enormous exploration potential. Yindongzi is the first three-dimensional block deployed by Changqing Oilfield for the complex piedmont structural belt. Based on high-resolution three-dimensional seismic data, the geometric characteristics, activity characteristics, formation and evolution mechanisms, and reservoir control mechanisms of the Yindongzi Huisha Fault are studied. The following understandings are obtained: (1) Huisha Fault has obvious segmentation characteristics, and can be divided into north, central, and south segments according to the intensity of fault activity, fault inclination, and dip angle. The central segment has the strongest fault activity intensity, followed by the north segment, and the south segment is the weakest. (2) The hanging wall of Huisha Fault mainly develops contraction structural styles under compression structural environment, while the foot wall of Huisha Fault mainly develops normal fault related structural styles under extension environment and reverse faults under later transformation. In addition to the development of extension related structural styles, compression structural styles, and reverse structural styles in the study area, strike slip structural style is observed locally. (3) Balanced cross-section restoration reveals that the Huisha Fault is a active fault with multiple stages, being a normal fault in the early stage and a large-scale thrust fault after multi-stage compression and transformation in the later stage. Tectonic movements in the Indosinian and Yanshanian periods have the greatest impact on its formation in multiple periods of tectonic movement. (4) Multiple sets of confirmed source rocks develop from top to bottom on the western margin of the Ordos Basin. Currently, oil and gas have been discovered in the Permian and Ordovician Wulalike source rocks. In the future, under the efficient migration and guidance of faults, faulted anticline, anticline, and faulted nose traps may become key exploration targets with the potential for breakthroughs.

The transparency of geological condition detection and precise exploration are critical challenges hindering safe and efficient coal mining operations. Coalfield seismic exploration technology plays a vital role as a means to address hidden, potentially hazardous geological issues during the mining process. This technology provides high-precision regional geological structures and coal formation patterns, offering reliable geological basis for coalfield development. The evolution of coalfield seismic exploration technology in China can be divided into three distinct phases: the early stage from the 1950s to the early 1970s marked by technological inception; the digital development era spanning the late 1970s through the 1980 s; and the current phase since the 1990 s, characterized by extensive promotion and practical application. Over the course of more than seventy years, this technology has made significant advancements in coalfield geological surveys and the precise detection of hidden factors contributing to geological hazards. It stands as an indispensable geological support for ensuring safe and efficient coal development. By integrating in-depth research on coalfield exploration techniques with typical engineering practices, the text systematically outlines its technical features and current status across the stages of data acquisition, processing, and interpretation. In response to the urgent need for transparent and intelligent mine construction, future developments in coalfield exploration technology will focus on advancing dense distributed data collection, intelligent data processing and interpretation, and the innovative application and development of multi-attribute integration techniques.

Microseismic monitoring technology is of application significance in fields such as unconventional oil and gas reservoir development and mine disaster monitoring. However, its signals are susceptible to noise interference, which results in a low signal-to-noise ratio (SNR), thus severely compromising the accuracy of subsequent seismic source localization and mechanism inversion. Traditional denoising methods such as the complete ensemble empirical mode decomposition (CEEMD) and wavelet modulus maxima (WMM) have limitations in processing non-stationary microseismic signals. To this end, this paper proposes a microseismic denoising method named SSA-VMD-CC-WT, which combines variational mode decomposition (VMD) optimized by the sparrow search algorithm (SSA) with the adaptive wavelet thresholding (WT). Firstly, SSA is employed to optimize key parameters of the VMD algorithm. Secondly, effective modal components are selected by utilizing the cross-correlation coefficient (CC) to suppress noise. Finally, adaptive WT is applied to perform secondary denoising on the effective components, reducing signal distortion. Simulation tests demonstrate that in strong noise conditions, the SSA-VMD-CC-WT method can separate noise from effective signals more accurately than the CEEMD and WMM methods. The processing of actual microseismic data reveals that the proposed method significantly suppresses both low-frequency and high-frequency noise while maintaining the fidelity of weak seismic sources, thereby improving data interpretability and SNR. Meanwhile, compared with the traditional genetic algorithm (GA), SSA demonstrates higher optimization efficiency.

Accurate estimation of coalbed methane (CBM) content plays a crucial role in assessing and efficiently exploiting CBM resources. Deep CBM is influenced by multiple controlling factors and complex genetic mechanisms. Currently, machine-learning approaches for CBM content prediction typically rely on either seismic or logging data. As a result, the complex geological conditions of deep coal seam are not fully accounted for. This study proposes an intelligent prediction method for CBM content, which achieves multi-source data fusion through a multi-scale modeling and deep integration strategy. The approach first extracts multi-scale sensitive attributes or features relevant to CBM content from geological, logging, and seismic sources. For each dataset of the same scale, adaptive modeling is performed using a Bayesian hyperparameter-optimized random forest (RF) algorithm, which enhances model robustness and prevents overfitting. The prediction results from individual scales are subsequently integrated through the least squares method to construct a multi-scale RF composite model. The proposed method is validated using a field dataset and compare its performance with that of conventional approaches, including single-scale RF and linear regression. The results show that, compared with these baseline methods, the proposed method reduces the mean relative error of CBM content prediction on test wells by 3.01% and 4.94%, respectively. This demonstrates that the proposed approach achieves higher accuracy and stronger generalization capability, enabling precise characterization of the spatial distribution of CBM content.

The reservoir space of the Lower Paleozoic carbonate rocks in the Jiyang Depression is mainly composed of fractures and dissolution pores. The reservoirs are highly heterogeneous both vertically and horizontally, which makes the prediction of the buried hill reservoirs quite challenging. Therefore, based on the development characteristics of the reservoir space and the correlation of seismic reflections, this paper proposes a seismic forward modeling method for the buried hill interior based on the dual-porosity medium model, which characterizes the seismic response features of the Lower Paleozoic buried hill interior. Meanwhile, a buried hill reservoir prediction method based on structure-oriented filtering is developed, achieving effective prediction of the favorable reservoir development zones in the Lower Paleozoic buried hill. The research results show that the development of the Lower Paleozoic carbonate reservoirs in the Jiyang Depression is the main factor causing seismic reflection anomalies, and the effective reservoir sections exhibit obvious seismic response anomalies. By comparing the data residuals before and after structure-oriented filtering, the development positions of the main inner reservoirs can be indicated. This method has achieved good results in the Pingnan buried hill and Dawangzhuang buried hill in the Dongying Sag.

Sand body are a kind of common reservoir unit, and their accurate identification and tracking are the key to discovering oil and gas fields and supporting the increase in oil and gas reserves and production. Existing methods such as attribute analysis and deep learning still face challenges such as low boundary identification accuracy, complex parameter selection, and poor noise resistance. To this end, this paper proposes a prompt prediction method based on the Segment Anything Model (SAM), a visual image segmentation foundation model. This method requires no model training, and by simply employing the boundary prompt points of target sand bodies, precise identification and tracking results of the boundaries of target sand bodies can be obtained. To address the prediction error of the prompt encoder in SAM when applied to seismic profiles, this paper proposes a KD-tree search method. By calculating the shortest distance from the prompt points to the potential sand body segmentation blocks, the optimal sand body prediction results are determined. After conducting verification with actual target area data and comparison with the customized training of a U-Net model on the target area data, it is demonstrated that the sand body tracking method based on SAM depicts more lateral changes of sand body boundaries and the boundaries are more consistent with seismic amplitude variations.

In recent years, the Keping fault uplift in Tarim Basin has emerged as a significant breakthrough zone for hydrocarbon exploration, where well-developed high-quality source rocks coexist with complex structural settings, creating a unique exploration environment. The combination of intense surface relief, pronounced velocity contrasts from shallow high-velocity rock masses, and multi-phase superimposed fault systems has led to technical bottlenecks in conventional seismic imaging methods, including substantial static correction errors, low velocity modeling accuracy, and structural distortion. This study systematically investigates key pre-stack depth migration (PSDM) technologies tailored for piedmont zones based on geological characteristics of the 3D seismic survey in the Kepingnan area. Three main innovations are presented: ① Integrated shallow layer modeling technology. To address near-offset data deficiencies caused by complex surface acquisition geometries, this paper develops a constrained tomographic inversion algorithm integrating uphole survey data with wide-azimuth seismic information was developed. This approach effectively enhances shallow velocity model accuracy while overcoming velocity-thickness coupling limitations inherent to conventional methods. ② True surface migration datum construction technology. By optimizing surface-consistent static correction schemes and implementing high-frequency static correction time-difference correction, the paper establishes a dynamic matching mechanism between true surface elevation and migration datum was established. This innovation significantly mitigates topographic effects on wavefield continuation and improves structural fidelity in steeply dipping stratigraphic imaging. ③ Multi-scale velocity iterative inversion technology. The paper develops a multi-azimuth grid tomography method incorporating structural constraints. Through azimuth-dependent hierarchical inversion strategies, this technique achieves synergistic improvements in both vertical and lateral velocity resolution, providing precise velocity models for deep nappe structures and buried fault systems. Field applications demonstrate that this technology system substantially enhances imaging quality of depth-migrated seismic data in Kepingnan area. Target horizons exhibit improved signal-to-noise ratios with clearer fault imaging and more accurate structural configurations. The proposed methodology provides an effective technical solution for hydrocarbon exploration in complex piedmont zones, offering practical value for advancing regional exploration progress.

Accurate underground velocity information is crucial for seismic imaging in complex area. While existing seismic waveform inversion techniques are highly accurate, they have shortcomings such as high computational amount and reliance on initial models. Currently, deep learning technology experiences rapid advancements in various fields and has successfully been applied to nonlinear seismic inversion. However, conventional end-to-end deep learning networks struggle to establish a multi-scale physical coupling relationship between velocity parameters and seismic records. To this end, this paper proposes a hybrid network AER-UNet, which reorganizes the encode and decoder structures and adds an attention mechanism-based jumping connection module on this basis. This approach effectively obtains key spatial information from seismic records and enhances the representation of the subtle structures in velocity fields, thus accurately capturing the characteristics of underground medium velocity parameters. An appropriate number of random velocity models should be built in the network training phase to simulate the true structure of the underground medium and thus obtain the accurate mapping relationship between velocity models and seismic records. Additionally, developing new loss functions can help improve the computational accuracy of velocity modeling. By carrying out numerical experiments using the SEG/EAGE thrust model, the effectiveness of the hybrid network for velocity modeling is evaluated. Compared to FWI and other deep learning networks, this method can more efficiently and accurately rebuild underground velocity models.

In response to high-resolution imaging in complex geological structures, surface imaging and vertical seismic profile (VSP) imaging mitigate illumination limitations inherent to either method alone, but are still subject to unbalanced image amplitude and structural distortion due to uneven illumination. To address this issue, a joint VSP-surface Gaussian beam depth migration method is proposed. First, a dynamic complementary mechanism is established for the surface and borehole wavefields, where illumination-based weighting factors are applied to balance their respective contributions and integrate their advantages. Second, to constrain the Gaussian beam propagation paths, a structural dip field is introduced, enabling dynamic adjustment of the initial beam direction for adaptive alignment with the local formation dip. Third, a dip-dependent weighting function is employed during the imaging stack to suppress scattering energy from non-geological directions, thereby enhancing the signal-to-noise ratio of the joint VSP-surface image. Applications to both synthetic and field data demonstrate that the joint VSP-surface Gaussian beam depth migration method produces superior images compared to those obtained using surface data alone and improves the imaging accuracy of complex structure.

Faults are among the key geological factors that induce major hazards such as water inrushes, gas outbursts, and roof falls in coal mines, and their accurate identification is a prerequisite for safe and efficient coal mining. Conventional fault interpretation methods based on post-stack seismic attributes have a relatively low identification rate for small faults with throws of 3~5 m, which has become a bottleneck severely restricting intelligent and precision coal mining. Therefore, a method for identifying small coal seam faults based on an improved U-Net is proposed. Residual structures, a convolutional block attention module (CBAM), and an atrous spatial pyramid pooling (ASPP) module are incorporated into the classical U-Net architecture. The residual structures alleviate the vanishing-gradient problem and improve the stability of model training. The CBAM enhances the network's ability to perceive key fault features. Moreover, the ASPP module enables the efficient capture and in-depth extraction of multi-scale information from seismic data, thus improving the accuracy and robustness of coal seam fault identification. The results obtained using a synthetic seismic test dataset, a coal seam fault model, and field seismic data demonstrate that the proposed improved U-Net model accurately identifies most coal seam faults, yielding clear fault boundaries, strong spatial continuity, and significantly improved identification accuracy. This method has important theoretical value and broad application prospects for promoting the intelligent upgrading of geological hazard prevention in coal mines.

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.

Shear-wave velocity is a key parameter for pre-stack seismic inversion and reservoir characterization. However, due to the technical and cost constraints of both direct and indirect measurement methods, it is quite difficult to obtain in practice. Therefore, a prediction method is proposed based on the broad learning system (BLS). First, appropriate well-log data are selected and pre-processed through denoising and correlation analysis. Second, a BLS neural network structure comprising mapping nodes and enhancement nodes is constructed to complete the BLS process. Finally, well-log data from the two typical wells Y301 and Y302 in the Y block of the Junggar Basin are used to construct a data set of machine learning. Two contrast experiments are designed and compared with curve fitting and deep learning system to verify the stability and generalization of BLS. The actual results show that the proposed BLS-based shear wave velocity prediction method can reduce training time while achieving prediction accuracy, providing a new neural network option for shear wave velocity, petroleum, and relevant reservoir parameter prediction.

Noise in post-stack seismic data can severely interfere with the identification and interpretation of reflection signals, making efficient denoising essential for improving the accuracy of seismic data interpretation. Owing to its superior multi-scale and multi-directional characteristics, curvelet transform has been widely adopted for post-stack data denoising. However, conventional curvelet transform tends to produce pseudo-Gibbs effects at boundaries, which leads to edge oscillations and spurious reflections. In addition, aggressive noise suppression often results in the loss of valid signals, which limits the practical applicability of these methods. To address these challenges, this paper proposes a curvelet transform approach incorporating multiscale adaptive block curvelet-domain thresholding (MABCDT). First, cycle spinning and MABCDT are introduced in the curvelet domain to enhance the preservation of weak signals and significantly mitigate pseudo-Gibbs phenomena. Then, a fast non-local mean filtering is applied to the data after inverse curvelet transform, which further retains valid signals while removing residual noise. Finally, a directional smoothing diffusion algorithm is introduced, which utilizes gradient direction information to perform directionally weighted smoothing and diffusion, thereby further suppressing noise and enhancing the continuity of valid signals. Both synthetic and field data tests demonstrate that the proposed method outperforms conventional post-stack denoising techniques in terms of signal-to-noise ratio enhancement and waveform fidelity. The method preserves the continuous structural features of seismic signals while suppressing noise and significantly reduces pseudo-Gibbs effects caused by high-frequency truncation.

Diffraction imaging is an important method to improve the imaging accuracy of underground small-scale geologic bodies. However, conventional seismic surveys are mainly based on reflection imaging, with weak-energy diffraction suppressed, as a result of which diffraction need to be separated and imaged separately. At present, the localized damped rank-reduction with adaptively chosen ranks (LDRRA) method widely used improves the separation accuracy of diffraction by damping the adaptively chosen singular value matrix with a damping operator, but its damping factor is mainly given manually, and all local window data use the same given damping factor. Different local windows contain different seismic data, and using the same damping factor will reduce the separation accuracy of diffraction. Therefore, a localized adaptive damped rank-reduction (LADRR) method for diffraction separation is proposed. First, based on the LDRRA framework, the Hankel matrix undergoes singular value decomposition (SVD) to truncate singular values. Second, a squared ratio of singular value is introduced to adaptively compute a damping factor for each localized data window, through which the optimal damping factor is selected to apply damping effects to the truncated singular values, thereby preserving the reflection components. Finally, the damped localized window data is subjected to inverse Hankelization and inverse Fourier transform, and then subtracted from the original wavefield to yield the separated diffraction. Theoretical simulation and field data test results demonstrate that the proposed method can separate diffraction with high accuracy, and the imaging results of the separated diffraction can get more accurate location of the underground small-scale geologic body.

Rock fabric contains abundant geological information， and lithology identification from rock thin sections is of great importance for oil and gas exploration， mineral extraction， and related fields. To address issues such as low accuracy and high labor costs in rock image classification， this paper proposes a ConvNext V2-VMamba-based method for lithology identification in rock thin sections. First， based on the ConvNext V2 model， grouped convolution and channel shuffle strategies are adopted to combine the ConvNext module with the VSS module. Then， a Conv-VMamba module is proposed to replace the original ConvNext module， enabling the model to possess both a global receptive field and excellent local feature extraction capabilities. Finally， a spatial attention EMA module is integrated into the model to enhance cross-channel spatial information aggregation and improve the capture of texture information in different rock images. Experimental results show that the model achieves an accuracy of 81.5%， precision of 81.1%， recall of 81.4%， specificity of 96.3%， and F1-score of 81.2% on the test set. Compared with the original model， the accuracy is improved by 5%. This method demonstrates the highest algorithmic accuracy and best classification performance， providing a new approach and method for the field of lithology identification.

The Marchenko method can reconstruct the Green's function at any subsurface position by utilizing the surface reflection response and the background velocity model, which can be applied to structural target-oriented imaging. Compared with traditional seismic interferometry, the imaging is not affected by the overburden medium. However, the imaging quality of the Marchenko method relies on high-density acquisition, which field data sets often fail to provide. To address this issue, this paper proposes the Marchenko imaging method based on CNN-POCS reconstruction. First, sparsely acquired seismic data is reconstructed using the convolutional neural network projection onto convex sets (CNN-POCS) method to generate high-accuracy data with dense spatial sampling. Marchenko imaging is then applied to the target zone using the reconstructed data. This approach not only reduces the high-density acquisition requirement of the Marchenko method but also preserves imaging accuracy. Finally, the method is extended to plane-wave imaging, which improves the computational efficiency of the conventional point-source approach and reduces the computational cost. The model test results show that the proposed method can reduce the receiver deployment by 60%, thus significantly lowering acquisition costs.

Denoising is a critical step in seismic data processing. Traditional dictionary learning methods lack constraints on the physical mechanism of signals， easily misidentifying coherent noise with linear or curvilinear features as effective signals and incorporating them into dictionary atoms， which results in distortion of effective signals or residual noise during reconstruction. To address these problems， this paper proposes a hybrid noise suppression method based on sparse basis learning with physical constraints. Firstly， the plane-wave destruction （PWD） filter is used to estimate and update the local dip field， ensuring that the dip field matches the iteratively optimized signal. Then， the local dip field estimated by PWD is taken as a constraint to guide the sparse basis learning process， which combines the signal representation ability of sparse transform with the local dip characteristics of effective signals. This enables dictionary atoms to match only effective signals with corresponding dip characteristics， avoiding the mislearning of coherent noise atoms. Finally， effective signals can be accurately sparsely represented on the sparse basis matching their dip characteristics， while various noises are difficult to be matched by the sparse basis due to the absence of corresponding dip characteristics. Thereby，high-fidelity effective signal reconstruction and effective suppression of multi-type noises are achieved. Tests on both complex synthetic data and field data show that the proposed method has superior suppression performance and wide applicability in suppressing hybrid noises such as strong abnormal amplitude interference， surface waves， random noise， and coherent noise， compared with conventional methods and industrial standard denoising workflows.

Seismic impedance inversion is a technical means to quickly estimate the thickness of underground rock layers and evaluate the quality of reservoirs by using post-stack seismic data. Howerver, limited by the inappropriateness of the inversion itself, the process is highly susceptible to the influence of the observed data and the model bias, which leads to the instability of the inversion results. Instability arising from noise or the forward model during the solution process can be suppressed by appropriately introducing smoothness, sparsity or structural assumptions in the objective function. The total-variation regularization optimization method can preserve formation edge features, but may also lead to blocky structures of the inversion results. To overcome this defect, this paper introduces hybrid fractional-order total-variation regularization constraints to characterize the inter- and intra-stratum variability characteristics in a more reasonable way. First, an objective function with hybrid total-variation regularization constraints is built. Then, the orthogonal finite-memory quasi-Newton method is used to solve this complex objective function, so as to improve the resolution and stability of seismic impedance inversion. Test results and real data application demonstrate that the hybrid fractional-order total-variation regularization can be better adapted to the inversion of lumped and smooth hybrid models than total-variation regularization and $ {\mathrm{L}}_{2} $ regularization. It is verified that the proposed inversion method has the features of high accuracy, strong noise immunity, and fast convergence speed.

During waterflooding, evaluating interwell connectivity is essential for reservoir management, production strategy optimization, and improved hydrocarbon recovery. Existing methods, those are based on graph neural networks fail to characterize the temporal and spatial relationships within well networks through graph structures, limiting the accurate description of delayed dynamic responses. To bridge this gap, this study proposes a multi-level graph-structured temporal network model to capture dynamic relationships from time-series well network data, and enable precise interwell connectivity evaluation. Specifically, considering the injection-production response delay of a well network, a multi-level graph-structured temporal-spatial dependence representation method is proposed, which integrates production time-series responses with the spatial structure information of the well network is introduced. Subsequently, a multi-level temporal graph neural network model is established; Next, an attention mechanism based hierarchical information interaction and update method is developed. These together enable the model to explore dynamic interactions between injection and production wells and achieve accurate inversion of interwell connectivity. Experimental results demonstrate that the proposed model exhibits superior accuracy over conventional temporal models, with a consistency of 93.8% with tracer test results. Moreover, it conforms to the gradual evolution of connectivity in accordance with physical principles, demonstrating the method's strong practical applicability in engineering.

In seismic exploration, owing to the viscoelastic properties of subsurface media, the seismic wavefield experiences substantial attenuation, loss of high-frequency components, and phase distortion during propagation. This is because energy is absorbed, which ultimately results in a reduction of the resolution and imaging accuracy of seismic data. Precisely acquiring the Q-value of the near-surface media and conducting inversion processing is a crucial approach to address this issue. Consequently, an adaptive near-surface Q-value estimation method based on the asymmetric wavelet spectrum is herein proposed. Initially, a synthetic wavelet is generated by taking the geometric mean of the input and output signals. Subsequently, an adaptive algorithm is introduced in light of the synthetic wavelet to determine the key parameters, thereby minimizing the errors arising from subjectively selecting constant parameters. Finally, a formula is derived to compute the near-surface Q-value, enabling precise estimation of the Q-value. Both theoretical model validation and practical application outcomes demonstrate that the method presented in this study can automatically adapt to the changes in spectral asymmetry during the absorption attenuation of seismic wave propagation. It can more accurately fit various asymmetric source amplitude spectra, enhance the accuracy of Q-value modeling, and offer technical support for high-precision seismic data processing.

By random sampling, compressive sensing seismic acquisition can significantly improve seismic acquisition efficiency and reduce seismic acquisition costs. Therefore, a design method and workflow are proposed for an offshore compressive sensing seismic acquisition geometry. Firstly, in response to the alias and difficulty in anti-alias reconstruction in the evaluation of offshore compressive sensing seismic acquisition geometry and field data processing, a frequency-wave number domain weighted anti-alias data reconstruction method is put forward. Then, the developed efficient offshore compressive sensing seismic acquisition technology and data reconstruction method are applied to a real survey of a volcanic buried hill in the South China Sea. A wide azimuth ocean bottom nodes compressive sensing acquisition geometry is also designed. The quality, feasibility, cost reduction and efficiency improvement of the compressive sensing acquisition geometry are evaluated by the methods including data reconstruction. The designed compressive sensing acquisition scheme is successfully implemented in the real survey in the South China Sea, obtaining the first independently designed and implemented offshore compressive sensing acquisition field data in China. The application results show that, under the premise of reducing costs and increasing efficiency by 25%, the compressive sensing acquisition in this survey achieves comparable results to the conventional acquisition, providing effective support for the subsequent research and promotion of efficient offshore compressive sensing acquisition technology.

The study on the sedimentary evolution process of the continental lacustrine basin in the Yanchang Formation of the Ordos Basin is of great significance for understanding the laws of generation, migration, and enrichment of petroleum resources. Based on the seismic data from the middle-upper part of the Yanchang Formation in the Ordos Basin, fractional derivative frequency-increasing processing is utilized to track the strike line of the slopes in angle-dependent horizon slice and seismic reflection event and realize isochronal classification and regional contrast. A total of 12 isochronal progradational sequences (fs1~fs12) are established within the work area. According to the size of the slopes and the depth of the lacustrine basin in different periods, a three-stage lacustrine basin evolution mode of "initial-development-shrinkage period" is proposed through the combination of the plane and section of 3D seismic data: ① The lacustrine basin during the initial period (fs1~fs3) was shallow to semi-deep water; ② in the development period (fs4~fs7), it was deep water, with sediment sources mainly from the west and southwest directions; ③in the shrinkage period (fs8~fs12), the lacustrine basin changed from deep to shallow water, and the paleocurrent underwent channel shift in the southwest sediment source area. Consequently, the filling of the lacustrine basin was completed. The progradation pattern mainly shows an S-shape or double S-shape in the lake area, and the migration angles (< 2°) show a slow upward trend. The filling of the lacustrine basin is composed of thick slope deposits and relatively thin topset deposits, belonging to normal lake regression-type sedimentation. Seismic inversion and drilling and logging data indicate that each progradational sequence exhibits positive grading sedimentary characteristics. The bottom is dominated by low-velocity dark mudstone, while the top by high-velocity sandstone. The alternating development of mudstone and sandstone indicates that during the evolution of the lacustrine basin, the paleoclimate alternated periodically between warm-humid and arid conditions. The research results provide theoretical support for the sedimentary dynamics evolution of continental lacustrine basins and the prediction of shale oil sweet spots.

Natural fracture systems serve as the storage space and seepage channel of shale gas reservoirs, and fracture identification of shale reservoirs is faced with such problems as high cost and low coverage of micro-resistivity imaging logging, as well as significant nonlinear characteristics of conventional logging data and strong subjectivity of manual interpretation. To this end, this paper screens four kinds of logging curves with high sensitivity to fractures of shale gas reservoirs, including the deep lateral resistivity, natural gamma ray, neutron porosity, and interval transit time via sensitivity analysis. Meanwhile, the first-order difference curve of resistivity and its product curve are introduced, and the complex temporal sequence relationship of conventional logging curves is transformed into a classifiable threshold discrimination problem. Subsequently, a BP neural network identification model for fractures in shale gas reservoirs is built based on the conjugate gradient descent optimization algorithm. The results demonstrate that this model can effectively eliminate the subjective bias in manual interpretation. Compared with the actual fractures, the recall ratio of fracture identification in shale gas reservoirs reaches 90%, with a precision rate of 87%. This research provides a novel approach for fracture identification of unconventional hydrocarbon reservoirs, effectively enhancing the identification efficiency of fractures in shale gas reservoirs.

In geological exploration, density and acoustic time difference curves can reflect key physical parameters, such as underground geological structure and reservoir porosity. However, due to the influence of complex geological conditions and other factors, logging data may be incomplete or missing. Therefore, this paper proposes a logging curve reconstruction method based on a Granger causality graph neural network (GCGNN). This method constructs a Granger causality graph by learning the Granger causality between logging curves and uses a graph convolutional network to process and predict missing data. The method is applied to the measured well data in the Gujing area and Jinjing area of the central depression of Songliao Basin in China. The correlation between the density and acoustic time difference curves of Well Gu204 and the original data is 71.70% and 83.76%, respectively, and that is 80.03% and 88.73%, respectively, for Well Gu432. The performance of GCGNN in the reconstruction experiment of the same well is better than that of the lightweight gradient boosting machine, time convolutional network, and long short-term memory network. The method is applied to the reconstruction experiment of different wells. The correlation between the density and acoustic time difference curves and the original data is 77.54% and 87.79%, respectively. Although the model obtained by GCGNN is not the optimal model, the reconstruction effect is still good. The application results on measured data show that the proposed method can effectively predict the missing logging data.

In seismic exploration， it is generally believed that a higher fold density means a greater number of seismic wave samplings of underground media， and theoretically， the signal-to-noise ratio and resolution of seismic imaging will be higher. However， in actual production， the disconnect between the setting of acquisition parameters and imaging quality indicators has become an increasingly prominent issue. The final imaging results of many seismic acquisition data with high fold density are unsatisfactory. Excessive pursuit of high fold density not only substantially increases the cost of seismic acquisition but also leads to the gradual neglect of basic parameters such as excitation energy and receiving sensitivity. Therefore， an innovative method for estimating the efficiency of seismic acquisition parameters is proposed. First， by constructing geophysical models and conducting quantitative calculations， the actual contribution of the observation system to imaging points and the contribution of excitation and receiving links to effective signal energy are clarified in sequence. Second， a multi-dimensional evaluation system focusing on fold density， excitation parameters and receiving parameters is established， and a comprehensive evaluation scheme for parameter efficiency covering the entire acquisition process is formulated. Finally， a proper balance is struck among the selection of acquisition parameters， the control of acquisition costs and the quality of seismic imaging， so as to achieve the dual improvement of the efficiency and quality of seismic acquisition. The results of case comparative analysis show that the method proposed in this paper breaks through the previous one-sided approach of evaluating imaging effects solely by relying on fold density， and provides technical support for the high-quality and high-efficiency precise exploration of the "wide-azimuth， wide-band and high-fidelity" seismic acquisition technology.

Significant velocity dispersion and energy attenuation may occur when seismic waves propagate in the gas-bearing reservoirs. Based on the time-frequency analysis techniques, the inversion of the seismic wave dispersion parameters can be used to predict the gas-bearing property of reservoir. In this paper, the time-frequency spectruml analysis method based on matching pursuit is improved by the wavelet shaping technique which can be used to compress the duration of the wavelet amplitude envelope and to enhance the time resolution of the spectral decomposition result. The proposed time-frequency analysis method is combined with dispersion parameter inversion to calculate the dispersion gradient of the reflection coefficient, which is corresponding to the gas bearing property. The test results of synthetic seismic data indicate that the proposed wavelet-shaping-based time-frequency spectrum analysis shows higher time-frequency resolution compared to the conventional time-frequency analysis methods. The dispersion parameter inversion based on the proposed spectrum analysis method is applied to the real seismic data from X exploration area in the Yinggehai Basin, which verifies the effectiveness of the proposed method in gas reservoir prediction. The practical application results provide important guidance for oil and gas exploration and development.

Overpressure is widely developed in the deep layers of the Bozhong Sag in the Bohai Bay Basin. The complex genesis of overpressure and limited pressure measurement data in some exploration well areas as well as the low prediction accuracy of the existing Eaton method pose significant risks to safe drilling. Targeting the deep overpressure development area in the Bozhong Sag, this paper proposes an innovative three-step method for quantitative prediction of overpressure, which includes determining the cause, building a model, and conducting strong training. Specific implementation steps are as follows: ① The genesis of overpressure is analyzed with pressure measurement, logging, and other data, and the contribution rates of overpressure arising from different causes are quantitatively evaluated. ② A new pressure prediction model is established for overpressure from composite genesis based on the acoustic velocity difference (ΔV) and the contribution rate of undercompaction overpressure (α). ③ The sample data calculated with the new model is input into the convolutional neural network to achieve the multi-parameter pre-drilling prediction of overpressure from composite genesis. The research results indicate the followings: ① The overpressure of deep mudstone in the Bozhong Sag is mainly caused by hydrocarbon generation and undercompaction. The contribution rate of undercompaction overpressure varies greatly among different structures and layers, and the difference in the same area and layer is relatively small. The overall contribution rate of undercompaction overpressure is between 20% and 98%. ② Based on the model and data-driven overpressure prediction technology, accurate prediction of overpressure from different causes can be achieved in the target area. The average relative error of the pressure factor prediction results of test wells does not exceed 6.2%. This research provides a new idea for quantitative prediction of overpressure from complex genesis in areas with few wells.

When dealing with complex seismic-geological condition data, Rayleigh wave exploration technology faces challenges in fast and accurate inversion of formation parameters because the inversion of dispersion curves involves multi-parameters, multi-extremes and nonlinearity. Therefore, a Rayleigh surface wave dispersion curve inversion method is proposed based on improved flying foxes optimization. First, in response to the local optimum of flying foxes optimization in the later stage of iteration, the Lévy flight strategy is introduced and its distribution is applied to replace the existing uniform distribution of flying foxes optimization, thus enhancing the global optimization ability of the algorithm. Second, local optimal perturbation is introduced during the iteration process to develop the improved flying foxes optimization and increase the population diversity, which helps the algorithm quickly approach the global optimal location and direction, avoid local extremum, and improve solution accuracy and approaching speed. Third, a Benchmark function test is carried out on the improved flying foxes optimization to verify its superior performance on issues with different complexities. A typical geologic model is also established to conduct the theoretical dispersion curve inversion with and without noise. Therefore, the effect and advantages of the algorithm in the Rayleigh surface wave dispersion curve inversion are confirmed. Performance test results of two commonly used Benchmark functions indicate that the improved flying foxes optimization is superior to the classical one in global optimization ability. The standard deviation and relative error are further introduced for quantitative analysis. The results show that the improved flying foxes optimization has higher inversion accuracy and stability than the traditional one, showing strong application potential.