This study represents the pioneering application of calcite U-Pb dating to directly date multiple deformation events in the eastern Sichuan Basin, China. This approach is crucial for advancing our understanding of the Meso-Cenozoic tectonic evolution of the Yangtze Block and the Tibetan Plateau. We present calcite U-Pb dating results from three reverse/strike-slip faults, which allow us to constrain four distinct deformation events. The data reveal that the first deformation event along the approximately west-east-striking reverse fault occurred at the end of the Middle Jurassic (similar to 164 Ma). This event is associated with northward subduction of the Bangong-Nujiang Ocean. Subsequent deformation event along the northeast-striking Qiyueshan Fault Zone occurred during the Early Cretaceous (similar to 135 Ma) and is interpreted as resulting from far-field oblique flat-slab subduction of the Paleo-Pacific Plate. Reactivation of the Qiyueshan Fault Zone during the Eocene (similar to 48 Ma) was driven by NW-SE compression associated with the thrusting of the Longmenshan fold-thrust belt, marking the first documented instance of this deformation event in the eastern Sichuan Basin. The youngest age, dating to the Oligocene (similar to 28 Ma), is linked to a near north-south-striking strike-slip fault. This final deformation event is attributed to the eastern expansion of the Tibetan Plateau and the subsequent counterclockwise rotation of the Sichuan Basin.

The Lintangchang area of the Sichuan Basin has experienced multiple phases of tectonic deformation, resulting in a complex development of faults and fracture systems. Drilling results indicate that the Upper Ordovician Wufeng Formation to the Lower Silurian Longmaxi Formation possesses favorable exploration and development potential. Therefore, clarifying the development characteristics of faults and their associated fractures within this interval is critical for subsequent exploration.In this study, faults in the Lintangchang area are classified into four hierarchical levels (A, B, C, and D), with three dominant strike directions: NE-trending, near E–W–trending, and near N–S–trending. A-level faults are mainly distributed on the flanks of the anticline, whereas fewer faults are developed at the plunging ends of the anticline. By integrating three-dimensional seismic attribute fusion with the Fault Damage Zone Index (FDI) method, a quantitative characterization of fault-related fractures in the Wufeng–Longmaxi formations was conducted.The results show that: (1) The width of fault-related fracture zones exhibits a significant positive correlation with fault displacement. The fracture zone widths of A-, B-, C-, and D-level faults are 510–660 m (average ~600 m), 160–280 m (average ~220 m), 130–200 m (average ~168 m), and 115–170 m (average ~150 m), respectively. Fracture zones in the hanging wall are generally wider than those in the footwall, and fractures are most intensely developed at fault intersection zones. Among faults of the same hierarchical level, near E–W–trending faults exhibit the widest fracture zones, followed by near N–S–trending faults, whereas NE-trending faults show the narrowest zones.(2) Fault-related fracture zones are closely associated with shale gas preservation conditions. Wells located within fracture zones display lower formation pressure coefficients and lower productivity, whereas wells farther away from fracture zones exhibit higher pressure coefficients and greater gas production. A significant positive correlation exists between formation pressure coefficient and gas productivity, indicating that fault-related fractures exert a strong control on pressure preservation and gas leakage.This study elucidates the development characteristics of faults and associated fractures in the Wufeng–Longmaxi formations of the Lintangchang area and provides an important geological basis for shale gas sweet-spot selection and well placement optimization in the region.

Automated machine learning (AutoML) methods, particularly AutoGluon, have demonstrated effectiveness for lithology identification by reducing reliance on expert experience through automated parameter determination. However, AutoGluon relies on a fixed set of base learners and a rigid two-layer structure. This paper proposes an AutoML stacking method (ASM) that integrates adaptive base learner selection and adaptive deep multi-layer stacking, providing a novel AutoML framework for lithology identification. Specifically, adaptive base learner selection is achieved through forward or backward search strategies, which are built on the statistical principles of stepwise regression. Adaptive deep multi-layer stacking is automatically optimized via feedforward propagation coupled with validity comparison. To validate the proposed method, experiments were conducted using lithology datasets from the Zagros Basin in the Middle East. The results demonstrated that the forward search strategy surpassed the backward strategy in both accuracy and efficiency. The ASM variant, ASM-B2, leveraging forward-adaptive base learner selection and a two-layer stacking architecture, improved the F1-score by 1.7

Sedimentary microfacies interpretation is fundamental to reservoir heterogeneity analysis but remains challenged by complex depositional geometries and geological consistency requirements. Generative models can effectively capture facies patterns, yet they suffer from limited stability, low inference efficiency, and insufficient integration of geological prior knowledge. To address these limitations, a knowledge-data hybrid generative pretrained model, termed HyFac, is proposed for sedimentary microfacies interpretation. HyFac separates unconditional diffusion pretraining from conditional generation. The diffusion model is first pretrained on sedimentary microfacies maps to learn intrinsic depositional representations. Geological priors, including constraints from well facies and sandstone thickness trends, are then fused through an attention-driven conditional fusion network and subsequently refined by few-step denoising diffusion implicit model (DDIM) sampling with the pretrained diffusion parameters frozen. Knowledge-guided multi-directional variogram and sandstone thickness trend constraints are further introduced to reinforce anisotropic spatial continuity and depositional trend consistency. To validate the proposed method, experiments were conducted using sedimentary microfacies datasets from the Songliao Basin together with open-source alluvial fan and braided river datasets. Unconditional diffusion pretraining yields more continuous and less fragmented facies representations, providing a reliable basis for conditional generation. Compared with conditional diffusion models and generative adversarial networks, HyFac achieves efficient few-step DDIM sampling while better preserving ground truth feature-space distributions, variogram characteristics, and geological consistency under well and sandstone thickness constraints. Attention-driven conditional fusion further improves reconstruction relative to direct feature concatenation. The results demonstrate that knowledge-guided geological priors effectively complement data-driven generative learning, enabling stable sedimentary microfacies interpretation across heterogeneous depositional systems.

The fault–fracture system plays a crucial role in the migration, accumulation, and preservation of shale oil. However, weak geophysical response of these fractures and the ambiguity in picking faults and natural fractures make their identification and recognition challenging. To address these issues, this paper proposes a set of methods for predicting the distribution of fault–fracture systems based on graph neural networks. The method set consists of three parts: (1) a logging-based fracture identification method via graph pooling and global graph construction, which uses logging data sampling points to create small-scale graphs that simulate fracture distribution in the wellbore and integrates lithology, stratigraphy, and mechanical stratigraphy information into the global graph structure and fracture identification model through depth sequences, spatial distance calculations, and global graph convolution; (2) a method combining semantic segmentation with graph convolution, which adds fault graph structures and graph computation modules to the U-Net network to expand the receptive field and clarify segmentation boundaries of the model, and constructs a joint loss function to simultaneously train and optimize the fault segmentation and fault continuity analysis modules; (3) a fracture intensity prediction method combining fault damage zone graph construction with ensemble learning, which effectively integrates the controlling effect of faults on fractures into the inter-well fracture distribution prediction model through information transfer between vertices. The proposed method set was applied to the Fengcheng Formation in the Mabei area of the Junggar Basin, achieving fracture identification accuracy of 91.96% through conventional logging data, a fault detection accuracy of 97.38%, and a mean absolute error of tectonic fracture distribution prediction of 9.38%, representing a 22.47% improvement over existing technologies. Overall, the research findings provide a new approach for predicting the distribution of fault–fracture systems in shale oil reservoirs.

Abstract: Natural fractures in tight sandstone reservoirs play an important role in hydrocarbon migration and accumulation. Fracture identification remains challenging due to the scarcity of labeled data and the complex logging responses of fractures. To address these problems, we propose a novel hybrid deep learning framework (CNN-Attention-BiLSTM). First, labeled fracture classification based on Full waveform sonic logs characteristics is employed to screen unlabeled data, replacing sampling algorithms for data balancing. This approach provides more fracture labels that align with authentic geological information. Subsequently, one-dimensional convolution is applied to construct multi-dimensional fracture logging response patterns that characterize fracture development. A Channel Self-Attention is introduced to assign optimal weights to response patterns across different dimensions, achieving an optimized pattern combination. A double-layer BiLSTM is then utilized to mitigate the impact of sedimentary cycles on logging identification, while capturing both short- and long-term dependencies of fracture responses across different network layers. The identification method is applied to the H1 member of the Lower Shihezi Formation in the Hangjinqi area, China. The test set accuracy is higher than 90%, and blind wells verification demonstrates an improvement of over 8% in accuracy compared to conventional methods. The identification results reveal that fractures are the most developed in H1-2, followed by H1-1 and H1-3, while H1-4 is the least developed layer. The fracture distribution pattern is evidently controlled by sedimentary rhythms, with fracture density decreasing in the order: interbedded sandstone and mudstone layers, thick sandstone and thick mudstone, thick mudstone and poorly developed sandstones. This trend is primarily attributed to the thickness of mechanical stratigraphy. Under equivalent tectonic stress conditions, thin sandstone layers are more prone to fracturing due to stress concentration, resulting in higher fracture density. Additionally, the proposed method deepens the correlation between the log response types of fractures and their development degree. It clarifies that fractures occur in varied patterns across different regions. In sandy conglomerates and gravel coarse sandstone intervals with high porosity and permeability, fractures tend to occur as single or multiple parallel fractures and are relatively less developed. fractures are more prevalent in the overlying and underlying intervals. Conversely, in tight sandstone intervals with poor porosity and permeability, the rock is more brittle, leading to the development of dense, interconnected fracture networks. And gas distribution shows correlate strongly with fracture-developed intervals. Therefore, it can be inferred that in intervals with high-quality sandstone reservoirs in the study area, fractures likely serve as vertical conduits connecting upper and lower gas-bearing zones, acting as preferential migration pathways. In contrast, within tight sandstone intervals, fractures primarily enhance matrix reservoir quality, thereby facilitating gas migration and accumulation. The intelligent fracture identification method proposed in this study can provide guidance for the migration, accumulation and efficient development of tight sandstone gas, Further, it can also offer a basis for the later carbon dioxide storage and the construction of underground gas storage of tight sandstone.Keywords: Fracture identification; Tight reservoirs; Full waveform sonic logs; Conventional logs; Deep learning

Natural fractures play an important role in controlling fluid flow and production performance, making accurate fracture zone prediction essential for petroleum exploration and development. However, prediction of fracture zones using seismic data remains difficult because fracture-related seismic responses are weak and labelled samples from wells are limited. To overcome these limitations, this study proposes a semi-supervised ladder neural network (SSLNN) for fracture zone prediction in the Asmari Formation of the A Oilfield, Iraq. Fracture development was identified from well logs, and fracture-sensitive seismic attributes were screened using support vector machine-based sensitivity analysis. Five attributes, namely variance, dip angle, curvature, azimuth angle, and dip deviation, were selected as the optimal input features. By integrating limited labelled data with abundant unlabelled seismic attribute data, the SSLNN effectively captures the nonlinear relationship between seismic responses and fracture development. The results show that the proposed method achieves an average test accuracy of 87.5%, exceeding that of the best-performing supervised model by 6.04%. Parameter analysis further indicates that model accuracy first increases, then decreases, and finally stabilises as the weight ratio between supervised and unsupervised losses increases. The prediction results reveal that fractures are more developed in Layer A than in Layer B, are stronger in the southern part of Layer A, and are mainly concentrated near structural highs and in the hanging walls of faults. These results demonstrate that the SSLNN is an effective method for fracture prediction in carbonate reservoirs with sparse labelled data.

Deep continental shale reservoirs of the Fengcheng Formation in the northern Mahu Sag have considerable hydrocarbon potential; however, their extremely low matrix porosity and permeability make reservoir quality highly dependent on fault-controlled tectonic fractures. In this study, core observations, thin section analyses, well logs, and seismic data are integrated with machine-learning-based fault and fracture prediction methods to quantitatively evaluate the control of faults on fracture development and to characterize fault-controlled fracture patterns under different structural styles. The results show that fracture development is strongly controlled by fault activity and fault damage zones. Both the cumulative width of fault damage zones and fracture intensity in the thrust nappe belt are significantly greater than those in the slope zone. Tectonic fractures are much more intensely developed within fault damage zones than in undamaged zones, and fracture intensity generally decreases with increasing distance from fault cores, whereas average fracture intensity is positively correlated with the width of fault damage zones. These relationships indicate that fault-related stress concentration is the primary mechanism governing fracture distribution and heterogeneity. Distinct structural styles exert a first-order control on the geometry and spatial organization of fault-controlled fracture belts. In the thrust nappe belt, imbricate and imbricate–backthrust structures are characterized by dense fault arrays, strong folding, and wide damage zones, resulting in high-intensity and spatially complex fracture networks. In contrast, the slope zone is dominated by simpler backthrust structures with fewer and smaller faults, where fracture development is mainly localized along individual faults and associated with narrower damage zones. The fault–fracture system experienced two main stages of tectonic development, which provide the temporal framework for the observed fracture patterns. NW–SE compression during the Late Permian generated NE–SW-trending thrust faults and associated fault-related fractures, whereas nearly S–N-oriented compression during the Late Triassic produced E–W-trending thrust faults and regionally oriented fractures. These findings provide a structural geological basis for identifying fracture-favorable zones and optimizing exploration strategies in deep continental shale oil systems.

The Tarim Basin, one of the few remaining underexplored large sedimentary basins worldwide, hosts abundant ultra-deep hydrocarbon resources. Among its key structural elements, strike-slip faults exert a fundamental control on the development of carbonate fracture-cavity reservoirs and the accumulation of hydrocarbons. Small- to medium-scale strike-slip faults exhibit pronounced segmentation in map view, comprising restraining stepovers, releasing stepovers, and pure strike-slip segments, while cross sections reveal stratified deformation expressed as sub-vertical shear zones, positive flower structures, and negative flower structures. The timing and intensity of strike-slip faulting vary across regions, with major episodes occurring during the Caledonian and Hercynian orogenies. Multiple reactivations of strike-slip faults, in concert with fluid-rock interactions, jointly governed the evolution, effectiveness, and heterogeneity of fracture-cavity reservoirs. Meteoric dissolution during the Middle Caledonian enhanced reservoir porosity, whereas marine cementation during the Late Caledonian exerted a destructive influence. Subsequently, hydrothermal alteration in the Late Hercynian further increased reservoir heterogeneity. Reservoir distribution differs among structural domains: on the Tabei and Tazhong Uplifts, stratiform reservoirs are primarily controlled by unconformities and paleogeomorphic surfaces, while vertically oriented fracture-cavity reservoirs are fault-controlled; in the Aman Transition Zone, reservoirs occur in a grid-like pattern along strike-slip faults, reflecting strong structural compartmentalization. Three main hydrocarbon charging events occurred during the Late Caledonian, Late Hercynian, and Late Himalayan periods. The spatial relationship between strike-slip faults and the Middle Cambrian gypsum layer critically governs hydrocarbon migration and productivity. Faults that penetrate the salt layer and maintain significant displacement provide efficient migration conduits and storage sites. Overall, the hierarchical rank, orientation, and segmentation of strike-slip faults collectively determine the spatial distribution and enrichment of hydrocarbons within the Tarim Basin.

To address the challenges of three-dimensional reservoir modeling with varying amounts of well data, this study introduces two image processing-inspired generative adversarial networks (GAN): three-dimensional handwritten digit generation GAN (3D-HDG-GAN) and image inpainting GAN (II-GAN). The 3D-HDG-GAN is particularly effective for reservoirs with limited well data, as it leverages three-dimensional convolution operations to learn spatial geological structures. This approach results in an accuracy improvement exceeding 10

Lithology identification underpins reservoir characterization and plays a pivotal role in hydrocarbon exploration. Considering vertical stratigraphic continuity is particularly important, as it reflects sedimentary patterns and enhances the reliability of lithology prediction. Existing methods based on statistical identifiability and K-Means clustering cannot fully capture the stratigraphic continuity. To address these limitations, a novel Simple Linear Iterative Clustering (SLIC) with Transformer (SLICTF) method is proposed that jointly captures vertical stratigraphic continuity and dependencies among petrophysical features in well logs to improve lithology identification. The SLIC algorithm is designed for well logs to extract stratigraphically coherent features that preserve continuity across transitional lithologic intervals. SLIC segmentation features are then fused with the original well logs and fed into Transformer model. It explicitly learns feature interactions. To validate the proposed method, experiments were conducted using complex lithological datasets from the Zagros Basin in Middle East. Empirical results demonstrate that SLICTF substantially improves lithology classification, achieving up to 92.5 % and outperforming K-Means based approaches by up to 4.1 %. By capturing vertical stratigraphic features and feature dependencies, SLICTF with attention heads matched to input feature dimensionality enhances discrimination among complex carbonate and mudstone lithologies. SLIC accelerates feature characterization by reducing mean computational time by 82.7 % and stabilizes Transformer accuracy four iterations earlier than K-Means. The method demonstrates reliability for engineering applications.

Natural fractures in tight sandstone reservoirs play an important role in hydrocarbon migration and accumulation. Fracture identification remains challenging due to the scarcity of labeled data and the complex logging responses of fractures. To address these problems, we propose a novel hybrid deep learning framework (CNN-Attention-BiLSTM) for labeled data balancing. First, labeled fracture classification based on full waveform sonic logs (FWS) characteristics is employed to screen unlabeled data, replacing sampling algorithms for data balancing. This approach provides conventional logging with more fracture labels that align with authentic geological information, thereby enhancing the reliability of fracture labels. Subsequently, one-dimensional convolution is applied to construct multi-dimensional fracture logging response patterns that characterize fracture development. A Channel Self-Attention (CSA) mechanism is introduced to assign optimal weights to response patterns across different dimensions, achieving an optimized pattern combination and thereby offering clearer response pattern guidance for subsequent identification models. A double-layer BiLSTM (DL-BiLSTM) is then utilized to mitigate the impact of sedimentary cycles on logging identification, while capturing both short- and long-term dependencies of fracture responses across different network layers. Ultimately, intelligent fracture identification is realized. The identification method is applied to the H1 member of the Lower Shihezi Formation in the Hangjinqi area, China. The test set accuracy is higher than 90%, and blind wells verification demonstrates an improvement of over 8% in accuracy compared to conventional methods. The identification results reveal that fractures are the most developed in H1-2 interval, followed by H1-1 and H1-3 intervals, while H1-4 interval is the least developed. The fracture distribution pattern is evidently controlled by both sedimentary rhythms and reservoir properties, resulting in complex storage and flow capabilities. The findings can provide guidance for the migration, accumulation and efficient development of tight sandstone gas.

In this study, the characteristics of different types of natural fractures were characterized through an integrated analysis of core, thin sections, imaging logs and laboratory tests. The main controlling factors and distribution patterns of natural fractures in lacustrine shale reservoirs were systematically investigated. The results indicate the presence of two main fracture types: structural fractures and diagenetic fractures. Structural fractures are predominantly high-angle to sub-vertical, with large scale and well-defined features. In contrast, diagenetic fractures are mostly sub-horizontal, with irregular and curved fracture surfaces. The development of structural fractures in these continental shales is controlled by multiple factors, including lithology, mineral composition and the mechanical stratigraphy. Among these, lithology and mineral composition form the material basis for fracturing, with brittle mineral content showing a significant positive correlation with structural fracture density. The mechanical stratigraphy plays a critical role in regulating the distribution and vertical extension of fractures. Within a mechanical unit of relatively uniform lithology, greater layer thickness correlates with larger fracture length, lower fracture density and increased vertical connectivity. The Palaeocene Second Member of the Funing Formation (E1 f 2) within the Qintong Sag, the interplay between sedimentation-controlled lithology and rock structure results in a distinct vertical zonation from Sub-members I to V: fracture types become simpler, average fracture length increases, while fracture density and filling degree decrease. This study provides valuable insights for understanding fracture distribution in lacustrine shale reservoirs and offers guidance for hydrocarbon exploration and development.

3D point clouds derived from digital photogrammetry have become a key data source for characterizing outcrop fracture networks and for constraining subsurface reservoir models and structural interpretations. In practice, however, low-texture rock surfaces and noisy surface normals frequently degrade the reconstructed point clouds, leading to incomplete and unreliable fracture extraction and limiting their effective use in these applications. This study aims to develop and evaluate a robust, largely automatic workflow, Im2Frac, for generating dense 3D point clouds from multi-angle 2D photographs and extracting individual fracture surfaces together with their geometric attributes under such challenging conditions. Im2Frac integrates Structure-from-Motion with a Multi-View Stereo Network (SfM-MVSNet) for dense point-cloud reconstruction, and a CIE L & lowast;a & lowast;b & lowast; (CIELAB) space density clustering algorithm (Lab-DBSCAN) for fracture delineation. SfM-MVSNet combines multi-view geometry and deep learning to produce high-fidelity depth maps, whereas Lab-DBSCAN transforms point normals into perceptually uniform CIELAB color space, groups points with similar orientations and applies DBSCAN to segment individual fracture planes. The workflow is tested on drone images of tight-sandstone outcrops from the Chang 6 Member in the Yan River area, Ordos Basin, China. Im2Frac reconstructs dense point clouds with an average reprojection error of similar to 1.45 pixels and recovers 96.1% of mapped fractures, with close agreement between automatic and manual statistics of fracture strike and dip. Compared with conventional Multi-View Stereo (MVS) or geometry-driven clustering approaches, the proposed workflow improves robustness in low-texture areas, reduces sensitivity to parameter choices and delivers fracture attributes with minimal human interaction, thereby enhancing the practical value of photogrammetric outcrop data for quantitative fracture characterization. Further work will focus on improving robustness to varying illumination and acquisition conditions and on validating the workflow in more heterogeneous lithologies.

The Tarim craton hosts an extensive network of ultra-deep strike-slip systems that have exerted primary control on reservoir architecture and hydrocarbon migration pathways. In this study, we integrate geological and geophysical methods to resolve the geometry, kinematics, and temporal evolution of these faults within the Tarim craton. In the Tabei area, conjugate NE- and NW-striking strike-slip faults are well developed; the eastern Shuntuoguole area is dominated by NE-striking faults, whereas NW-striking faults prevail in the western Shuntuoguole area. In map views, faults comprise pure strike-slip segments punctuated by contractional stepovers and extensional stepovers, with en echelon normal fault splays further deforming the overlying stratigraphy. Cross-sectional profiles reveal typical composite flower structures, indicating that progressive slip along deep-seated strike-slip faults was transferred upward through strain accommodation, thereby affecting the overlying strata and reactivating pre-existing faults. Fault activity occurred during the Middle Caledonian, Late Caledonian, and Late Hercynian stages. The F5 fault exemplifies progressive segment linkage and hybridization, evolving from isolated strands into a throughgoing strike-slip corridor. Initial transpressional faults formed during the Ordovician collision between the Tarim Block and the South Kunlun Terrane. Subsequent Silurian back-arc extension in the South Tianshan region and collision between the Tarim and Qaidam blocks led to reactivation of subvertical basement strike-slip faults. During the Devonian, the opening of the eastern branch of the Paleo-Tethys Ocean, followed by Permian lateral extrusion associated with final convergence between the Tarim Block and the Tianshan Orogen, further reactivated these fault systems. The evolution of the strike-slip fault system was controlled by Paleozoic plate convergence, subduction rollback, and subsequent intracontinental lateral extrusion, driving multi-stage fault reactivation and segment linkage within the Tarim craton.

Lithology identification in complex carbonate reservoirs is essential for high-resolution characterization and quantitative assessment, prompting the introduction of an advanced Deep Forest (DF) algorithm to improve recognition accuracy. The DF method leverages the advantages of Random Forests (RF) with deep neural network architectures, preserving RF's robust generalization and efficiency with small labeled datasets. A cascading neural network architecture further improves RF's capacity to extract nonlinear features crucial for differentiating between various lithologies. Furthermore, DF automatically determines the network depth, eliminating the need for manual hyperparameter tuning common in neural networks. To validate this approach, a dataset featuring eight lithologies from the Paleogene to Neogene Miocene formations in the B oilfield of the Zagros Basin was employed. The lithology label of each sample was characterized by nine conventional logging features, including gamma ray (GR), compensated density (DEN), acoustic (AC), compensated neutron (CNL), caliper (CAL), photoelectric absorption cross-section index (PE), invasion zone resistivity (RXO), deep lateral resistivity (RD), and shallow lateral resistivity (RS). Comparative analyses demonstrated that DF outperforms traditional RF, especially in recognizing thin and transitional lithologies, achieving nearly a 2% increase in accuracy and consistently exceeding 91% accuracy with default parameters across various sample sizes.

In the context of complex tectonic evolution, due to the control of tectonic compression stress and faults on tectonic fractures, the formation and development of tectonic fractures in the T3x2 tight reservoirs present significant variations across different tectonic segments in the Western Sichuan Foreland Basin. We clarified the control of differential tectonic evolution on the formation and development of tectonic fractures in different tectonic segments through field-based observations, core samples, image logging, as well as fluid inclusion petrography and temperature determinations of fracture-filling materials, combined with 2D balanced cross-section restoration. The study area primarily manifests two types of tectonic fractures in the tight reservoirs: orogen-related fractures (regional fractures) and fault-related fractures. The orientations of these fractures are predominantly E-W, nearly N-S, NE, and NW. Specifically, the northern segment area only shows the development of regional fractures, while the southern and middle segments exhibit the development of both regional and tectonic fractures. There are three phases of tectonic fractures in different tectonic segments, and their formation times are relatively consistent. The Mesozoic tectonic events had a significant impact on the northern and central segments, with the amount of tectonic shortening and the rate of stratigraphic shortening gradually decreasing from the northeast to the southwest. The compressional stress resulting from tectonic compression also decreases from the northeast to the southwest. As a result, the development of first-phase and second-phase tectonic shear fractures is more pronounced in the northern and middle segments compared to the southern segment. Under the significant control of faults, the development of N-S- and NE-oriented fault-related fractures is more pronounced in the southern segment, while the development of NE-oriented fault-related fractures is relatively higher in the middle segment. Overall, there is an increased density of fractures and an increasing trend in fracture scale from the northern to the middle and then to the southern segment.

Current fault detection methods mainly take advantage of a convolutional neural network, simplified U-Net, for seismic image semantic segmentation, which is a computer vision task aimed at generating a dense pixel-wise segmentation map of an image, where each pixel is assigned to a specific class or object. However, these methods face challenges such as unclear segmentation boundaries and limited receptive fields, which hinder the model's capability to detect all features of the fault. To address these issues, this paper proposes a method based on the integration of Graph Neural Network (GNN) and U-Net, referred to as GNU-Net, which incorporates fault graph structures and graph computation modules into U-Net. This approach aims to expand the receptive field, clarify segmentation boundaries, and optimize fault segmentation and continuity analysis through a joint loss function. The method consists of five modules: graph construction, data augmentation, fault segmentation, fault continuity analysis, and joint training. The graph construction module grids the fault labels with a grid edge length of 8 voxels, sets vertices at the locations where the fault intersects the grid, and selects representatives from the voxels labeled as faults within each grid as vertices. The dilation function is applied to expand the fault voxels in the labels and construct a velocity field, in which the geodesic distance between vertices is computed to establish connections between vertices. Data augmentation of the 3D seismic data and fault labels is achieved by rotating the images. A custom rotation matrix is applied to transform the vertex feature coordinates of the graph structure, significantly increasing the size and diversity of the training dataset. The fault segmentation module employs a simplified U-Net with square convolution kernels, and the segmentation loss includes both Dice and binary cross-entropy losses. The fault continuity analysis module leverages a graph neural network with irregular convolution kernels to capture macroscopic fault features and enhance fault continuity, employing a connection loss based on binary cross-entropy. Compared to the original U-Net, the modified GNU-Net achieves a fault detection accuracy of 97.39 % on the testing set, an improvement of 3.96 % over the original U-Net.

Shallow-water deltas are not only a hot spot for sedimentological research but also a key target for oil and gas exploration. In this paper, taking the third member (E1f3) of the Funing Formation in the Upper Jurassic as an example, based on observations made from core samples, well logging, cathode luminescence characteristics, and analytical assays, the development conditions, sedimentary characteristics, and sedimentary models of shallow-water deltas are summarized. These shallow-water deltas were deposited in conditions with the following characteristics: a gentle terrain platform, a subtropical climate with ample rainfall, an abundant source supply, strong hydrodynamic forces, shallow water bodies, and a frequently eustatic lake level. Shallow-water deltas are characterized by sediment deposition from traction currents, numerous underwater distributary channel scour structures, overlapping scouring structures, sand body distribution with planar features, underwater distributary channels as skeletal sand bodies, and undeveloped mouth bars. Based on these, it is believed that during the deposition period of E1f3, the Gaoyou Sag in the Subei Basin had favorable geological conditions for the development of shallow-water delta deposition. The shallow-water delta deposition that occurred during the sedimentary periods of the five major sand units in the Funing Formation is characterized by front subfacies, with underwater distributary channels as the framework for sand bodies, and multiple intermittent positive rhythms overlapping vertically with the Jianhu Uplift as the source of material supply. In this paper, a depositional model for shallow-water delta deposition during the E1f3 deposition period in the Gaoyou Sag is established, expanding the scope of oil reservoir exploration in the north slope region of the Gaoyou Sag and providing important geological evidence for the selection of favorable subtle zones.