A conditional generative hydrate digital core construction method

Abstract

This invention relates to the interdisciplinary field of natural gas hydrate resource exploration and development and digital rock physics, and discloses a conditionally generated digital core construction method for hydrates. The method includes the following steps: performing three-phase segmentation on CT images of hydrate-bearing sediments to obtain labeled data for the skeleton, pore fluid, and hydrates; identifying and labeling the microscopic occurrence types of hydrates based on the spatial characteristics of hydrate clusters and the skeleton, constructing a labeled dataset containing type labels and measured saturation; extracting the microscopic spatial distribution characteristics of hydrates through an encoder to establish a feature database; training to obtain a saturation-type mapping function, constructing a conditional generative adversarial network, and using the target saturation and feature vector as conditional inputs. After training, inputting the target saturation can generate a three-dimensional digital core of hydrates. This invention provides a conditionally generated digital core construction method for hydrates, which can solve the long-standing problem of the disconnect between hydrate characterization and hydrate sediment modeling.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of natural gas hydrate resource exploration and development and digital rock physics, and particularly to a condition-generated hydrate digital core construction method. Background Technology

[0002] Natural gas hydrates, as one of the most promising clean energy sources of the 21st century, present a critical challenge in the energy sector for safe and efficient extraction. The key to achieving this goal lies in accurately revealing and predicting the intrinsic relationship between the microstructure and macroscopic physical properties (such as seepage, acoustic, electrical, and mechanical properties) of hydrate reservoirs. Driven by this need, digital core technology has emerged and rapidly developed into a bridge connecting microscopic characterization and macroscopic property simulation. However, for the unique and complex object of hydrate sediments, existing technologies still face a series of theoretical and technical bottlenecks in constructing digital core models that can realistically reflect their microscopic structure and accurately respond to changes in saturation.

[0003] Digital core technology provides a revolutionary means to study physical processes at the pore scale by reconstructing and numerically simulating high-resolution CT scan images in three dimensions. This technology has evolved from early simple three-dimensional visualization to today's quantitative analysis of microstructures and multiphysics coupling simulations. In the conventional oil and gas sector, digital core technology is relatively mature and has spawned core patents such as "A method for segmenting pores and throats in three-dimensional core images" (ZL201510170889.2) and "Super-resolution reconstruction of three-dimensional CT core images" (ZL201510388670.X).

[0004] However, when the technology is applied to the study of hydrate sediments, their complexity and unique characteristics pose significant challenges to traditional digital core methods. Hydrates are not static sedimentary skeletons but exist in various forms within the pore spaces of sediments, and their occurrence types have a substantial impact on reservoir properties. For example, in hydrate-bearing sediments in the Shenhu area of ​​the South China Sea, hydrates tend to occupy the interior of foraminiferal shells and block connecting throats, significantly altering permeability. Existing general digital core construction methods, whether based on physical experiments (such as CT scan reconstruction) or numerical reconstruction (such as process simulation and stochastic methods), struggle to systematically and quantitatively integrate prior knowledge of these microscopic occurrence characteristics. Most methods focus on modeling pores and the sedimentary skeleton, or simplify hydrates as homogeneous pore-filling materials, neglecting the differentiated control mechanisms of their diverse occurrence types on multiple physical fields such as acoustics, electricity, force, and permeability, leading to significant biases in the simulation and prediction of the constructed models.

[0005] In recent years, advanced imaging technologies such as microfocal CT and cryo-scanning electron microscopy (Cryo-SEM) have made significant progress, providing direct means to observe the spatial distribution of hydrates. To address core challenges such as insufficient resolution in CT scan images, researchers have developed image processing techniques based on deep learning. For example, super-resolution reconstruction algorithms can improve the spatial resolution of CT scan images by 2 to 4 times, thus more clearly identifying hydrate boundaries; 3D U-Net combined with deep segmentation networks such as attention mechanisms can improve the segmentation accuracy of the hydrate-water-sediment three-phase boundary to the sub-micrometer level. These technologies provide a higher-quality data foundation for microscopic studies. Nevertheless, current characterization work still suffers from two key shortcomings: first, a disconnect between characterization and modeling. Most studies remain at the level of qualitative description and static quantitative statistics for single or a few samples. For example, studies can observe the redistribution pattern of hydrates "from bottom to top and from center to side" during depressurization mining, or automatically classify hydrate types and their proportions using self-organizing map (SOM) neural networks. However, a standardized process has not yet been established to extract the spatial distribution patterns of hydrates observed in massive images (such as coexistence patterns of different morphologies and spatial correlations) into a quantifiable and portable feature database to directly drive or constrain the generation of digital cores. Secondly, there is the subjectivity and one-sidedness of type classification. Although there is a consensus on the classification of hydrate types, in practice, it heavily relies on researchers' experience and judgment, lacking objective, automated, and repeatable classification and labeling standards based on the fusion of multiple features such as image grayscale, shape, and topological relationships. This makes it difficult to compare and integrate different research results.

[0006] Hydrate saturation is a core parameter for reservoir evaluation and exploitation design. Existing technologies can obtain total saturation or identify local types relatively independently through well logging and experiments, but studies that systematically and quantitatively correlate these two factors are extremely scarce. In other words, it remains unclear how the relative proportions of different hydrate types change as total hydrate saturation increases. This quantitative mapping relationship between hydrate saturation and type distribution is key to understanding the mechanism of hydrate formation and evolution and predicting dynamic changes in physical properties during exploitation. However, there is currently a lack of a method to automatically establish and validate this crucial relationship model based on large amounts of sample data. This makes existing digital core models often static representations, unable to reflect the control effect of saturation, a key state parameter, on microstructure, and difficult to predict how microstructure evolves under exploitation disturbances and how it reacts to macroscopic physical properties.

[0007] However, a comprehensive analysis of existing technologies reveals that current technologies primarily focus on the standardized preparation of front-end physical models and specific back-end physical property simulation algorithms. A crucial, pivotal link is lacking. Existing technologies mainly include: Step 1, relying on low-temperature, high-pressure chambers to perform CT scans on finitely pressurized core samples, resulting in high sample costs, limited data volume, and insufficient coverage of saturation ranges, leading to data starvation and difficulty in learning the statistical regularities of microstructures; Step 2, using thresholding methods and conventional machine learning for phase segmentation, which can only identify hydrate phases but cannot automatically and precisely quantify their occurrence types, losing key typological information; Step 3, directly reconstructing or simply simulating modeling the segmented images without using saturation as a control variable, failing to synthesize the hydrate microstructure distribution based on the target saturation, resulting in models that are merely replicas of specific samples rather than flexible digital twins; Step 4, although physical property simulation algorithms are mature, the upstream model fails to reflect the diversity of hydrate occurrence types and the quantitative relationship with saturation, leading to fundamental biases in the input basis and making it difficult for simulation results to accurately reflect the dynamic changes in reservoir exploitation. Overall, the existing algorithmic step chain has a break between "microscopic feature quantification" and "macroscopic condition constraints", and cannot learn the "saturation-microstructure" law and synthesize three-dimensional core structures. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a condition-generated digital core construction method for hydrates, which can solve the long-standing problem of disconnect between hydrate characterization and hydrate sediment modeling.

[0009] This invention provides a method for constructing conditionally generated hydrate digital cores, comprising the following steps: Preprocessing and three-phase segmentation of CT scan images of hydrate-bearing sediments were performed to obtain three-phase tag data of solid sediment skeleton, pore fluid and hydrate; Based on three-phase tag data, the microscopic occurrence type of hydrates is automatically identified and labeled by calculating the spatial geometric and topological characteristics of hydrate clusters and surrounding sediment skeleton particles. Based on the labeling results, a labeled dataset containing 3D images, microscopic storage type labels, and measured total saturation is constructed. We construct and train a 3D convolutional autoencoder, use the trained autoencoder to extract feature vectors representing the microscopic spatial distribution patterns of hydrates from the labeled dataset, and establish a microstructural feature database based on the feature vectors, which includes associated features, macroscopic attributes, and location information. Based on the labeled dataset, the data relationship between the total saturation of hydrates and the volume content of each microscopic occurrence type was statistically analyzed. Using total saturation as input and the relative content of each microscopic occurrence type as output, a mapping function is trained to establish a quantitative relationship model between hydrate saturation and microscopic occurrence type. A conditional generative adversarial network is constructed, whose generator input includes random noise and a conditional vector consisting of the target saturation value and feature vectors retrieved from a microstructure feature database; Based on the input target hydrate saturation value, the expected microstructure content is determined using a quantitative relationship model. Feature vectors are retrieved from the microstructure feature database. The retrieval results and the target saturation value are used together as conditional vectors and input into a trained conditional generative adversarial network to generate a three-dimensional hydrate digital core that satisfies the target saturation constraint and whose microstructure conforms to geological laws.

[0010] Specifically, the micro-occurrence types include particle cementation type, particle encapsulation type, and pore filling type; the characteristic parameters include: the number of different sedimentary skeleton particles directly adjacent to the hydrate cluster, denoted as the contact particle number N_g, the ratio of the surface area of ​​the hydrate cluster to the surface area of ​​the contacted sedimentary skeleton particles, denoted as the encapsulation index C_i, and the shape factor describing the macroscopic morphology of the hydrate cluster.

[0011] Specifically, feature parameters are identified based on preset thresholds and rule sets to automatically assign type labels: When N_g≥2 and the shape factor indicates that it is in a narrow region, it is determined to be a particle cementation type; When N_g=1 and C_i is greater than the first threshold, it is determined to be a particle-encapsulated type; When N_g=0, or N_g≤1 and C_i is less than the second threshold, it is determined to be a pore-filling type.

[0012] Specifically, the three-phase segmentation is performed using the 3DU-Net deep learning network with an embedded channel attention mechanism.

[0013] Specifically, the input to the 3D convolutional autoencoder is a multi-channel local 3D image block containing a uniquely thermally encoded channel for sediment skeleton / pore fluid and various types of hydrates.

[0014] Specifically, the mapping function is implemented through a multi-objective regression model, such as support vector regression, gradient boosting decision tree, or shallow neural network.

[0015] Specifically, the generator of the conditional generative adversarial network adopts a conditional generative network with 3DU-Net as the backbone, and injects conditional vectors through a spatial adaptive normalization module; the discriminator adopts the 3DPatchGAN structure.

[0016] Specifically, the loss functions used when training conditional generative adversarial networks include: adversarial loss, saturation and type content constraint loss. The constraint loss is obtained by inputting the generated image into a pre-trained segmentation and classification network and calculating the difference between its predicted saturation and type content and the target value.

[0017] Specifically, the conditional vector is constructed by concatenating or weighting the expected type content vector calculated by the mapping function based on the target saturation with one or more of the best-matching microstructure feature vectors retrieved from the microstructure feature database based on Euclidean distance.

[0018] Specifically, the first and second thresholds are determined by the following method: constructing a training set containing manually labeled hydrate clusters of various types, calculating the characteristic parameters of each cluster and statistically analyzing the characteristic distribution of each category, using receiver operating characteristic curve analysis, and selecting the value that maximizes the Youden index as the optimal threshold.

[0019] The technical solution provided by this invention has the following advantages compared with the prior art: By systematically mining the microscopic occurrence types in CT images, establishing a quantitative relationship model of "saturation-type distribution," and introducing a conditional generative adversarial network, it achieves a leap from "static replication" to "on-demand intelligent generation." This method automatically identifies and quantifies various microscopic occurrence types of hydrates through intelligent means, constructs a microscopic structure feature database, and transforms the traditional subjective judgment relying on human experience into an automated standard process based on multi-feature fusion, effectively alleviating the problem of insufficient data caused by the scarcity of pressurized core samples; by establishing a quantitative relationship model between the total saturation of hydrates and the content of each microscopic occurrence type for the first time, it reveals the evolution law of hydrate accumulation, enabling the generated digital cores to have "saturation sensitivity," and can... This model can reflect the dynamic evolution of microstructure under different saturation conditions. By utilizing conditional generative adversarial networks, macroscopic saturation constraints and microscopic feature priors are integrated into a generative condition vector. This enables the intelligent synthesis of 3D digital cores that meet both macroscopic parameter requirements and have realistic microscopic structures based on any specified target saturation. This breaks the limitation of traditional methods that can only reconstruct specific samples, providing a high-quality input foundation for subsequent accurate simulations of multiple physical fields such as acoustics, electricity, force, and permeability. Ultimately, by accurately reproducing the diverse occurrence types of hydrates and their quantitative relationship with saturation, the macroscopic physical property simulation based on this model can more accurately reflect the real changes in reservoir characteristics. It effectively corrects the simulation bias caused by the simplification of hydrate morphology in traditional models, providing more reliable theoretical support for hydrate exploitation potential assessment and scheme optimization. Attached Figure Description

[0020] Figure 1 This is a flowchart of a conditionally generated digital core construction method for hydrates. Detailed Implementation

[0021] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0023] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0024] Figure 1 This is a flowchart of a conditionally generated digital core construction method for hydrates.

[0025] like Figure 1As shown, this invention provides a conditionally generated digital core construction method for hydrates, comprising the following steps: Preprocessing and three-phase segmentation of CT scan images of hydrate-bearing sediments to obtain three-phase tag data of the solid sediment skeleton, pore fluid, and hydrates; Based on the three-phase tag data, automatically identifying and labeling the microscopic occurrence type of hydrates by calculating the spatial geometric and topological characteristic parameters of hydrate clusters and surrounding sediment skeleton particles; Based on the labeling results, constructing a labeled dataset containing three-dimensional images, microscopic occurrence type labels, and measured total saturation; Constructing and training a 3D convolutional autoencoder, using the trained autoencoder to extract feature vectors characterizing the microscopic spatial distribution pattern of hydrates from the labeled dataset, and establishing a microscopic structural feature database based on the feature vectors, including associated features, macroscopic attributes, and location information; Based on... A labeled dataset was used to statistically analyze the data relationship between total hydrate saturation and the volume content of each microstructure occurrence type. Using total saturation as input and the relative content of each microstructure occurrence type as output, a mapping function was trained to establish a quantitative relationship model between hydrate saturation and microstructure occurrence types. A conditional generative adversarial network (GAN) was constructed, with its generator input including random noise and a conditional vector consisting of the target saturation value and feature vectors retrieved from a microstructure feature database. Based on the input target hydrate saturation value, the expected content of each microstructure occurrence type was determined according to the quantitative relationship model. Feature vectors were retrieved from the microstructure feature database, and the retrieval results, along with the target saturation value, were used as the conditional vector input to the trained conditional GAN ​​to generate a three-dimensional digital hydrate core that satisfies the target saturation constraint and whose microstructure conforms to geological laws.

[0026] Specifically, the preprocessing and high-precision three-phase segmentation of hydrate sediment CT scan images were performed. First, denoising and contrast equalization were preprocessed on the acquired CT scan images of hydrate-bearing sediments with different saturations. Then, a modified 3DU-Net deep learning segmentation network was used for pixel-level segmentation. This network embeds a channel attention mechanism and residual connections in its encoder-decoder structure, effectively distinguishing regions with similar grayscale values ​​but different textures. The network output classifies each voxel into three phases: solid sediment framework, pore fluid, and hydrate, obtaining high-quality three-phase labeled data.

[0027] Specifically, a microstructure-feature vector database is constructed. A trained feature extractor is used to process the entire labeled dataset D_labeled. For each sample, multiple local image patches are extracted using a sliding window, and their feature vectors are obtained. All feature vectors are associated with and stored along with the location information of their corresponding original image patches and the macroscopic attributes of the sample (total saturation S_h, content of each type C_type), forming a large-scale "hydrate microstructure feature database" DB_feature.

[0028] Specifically, the saturation of the statistical samples and the content of hydrate microstructure types are analyzed. For each sample i in D_labeled, the volume V_type_i of each hydrate phase is accurately calculated based on its type label map L_i, and then its relative content C_type_i = V_type_i / ΣV_type_i * 100% is calculated.

[0029] Model training and digital core generation of hydrate sediments. During training, real three-phase CT images and their corresponding saturation and type content are used as a set of training samples. The generator G(z,c) attempts to generate fake images with condition c. The discriminator D(x,c) not only judges whether image x is real or fake, but also whether image x satisfies condition c (i.e., whether its calculated saturation and type content are consistent with c). Through adversarial training, the generator learns how to synthesize highly realistic 3D images that meet specific microstructural constraints based on the complex condition c. After training, the user only needs to input the target hydrate saturation value S_target, and the system will automatically call F_model and DB_feature to form condition c, driving the generator to generate a brand new, geologically consistent 3D hydrate digital core within seconds to minutes.

[0030] Furthermore, this invention provides a refined classification of hydrate facies, defining at least three microscopic occurrence types with clear geophysical significance: grain-cemented, grain-encapsulated, and pore-filled. To achieve automated identification, this invention proposes a quantification method based on multi-feature fusion. The calculated feature parameters include the number of different sedimentary framework particles directly adjacent to the hydrate cluster, denoted as the contact particle number N_g; the ratio of the surface area of ​​the hydrate cluster to the surface area of ​​the contacting sedimentary framework particles, denoted as the encapsulation index C_i; and a shape factor describing the macroscopic morphology of the hydrate cluster. By transforming these qualitative descriptions into computable geometric and topological features, a quantitative foundation is laid for subsequent automated type labeling.

[0031] Specifically, the micro-occurrence types include particle cementation type, particle encapsulation type, and pore filling type; the characteristic parameters include: the number of different sedimentary skeleton particles directly adjacent to the hydrate cluster, denoted as the contact particle number N_g, the ratio of the surface area of ​​the hydrate cluster to the surface area of ​​the contacted sedimentary skeleton particles, denoted as the encapsulation index C_i, and the shape factor describing the macroscopic morphology of the hydrate cluster.

[0032] After determining the aforementioned characteristic parameters, it is still necessary to address the problem of type classification relying on subjective experience and lacking objective standards. To this end, based on statistical analysis of a large number of known type samples, this invention establishes a set of executable automated identification rules: when N_g ≥ 2 and the shape factor indicates it is in a narrow region, it is determined to be a granular cementation type; when N_g = 1 and C_i is greater than the first threshold α, it is determined to be a granular encapsulation type; when N_g = 0, or N_g ≤ 1 and C_i is less than the second threshold β, it is determined to be a pore-filling type. This rule set frees the type identification process from manual intervention, achieving large-scale, standardized automated labeling, and providing a repeatable technical means for subsequently constructing labeled datasets.

[0033] Specifically, feature parameters are identified based on preset thresholds and rule sets to automatically assign type labels: when N_g≥2 and the shape factor indicates that it is in a narrow region, it is determined to be particle cementation type; when N_g=1 and C_i is greater than the first threshold (α), it is determined to be particle encapsulation type; when N_g=0, or N_g≤1 and C_i is less than the second threshold (β), it is determined to be pore filling type.

[0034] Precise definition of hydrate microstructure occurrence type based on multi-feature fusion. Building upon the three-phase segmentation results, not only are hydrate phases identified, but their internal structures are further subdivided. The definition is primarily based on the spatial geometry and topological relationship between the hydrate connectivity region and the surrounding sedimentary framework, achieved through the calculation of the following set of characteristic parameters: ① Number of contacting particles (N_g): The number of different sedimentary framework particles directly adjacent (surface contact) to the hydrate connectivity region. ② Encapsulation index (C_i): The ratio of the surface area of ​​the hydrate region to the main sedimentary framework particles in contact, quantifying the degree of coverage. ③ Shape factor (S_f): Including elongation (major axis / minor axis), flatness, and sphericity, describing the macroscopic morphology of the hydrate aggregate. ④ Local location characteristics: The position of the hydrate aggregate's centroid relative to the nearest pore throat or intergranular space. Based on the above parameters, the following automated identification rules are established: ① Particle cementation type: N_g≥2, and the shape factor indicates that it is located in a narrow region (such as a throat), its main function being to connect multiple particles. ② Particle-encapsulated type: N_g=1 and C_i>threshold α, the hydrate covers the surface of a single particle like a "coating". ③ Pore-filled type: N_g=0 or (N_g≤1 and C_i<threshold β), the hydrate is suspended in isolation in the center of the pore.

[0035] For a specific hydrate cluster H_k and its only contacting framework particle (when N_g=1), the algorithm calculates the following steps: Locate the contact interface. Traverse all boundary voxels of the hydrate cluster H_k. For each boundary voxel, check its six directly adjacent spatial locations. Discrimination and statistics. If a neighborhood location is marked as a "sediment skeleton" voxel by I_seg, then the hydrate voxel is determined to be in contact with the skeleton at that location. Record this boundary voxel as a "contact voxel". Area calculation. The value of A_contact is equal to the total number of "contact voxels" obtained from the statistics. This is because in regularly meshed volume data, each voxel can be considered as a tiny cube, and its contact surface with adjacent voxels is a small square with a fixed area. The total number of contact voxels is numerically approximately equal to the total surface area of ​​the contact surfaces between the hydrate cluster and the skeleton particles (in units of "voxel faces"). To convert this to actual physical area (e.g., square micrometers), simply multiply this number by the actual physical area of ​​a single voxel face (i.e., the square of the spatial resolution of the CT image).

[0036] Specifically, the threshold used to identify the microscopic occurrence type of hydrates (particle cementation type, particle encapsulation type, pore filling type) was obtained by statistical distribution analysis of a dataset of manually labeled "known type samples".

[0037] First, a "gold standard" training set is constructed, containing a certain number of independent hydrate clusters extracted from CT images. Each cluster has been manually identified as an ideal type by domain experts based on its morphology, location, and interaction with particles. Next, characteristic parameters of all clusters in the set are automatically calculated, including the number of contacting particles N_g, encapsulation index C_i, elongation, and sphericity. Then, statistical analysis of the feature distribution is performed by type.

[0038] It should be noted that the reliability of all subsequent analyses fundamentally depends on the accuracy of the three-phase segmentation. Addressing the technical challenge of similar grayscale and blurred boundaries between hydrates and pore fluids in CT images, this invention employs a 3D U-Net deep learning network with an embedded channel attention mechanism for segmentation. This network introduces a channel attention module into the encoder-decoder structure, enabling the network to adaptively enhance its response to key features distinguishing the three phases, while simultaneously promoting the fusion of shallow detail information and deep semantic information through residual connections. This improvement overcomes the shortcomings of traditional thresholding methods or conventional machine learning algorithms in reliably distinguishing blurred boundaries, ensuring the accuracy of subsequent type recognition, feature extraction, and other analyses from the outset.

[0039] Specifically, in this embodiment of the invention, the improved 3DU-Net deep learning segmentation network achieves high-precision automated segmentation of the three phases—solid sediment skeleton, pore fluid, and hydrate—in CT scan images of hydrate-bearing sediments. Because the X-ray absorption coefficients of each phase in the CT image are similar, the grayscale contrast is low and the boundaries are blurred, making it difficult for traditional thresholding or simple segmentation algorithms to reliably distinguish them. This algorithm automatically extracts deep features through a deep learning model, achieving accurate identification of blurred boundaries and providing an accurate and reliable digital foundation for subsequent quantitative analysis of the microstructure. The input to this algorithm is a preprocessed sequence of three-dimensional CT grayscale images, and the output is three-dimensional segmentation label data of the same size as the input, with each voxel labeled with a category (0-background / skeleton, 1-pore fluid, 2-hydrate). The specific execution steps include: First, training the network using a small number of manually annotated CT image slices to learn three-phase texture and boundary features; during training, a weighted cross-entropy loss function is used to alleviate class imbalance by increasing the weight of hydrate phases, ensuring the recognition accuracy of small target phases; after training, the network is deployed on all CT data for forward inference to obtain preliminary segmentation results; subsequently, connected component analysis is used to remove extremely small isolated regions caused by image noise, and morphological closing operations are used to smooth the segmentation boundaries, finally outputting high-quality segmentation volume data I_seg. This output result serves as the cornerstone of the entire invention process and is directly used in the following two core stages: First, in the precise definition of micro-endowment types, hydrate phases in I_seg are extracted as computational geometric and topological features and then classified into micro-types such as cemented and encapsulated types; Second, in the statistical type content, based on I_seg and its derived type label map, the volume of various hydrates can be accurately counted, providing data support for establishing a saturation-type relationship model.

[0040] Regarding the network improvements, this solution makes two key optimizations based on the standard 3DU-Net architecture to specifically address the unique challenges of hydrate CT image segmentation: First, it embeds a channel attention mechanism, introducing a channel attention structure similar to the SE module into the deep features of the encoder. This allows the network to adaptively calibrate the weights of each feature channel, enhancing its response to key features that distinguish three phases (especially hydrates and porous fluids with similar gray levels), while suppressing unimportant background features. This improves the model's feature selection ability and segmentation accuracy in complex scenes. Second, it introduces residual connections, employing residual structures in each convolutional block of the encoder to alleviate the gradient vanishing problem in deep networks. This promotes the fusion of shallow detail information and deep semantic information, improving the accuracy of category discrimination while maintaining clear contours in the segmentation results. These improvements bring significant benefits: First, the segmentation accuracy is substantially improved, especially in distinguishing the fuzzy boundaries between hydrates and pore fluids, reducing misjudgments and omissions, and ensuring the reliability of subsequent analysis from the source; second, the model robustness is enhanced, showing better generalization ability for CT images with different geological backgrounds and scanning conditions; finally, high-precision automatic segmentation replaces tedious and subjective manual interpretation, making it possible to process massive amounts of CT data and construct large-scale standardized datasets, laying a solid foundation for the "data-driven" and "intelligent" nature of the overall method.

[0041] Specifically, residual connections are located at the front and middle of the 3DU-Net architecture. They mainly address the gradient propagation and training stability issues caused by network depth, ensuring that information and gradients can flow effectively, making the network easier to optimize and thus enabling it to learn features more fully.

[0042] The channel attention module, located at the end of the encoder, acts as an "intelligent feature filter." Building upon the rich features extracted in the preceding layers (including residual blocks), it performs global-level analysis and reweighting, dynamically increasing the contribution of feature channels crucial for segmentation. This is essential for distinguishing between hydrates and pore fluid boundaries that have similar grayscale levels and complex textures.

[0043] After obtaining 3D data with type labels, this embodiment of the invention needs to further address the problem of how to effectively characterize the microscopic spatial distribution patterns of hydrates. To this end, the constructed 3D convolutional autoencoder employs a unique input design: its input is a multi-channel local 3D image patch containing sediment skeleton / pore fluid channels and uniquely thermally encoded channels for various types of hydrates. This design enables the feature extractor to learn not only geometric morphology but also semantic information including the relative positional relationships between different types of hydrates and their configuration with the pore network, rather than just grayscale texture. After training, the encoder can compress complex microstructures into low-dimensional feature vectors, providing a high-quality "structural fingerprint" for the subsequent construction of a microstructure feature database.

[0044] Specifically, the input to the 3D convolutional autoencoder is a multi-channel local 3D image block containing a uniquely thermally encoded channel for sediment skeleton / pore fluid and various types of hydrates.

[0045] An automated hydrate microstructure type labeling and annotation dataset generation method was developed. An automatic hydrate microstructure type labeling algorithm was developed, traversing all connected regions of hydrates, calculating their multi-dimensional features, and automatically assigning type labels (e.g., 1, 2, 3, 4) according to the aforementioned rule set. Finally, a series of "3D image samples with microstructure type labels" was generated, forming the subsequent machine learning annotation dataset D_labeled={I_i,L_i,S_h_i}, where I is the image, L is the type label map, and S_h is the measured total saturation of the sample.

[0046] Specifically, an automatic labeling algorithm for hydrate microstructure types was developed. The processed 3D three-phase segmentation data was converted into binary volumetric data I_seg, where each voxel was labeled as either the skeleton, pore fluid, or hydrate input. The output is a novel 3D labeled volumetric data L_i of the same size as the input. Within the original "hydrate phase" region, different integer values ​​are further labeled according to the type (e.g., 1 represents cementation, 2 represents encapsulation, and 3 represents pore filling), while the skeleton and pore fluid regions retain background values. All samples' I_i, L_i, and their measured saturation S_h_i together constitute the labeled dataset D_labeled.

[0047] Hydrate Microstructure Feature Extraction Based on 3D Convolutional Autoencoder. A deep 3D convolutional autoencoder is constructed as the feature extractor. The encoder consists of multiple 3D convolutional and pooling layers, used to compress a local 3D image patch (containing sediment skeleton, pore fluid, and labeled hydrate types) into a low-dimensional, dense feature vector. This vector potentially encodes various information within the region, including the volume percentage of different hydrate types, spatial distribution patterns, relative positional relationships between different hydrate types, and their configuration with the pore network.

[0048] In a specific embodiment, feature extraction and database construction are based on a 3D convolutional autoencoder. Feature extractor training: A 3D convolutional autoencoder (CAE) is constructed. Encoder: The input is a 64x64x64 image patch (containing multi-channel information obtained from the L_i mapping, e.g., channel 1 is the sediment skeleton / pore fluid, and channels 2-4 are one-hot encoded different types of hydrates). After four layers of 3D convolution (with filter numbers of 32, 64, 128, and 256 respectively) and pooling, it is compressed into a 512-dimensional feature vector. Decoder: Symmetric. The loss function is the mean squared error (MSE) between the input image patch and the reconstructed image patch. The training aims to teach the encoder to extract key features that can effectively reconstruct the original structure. Feature database DB_feature construction: Using the trained encoder, dense sliding window sampling (stride can be set to 32) is performed on all samples in D_labeled to extract the feature vector v of each image patch. Each v is associated with the following metadata stored in a database (such as SQLite or HDF5 file), including the source sample ID and image patch center coordinates, the total volume percentage of hydrates, the volume percentage of various types of hydrates, and the global saturation S_h of the sample to which the image patch belongs.

[0049] After characterizing the microstructure, a multi-objective regression model is trained using the total saturation S_h as input and the relative content [C_cem, C_coat, C_fill] of each microstructure type as output. This model can be implemented using algorithms such as support vector regression, gradient boosting decision trees, or shallow neural networks. Essentially, it formalizes the geological law of "saturation-type distribution" into a computable mathematical function F_model(S_h). Through this mapping function, saturation is transformed from a passive measurement parameter into an active model control variable, making it possible to predict the proportion of corresponding microstructure types based on any given saturation.

[0050] Specifically, the mapping function is implemented through a multi-objective regression model, such as support vector regression, gradient boosting decision tree, or shallow neural network.

[0051] Perform data statistics, iterate through D_labeled, and for each sample i, directly calculate the total volume of each type of hydrate V_cem_i, V_coat_i, and V_fill_i based on its L_i. Calculate relative content C_cem_i=V_cem_i / (V_cem_i+V_coat_i+V_fill_i), And so on. Model selection and training are carried out, using S_h as the independent variable and [C_cem, C_coat, C_fill] as the dependent variable, to construct a multi-output regression model. After training, the mapping function F_model is obtained.

[0052] Train a predictive model from hydrate saturation to hydrate content based on hydrate presence type. Using total saturation (Sh) as input and the relative content of various hydrate types [Ccem cemented hydrate, Ccoat skeleton hydrate, Cfill filled hydrate] as output, construct a multi-objective regression model. This model can be support vector regression, gradient boosting decision tree, or a shallow neural network. Train the model by minimizing the error between the predicted and actual contents to obtain a mapping function:

[0053] [Ccem,Ccoat,Cfill]=Fmodel(Sh), This function, Fmodel, can predict the most likely proportions of various hydrates in the microstructure for a given total saturation value.

[0054] This invention features a specially designed generator and discriminator: the generator employs a conditional generation network with 3D U-Net as its backbone, and injects conditional vectors into each layer of the network through a Spatial Adaptive Normalization (SPADE) module, achieving fine-grained control over the generated content layer by layer; the discriminator uses a 3D PatchGAN structure to independently determine the authenticity of each local region of the input 3D image, forcing the generator to maintain high realism not only in overall statistics but also in every local detail. This design solves the technical problems of uncontrollable generation process and generated results that "look like something but are physically unreasonable."

[0055] Specifically, the generator of the conditional generative adversarial network adopts a conditional generative network with 3DU-Net as the backbone, and injects conditional vectors through a spatial adaptive normalization module; the discriminator adopts the 3DPatchGAN structure.

[0056] A conditional generative adversarial network (cGAN) is constructed. The generator (G) adopts a conditional generative network with 3DU-Net as its backbone. The discriminator (D) adopts a 3DPatchGAN structure to determine the authenticity of local image patches. The key lies in the injection method of conditional information: in addition to random noise z, the generator input also includes a conditional vector c. c is composed of two parts: (a) the expected hydrate type content vector [C_i]_pred, calculated from the target hydrate saturation S_target through F_model; (b) the microstructure feature vector v, retrieved from the knowledge database DB_feature, that best matches [C_i]_pred. In this way, condition c constrains the generation process from both the macroscopic proportion and microscopic morphology levels.

[0057] Specifically, 3DPatchGAN performs dense sliding window discrimination on the input 3D image, ultimately outputting a 3D probability tensor. Each value in this tensor corresponds to the probability that a specific local 3D region (i.e., a "patch") in the input image is classified as "real". The 3DPatchGAN structure acts as an extremely sensitive "microscopic quality inspector," capable of performing pixel-level (voxel-level) fidelity evaluation and condition compliance verification on the generated 3D digital core. This forces the generator to synthesize hydrate microstructures that are highly reliable not only statistically overall but also in every local detail, and that conform to geophysical laws, thus ensuring that the final generated digital core model has excellent physical realism and reliability as the basis for subsequent simulation experiments. This is an indispensable technical step in achieving "on-demand generation" of high-fidelity models.

[0058] In one specific embodiment, an improved 3DU-Net is used as the segmentation network backbone. Specifically, the encoder contains four downsampling stages, each using two 3x3x3 convolutional layers (each convolution followed by batch normalization and ReLU activation) and a 2x2x2 max-pooling layer. After the penultimate downsampling layer, a channel attention module (SEBlock) is embedded, enabling the network to adaptively focus on feature channels of different lithofacies. The decoder path is symmetrical, using transposed convolutions for upsampling and skip connections with the high-resolution features of the corresponding encoder layers. The network finally outputs the probability of each voxel belonging to one of the three classes through a 1x1x1 convolutional layer and a Softmax activation function. Preparing the training data requires manual, detailed three-phase annotation of some typical CT slices. The annotated data is divided into training, validation, and test sets in an 8:1:1 ratio. A weighted cross-entropy loss function is used, and to address the issue that hydrated voxels typically account for a small proportion, their class weights can be appropriately increased. The optimizer uses Adam, with an initial learning rate set to 1e-4. Real-time data augmentation strategies, such as random rotations of 90°, 180°, and 270°, and random mirror flips along each axis, were employed during training to improve model robustness. The trained model was then applied to all CT data to obtain initial segmentation results. Subsequently, connected component analysis was used to remove small noise points (such as isolated regions with a volume of less than 100 voxels), and morphological closing operations (such as 3x3x3 spherical kernels) were used to smooth phase boundaries, ultimately obtaining high-quality three-phase binary volumetric data I_seg.

[0059] During the training of the generative network, this embodiment of the invention further introduces a physical attribute constraint mechanism to address the problem of focusing solely on visual realism while neglecting the consistency of physical parameters. Specifically, the loss function used in training not only includes the conventional adversarial loss but also introduces saturation and type content constraint losses. This constraint loss calculates the predicted total saturation S_gen and type content C_gen in real time by inputting the generated image into a pre-trained segmentation and classification network, and compares the difference with the target values ​​S_target and C_pred in the conditional vector. This mechanism ensures that the generated results are not only visually realistic but also strictly satisfy the constraints of the input conditions in terms of macroscopic (saturation) and microscopic (type ratio) physical quantities, thereby guaranteeing the physical credibility of the generative model.

[0060] Specifically, the loss functions used when training conditional generative adversarial networks include: adversarial loss, saturation and type content constraint loss. The constraint loss is obtained by inputting the generated image into a pre-trained segmentation and classification network and calculating the difference between its predicted saturation and type content and the target value.

[0061] Loss Function: The total loss L_total is a weighted sum of multiple losses: Adversarial Loss (L_adv): Uses least squares GAN (LSGAN) loss for more stable training. Saturation and Type Content Constraint Loss (L_const): Inputs the generated image into a fixed "segmentation and classification network" pre-trained in steps S1 and S2, calculates its predicted total saturation S_gen and type content C_gen, and calculates the L1 loss with the target values ​​S_target and C_pred in condition c: L_const=λ1*|S_gen-S_target|+λ2*Σ|C_gen-C_pred|. Perceptual Loss (L_percep, optional): Utilizes a pre-trained feature extractor (such as the CAE encoder in S4) to compare the distance between the generated image and the real image in the feature space to improve structural realism. Training and inference are performed, alternating between training G and D. The real image I_real and its corresponding saturation S_h are taken from D_labeled to generate the conditional vector c, which is input into G to obtain I_fake. D distinguishes between (I_real,c) and (I_fake,c) respectively. Network weights are updated via backpropagation. Example hyperparameters: learning rate 2e-4, batch size 2 (limited by GPU memory), 200,000 training iterations. Inference and generation are performed. After training, the user inputs an arbitrary target saturation S_target (which should be within the saturation range of the training data). The system automatically calculates C_pred = F_model(S_target), retrieves K feature vectors from DB_feature that are closest to C_pred (using Euclidean distance), and averages them to obtain v. C_pred and v are concatenated as condition c, and along with random noise z, are input into the generator G to synthesize a novel 3D hydrate digital core I_gen.

[0062] In constructing the condition vector, the condition vector c is constructed using a concatenation method: one part is the expected type content vector [C_i]_pred calculated by the mapping function F_model based on the target saturation S_target, reflecting the constraint on the macroscopic type ratio; the other part is the microscopic structure feature vector v retrieved from the microscopic structure feature database DB_feature that best matches [C_i]_pred, reflecting the constraint on the specific spatial distribution pattern. Through this design, the condition vector carries two layers of information simultaneously: "how many types of hydrates there are" and "how these hydrates are distributed in space," achieving multi-level and multi-scale control over the generated results. It serves as a bridge connecting the learning outcomes of the "construction phase" with the practical applications of the "generation phase."

[0063] Specifically, the conditional vector is constructed by concatenating or weighting the expected type content vector calculated by the mapping function based on the target saturation with one or more of the best-matching microstructure feature vectors retrieved from the microstructure feature database based on Euclidean distance.

[0064] In the technical solution of this invention, the feature vector that best matches the target type content vector Ci_pred is retrieved from the microstructure feature database DB_feature. The standard is based on a distance metric in the feature space, specifically defined as: Search input: Target type content vector The saturation is calculated by the mapping function F_model based on the target saturation S_target. Each record in the database DB_feature contains: a feature vector v (a 512-dimensional vector extracted by a 3D convolutional autoencoder); and the actual type content vector of the local image patch corresponding to this feature vector. (That is, the volume percentage of the three types of hydrates within the image block).

[0065] "Best match" means retrieving the actual type content vector from DB_feature. With target type content vector The K feature vectors with the smallest Euclidean distance between each other (K ​​is a preset positive integer, usually 1, 3, or 5). The distance calculation formula is as follows:

[0066] , If K=1, then the single feature vector with the smallest distance is taken as v; If K>1, the average (or weighted average) of the K feature vectors with the smallest distance is taken as the final v to enhance the robustness and representativeness of the retrieval results.

[0067] When averaging multiple feature vectors, a distance-inverse weighting method can be used, where feature vectors with smaller distances have larger weights, to strengthen the dominant role of the "best match". , Here, ϵ is a very small positive number to prevent division by zero.

[0068] This invention further provides a threshold optimization method based on statistical analysis. First, a training set containing manually labeled hydrate clusters of various types is constructed. The characteristic parameters of each cluster are automatically calculated, and the feature distribution of each category is statistically analyzed. Then, receiver operating characteristic (ROC) curve analysis is used to calculate the true positive rate and false positive rate under different candidate thresholds, selecting the value that maximizes the Youden index as the optimal threshold. This method overcomes the shortcomings of threshold setting relying on subjective experience or a "trial and error" approach, making the classification rules statistically grounded and ensuring the comparability and reproducibility of classification results between different researchers and different datasets.

[0069] Specifically, the first threshold (α) and the second threshold (β) are determined by the following method: constructing a training set containing manually labeled hydrate clusters of various types, calculating the characteristic parameters of each cluster and statistically analyzing the characteristic distribution of each category, using receiver operating characteristic curve analysis, and selecting the value that maximizes the Youden index as the optimal threshold.

[0070] The threshold point is typically chosen at the minimum of the intersection or overlap region of the two distribution curves to maximize classification discrimination. In practice, receiver operating characteristic (ROC) curve analysis can be used to calculate the true positive rate and false positive rate under different candidate thresholds, selecting the point that maximizes the Youden index as the optimal threshold. For the shape factor, the threshold is set by observing the clustering behavior of cemented clusters and other types of clusters on these geometric indices.

[0071] Specific embodiment: A digital core of a certain type of silty clay hydrate sediment in the Shenhu area of ​​the South China Sea was generated at a saturation of 41.3%.

[0072] Data: CT scan data of 10 pressurized sediment cores from the Shenhu area of ​​the South China Sea were used. The saturation measured by nuclear magnetic resonance were 11.3%, 22.6%, 28.1%, 32.7%, 37.2%, 41.3%, 47.8%, 51.5%, 56.7%, and 61.3%, respectively. The resolution was uniformly set to 3 μm / voxel, and the image size was 800x800x800 voxels.

[0073] Implementation process: Steps S1-S3: Segment all samples using the pre-trained 3DU-Net and run the automatic classification algorithm. Statistical analysis revealed that in samples with low saturation (e.g., 11.3%), pore-filling hydrates accounted for over 70%; as saturation increased, the proportions of cemented and encapsulated hydrates gradually rose.

[0074] Steps S4-S5: Construct and train a 3D-CAE, extract feature vectors of approximately 500,000 local image patches from 10 samples, and establish a feature database.

[0075] Steps S6-S7: Based on the saturation and type content data of 10 samples, a Gaussian process regression model F_model is trained. This model predicts that when S_h=41.3%, the typical contents of the three types of hydrates are: cemented type about 26.3%, encapsulated type about 21.4%, and pore-filling type about 52.3%.

[0076] Steps S8-S9: Construct and train the cGAN model. After training, input S_target=41.3%. The system queries F_model to obtain the expected content [25.8%, 21.1%, 53.1%], and retrieves multiple vectors with the best feature matching from the database, which are then fused to obtain condition c.

[0077] Generation Results: The generator synthesized an 800x800x800 voxel 3D digital core I_gen within 30 minutes. Post-processing analysis showed that the total volume fraction of hydrates in I_gen was 40.9%, with the relative contents of cemented, encapsulated, and pore-filling hydrates being 25.8%, 21.1%, and 53.1%, respectively, highly consistent with the predicted values. Furthermore, I_gen was imported into the Pore Network Model (PNM) software for absolute permeability simulation, yielding a result of 85 mD. This result is on the same order of magnitude and shows a consistent trend with the permeability measured in a real core experiment at the same saturation level (78 mD), validating the effectiveness and physical rationality of the generated model.

[0078] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for constructing conditionally generated hydrate digital cores, characterized in that, Includes the following steps: Preprocessing and three-phase segmentation of CT scan images of hydrate-bearing sediments were performed to obtain three-phase tag data of solid sediment skeleton, pore fluid and hydrate; Based on the three-phase tag data, the microscopic occurrence type of hydrates is automatically identified and labeled by calculating the spatial geometric and topological characteristic parameters of hydrate clusters and surrounding sediment skeleton particles. Based on the labeling results, a labeled dataset containing 3D images, microscopic storage type labels, and measured total saturation is constructed. A 3D convolutional autoencoder is constructed and trained. The trained autoencoder is used to extract feature vectors representing the microscopic spatial distribution patterns of hydrates from the labeled dataset. A microstructural feature database of associated features, macroscopic attributes and location information is established based on the feature vectors. Based on the labeled dataset, the data relationship between the total saturation of hydrates and the volume content of each microscopic occurrence type was statistically analyzed. Using total saturation as input and the relative content of each microscopic occurrence type as output, a mapping function is trained to establish a quantitative relationship model between hydrate saturation and microscopic occurrence type. A conditional generative adversarial network is constructed, the generator of which is input to random noise and a conditional vector consisting of a target saturation value and a feature vector retrieved from the microstructure feature database; Based on the input target hydrate saturation value, the expected microstructure content is determined according to the quantitative relationship model. Feature vectors are retrieved from the microstructure feature database. The retrieval results and the target saturation value are used together as conditional vectors and input into the trained conditional generative adversarial network to generate a three-dimensional hydrate digital core that satisfies the target saturation constraint and whose microstructure conforms to geological laws.

2. The condition-generating hydrate digital core construction method as described in claim 1, characterized in that, The microstructure types include particle cementation, particle encapsulation, and pore filling; the characteristic parameters include: the number of different sedimentary skeleton particles directly adjacent to the hydrate cluster, denoted as the contact particle number N_g; the ratio of the surface area of ​​the hydrate cluster to the surface area of ​​the contacted sedimentary skeleton particles, denoted as the encapsulation index C_i; and the shape factor describing the macroscopic morphology of the hydrate cluster.

3. The condition-generating hydrate digital core construction method as described in claim 2, characterized in that, The feature parameters are identified based on a preset threshold and rule set to automatically assign type labels: When N_g≥2 and the shape factor indicates that it is in a narrow region, it is determined to be a particle cementation type; When N_g=1 and C_i is greater than the first threshold, it is determined to be a particle-encapsulated type; When N_g=0, or N_g≤1 and C_i is less than the second threshold, it is determined to be a pore-filling type.

4. The condition-generating hydrate digital core construction method as described in claim 1, characterized in that, The three-phase segmentation is performed using a 3DU-Net deep learning network with an embedded channel attention mechanism.

5. The condition-generating hydrate digital core construction method as described in claim 1, characterized in that, The input to the 3D convolutional autoencoder is a multi-channel local three-dimensional image block containing a uniquely thermally encoded channel for sediment skeleton / pore fluid and various types of hydrates.

6. The method for constructing conditionally generated hydrate digital cores as described in claim 1, characterized in that, The mapping function is implemented through a multi-objective regression model, which can be a support vector regression, gradient boosting decision tree, or shallow neural network.

7. The method for constructing conditionally generated hydrate digital cores as described in claim 1, characterized in that, The generator of the conditional generative adversarial network adopts a conditional generative network with 3DU-Net as the backbone, and injects the conditional vector through a spatial adaptive normalization module; the discriminator adopts the 3DPatchGAN structure.

8. The method for constructing conditionally generated hydrate digital cores as described in claim 1, characterized in that, The loss functions used when training the conditional generative adversarial network include: adversarial loss, saturation and type content constraint loss. The constraint loss is obtained by inputting the generated image into a pre-trained segmentation and classification network and calculating the difference between its predicted saturation and type content and the target value.

9. The method for constructing conditionally generated hydrate digital cores as described in claim 1, characterized in that, The conditional vector is constructed by concatenating or weighting the expected type content vector calculated by the mapping function based on the target saturation with one or more of the best-matching microstructure feature vectors retrieved from the microstructure feature database based on Euclidean distance.

10. The method for constructing conditionally generated hydrate digital cores as described in claim 3, characterized in that, The first threshold and the second threshold are determined by the following method: constructing a training set containing manually labeled hydrate clusters of various types, calculating the characteristic parameters of each cluster and statistically analyzing the characteristic distribution of each category, using receiver operating characteristic curve analysis, and selecting the value that maximizes the Youden index as the optimal threshold.

Images