Method and system for constructing pore throat network model of weakly sanded dolomite based on CT (Computed Tomography) scanning

Abstract

The invention provides a weakly-sandy dolomite pore throat network model construction method and system based on CT scanning, and the method comprises the steps: carrying out the macroscopic, mesoscopic and microscopic CT scanning of a rock sample of weakly-sandy dolomite, and obtaining three-dimensional image data with different resolutions; performing multi-mineral phase intelligent identification and segmentation on the three-dimensional image data by adopting a multi-modal fusion and boundary perception progressive 3D segmentation network to obtain a semantic segmentation result; carrying out registration and fusion on the semantic segmentation results, and constructing a multi-scale digital core model; and based on the multi-scale digital core model, a pore throat network structure is extracted by adopting a maximum sphere algorithm, a pore throat network model is obtained, and quantitative characterization of the micro-pore structure of the weakly sanded dolomite is realized. The high-precision pore throat network model constructed by the method can provide a crucial digital basis and theoretical basis for seepage mechanism research, reserve evaluation, well drilling and completion scheme optimization and recovery ratio improvement strategy formulation of the weakly sandy dolomite reservoir.

Description

Technical Field

[0001] This invention belongs to the field of geological engineering technology, specifically relating to a method and system for constructing a pore-throat network model of weakly sandy dolomite based on CT scanning. Background Technology

[0002] Weakly sandy dolomite is a special type of carbonate reservoir characterized by a large amount of terrigenous clastic material (such as quartz and feldspar) mixed within the dolomite matrix. This complex mineral composition and genetic mechanism result in extremely heterogeneous reservoir space, typically exhibiting various types of pores, including intergranular pores, dissolution pores, microfractures, and primary pores between grains. Furthermore, the pore throat sizes vary widely, with significant differences in connectivity. Current conventional pore-throat network modeling methods are primarily designed for relatively homogeneous sandstone or single pore types. When applied to weakly sandy dolomite, the following prominent technical challenges arise: 1. Difficulty in mineral segmentation: The grayscale values ​​of minerals such as dolomite, quartz, and clay may overlap in CT images, making it impossible to accurately segment different mineral phases and pores using a single threshold method, severely impacting the accuracy of subsequent analysis. 2. Incomplete identification of multi-scale pores: Micron-level CT scans are insensitive to nanoscale throats and micropores, while focused ion beam scanning electron microscopy (FEM) has too small a field of view to represent macroscopic-scale heterogeneity. 3. Distortion in pore-throat network extraction: Traditional maximum sphere algorithms are prone to generating incorrect pore-throat connections when dealing with complex structures containing multiple minerals and porous states, failing to accurately reflect the fluid seepage path. 4. Limited model validation methods: Constructed models are typically compared only with macroscopic parameters such as porosity and permeability, lacking effective validation of the model's internal structural rationality.

[0003] Therefore, there is an urgent need for a new method that can accurately, multi-scale, and with high fidelity construct a pore-throat network model of weakly sandy dolomite. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for constructing a pore-throat network model of weakly sandy dolomite based on CT scanning. By combining multi-scale CT scanning, artificial intelligence mineral segmentation, digital core fusion, and improved pore-throat network extraction technology, high-precision and quantitative characterization of the micropore structure of weakly sandy dolomite is achieved.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for constructing a pore-throat network model of weakly sandy dolomite based on CT scans includes: CT scans were performed on samples of weakly sandy dolomite at macroscopic, mesoscopic, and microscopic scales to obtain three-dimensional image data at different resolutions. A progressive 3D segmentation network with multimodal fusion and boundary awareness is used to perform intelligent multi-mineral phase identification and segmentation on the three-dimensional image data to obtain semantic segmentation results; The semantic segmentation results are registered and fused to construct a multi-scale digital core model. Based on the multi-scale digital core model, the maximum sphere algorithm is used to extract the pore throat network structure, obtain the pore throat network model, and realize the quantitative characterization of the micropore structure of weakly sandy dolomite.

[0006] Preferably, methods for acquiring 3D image data at different resolutions include: Standard core plungers were drilled from the target reservoir to prepare rock samples of weakly sandy dolomite; The rock sample was scanned using an industrial CT scanner to obtain a pre-defined low-resolution macroscopic three-dimensional image that covers the heterogeneity of the entire core scale. A region of interest (ROI) is selected from the macroscopic three-dimensional image, and a preset high-resolution scan of the ROI is performed using micron-scale CT to obtain a mesoscopic three-dimensional image. Based on mesoscopic 3D images, micron-sized rock cores are drilled and scanned using focused ion beam scanning electron microscopy to obtain microscopic 3D images.

[0007] Preferably, the progressive 3D segmentation network with multimodal fusion and boundary awareness includes: A multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data; wherein, the multimodal encoder branch includes a main encoder branch, an auxiliary encoder branch, and a gated fusion unit; A progressive decoder is used to obtain depth features of three-dimensional image data based on the cross-modal fusion features; the progressive decoder includes a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer; The boundary-aware attention module, embedded in the skip connection between the encoder and the decoder, is used to concatenate the cross-modal fusion features with the deep features to predict a 3D boundary probability map, and multiply the 3D boundary probability map with the feature map in the skip connection to obtain a boundary enhancement feature map.

[0008] Preferred methods for constructing multi-scale digital core models include: Based on the semantic segmentation results, automatic registration and localization of macroscopic and mesoscopic scales are performed by matching the spatial distribution and geometric morphology of pore groups at macroscopic and mesoscopic scales to obtain registration results. Based on the registration results, irregular three-dimensional fusion transition zones are dynamically generated using morphological constraint functions, with the distance of the microscale boundary and local porosity as variables. Within the three-dimensional fusion transition zone, a tapered virtual throat with a smooth radius transition is created to geometrically and topologically connect the macroscopic and mesoscopic pore networks, thereby obtaining a multi-scale digital core model.

[0009] Preferred methods for obtaining pore-throat network models include: The multi-scale digital core model is subjected to median filtering to calculate the Euclidean distance field from each voxel in the pore space to the nearest solid skeleton. An improved maximum sphere algorithm is used to identify and extract an initial set of maximum spheres representing the center and radius of pores and throats in the distance field; Based on the initial set of maximum spheres, when determining the topological connection relationship between the maximum spheres, mineral type is introduced as a constraint condition, and the throat connectivity through different mineral phases is differentiated and verified to obtain the optimized set of maximum spheres. The optimized set of maximum spheres is parameterized into a pore-throat network model, where spheres represent pores, cylinders connecting pores represent throats, and geometric and mineralogical properties are assigned to each element.

[0010] This invention also provides a pore-throat network model construction system for weakly sandy dolomite based on CT scans, used to implement the method, including: The CT scanning module is used to perform macroscopic, mesoscopic and microscopic CT scans on samples of weakly sandy dolomite to acquire three-dimensional image data at different resolutions. The semantic segmentation module is used to perform multi-mineral phase intelligent identification and segmentation on the three-dimensional image data using a progressive 3D segmentation network with multimodal fusion and boundary awareness, so as to obtain semantic segmentation results. The core model construction module is used to register and fuse the semantic segmentation results to construct a multi-scale digital core model. The pore-throat network construction module is used to extract the pore-throat network structure based on the multi-scale digital core model using the maximum sphere algorithm, obtain the pore-throat network model, and realize the quantitative characterization of the micropore structure of weakly sandy dolomite.

[0011] Preferably, the CT scanning module includes: The rock sample preparation unit is used to drill standard core plungers from the target reservoir to prepare rock samples of weakly sandy dolomite; The macroscopic scanning unit is used to scan the rock sample using industrial CT to obtain a preset low-resolution macroscopic three-dimensional image that covers the heterogeneity of the entire core scale; The mesoscopic scanning unit is used to select a region of interest (ROI) from the macroscopic three-dimensional image and perform a preset high-resolution scan on the ROI using micron-scale CT to obtain a mesoscopic three-dimensional image. The microscopic scanning unit is used to drill micron-sized rock cores based on mesoscopic three-dimensional images and scan them using a focused ion beam scanning electron microscope to obtain microscopic three-dimensional images.

[0012] Preferably, in the semantic segmentation module, the progressive 3D segmentation network with multimodal fusion and boundary awareness includes: A multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data; wherein, the multimodal encoder branch includes a main encoder branch, an auxiliary encoder branch, and a gated fusion unit; A progressive decoder is used to obtain depth features of three-dimensional image data based on the cross-modal fusion features; the progressive decoder includes a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer; The boundary-aware attention module, embedded in the skip connection between the encoder and the decoder, is used to concatenate the cross-modal fusion features with the deep features to predict a 3D boundary probability map, and multiply the 3D boundary probability map with the feature map in the skip connection to obtain a boundary enhancement feature map.

[0013] Compared with existing technologies, the advantages of this invention are as follows: Utilizing a deep learning-based intelligent segmentation method, the accuracy of identifying pores and different mineral phases in complex mineral backgrounds is significantly improved, laying the foundation for constructing reliable models. By combining multi-scale CT scanning with digital core fusion technology, the limitations of single resolution are overcome, achieving a complete characterization from nanothroat channels to the macroscopic pore system, making the model closer to real geological conditions. Introducing mineral constraint factors into the pore-throat network extraction allows the constructed model to not only reflect geometric morphology but also the control effect of mineral composition on seepage capacity, thus enhancing the geological significance of the model. The high-precision pore-throat network model constructed by this method can provide a crucial digital foundation and theoretical basis for the study of seepage mechanisms, reserve assessment, drilling and completion scheme optimization, and the formulation of enhanced oil recovery strategies in weakly sandy dolomite reservoirs. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the method for constructing a pore-throat network model of weakly sandy dolomite based on CT scanning, according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1: like Figure 1 As shown, the method for constructing a pore-throat network model of weakly sandy dolomite based on CT scans includes: S1: CT scans were performed on weakly sandy dolomite samples at macroscopic, mesoscopic, and microscopic scales to obtain three-dimensional image data at different resolutions.

[0019] A further implementation method for acquiring three-dimensional image data at different resolutions includes: Standard core plungers were drilled from the target reservoir to prepare samples of weakly sandy dolomite. Specifically, before CT scanning, the plungers were pre-processed: rapid XRF scanning was performed along the plunger surface to obtain longitudinal distribution profiles of key elements such as calcium, magnesium, and silicon, preliminarily identifying sandy and dolomite-rich sections. The P-wave and S-wave velocities of the plungers were measured to preliminarily assess their mechanical properties and macroscopic heterogeneity in terms of density. The XRF and ultrasonic data were integrated to generate a macroscopic "lithofacial-physical property guide map," which will serve as the primary scientific basis for subsequent selection of the Region of Interest (ROI), rather than relying solely on random selection from CT grayscale images.

[0020] Industrial CT was used to scan rock samples to obtain a pre-set low-resolution macroscopic 3D image covering the entire core-scale heterogeneity. Specifically, the prepared core plunger was placed in the industrial CT scanning system. While ensuring a balance between radiation dose and signal-to-noise ratio, optimized scanning parameters (such as voltage, current, and filters) were used to obtain a macroscopic 3D image covering the entire plunger with a pre-set resolution of 10-25 micrometers / voxel. This resolution was designed to effectively capture fractures, large dissolution cavities, and significant lithological stratification. The "lithofiform-physical property guide map" obtained in the first step was fused and spatially registered with the 3D grayscale image obtained from the macroscopic CT scan. Based on the fused information, three representative ROIs were initially selected: Characteristic ROIs, clearly located within sandy bands, obvious fractures or cavities, or mineral aggregates; Transitional ROIs, located at the contact boundary between dolomite and sandy minerals or in lithological gradient zones; and Matrix ROIs, seemingly homogeneous, dense dolomite matrix regions without obvious characteristics. This classification ensures that subsequent sampling can comprehensively cover all key pore structure types of the reservoir.

[0021] Regions of Interest (ROIs) are selected from macroscopic 3D images, and micron-scale CT is used to perform high-resolution scans of these ROIs to obtain mesoscopic 3D images. Specifically, small core samples (approximately 1-2 mm in diameter) are drilled from the core plunger at the pre-selected target centers of various ROIs using a precision coring drill. These small core samples are then scanned using a micron-scale CT system (such as micro-CT). By optimizing the scanning conditions, the resolution is increased to approximately 1-2 micrometers per voxel to accurately characterize intergranular pores, intragranular pores, and microfracture systems. Preliminary thresholding and pore network analysis are performed on the obtained mesoscopic 3D images to calculate parameters such as local porosity and pore size distribution. On these high-resolution images, "target points" for subsequent nanoscale scanning are further identified. Target points should focus on two key structures: regions where mesoscopic pores connect to suspected nanothroats, and mineral aggregates (such as clay mineral clusters or microcrystalline dolomite regions) that appear dense at the micron scale but are inferred to have nanoscale pores based on geological background.

[0022] Based on mesoscopic 3D images, micron-sized core samples are drilled and scanned using focused ion beam scanning electron microscopy (FIB-SEM) to obtain microscopic 3D images. Specifically, using a dual-beam FIB-SEM system, ultramicrocore samples of approximately 10x10x10 micrometers are precisely cut and peeled from 1-2 mm core samples that have undergone micron-level CT scanning at pre-marked "target points." These ultramicrocore samples are then subjected to automated 3D slicing and imaging using FIB-SEM. FIB is used for layer-by-layer peeling (layer thickness can reach 5-10 nanometers), while SEM is used to perform high-resolution imaging of each newly exposed section (resolution can reach ~5-10 nanometers / voxel). Finally, hundreds to thousands of consecutive 2D SEM images are aligned, stacked, and 3D reconstructed to obtain microscopic 3D image data that clearly reveals structures such as nanopore throats and interlayer pores in clay.

[0023] All 3D image data at all scales (macro, meso, and micro) along with scanning parameters and their spatial location information within the original core were systematically entered into a unified database. A traceable "data chain" was creatively established to ensure that any nanoscale FIB-SEM data block can be accurately located back to its corresponding micron-scale CT core and further correlated to a specific geological facies zone within the macroscopic plunger. This provides an indispensable spatial coordinate foundation for subsequent multi-scale fusion modeling.

[0024] S2: Employs a progressive 3D segmentation network with multimodal fusion and boundary awareness to perform intelligent multi-mineral phase identification and segmentation on 3D image data, obtaining semantic segmentation results.

[0025] A further implementation method involves a progressive 3D segmentation network for multimodal fusion and boundary awareness, comprising: The multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data. This branch includes a main encoder branch, an auxiliary encoder branch, and a gated fusion unit. Specifically, the main encoder branch uses a standard 3D U-Net or ResNet encoder portion as its backbone. It consists of multiple cascaded downsampling blocks, each containing 3D convolutions, normalization layers, and activation functions. This branch uses CT grayscale images as its primary input, responsible for extracting their inherent geometric structure, density contrast, and macroscopic pore space features. As the number of layers increases, its receptive field gradually expands to capture the contextual semantic information of the image.

[0026] The auxiliary encoder branch is a relatively lightweight encoder (e.g., consisting of a few 3D convolutional layers) parallel to the main encoder hierarchy. This branch takes a multimodal fusion feature block as input. This feature block is constructed by concatenating a registered SEM texture image (which can be converted into multichannel features) with an EDS elemental distribution map along the channel dimension. This branch is specifically designed to extract lithological features related to mineral composition and microscopic surface texture.

[0027] The gated fusion unit is located after each corresponding layer of the main and auxiliary encoders. It is a small neural network, typically consisting of convolutional layers and a sigmoid activation function, used to generate a gated signal map with values ​​between 0 and 1. This is crucial for achieving adaptive fusion. It receives feature maps from the main and auxiliary branches and learns to generate an adaptive weight matrix.

[0028] A progressive decoder is used to obtain depth features from 3D image data based on cross-modal feature fusion. The progressive decoder consists of a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer. Specifically, the decoder is responsible for upsampling the fused and enhanced features, progressively restoring the spatial resolution, and outputting the final segmentation result. It consists of multiple upsampling blocks, each containing a transposed convolutional or interpolated upsampling layer followed by a convolutional layer. The feature selection upsampling layer is the core operation layer of the decoder. After upsampling, it does not simply concatenate the features but fuses the upsampling result with the boundary enhancement feature map output from the boundary-aware attention module. This ensures that high-resolution boundary details are preferentially and accurately utilized during reconstruction.

[0029] The first segmentation output layer is connected to the middle layer of the decoder. It outputs a lower-resolution, preliminary segmentation result. This output is mainly used to assist training, providing a mid-range gradient to the network through an additional loss function, accelerating convergence, and providing preliminary morphological evaluation.

[0030] The second segmentation output layer is the final output layer of the decoder. It integrates the optimized features from all levels and outputs a high-precision final 3D semantic segmentation map with the same resolution as the input CT image, in which each voxel is precisely classified into a specific category such as pore, dolomite, or quartz.

[0031] The network takes the aforementioned multi-scale CT image data and registered SEM / EDS data as input. The data is first refined into a series of cross-modal fusion features via a multimodal encoder branch. During the feature transmission to the decoder, a boundary-aware attention module processes these features, highlighting boundary information. The progressive decoder then utilizes these enhanced features to first generate a coarse segmentation to grasp the overall structure, and then gradually optimizes and outputs the final high-fidelity segmentation result by fusing fine boundary information. The entire process achieves end-to-end computation from multi-source information input to accurate structural resolution, laying a solid foundation for building a reliable pore-throat network model.

[0032] The boundary-aware attention module, embedded in the skip connection between the encoder and decoder, concatenates cross-modal fused features with deep features to predict a 3D boundary probability map. This 3D boundary probability map is then multiplied with the feature map in the skip connection to obtain a boundary-enhanced feature map. Specifically, this module first concatenates shallow fused features from the encoder, rich in harmful boundary details, with deep features from the decoder, rich in semantic information. Then, a 1x1x1 convolutional layer and a sigmoid activation function are applied to output a 3D boundary probability map with the same resolution as the input feature map. This boundary map visually displays the uncertain regions of pore-mineral and mineral-mineral boundaries as perceived by the model. This boundary probability map is then multiplied element-wise with the feature map originally intended to be passed in the skip connection. This operation is equivalent to an attention mechanism that significantly amplifies the weight of boundary region features while suppressing non-boundary regions. As a result, the decoder, when recovering details, obtains a "highlighted" boundary region, driving it to generate sharper and more accurate segmentation results.

[0033] S3: Register and fuse the semantic segmentation results to construct a multi-scale digital core model.

[0034] A further implementation method involves constructing a multi-scale digital core model, including: Based on semantic segmentation results, automatic registration and localization between macroscopic and mesoscopic scales are performed by matching the spatial distribution and geometric morphology of pore clusters at both macroscopic and mesoscopic scales to obtain registration results. Specifically, in the semantically segmented macroscopic and mesoscopic 3D models, individual pores are no longer the focus; instead, "pore communities" are identified and extracted as feature units. For each community, a series of morphological and topological descriptors are calculated, including but not limited to: equivalent diameter distribution, eccentricity, surface-to-volume ratio, and coordination number distribution. These descriptors are used to construct a high-dimensional feature vector. Subsequently, a robust point set registration algorithm (such as the Coherent Point Drift algorithm combined with RANSAC) is used to search for the corresponding region in the macroscopic model that has the most similar feature vector distribution to the mesoscopic model. This process outputs the optimal rigid and non-rigid spatial transformation matrices, thereby accurately "anchoring" the mesoscopic model to the real geological context of the macroscopic model.

[0035] Based on the registration results, an irregular 3D fusion transition zone is dynamically generated using the distance to the microscale boundary and local porosity as variables through a morphological constraint function. Specifically, based on the registration results, a buffer surrounding the mesoscopic model is defined within the macroscopic model. Within this buffer, a normalized boundary distance field and a local porosity difference field are calculated. The normalized boundary distance field is obtained by calculating and normalizing the distance from each voxel to the boundary of the mesoscopic model. The local porosity difference field is obtained by calculating the difference in the absolute value of local porosity of each voxel at the macroscopic scale and the (upsampled) mesoscopic scale. Based on the normalized boundary distance field, the local porosity difference field, and adjustable weights, a morphological constraint function is constructed. The output value of this function defines the probability that each voxel belongs to the fusion transition zone. By setting a threshold, an irregular 3D transition zone shell highly correlated with the complexity of the pore structure can be generated. In regions with complex and highly variable pore structures, the transition zone automatically widens; in homogeneous regions, it narrows.

[0036] Within the three-dimensional fusion transition zone, a multi-scale digital core model is obtained by creating a conical virtual throat with a smoothly transitioning radius, geometrically and topologically connecting the macroscopic and mesoscopic pore networks. Specifically, all "boundary pores" located within the transition zone are identified from both the mesoscopic and macroscopic networks. Instead of simply connecting the nearest pore pairs, a multi-objective optimization matching is established, considering the center distance between pore pairs, the similarity of their radii, and the angle between their principal axes. This process finds the most logically connected "partner" in the macroscopic network for each mesoscopic boundary pore. Between the matched pore pairs, a double-conical or smoothly variable-diameter cylinder is created as a virtual throat. Its radius smoothly transitions from the minimum at the midpoint to the radius of the connected pores at both ends. This virtual throat is then formally added as a new topological element to the unified pore network model, completing the geometric and topological connection. Based on the geometry of the virtual throat (such as volume average radius and length) and the mineral phases it passes through (queried from the digital core model), flow properties such as absolute permeability are calculated to ensure that it can accurately reflect the transport capacity of the area spanned at this scale in subsequent flow simulations.

[0037] For the unified digital core model after fusion, properties such as porosity and permeability may exhibit abrupt changes in the transition zone. An anisotropic diffusion algorithm is used to smooth these properties, ensuring a natural and continuous gradient change from the mesoscopic to the macroscopic boundary, thus eliminating potential abrupt changes in the physical field during numerical simulation. The final verification goes beyond geometric comparison. By running virtual seepage experiments (such as the Lattice Boltzmann method) and virtual high-pressure mercury intrusion on the fused model, the simulated macroscopic permeability and capillary pressure curves are compared with the experimental data. Their high consistency is the final criterion proving that the fusion process successfully maintained the rock transport function.

[0038] S4: Based on the multi-scale digital core model, the maximum sphere algorithm is used to extract the pore throat network structure and obtain the pore throat network model, so as to realize the quantitative characterization of the micropore structure of weakly sandy dolomite.

[0039] A further implementation method for obtaining the pore-throat network model includes: Median filtering is applied to the multi-scale digital core model to calculate the Euclidean distance field from each voxel in the pore space to the nearest solid skeleton. Specifically, anisotropic diffusion filtering is performed on the multi-scale digital core model, which is superior to simple median filtering. It can smooth noise while effectively preserving the sharp boundaries between pores and the skeleton, avoiding distortion in subsequent distance field calculations. The exact Euclidean distance transform (EDT) from each voxel in the pore space to the nearest solid skeleton is calculated. This distance field is the basis for all subsequent analyses, and the value of each voxel defines the radius of the largest inscribed sphere that can be accommodated within it.

[0040] An improved maximum sphere algorithm is employed to identify and extract an initial set of maximum spheres representing the centers and radii of pores and throats in the distance field. Specifically, all local maxima are identified in the distance field, which are the centers of potential pores or throats. The initial set of maximum spheres is generated with their corresponding distance values ​​as radii. A hierarchical extraction process is introduced. First, all "main spheres" with radii greater than a set threshold (e.g., 3 voxels) are extracted to represent the main pore space. Subsequently, smaller "secondary spheres" are searched and extracted again in the residual space not fully occupied by the main spheres. This strategy can more completely capture the entire series of pores and throats from large pores to micro-throats, while avoiding the excessive submergence of large pore structures by small spheres, laying a clear foundation for subsequent network simplification.

[0041] Based on the initial set of maximum spheres, mineral type is introduced as a constraint when determining the topological connectivity between maximum spheres. The connectivity of throats passing through different mineral facies is differentiated and verified to obtain an optimized set of maximum spheres. This is a crucial step in integrating geological information into the network topology and serves as a bridge for achieving coupled simulation of "geomechanics-seepage". Specifically, any two maximum spheres that spatially overlap or whose distance is less than the sum of their radii are marked as potential connection candidate pairs. For each connection candidate pair, instead of directly identifying it as a throat, "mineral constraint verification" is performed, including path exploration and connectivity differentiation. Path exploration involves constructing a shortest path between the two sphere centers and querying the mineral facies labels (from the aforementioned semantic segmentation results) of the voxels traversed by this path.

[0042] Connectivity differentiation includes: if the path lies entirely within the porous phase or passes through microfractures in rigid minerals (such as quartz or dolomite), a throat connection is directly established. If the path passes through plastic or adsorbent mineral (such as clay) aggregates, throat validity is assessed. By querying the local porosity of the clay region, if it is below a critical threshold (indicating it is dense interstitial material rather than porous clay), the connection is rejected because it may have been sealed in geological history or contributes very little to seepage. For throats deemed valid but passing through different mineral phases, an equivalent flow resistance coefficient is assigned based on the proportion of different minerals in the path. This coefficient will be used to correct the conductivity in subsequent seepage simulations.

[0043] Based on the optimized connectivity, "dead end" pores that have only one or two connections to the network (unless they are large in size and may represent dissolution pores) are identified and removed, thereby simplifying the network and making it more focused on connected seepage paths.

[0044] The optimized set of maximum spheres is parameterized into a pore-throat network model, where spheres represent pores, cylinders connecting pores represent throats, and geometric and mineralogical properties are assigned to each element.

[0045] Specifically, the optimized set of maximum spheres is abstracted as network nodes (pores), whose volume and radius are determined by the corresponding maximum sphere. The identified throats are abstracted as cylinders, whose radius is determined by the distance field value at the narrowest point of the connecting path (i.e., the throat radius). Each pore and throat is assigned a series of physical properties, forming a rich property network: Geometric properties: coordinates, volume, radius of the inscribed sphere, shape factor.

[0046] Mineralogical properties: the dominant mineral phase corresponding to its spatial location, and the adjacent mineral environment.

[0047] Mechanical properties: Based on the mineral phase in which it is located, an estimated Young's modulus is assigned to it, which is crucial for subsequent research on stress sensitivity and geomechanical deformation.

[0048] This embodiment also provides multiphysics coupling and model validation. Based on the constructed pore-throat network model, the absolute permeability of the model is calculated by solving the Poisson equation or simulating network flow, and then compared with the permeability measured in core experiments. The capillary force-driven displacement process of the unwetting phase (mercury) is simulated in the pore-throat network model. The simulated capillary pressure curve is compared with the experimental MIP curve to verify the model's accuracy in terms of pore-throat radius distribution and connectivity. The porosity, specific surface area, pore-throat coordination number distribution, topology, and other parameters calculated by the model are cross-validated with results obtained directly from CT images and SEM image analysis.

[0049] Example 2 This invention also provides a system for constructing a pore-throat network model of weakly sandy dolomite based on CT scans, and a method for implementing this method, including: The CT scanning module is used to perform macroscopic, mesoscopic and microscopic CT scans on samples of weakly sandy dolomite to acquire three-dimensional image data at different resolutions. The semantic segmentation module is used to perform multi-mineral phase intelligent identification and segmentation on three-dimensional image data using a progressive 3D segmentation network with multimodal fusion and boundary awareness, and to obtain semantic segmentation results. The core model construction module is used to register and fuse semantic segmentation results to construct a multi-scale digital core model. The pore-throat network construction module is used to extract the pore-throat network structure based on a multi-scale digital core model and the maximum sphere algorithm, thereby obtaining the pore-throat network model and realizing the quantitative characterization of the micropore structure of weakly sandy dolomite.

[0050] A further embodiment of the invention includes a CT scanning module comprising: The rock sample preparation unit is used to drill standard core plungers from the target reservoir to prepare rock samples of weakly sandy dolomite; The macroscopic scanning unit is used to scan rock samples using industrial CT to obtain a preset low-resolution macroscopic three-dimensional image that covers the heterogeneity of the entire core scale; The mesoscopic scanning unit is used to select a region of interest (ROI) from a macroscopic three-dimensional image and perform a preset high-resolution scan on the ROI using micron-level CT to obtain a mesoscopic three-dimensional image. The microscopic scanning unit is used to drill micron-sized rock cores based on mesoscopic three-dimensional images and scan them using a focused ion beam scanning electron microscope to obtain microscopic three-dimensional images.

[0051] A further implementation method involves a semantic segmentation module where the progressive 3D segmentation network for multimodal fusion and boundary awareness includes: The multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data; the multimodal encoder branch includes the main encoder branch, the auxiliary encoder branch, and the gated fusion unit. A progressive decoder is used to obtain depth features of 3D image data based on cross-modal fusion features; the progressive decoder includes a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer; The boundary-aware attention module, embedded in the skip connection between the encoder and decoder, is used to concatenate cross-modal fusion features with deep features to predict a 3D boundary probability map, and then multiply the 3D boundary probability map with the feature map in the skip connection to obtain a boundary enhancement feature map.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for constructing a pore-throat network model of weakly sandy dolomite based on CT scanning, characterized in that, include: CT scans were performed on samples of weakly sandy dolomite at macroscopic, mesoscopic, and microscopic scales to obtain three-dimensional image data at different resolutions. A progressive 3D segmentation network with multimodal fusion and boundary awareness is used to perform intelligent multi-mineral phase identification and segmentation on the three-dimensional image data to obtain semantic segmentation results; The semantic segmentation results are registered and fused to construct a multi-scale digital core model. Based on the multi-scale digital core model, the maximum sphere algorithm is used to extract the pore throat network structure, obtain the pore throat network model, and realize the quantitative characterization of the micropore structure of weakly sandy dolomite.

2. The method according to claim 1, characterized in that, Methods for acquiring 3D image data at different resolutions include: Standard core plungers were drilled from the target reservoir to prepare rock samples of weakly sandy dolomite; The rock sample was scanned using an industrial CT scanner to obtain a pre-defined low-resolution macroscopic three-dimensional image that covers the heterogeneity of the entire core scale. A region of interest (ROI) is selected from the macroscopic three-dimensional image, and a preset high-resolution scan of the ROI is performed using micron-scale CT to obtain a mesoscopic three-dimensional image. Based on mesoscopic 3D images, micron-sized rock cores are drilled and scanned using focused ion beam scanning electron microscopy to obtain microscopic 3D images.

3. The method according to claim 1, characterized in that, Multimodal fusion and boundary-aware progressive 3D segmentation networks include: A multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data; wherein, the multimodal encoder branch includes a main encoder branch, an auxiliary encoder branch, and a gated fusion unit; A progressive decoder is used to obtain depth features of three-dimensional image data based on the cross-modal fusion features; the progressive decoder includes a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer; The boundary-aware attention module, embedded in the skip connection between the encoder and the decoder, is used to concatenate the cross-modal fusion features with the deep features to predict a 3D boundary probability map, and multiply the 3D boundary probability map with the feature map in the skip connection to obtain a boundary enhancement feature map.

4. The method according to claim 1, characterized in that, Methods for constructing multi-scale digital core models include: Based on the semantic segmentation results, automatic registration and localization of macroscopic and mesoscopic scales are performed by matching the spatial distribution and geometric morphology of pore groups at macroscopic and mesoscopic scales to obtain registration results. Based on the registration results, irregular three-dimensional fusion transition zones are dynamically generated using morphological constraint functions, with the distance of the microscale boundary and local porosity as variables. Within the three-dimensional fusion transition zone, a tapered virtual throat with a smooth radius transition is created to geometrically and topologically connect the macroscopic and mesoscopic pore networks, thereby obtaining a multi-scale digital core model.

5. The method according to claim 1, characterized in that, Methods for obtaining pore-throat network models include: The multi-scale digital core model is subjected to median filtering to calculate the Euclidean distance field from each voxel in the pore space to the nearest solid skeleton. An improved maximum sphere algorithm is used to identify and extract an initial set of maximum spheres representing the center and radius of pores and throats in the distance field; Based on the initial set of maximum spheres, when determining the topological connection relationship between the maximum spheres, mineral type is introduced as a constraint condition, and the throat connectivity through different mineral phases is differentiated and verified to obtain the optimized set of maximum spheres. The optimized set of maximum spheres is parameterized into a pore-throat network model, where spheres represent pores, cylinders connecting pores represent throats, and geometric and mineralogical properties are assigned to each element.

6. A pore-throat network model construction system for weakly sandy dolomite based on CT scans, used to implement the method described in any one of claims 1-5, characterized in that, include: The CT scanning module is used to perform macroscopic, mesoscopic and microscopic CT scans on samples of weakly sandy dolomite to acquire three-dimensional image data at different resolutions. The semantic segmentation module is used to perform multi-mineral phase intelligent identification and segmentation on the three-dimensional image data using a progressive 3D segmentation network with multimodal fusion and boundary awareness, so as to obtain semantic segmentation results. The core model construction module is used to register and fuse the semantic segmentation results to construct a multi-scale digital core model. The pore-throat network construction module is used to extract the pore-throat network structure based on the multi-scale digital core model using the maximum sphere algorithm, obtain the pore-throat network model, and realize the quantitative characterization of the micropore structure of weakly sandy dolomite.

7. The system according to claim 6, characterized in that, The CT scanning module includes: The rock sample preparation unit is used to drill standard core plungers from the target reservoir to prepare rock samples of weakly sandy dolomite; The macroscopic scanning unit is used to scan the rock sample using industrial CT to obtain a preset low-resolution macroscopic three-dimensional image that covers the heterogeneity of the entire core scale; The mesoscopic scanning unit is used to select a region of interest (ROI) from the macroscopic three-dimensional image and perform a preset high-resolution scan on the ROI using micron-scale CT to obtain a mesoscopic three-dimensional image. The microscopic scanning unit is used to drill micron-sized rock cores based on mesoscopic three-dimensional images and scan them using a focused ion beam scanning electron microscope to obtain microscopic three-dimensional images.

8. The system according to claim 6, characterized in that, The semantic segmentation module includes a progressive 3D segmentation network with multimodal fusion and boundary awareness, comprising: A multimodal encoder branch is used to acquire cross-modal fusion features of 3D image data; wherein, the multimodal encoder branch includes a main encoder branch, an auxiliary encoder branch, and a gated fusion unit; A progressive decoder is used to obtain depth features of three-dimensional image data based on the cross-modal fusion features; the progressive decoder includes a first segmentation output layer, a feature selection upsampling layer, and a second segmentation output layer; The boundary-aware attention module, embedded in the skip connection between the encoder and the decoder, is used to concatenate the cross-modal fusion features with the deep features to predict a 3D boundary probability map, and multiply the 3D boundary probability map with the feature map in the skip connection to obtain a boundary enhancement feature map.

Images