A three-dimensional lamellar multi-component digital rock modeling method

Abstract

The application belongs to the technical field of oil and gas reservoir rock data identification, and discloses a three-dimensional laminated multi-component digital rock modeling method. A GAN-MPS hybrid method fuses a generative adversarial network (GAN) and multiple point geostatistics (MPS) technology, and reconstructs a large-size three-dimensional laminated multi-component digital rock of a shale sample relying on a single representative two-dimensional image, so that synchronous upgrading of the digital rock in dimension and size is realized. The application takes a high-resolution two-dimensional mineral scanning image as training data, firstly reconstructs a three-dimensional multi-component digital rock model of a single lamina through a GAN deep learning method, then uses MPS for model scale expansion, finally superimposes and combines digital models of two or more lamina types, and realizes construction of a multi-component, high-precision and high-efficiency three-dimensional laminated multi-component shale digital rock. The constructed model lays a solid foundation for subsequent research on reservoir physical properties and mechanical properties based on the digital rock.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas reservoir rock data identification technology, and particularly relates to a three-dimensional lamellar multi-component digital rock modeling method. Background Technology

[0002] Digital rock technology has long been a core technique for studying the physical and mechanical properties of reservoir rocks. Previous tests on the physical and mechanical properties of shale reservoirs primarily relied on rock physics experiments at the core plunger scale. However, continental shale exhibits strong heterogeneity, with shale laminae within a single plunger often showing significant variations in type, thickness, and combination. Conventional rock physics experiments struggle to clarify the controlling factors and cannot reuse valuable core samples. Digital rock technology, as an effective numerical method for rock physics research, offers numerous advantages. Based on a single digital rock sample, multiple rock physical properties, such as porosity, permeability, electrical properties, and strength, can be simulated simultaneously at the micrometer-millimeter scale. Using SEM and CT scanning equipment to acquire core images, combined with deep learning and multi-point statistical methods, multi-scale, multi-mineral models can be constructed. This allows for the control of influencing factors, thereby clarifying the impact of laminae type and combination on the physical properties of shale rocks. However, the established multi-mineral digital rock model needs to be sufficiently accurate to reflect the mineral composition, distribution, laminae type, laminae thickness, and laminae combination types in real rocks.

[0003] Physical experimental methods are currently the most commonly used approach for establishing digital rock models of laminated rocks. The essence of physical modeling is to directly image the sample using physical experimental instruments and then reconstruct it into a three-dimensional digital rock. Commonly used physical modeling methods include sequential slice imaging, X-ray computed tomography (CT), and FIB-SEM. Using multi-scale CT scans, millimeter- to micrometer-scale digital rock models can be established. The study analyzed the seepage differences between silty mudstone and laminated shale. It was found that silty mudstone has good connectivity and strong seepage capacity at different scales, but laminated shale and fractured laminated mudstone can only seep at large scales, while laminated mudstone and silty banded mudstone only have good connectivity at small scales. Because different minerals exhibit overlapping grayscale values ​​in CT scans, the accuracy of multi-mineral digital rock models obtained from CT scans is significantly affected. Yang Yuxuan et al. obtained high-precision digital rock models of organic and inorganic materials in the layered shale of the Jiyang Depression using FIB-SEM. The established 3D spatial dimensions of the digital rock were 6.6 nm × 6.5 nm × 10 nm. Based on this, they constructed a pore network model of organic and inorganic materials and analyzed the influence of laminar proportion and seepage direction on the seepage capacity of shale oil. Liu et al. also used FIB-SEM to obtain a digital rock model of shale in a certain region of China. Based on this model, they conducted uniaxial compression and uniaxial tension simulation experiments to explore the influence of pyrite laminarization on the physical and mechanical properties of shale and the mechanism of micro-crack propagation. This method has high modeling accuracy, but it is limited by the shale sample itself, making comparative experiments difficult and unable to flexibly match various laminar structure parameters. Furthermore, the modeling cost is high. Meng et al. used QEMSCAN to obtain a two-dimensional digital rock model of a section of shale in the Qingshankou Formation of the Songliao Basin. Based on this model, they conducted finite element simulations. This modeling method has a lower cost and can characterize information such as laminarization, mineral morphology, and distribution. However, due to the limitations of the samples themselves, experiments can only be conducted based on the actual models obtained. When conducting comparative experiments, new variables are often introduced.

[0004] Hybrid modeling is also a commonly used modeling method, which emerged to avoid the limitations of single modeling methods. Li et al. used a multi-scale discrete structure modeling method based on statistical regularities and Monte Carlo simulation. They first statistically analyzed the development of laminae at multiple scales, then used the Monte Carlo method to randomly generate horizontal laminar structures according to the statistical regularities of each scale, and finally superimposed them to generate a multi-scale composite model. This modeling method has low accuracy and resolution, and can only reflect the macroscopic seepage characteristics of the reservoir, but cannot study the micromechanics and seepage characteristics of the reservoir. Wang et al. established a multi-mineral digital rock model of shale using the discrete element method. They continuously adjusted the mechanical parameters of the PFC model particles, matched the simulation results with the data obtained from physical experiments, and finally obtained a set of suitable parameters. Based on this set of parameters, they investigated the influence of the elastic modulus and laminar angle of carbonate laminae in shale on the number and propagation mode of shale fractures. Yan et al. used high-resolution CT scans and geological regularity statistics to extract shale information, and then combined it with the discrete element method to establish a digital rock model, studying the influence of laminar dip angle, thickness, laminar type, and laminar linear density on the mechanical properties of shale rocks. The discrete element method (DEM) model shows only one type of mineral within the laminae, which is inconsistent with real shale. Hard minerals within the laminae have been shown to significantly influence the propagation of microscopic fractures in shale. This method struggles to distinguish between different minerals within the laminae and cannot study the more microscopic mechanisms of intralayer fracture propagation.

[0005] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0006] (1) Existing physical experimental methods for constructing digital rock models require high costs and are time-consuming. The grayscale images obtained from the experiments are difficult to accurately distinguish different mineral components. Due to the complexity of the images, they cannot effectively present the spatial distribution characteristics of multiple components. At the same time, the processing procedures are cumbersome, which limits their efficiency and accuracy in practical applications.

[0007] (2) When constructing a three-dimensional digital rock model that conforms to the actual geological characteristics, the existing numerical simulation methods mostly adopt the assumption of a single mineral composition, which makes it difficult to simultaneously take into account the authenticity of the rock structure and the diversity of mineral composition. As a result, the constructed three-dimensional digital rock model has a single mineral composition and cannot support the simulation study of reservoir electrical, acoustic and other properties.

[0008] (3) Existing numerical simulation technologies for constructing three-dimensional multi-component digital rock models are mostly based on idealized model assumptions, which simplify the distribution morphology and spatial combination of minerals. This results in significant differences from the heterogeneity and complex structure of minerals in real geological environments, thus limiting the geological applicability and engineering reference value of the models.

[0009] (4) Existing methods struggle to achieve a balance between model size and microstructural accuracy. Either the model size is too small to represent the macroscopic core, or key pore structure and mineral spatial distribution characteristics are sacrificed during the scaling process. Furthermore, for reservoirs with significant bedding structures, such as shale, how to construct digital rocks that reflect both macroscopic laminar features and microscopic multi-component heterogeneity remains a current technical challenge. Summary of the Invention

[0010] To overcome the problems existing in the prior art, the present invention provides a three-dimensional layered multi-component digital rock modeling method. To achieve the above objective, the present invention includes the following steps:

[0011] S1. Two-dimensional mineral scanning images of rock samples were obtained through AMICS experiments, and representative local areas were selected as the original training images TI1.

[0012] S2. Convert the selected original training image TI1 from an RGB color image to a grayscale image, and assign continuous grayscale encoding values ​​to different mineral phases in a preset order. Perform scaling processing while ensuring that the image is not distorted to obtain a standardized training image TI2 that meets the input requirements of the deep learning model.

[0013] S3. Input the standardized training image TI2 into the GAN deep learning framework to achieve dimensional upgrade from two-dimensional image to three-dimensional multi-component digital rock, and generate small-sized three-dimensional multi-component digital rock.

[0014] The specific process of GAN is as follows: extract real image slices as samples from the standardized training image TI2; the generator maps the latent vectors to generate a three-dimensional voxel grid, and extracts two-dimensional slices along the X, Y, and Z directions to obtain generated samples; form matching data pairs between real samples and generated samples, and input the two sets of data into the two-dimensional discriminator simultaneously to complete feature comparison and discrimination processing.

[0015] S4. Perform qualitative analysis and quantitative comparison on all small-sized three-dimensional multi-component digital rocks generated in step S3, and select the small-sized three-dimensional multi-component digital rocks with the highest fidelity as the basic qualified model for subsequent scale upgrades.

[0016] The dual verification includes: ① Qualitative verification of microstructure, observing the mineral distribution morphology, particle contact relationship, and pore structure characteristics of the small-sized 3D model, and combining the diagenetic characteristics of the original rock sample for qualitative evaluation; ② Quantitative verification of mineral composition, statistically analyzing the volume fraction of each mineral phase and pore in the small-sized 3D model, comparing it one by one with the percentage of the corresponding mineral components in the training image TI2, calculating the relative error of each component content and the overall average relative error, and selecting models with consistent mineral composition.

[0017] S5. The qualified small-sized three-dimensional multi-component digital rocks selected in step S4 are used as training images for the MPS algorithm. The model size is upgraded by adopting a modeling method based on cross-correlation function, and multiple large-sized three-dimensional multi-component digital rocks are reconstructed.

[0018] The specific steps are as follows: ① Scan the original training image using a search template, randomly select a pattern, and insert it into the simulation domain Ω of the fitted image. Set this pattern as pattern I, and set the search template size to 1 / 8 to 1 / 4 of the entire training image; ② Select 1 / 30 to 1 / 10 of pattern I as the overlapping area between the current pattern and the pattern to be determined. Use the voxels in the overlapping area as known data to constrain the selection of the next pattern. Further scan the training image from left to right and from bottom to top using the search template to extract n selectable patterns P. n ④ Find n patterns P n The correlation between the simulation domain and the blank pattern containing the overlapping area is calculated and sorted. The pattern with the strongest correlation is selected and set as the next pattern. Steps 2 to 4 are repeated until the entire simulation domain Ω is filled to obtain a large-size three-dimensional multi-component digital rock.

[0019] S6. Perform a second accuracy verification on the large-size three-dimensional multi-component digital rock reconstructed in step S5. The verification process is the same as in S4. Select the large-size three-dimensional multi-component digital rock with the highest fidelity.

[0020] S7. The mineral phases of the large-size three-dimensional multi-component digital rock obtained in step S6 are re-encoded and the mineral colors are labeled to obtain a standardized three-dimensional multi-component digital rock model S3D1.

[0021] S8. Based on the standardized three-dimensional multi-component digital rock S3D1, three-dimensional multi-component digital rock S3D2 with real geological characteristics is constructed by extracting the volume of different types of laminar digital rock particles, preparing slices, and orderly superimposing and recombining them.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] (1) This invention combines the advantages of accurate mineral identification in AMICS experiments with the advantages of deep learning in GANs, which not only ensures the authenticity and effectiveness of training images, but also minimizes the computational load of deep learning, providing a high-quality data foundation for the accurate construction of multi-component digital rocks.

[0024] (2) Before upgrading to MPS scale, the present invention performs qualitative and quantitative small-size model accuracy verification steps, and selects high-fidelity models from two core dimensions: microstructure and mineral composition. This completely solves the technical problem that direct upgrading of existing technologies easily amplifies model errors and leads to distortion of large-size models, and greatly improves the overall fidelity of the final digital rock model.

[0025] (3) This invention selects the MPS algorithm based on the cross-correlation function to scale up qualified small-sized models. By reasonably setting the search template and overlapping region parameters, it can realize the rapid construction of large-sized models while ensuring the fidelity of the model structure, thus solving the problems of low efficiency and large amount of computation in the scale upgrading of existing algorithms.

[0026] (4) Based on three-dimensional digital rocks with different types of laminae, this invention extracts the volume of laminae along the vertical bedding direction, designs personalized thickness ratio schemes, and alternately superimposes and recombines them in three-dimensional space. This enables the accurate reproduction of laminae structures with different types, thicknesses and configuration relationships, making the model more consistent with the laminae geological characteristics of real reservoirs and filling the gap in existing technologies that cannot effectively construct laminae multi-component digital rocks.

[0027] (5) This invention overcomes the disadvantages of traditional physical experimental methods, such as high cost, poor economy, inability to identify multiple components, difficulty in balancing multiple components and accuracy in current numerical simulation methods, and the incompatibility of macroscopic laminar features and microscopic multi-component heterogeneity. It integrates the advantages of AMICS and deep learning. The constructed digital rock model contains multiple mineral components and can quickly generate multiple laminar multi-component digital rock models in batches. Compared with ideal models and single-mineral digital rock models, the accuracy and laminar multi-component digital rock model can more accurately predict the porosity, permeability, elastic parameters and mechanical properties of rocks. It can truly characterize the features of laminar reservoir rocks, while taking into account both accuracy and economy. It lays a solid foundation for subsequent simulation of diagenesis based on laminar multi-component digital rock models, as well as the study of reservoir physical and mechanical properties. Attached Figure Description

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0029] Figure 1 This invention provides a flowchart for constructing a three-dimensional lamellar multi-component digital rock using the "AMICS+GAN-MPS" method.

[0030] Figure 2 This is a flowchart of the GAN deep learning method provided in the embodiments of the present invention;

[0031] Figure 3This invention provides a flowchart for constructing a three-dimensional lamellar multi-component digital rock using the "AMICS+GAN-MPS" method.

[0032] Figure 4 This invention provides a three-dimensional multi-component digital rock of sample S1 reconstructed using a deep learning GAN algorithm based on two-dimensional AMICS images, as provided in this embodiment.

[0033] Figure 5 This invention provides a large-scale three-dimensional multi-component digital rock model of the S1 sample, constructed based on the multi-point geostatistical method MPS, which includes felsic, dolomitic, mixed, and clayey laminae.

[0034] Figure 6 This invention provides a comparison chart of the mineral content of a three-dimensional multi-component digital rock sample S1 generated by the GAN algorithm and a large-size three-dimensional digital rock constructed by the multi-point geostatistical method MPS with a two-dimensional training image.

[0035] Figure 7 The present invention provides a large-size three-dimensional multi-component digital rock with different types of laminae, which is constructed by extracting sub-volumes and orderly superimposing them. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] Taking a shale sample S1 from an oilfield as an example, the specific implementation steps of this invention are described in detail with reference to the accompanying drawings. This invention constructs a three-dimensional lamellar multi-component digital rock using the "AMICS+GAN-MPS" method. AMICS is used to acquire high-resolution, high-fidelity two-dimensional training images, the GAN algorithm is used to generate the three-dimensional multi-component digital rock, and MPS is used for size upgrading of the three-dimensional multi-component digital rock. Finally, by extracting and combining different types of lamellar sub-volumes, the lamellar multi-component digital rock model is constructed. The specific implementation steps are as follows:

[0038] S1: Acquire two-dimensional training images. Samples are scanned using AMICS experiments to obtain high-resolution initial two-dimensional mineral distribution images. Regions with typical mineral assemblage features and representative texture structures are selected from these images as training images (TI1) for the deep learning model. This method effectively reduces the number of training samples while fully preserving key features of mineral distribution, laying the foundation for subsequent training and improved recognition accuracy of multi-component digital rock models.

[0039] Mineral scan images are obtained using AMICS, with a pixel count of N1*N2. Representative regions are selected from these scan images using the REV method as the original training images. If the mineral content of a region differs from that of the scan image by less than 10%, the region is considered a REV image with a pixel count of N32.

[0040] S2. Training Image Preprocessing: Standardized multi-component training images are obtained. The TI1 training image selected in step S1 is preprocessed. In MATLAB, the original RGB format training image is converted to a grayscale format image. Then, according to the category information of different mineral phases, the grayscale image is encoded in order starting from 0, so that each mineral phase is distinguished in the image with different numerical labels, ensuring the uniqueness and identifiability of mineral phase information. Distortionless scaling of the image is performed in MATLAB to obtain the final training image TI2 with N42 pixels, providing standardized training data for the subsequent construction of multi-component digital rocks.

[0041] S3. Dimensional upgrade based on deep learning GAN: The standardized training image TI2 obtained in S2 is used as input data. The GAN deep learning algorithm is introduced for model training and generation. GAN learns the spatial distribution features, morphological features and correlations between different mineral phases in the two-dimensional training image. Under the adversarial generative network framework, it realizes the mapping and reconstruction from two-dimensional slices to three-dimensional multi-component structures. After sufficient training, multiple three-dimensional multi-component digital rock models with statistical consistency and spatial coherence are generated using the trained GAN model.

[0042] Figure 2 As shown, the training image TI2 is input into the GAN deep learning algorithm to generate multiple three-dimensional multi-component digital rocks. The steps include generating three-dimensional volumes and extracting slices, comparing the judge and optimizing the generator, preserving multi-component features and outputting high-fidelity data.

[0043] ① Generation of 3D Volume and Slice Extraction: The generator G maps the input latent vector into a 3D cube volume through multiple transposed convolutions, ensuring uniform information distribution. To achieve effective comparison of 2D training images, 2D slices are extracted along the X, Y, and Z directions of the generated volume to avoid image artifacts. This is crucial for the accuracy of the spatial distribution of mineral grains and pore structure in multi-mineral digital rocks.

[0044] ② Discriminator comparison and generator optimization: The extracted 2D slices are input into the discriminator D and compared with random cropping of real 2D training images. The discriminator parameters are updated to accurately identify the differences between real slices and generated slices. The results are fed back to the generator to guide the generator to continuously optimize the 3D output, ensuring that the final generated digital rock is highly close to the real mineral scan image.

[0045] ③ Multi-component feature preservation and high-fidelity output: The trained generator can quickly generate high-resolution three-dimensional digital rocks. Each voxel accurately corresponds to the mineral type and pore size, preserving the mineral characteristics and the connectivity and distribution patterns of the pore space. It truly reflects the microstructural characteristics of the reservoir, providing reliable and repeatable digital rock data for subsequent numerical simulation, pore connectivity analysis and reservoir property analysis.

[0046] Figure 4 The three-dimensional multi-component digital rock of the S1 sample is reconstructed using the GAN algorithm based on two-dimensional AMICS images. The four two-dimensional training images of the S1 sample are obtained after the two-dimensional high-precision mineral scanning images obtained from the AMICS experiment are screened, cropped and scaled for typical regions. The size of each image is 360×360 pixels. The three-dimensional multi-component digital rock of the S1 sample generated by the GAN algorithm is 312×312×312 voxels, which truly reflects the microstructure characteristics of the reservoir.

[0047] S4. Accuracy Verification of Small-Scale 3D Multi-Component Digital Rocks: Screening Qualified Models. All small-scale 3D multi-component digital rocks generated in step S3 undergo dual verification through qualitative analysis and quantitative comparison: ① Qualitative Verification of Microstructure: Observe the mineral distribution morphology, particle contact relationships (point contact, line contact, mosaic contact, etc.), and pore structure characteristics (pore morphology, connectivity, pore size distribution, etc.) of the small-scale 3D models. Combine this with the diagenetic characteristics (dissolution, cementation, etc.) of the original rock samples for qualitative evaluation. Screen models with highly consistent microstructures to ensure that each feature is highly consistent with the core features of the real reservoir and the 2D training image, without obvious image artifacts or structural distortions. ② Quantitative Verification of Mineral Composition: Statistically calculate the volume fraction of each mineral phase and pore in the small-scale 3D model. Compare this with the percentage of the corresponding mineral components in the 2D training image TI2, calculate the relative error of each component content and the overall average relative error, and screen models with highly consistent mineral composition to ensure consistency in mineral composition across the generated volume. Based on the above evaluation indicators, a comprehensive accuracy verification was performed on all small-sized three-dimensional multi-component digital rocks generated in step S3. The small-sized three-dimensional multi-component digital rocks with the highest fidelity were selected as the basic qualified models for subsequent scale upgrades, thus avoiding error amplification during the scale upgrade process from the source.

[0048] Figure 4 As shown, the felsic laminae are mainly composed of quartz and feldspar, with relatively well-developed primary intergranular pores. Unstable particles such as feldspar are easily dissolved, forming dissolution pores at the grain edges of albite and numerous intragranular dissolution pores, and a small amount of intragranular dissolution pores and grain edge dissolution pores in potassium feldspar. The intergranular contact is mainly point-line. The dolomitic laminae are mainly composed of carbonate minerals, with the most developed intercrystalline pores. The pore size is small, and the particles exhibit mosaic contact. In the mixed laminae, the content of terrigenous clastic minerals (quartz, feldspar) and carbonate minerals is close. Intergranular pores, intercrystalline pores, and dissolution pores are all well-developed, with complex contact relationships, including both point and mosaic contacts. The clayey laminae are mainly composed of clay minerals and serve as the matrix support structure. This indicates that the generated three-dimensional multi-component digital rock maintains the mineral shapes and combination patterns of the two-dimensional image, with good consistency, and also possesses high structural realism and rationality.

[0049] Figure 6As shown, in the two-dimensional training images, the porosity, quartz, calcite, albite, potassium feldspar, organic matter, and chlorite volume percentages in the felsic, dolomitic, mixed, and clayey lamellar layers are 7.21%, 6.33%, 31.90%, 27.22%, 14.81%, 3.38%, and 9.15%; 5.38%, 4.60%, 55.16%, 11.91%, 13.92%, 3.15%, and 5.88%; 4.48%, 2.37%, 33.24%, 39.87%, 3.83%, 1.14%, and 15.06%; and 3.06%, 11.91%, 4.38%, 6.09%, 0.15%, 2.76%, and 71.65%, respectively. The porosity, volume percentages of quartz, calcite, albite, potassium feldspar, organic matter, and chlorite in small-sized three-dimensional multi-component digital rocks constructed using GAN deep learning were 6.56%, 8.60%, 29.80%, 26.01%, 17.73%, 1.75%, and 9.54% in felsic, dolomitic, mixed, and clayey laminae; 5.84%, 8.92%, 52.59%, 9.17%, 15.92%, 2.17%, and 5.39% in quartz, dolomitic, mixed, and clayey laminae; 2.57%, 2.96%, 35.23%, 40.44%, 3.70%, 0.86%, and 14.23% in quartz, dolomitic, mixed, and clayey laminae; and 2.86%, 14.09%, 3.25%, 5.82%, 0.06%, 3.38%, and 70.55% in quartz, dolomitic, mixed, and clayey laminae. The calculated average relative errors were 0.1830, 0.2632, 0.1555, and 0.1979, respectively. Overall, the average relative errors of both groups of samples were at a low level, indicating that the small-sized three-dimensional multi-component digital rocks generated by GAN are quite close to the two-dimensional images in terms of mineral content, with small differences between the mineral components, and the generated results have good compositional consistency and reliability.

[0050] S5. Scale upgrade of three-dimensional multi-component digital rocks based on multi-point geostatistics (MPS): The qualified small-sized three-dimensional multi-component digital rock model obtained in step S4 is used as input data. The algorithm based on cross-correlation function in multi-point geostatistics is used to augment and synthesize the image. This algorithm synthesizes simulation images with large simulation size and fast execution speed.

[0051] The specific steps are as follows: ① Scan the original training image using a search template, randomly select a pattern, and insert it into the simulation domain Ω of the fitted image. Set this pattern as pattern I. The search template size is usually set to 1 / 8 to 1 / 4 of the entire training image. ② Select 1 / 30 to 1 / 10 of pattern I as the overlapping area between the current pattern and the pattern to be determined. Use the voxels in the overlapping area as known data to constrain the selection of the next pattern. Further scan the training image from left to right and from bottom to top using the search template to extract n selectable patterns Pn. ④ Calculate the correlation between the n patterns Pn and the blank pattern containing the overlapping area, sort the correlations, select the pattern with the strongest correlation, and set it as the next pattern. ⑤ Repeat steps ② to ④ until the entire simulation domain Ω is filled to obtain a large-size three-dimensional multi-component digital rock.

[0052] like Figure 5 As shown, the large-scale three-dimensional multi-component digital rock models of the S1 sample, including felsic, dolomitic, mixed, and clayey laminae, were reconstructed based on the multi-point geostatistical method MPS. The models are all 400×400×1000 voxels in size.

[0053] S6. Accuracy Verification of Large-Scale Three-Dimensional Multi-Component Digital Rocks: Screening High-Fidelity Models. A second accuracy verification was performed on the large-scale three-dimensional multi-component digital rocks reconstructed in S5. The verification process was consistent with S4, including qualitative analysis and quantitative comparison. The focus was on verifying whether the mineral composition and distribution, and pore structure of the model after scale upgrade remained consistent with the original sample. The large-scale three-dimensional multi-component digital rocks with the highest fidelity were screened to ensure the high fidelity of the large-scale model and provide an accurate single-layer multi-component model basis for the subsequent construction of laminated digital rocks.

[0054] Figure 5 It can be seen that the large-scale three-dimensional multi-component digital rock model maintains the spatial continuity of minerals, the connectivity and distribution pattern of pore space, and reproduces the characteristics of real shale reservoirs in a good manner in terms of microstructure.

[0055] Figure 6The histogram shows the porosity, quartz, calcite, albite, potassium feldspar, organic matter, and chlorite volume percentages in the large-scale three-dimensional digital rock, including felsic, dolomitic, mixed, and clayey lamellar structures. These percentages are: 6.48%, 8.29%, 31.97%, 23.23%, 18.66%, 1.72%, 9.65%; 5.66%, 8.09%, 53.35%, 10.13%, 14.79%, 2.01%, 5.97%; 2.80%, 2.59%, 35.66%, 41.80%, 2.78%, 0.82%, 13.55%; and 2.84%, 14.01%, 3.94%, 6.64%, 0.06%, 3.52%, 69.00%. The average relative errors calculated from the two-dimensional training images were 0.1953, 0.2043, 0.1780, and 0.1949, respectively. Overall, the average relative errors of each layer remained within a low range, indicating that the generated large-size three-dimensional multi-component digital rock has a high degree of consistency with the two-dimensional images in terms of mineral content, and the differences between different mineral components are not significant. The results show that it has good compositional consistency and reliability.

[0056] S7. Post-processing enables standardized characterization of multi-component minerals. To address the issue of disordered mineral numbering in the 3D digital rock model obtained by deep learning, the 3D voxel data is renumbered one by one according to the grayscale labeling rules set in the preprocessing stage, restoring the correct mineral category mapping relationship. This ensures the consistency of mineral categories from the input image to the final digital rock model, providing reliable data support for subsequent pore structure analysis, diagenetic simulation, and physical property prediction.

[0057] Based on the RGB color values ​​of each mineral in the mineral scan image, the minerals in the generated digital rock model are assigned corresponding colors to ensure the consistency of visual representation of different images. This facilitates intuitive identification of mineral types and makes subsequent visualization analysis and model display easier.

[0058] Through the above processing, a three-dimensional multi-component digital rock model S3D1 with clear mineral phase identification, distinct colors, and complete structural information is obtained.

[0059] S8. Construct a three-dimensional lamellar multi-component digital rock. Based on the standardized three-dimensional multi-component digital rock S3D1, firstly, from large-size three-dimensional multi-component digital rocks corresponding to different types of lamellarity, a series of lamellar slices of specific thickness are cut along the direction perpendicular to the bedding. The slice thickness can be flexibly set according to the lamellar thickness characteristics of the real reservoir. Then, based on the lamellar distribution law of the real reservoir, various lamellar thickness ratios and arrangement schemes are designed. Subsequently, according to the alternating arrangement order, different types of lamellar slices are superimposed and recombined in three-dimensional space according to the preset scheme. Finally, a three-dimensional lamellar multi-component digital rock S3D2 with real geological characteristics containing different lamellar types, degrees and configuration relationships is constructed.

[0060] Figure 7 The image shows the final large-size three-dimensional multi-component digital rocks based on different types of laminae. These rocks were constructed by extracting sub-volumes and orderly superimposing them to form three-dimensional laminar multi-component digital rocks of different types and thicknesses. The dimensions of each rock are 400×400×2400 voxels.

[0061] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the protection scope of the claims and specification of the present invention.

Claims

1. A method for modeling three-dimensional layered multi-component digital rocks, characterized in that, Includes the following steps: S1. Two-dimensional mineral scanning images of rock samples were obtained through AMICS experiments, and representative local areas were selected as the original training images TI1. S2. Convert the selected original training image TI1 from an RGB color image to a grayscale image, and assign continuous grayscale encoding values ​​to different mineral phases in a preset order. Scale the image without distortion to obtain a standardized training image TI2 that meets the input requirements of the deep learning model. S3. Input the standardized training image TI2 into the GAN deep learning framework to realize the dimensional upgrade from two-dimensional image to three-dimensional multi-component digital rock, and generate small-sized three-dimensional multi-component digital rock. S4. Perform qualitative analysis and quantitative comparison on all small-sized three-dimensional multi-component digital rocks generated in step S3, and select the small-sized three-dimensional multi-component digital rocks with the highest fidelity as the basic qualified model for subsequent scale upgrades. S5. The qualified small-sized three-dimensional multi-component digital rocks selected in step S4 are used as training images for the MPS algorithm. The model size is upgraded by adopting a modeling method based on cross-correlation function, and multiple large-sized three-dimensional multi-component digital rocks are reconstructed. S6. Perform a second accuracy verification on the large-size three-dimensional multi-component digital rock reconstructed in step S5. The verification process is the same as in S4. Select the large-size three-dimensional multi-component digital rock with the highest fidelity. S7. The mineral phases of the large-size three-dimensional multi-component digital rock obtained in step S6 are re-encoded and the mineral colors are labeled to obtain a standardized three-dimensional multi-component digital rock model S3D1. S8. Based on the standardized three-dimensional multi-component digital rock S3D1, three-dimensional layered multi-component digital rock S3D2 with real geological characteristics is constructed by extracting the volume of different types of layered digital rock particles, preparing slices, and orderly superimposing and recombining them. Step S5 is performed as follows: ① Scan the original training image using a search template, randomly select a pattern, and insert it into the simulation domain Ω of the fitted image. Set this pattern as pattern I, and set the search template size to 1 / 8 to 1 / 4 of the entire training image; ② Select 1 / 30 to 1 / 10 of pattern I as the overlapping area between the current pattern and the pattern to be determined. Use the voxels in the overlapping area as known data to constrain the selection of the next pattern. Further scan the training image from left to right and from bottom to top using the search template to extract n selectable patterns P. n ④ Find n patterns P n The correlation between the simulation domain and the blank pattern containing the overlapping area is calculated and sorted. The pattern with the strongest correlation is selected and set as the next pattern. Steps 2 to 4 are repeated until the entire simulation domain Ω is filled to obtain a large-size three-dimensional multi-component digital rock.

2. The method for modeling three-dimensional layered multi-component digital rocks according to claim 1, characterized in that, In step S3, the specific process of the GAN is as follows: extract real image slices from the standardized training image TI2 as real samples; the generator maps the latent vectors to generate a three-dimensional voxel grid, and extracts two-dimensional slices along the X, Y, and Z directions to obtain generated samples; form matching data pairs between real samples and generated samples, and input the two sets of data synchronously into the two-dimensional discriminator to complete feature comparison and discrimination processing.

3. The method for modeling three-dimensional layered multi-component digital rocks according to claim 1, characterized in that, In step S4, the dual verification includes: ① qualitative verification of microstructure, observing the mineral distribution morphology, particle contact relationship, and pore structure characteristics of the small-sized three-dimensional model, and performing a qualitative evaluation in combination with the diagenetic characteristics of the original rock sample; ② quantitative verification of mineral composition, statistically analyzing the volume fraction of each mineral phase and pore in the small-sized three-dimensional model, comparing it one by one with the percentage of the corresponding mineral components in the training image TI2, calculating the relative error of each component content and the overall average relative error, and selecting models with consistent mineral composition.

4. The method for modeling three-dimensional layered multi-component digital rocks according to claim 1, characterized in that, Step S8 is specifically implemented as follows: Based on the standardized three-dimensional multi-component digital rock S3D1, a series of laminar slices of specific thickness are first cut from the large-size three-dimensional multi-component digital rock corresponding to different types of laminar layers along the direction perpendicular to the bedding. The slice thickness is flexibly set according to the laminar thickness characteristics of the real reservoir. Based on the distribution pattern of the real reservoir's laminae, various laminae thickness ratios and arrangement schemes are designed. Then, in an alternating arrangement order, different types of laminae slices are superimposed and recombined in three-dimensional space according to the preset scheme, and finally a three-dimensional laminated multi-component digital rock S3D2 with real geological characteristics containing different laminae types, thicknesses and configuration relationships is constructed.

Images