A method for generating three-dimensional digital core

Abstract

The present application belongs to the field of digital core technology. The present application provides a three-dimensional digital core generation method. The disclosed embodiment obtains three-dimensional digital data of a real core and pre-processes it to obtain a core training data set; based on the DDIM diffusion model and the generation model architecture of CSAUnet, a digital core generation model is constructed, and the training data set is input for training, and the network parameters are gradually optimized to finally obtain the optimal model. During the generation process, by inputting noise and geological condition information, a target digital core image that meets the target geological characteristics is generated; the generated target digital core image has the geological characteristics and visual effects of a real core, and can accurately simulate the porosity and bedding structure inside the core.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of digital core technology, and in particular to a method for generating a three-dimensional digital core. Background Art

[0002] With the continuous advancement of oil and gas exploration and development, digital core technology, as a key tool, plays a vital role in oil and gas field development, reservoir evaluation, and fluid simulation. Digital core technology simulates rock microstructure, porosity, and formation characteristics through high-precision 3D modeling and simulation, providing authentic and reliable digital support for oil and gas exploration. Existing digital core generation methods often use methods such as CT scanning and nuclear magnetic resonance imaging (NMR) to acquire core image data, followed by modeling and reconstruction using image processing algorithms. However, traditional digital core generation methods still face many challenges, such as long processing time, high cost, and low generation accuracy. This is particularly true when it comes to meticulously simulating rock microstructure, nanoscale pores, and the connections between pores. Existing technologies often struggle to meet the high-precision and high-efficiency requirements.

[0003] For example, the patent "A Method and System for Constructing Digital Cores to Simulate Nano- and Micro-Porosity in Shale Matrix" (Patent No. CN 115115783B) proposes a digital core construction method based on a dual-discriminator generative adversarial network (GAN). Shale matrix core data is preprocessed using CT scans to generate a training dataset, which is then trained using a generative adversarial network to generate high-precision three-dimensional digital cores. By introducing a dual-discriminator optimization process, this method overcomes the time and cost issues associated with traditional methods in constructing digital cores, significantly improving the accuracy of three-dimensional reconstruction of nanoscale pores in the shale matrix. However, GAN methods are often susceptible to training instability and model collapse, and have limitations in capturing detailed rock features and pore structure.

[0004] Furthermore, the patent "Multi-Condition Constrained 3D Digital Core Generation Method Based on Diffusion Model" (Patent No. CN117830510B) utilizes a multi-condition constrained generation method based on the 3DUnet network. This method obtains multiple 3D digital core samples and their label information, and trains a generation model using various geological conditions (such as porosity and pore size standard deviation). This method generates high-fidelity 3D digital cores under these constraints. This method faithfully reproduces the internal pore structure of the rock, improving the quality of the generated cores. However, while the 3DUnet method performs well when processing high-dimensional images, it still faces difficulties in modeling complex geological features (such as bedding structure and pore distribution), and its performance in noise suppression and refined modeling remains limited.

[0005] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0006] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the Invention

[0007] The purpose of the embodiments of the present disclosure is to provide a three-dimensional digital core generation method, thereby overcoming one or more problems caused by the limitations and defects of related technologies, at least to a certain extent.

[0008] According to an embodiment of the present disclosure, a method for generating a three-dimensional digital core is provided, the method comprising:

[0009] The noise data and geological conditions are input into the trained CSAUnet model for processing; the CSAUnet model includes a dual-channel convolution block, a downsampling module, an upsampling module, a dual attention module, and a convolution module;

[0010] The two-channel convolution block extracts the features of the noise data to obtain the initial feature map;

[0011] The downsampling module performs three downsampling operations on the initial feature map, and then processes it using the dual attention module to obtain the first feature image block, the second feature image block, and the third feature image block. Each downsampling operation is processed by two first dual convolution blocks, and then auxiliary features such as time embedding, porosity information, and lithology category are introduced for integration.

[0012] The upsampling module performs three upsampling operations on the third feature image block. Each upsampling operation is processed by two second double convolution blocks, and then auxiliary features such as time embedding, porosity information, and lithology category are introduced. After splicing and fusing with the first and second feature image blocks, they are processed by the dual attention module to obtain the fused features.

[0013] The convolution module processes the fused features to generate an initial three-dimensional digital core image;

[0014] The DDIM diffusion model is used to gradually denoise the noise in the initial three-dimensional digital core image to generate the denoised target three-dimensional digital core image.

[0015] Furthermore, the downsampling module includes a first downsampling block, a second downsampling block and a third downsampling block, and the upsampling module includes a first upsampling block, a second upsampling block and a third upsampling block.

[0016] Furthermore, the downsampling module performs three downsampling operations on the initial feature map, and then processes it using the dual attention module to obtain a first feature image block, a second feature image block, and a third feature image block, respectively, including:

[0017] The first downsampling block downsamples the initial feature map. After two 3D convolutions and three activation functions in the two first biconvolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The dual attention module is used for processing to obtain the first feature image block with a size of 40×40×40 and 16 channels.

[0018] The second downsampling block downsamples the first feature image block. After two 3D convolutions and three activation functions in the two first double convolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The block is then processed using a dual attention module to obtain a second feature image block with a size of 20×20×20 and 32 channels.

[0019] The third downsampling block performs a downsampling operation on the two feature image blocks. After two 3D convolutions and three activation functions in the two first double convolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features of time embedding, porosity information and lithology category are introduced, and the dual attention module is used for processing to obtain a target feature image block with a size of 10×10×10 and 128 channels.

[0020] Furthermore, the upsampling module performs three upsampling operations on the third feature image block, and each upsampling operation is processed by two second dual convolution blocks, and auxiliary features of time embedding, porosity information and lithology category are introduced, and after splicing and fusing with the first feature image block and the second feature image block, the dual attention module is used for processing to obtain the fused feature, including:

[0021] The first upsampling block performs an upsampling operation on the third feature image block. After two three-dimensional convolutions and a cubic activation function in the two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The dual attention module is used for processing to obtain the first fused feature image block with a size of 20×20×20 and 32 channels.

[0022] The second upsampling block performs an upsampling operation on the first fused feature image block. After two 3D convolutions and a cubic activation function in two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. After splicing and fusing with the second feature image block, it is processed using the dual attention module to obtain a second fused feature image block with a size of 40×40×40 and 16 channels.

[0023] The third upsampling block performs an upsampling operation on the second fused feature image block. After two three-dimensional convolutions and three activation functions in the two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features of time embedding, porosity information and lithology category are introduced. After splicing and fusing with the first feature image block, it is processed using the dual attention module to obtain a fused feature with a size of 80×80×80 and 8 channels.

[0024] Furthermore, the convolution module processes the fused features to generate an initial three-dimensional digital core image, including:

[0025] The convolution module performs a convolution operation with a convolution kernel of 1×1×1 on the fused features with a size of 80×80×80 and a channel number of 8 to obtain an initial three-dimensional digital core image with a size of 80×80×80 and a channel number of 1.

[0026] Furthermore, the step of gradually denoising the initial three-dimensional digital core image using the DDIM diffusion model to generate a denoised target three-dimensional digital core image includes:

[0027] The initial three-dimensional digital core image was sampled using the DDIM diffusion model;

[0028] By iteratively applying the denoising equation, a real image is generated from the noise and optimized through conditional control to generate the target three-dimensional digital core image:

[0029]

[0030] in, is the sample at time t, Indicates category condition information, Represents the porosity condition information, is the parameter that controls the degree of noise diffusion at time t, is the parameter that controls the noise diffusion degree at time t-1, is the noise function, for The cumulative product of for The cumulative product of .

[0031] Furthermore, the method further comprises:

[0032] Acquire the original image of the real core and perform preprocessing to obtain a three-dimensional core data set;

[0033] Based on the CSAUnet model and the DDIM diffusion model, a digital core generation model was constructed;

[0034] The three-dimensional core data set and the corresponding geological conditions are input into the digital core generation model for training, and a weighted loss function is introduced in the training process to obtain a trained digital core generation model.

[0035] Furthermore, the weighted loss function is:

[0036]

[0037] in, To generate pixel values in an image, is the pixel value in the real image, N is the total number of pixels in the image, is the weighting factor:

[0038]

[0039] in, is the geological feature weight, is the noise weight, is the first hyperparameter, is the second hyperparameter.

[0040] The technical solutions provided by the embodiments of the present disclosure may have the following beneficial effects:

[0041] In the embodiments of the present disclosure, the above-mentioned three-dimensional digital core generation method, on the one hand, obtains three-dimensional digital data of a real core and preprocesses it to obtain a core training data set; based on the DDIM diffusion model and the CSAUnet generation model architecture, a digital core generation model is constructed, and is trained by inputting the training data set, gradually optimizing the network parameters, and finally obtaining the optimal model. During the generation process, by inputting noise and geological condition information, a target digital core image that meets the target geological characteristics is generated; the generated target digital core image has the geological characteristics and visual effects of a real core and can accurately simulate the porosity and bedding structure inside the core. On the other hand, the self-attention mechanism of the DDIM model and the CSAUnet model can more finely restore the geological characteristics in the core, including porosity, pore distribution, and rock type. In particular, the generation accuracy is significantly improved in capturing complex bedding structures and subtle geological features. By introducing an enhanced MSE loss function that factors in noise level and geological features, the noise processing process is further optimized, reducing the interference of noise on core image generation while improving the quality of detail recovery. This is particularly significant when simulating subtle pore structures and small geological differences. Combining the advantages of the DDIM model and the CSAUnet model, the model is capable of generating more accurate 3D digital cores under complex geological conditions, demonstrating greater adaptability and generation accuracy, especially in simulating complex geological features such as bedding structures and micropores. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification, are used to explain the principles of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0043] Figure 1 A diagram showing the steps of a method for generating a three-dimensional digital core in an exemplary embodiment of the present disclosure;

[0044] Figure 2 A specific flow chart showing a method for generating a three-dimensional digital core in an exemplary embodiment of the present disclosure is provided;

[0045] Figure 3 A three-dimensional example diagram of a digital core sample in an exemplary embodiment of the present disclosure is shown;

[0046] Figure 4 shows a porosity distribution diagram of a sample in an exemplary embodiment of the present disclosure;

[0047] Figure 5 A structural diagram of the CSAUNet model in an exemplary embodiment of the present disclosure is shown;

[0048] Figure 6 A structural diagram of a dual convolution module in an exemplary embodiment of the present disclosure is shown;

[0049] Figure 7 A structural diagram of a dual attention module in an exemplary embodiment of the present disclosure is shown;

[0050] Figure 8 A structural diagram of a downsampling module in an exemplary embodiment of the present disclosure is shown;

[0051] Figure 9 A structural diagram of an upsampling module in an exemplary embodiment of the present disclosure is shown;

[0052] Figure 10 A fully convolutional self-attention network in an exemplary embodiment of the present disclosure is shown;

[0053] Figure 11 shows the training loss curve of the CSAUnet model in an exemplary embodiment of the present disclosure;

[0054] Figure 12 shows two-dimensional slices of real samples and generated carbonate and sandstone cores in an exemplary embodiment of the present disclosure;

[0055] Figure 13 A comparison diagram of real samples, digital core samples generated by CGAN and DDIM in an exemplary embodiment of the present disclosure is shown;

[0056] Figure 14 A comparison diagram showing a DDIM-generated sample and a real sample of the pore space of carbonate rock in an exemplary embodiment of the present disclosure;

[0057] Figure 15 A comparison diagram showing a DDIM-generated sample and a real sample of the pore space of sandstone in an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0058] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0059] In addition, the accompanying drawings are merely schematic illustrations of embodiments of the present disclosure and are not necessarily drawn to scale. Like reference numerals in the figures represent like or similar parts, and thus repeated descriptions thereof will be omitted. Some of the blocks shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically separate entities.

[0060] This example embodiment provides a method for generating a three-dimensional digital core. Figure 1 As shown in , the three-dimensional digital core generation method may include:

[0061] Step S101: Input the noise data and geological conditions into the trained CSAUnet model for processing; wherein the CSAUnet model includes a dual-channel convolution block, a downsampling module, an upsampling module, a dual attention module and a convolution module;

[0062] Step S102: A dual-channel convolution block extracts features of the noise data to obtain an initial feature map;

[0063] Step S103: The downsampling module performs three downsampling operations on the initial feature map, and then processes it using the dual attention module to obtain the first feature image block, the second feature image block, and the third feature image block respectively; wherein each downsampling operation is processed by two first dual convolution blocks, and auxiliary features of time embedding, porosity information, and lithology category are introduced for integration;

[0064] Step S104: The upsampling module performs three upsampling operations on the third feature image block. Each upsampling operation is processed by two second dual convolution blocks, and auxiliary features such as time embedding, porosity information, and lithology category are introduced. After splicing and fusing with the first feature image block and the second feature image block, the upsampling module is processed by the dual attention module to obtain a fused feature.

[0065] Step S105: The convolution module processes the fused features to generate an initial three-dimensional digital core image;

[0066] Step S106: using the DDIM diffusion model to gradually denoise the noise in the initial three-dimensional digital core image to generate a denoised target three-dimensional digital core image.

[0067] The above-mentioned 3D digital core generation method, firstly, obtains 3D digital data of real rock cores and preprocesses it to generate a core training dataset. Then, based on the generative model architecture of the DDIM diffusion model and CSAUnet, a digital core generation model is constructed. This model is trained using the training dataset, and the network parameters are gradually optimized to ultimately obtain the optimal model. During the generation process, noise and geological condition information are input to generate a target digital core image that matches the target geological characteristics. The generated target digital core image possesses the geological characteristics and visual effects of a real rock core, accurately simulating the porosity and bedding structure within the core. Furthermore, the self-attention mechanism of the DDIM and CSAUnet models enables more detailed restoration of geological features within the core, including porosity, pore distribution, and rock type. Generation accuracy is significantly improved, particularly in capturing complex bedding structures and subtle geological features. By introducing an enhanced MSE loss function that factors in noise level and geological feature weighting, the noise processing process is further optimized, reducing the interference of noise on the core image generation while improving the quality of detail restoration, particularly when simulating subtle pore structures and minor geological variations. Combining the advantages of the DDIM model and the CSAUnet model, the model can generate more accurate three-dimensional digital cores under complex geological conditions, especially in the simulation of complex geological features such as bedding structure and micropores, showing higher adaptability and generation accuracy.

[0068] Below, we will refer to Figures 1 to 15 Each step of the above-mentioned three-dimensional digital core generation method in this example embodiment is described in more detail.

[0069] In one embodiment, Figure 2 As shown in FIG, a specific flow chart of the three-dimensional digital core generation method.

[0070] 1. Obtain the original image of the real core obtained by CT scanning and obtain a three-dimensional core data set through preprocessing.

[0071] Specifically, the original image is obtained from the existing three-dimensional digital core dataset and binarized. The Otsu method is used to analyze the grayscale histogram of the image to determine the optimal threshold and calculate the porosity of the core image. Different degrees of noise are introduced into the dataset, and the rock type is marked at the same time.

[0072] More specifically, the 3D core dataset used in this example is the Digital Core Super-Resolution Dataset 1 (DRSRD1), which contains 2,000 high-resolution micro-CT 3D images covering the Bentheimer sandstone and Estaillades carbonate rocks. The sandstone images have a resolution of 3.8 microns, and the carbonate images have a resolution of 3.1 microns, which can reveal complex pore structures. Figure 3 As shown, sandstone is primarily composed of quartz sand grains, with minor amounts of feldspar and rock fragments, and exhibits good porosity and permeability. Carbonate rocks are calcareous, monomineralic rocks with both macropore and micropore porosity. The image provides three-phase segmentation data, including pore voxels, unresolved voxels, and solid voxels.

[0073] The image porosity is obtained through binarization and subsequent calculation. The Otsu method is used to automatically calculate the optimal threshold value and divide the core image into the porous part and the solid part. The pixel ratio of the porous part in the binary image is calculated to obtain the porosity value. According to the calculation, the porosity of the carbonate rock samples in the dataset ranges from 13% to 35%, and the porosity of the sandstone samples ranges from 14% to 32%. Figure 4 shown.

[0074] To increase the diversity of the dataset, data augmentation technology was used to rotate the images and expand the three-dimensional core dataset to 8,000 samples, thereby improving the robustness of the model on core images at different angles.

[0075] 2. The digital core generation model framework is based on the CSAUnet architecture, combining the feature extraction capabilities of the UNet structure and the detail optimization function of the self-attention mechanism. The UNet structure captures the spatial hierarchical features of the image through an encoder-decoder architecture and restores image details through skip connections. The self-attention mechanism uses a fully convolutional self-attention network to optimize feature focus and detail extraction, improving the accuracy and sophistication of the denoising process.

[0076] In this example, a 3D digital core generation model framework based on CSAUnet was constructed. This framework learns features from pre-processed core data and optimizes denoising capabilities. The model architecture is based on the traditional UNet structure and incorporates a self-attention mechanism to improve the model's performance when processing long sequence inputs, especially in enhancing the ability to identify and focus on important features during the denoising process. Figure 5 As shown in the figure, the model predicts and removes noise based on conditions such as time, category, and porosity. The details are as follows:

[0077] First, the core image Xt with a size of 1×80×80×80 is input into the dual-channel convolution block DoubleConvBlock (i.e., dual-channel convolution block) to extract the initial feature map, and the number of channels is gradually reduced to 8.

[0078] Through three layers of downsampling blocks DownBlock (i.e., downsampling module), each layer contains double convolution or attention blocks, extracting the first feature image block, the second feature image block, and the third feature image block respectively. The output channels increase from 16 → 32 → 64 → 128, and the spatial size decreases from 80×80×80 → 40×40×40 → 20×20×20 → 10×10×10;

[0079] In each downsampling stage, auxiliary features such as time embedding, porosity information, and lithology category are introduced and then spliced through MLP embedding to improve feature representation capabilities;

[0080] After the deepest convolution, the upsampling block UpBlock (i.e., upsampling module) is used for decoding and reconstruction. The number of channels is gradually restored to 128 → 64 → 32 → 16, and the spatial size is gradually enlarged to 10×10×10 → 20×20×20 → 40×40×40 → 80×80×80.

[0081] In each upsampling block, skip-connected features from the corresponding layer of the downsampling block are received and concatenated and fused. At the end of the decoding path, a 1×1×1 convolution is performed to generate a predicted "noise" map with the same size as the input (1×80×80×80). This map is used to guide noise estimation during the diffusion process.

[0082] like Figures 6 to 9 As shown in the figure, it is the structural diagram of the main functional units in the CSAUnet model; Figure 6 It is the structural diagram of the double convolution module; Figure 7 This is the structural diagram of the dual attention module; Figure 8 It is the structural diagram of the downsampling module; Figure 9 It is the structural diagram of the upsampling module.

[0083] Specifically, a) Double Convolution Block (DoubleConvBlock), consisting of two groups of Conv3d + GroupNorm + ReLU, is responsible for local feature extraction.

[0084] b) DoubleAttentionBlock, which consists of two layers: ConvAttentionBlock + ConvBlock, and introduces a spatial attention mechanism to enhance information selectivity.

[0085] c) The downsampling module (DownBlock) fuses the input features with information such as time, porosity, and category, and then encodes them. It is first activated through SiLU before entering the fully connected layer and broadcasting. It is then concatenated with the input features and then enters two convolutional modules.

[0086] d) The upsampling module (UpBlock) not only receives the feature map from the previous step, but also integrates the skip connection features Xdown from the downsampling path. At the same time, all embedded information and skip features are concatenated and input into two consecutive convolutional blocks to achieve high-resolution feature reconstruction.

[0087] In order to improve the accuracy of denoising, the model not only relies on the basic convolutional network structure, but also adopts an improved self-attention mechanism. Figure 10 As shown in the figure, the traditional self-attention mechanism is improved into a fully convolutional self-attention network, which enables the model to flexibly learn complex features in long time series and improve the ability of feature focusing.

[0088] 3. Input the preprocessed core data set and geological condition information into the generated CSAUnet model for model training. The training process includes:

[0089] 3-1. Input the preprocessed core data set and corresponding geological condition information into the generative model, and add noise data to enhance the robustness of the training. The geological condition information includes rock type and porosity.

[0090] 3-2. Set the training parameters, including the learning rate, batch size, number of training rounds, and the input noise level and output image size that match the model's generation task; the parameter settings in this embodiment are shown in Table 1.

[0091] Table 1 Training parameter settings

[0092]

[0093] 3-3. In the CSAUnet module, a fully convolutional architecture and self-attention mechanism are used to extract geological features and bedding structures from the input data to optimize the input data, accurately remove noise, and enhance image details;

[0094] 3-4. Introduce a weighted loss function to optimize the model training process by minimizing the difference between the generated image and the real image. The formula is as follows:

[0095]

[0096] in, To generate pixel values in an image, is the pixel value in the real image, and N is the total number of pixels in the image; is a weighting factor, which is weighted by combining the noise level and geological characteristics. Its calculation formula is as follows:

[0097]

[0098] in, represents the weight of geological features; represents the noise weight; and It is a hyperparameter that adjusts the importance of geological features and noise, and is used to balance their influence on the weighted loss function.

[0099] 3-5. Through backpropagation and gradient update, the model parameters are gradually adjusted to make the generated noise prediction result closer to the target denoised image until the model converges.

[0100] After training, the model finally converges successfully, as shown in Figure 11 As shown in the figure, CSAUnet successfully reduces the MSE loss to about 0.026, reaching a relatively accurate benchmark and laying the foundation for subsequent DDIM iterative generation.

[0101] 4. Input the noise data and geological conditions into the trained CSAUnet model, use the DDIM sampling method to perform step-by-step denoising, and generate the denoised target 3D digital core image.

[0102] The denoising process generates a real image from noise by iteratively applying the denoising equation and further optimizes the generated result through conditional control. The specific formula is:

[0103]

[0104] in, is the sample at time t, Indicates category condition information, Represents the porosity condition information, Parameters that control the degree of noise diffusion, is a noise function that gradually removes noise and restores core structural characteristics.

[0105] In this example, the CSAUnet model was used to perform an iterative denoising process (DDIM) to generate three-dimensional digital cores of two rock types. Figure 12 As shown in the figure, compared to images generated by traditional generative adversarial networks (CGANs), DDIM exhibits superior detail, texture, and continuity, and the generated images are more similar to real samples. The slices generated by CGANs capture some basic granular textures, but are generally blurry and contain noise in some slices (as indicated by the boxes). Furthermore, the CGAN generation process also exhibits mode collapse, further affecting the generation performance.

[0106] Figure 13Two sets of 3D digital rock core samples were presented, each consisting of carbonate and sandstone rock types. The DDIM-generated samples resembled actual rock cores. Compared to samples generated using CGAN (Computer Aided Image Network), the DDIM-generated carbonate rock samples displayed finer details in porosity and texture, with grayscale variations and spatial distribution more accurately reflecting rock heterogeneity. The DDIM-generated images of the sandstone samples revealed a more complex grain texture and pore structure, demonstrating higher accuracy and realism than the CGAN-generated samples.

[0107] Figure 14 The comparison between the carbonate pore space generated by DDIM and the real sample is shown, and the generated sample accurately reproduces the pore structure of the actual core. Figure 15 The pore space comparison of sandstone is shown, and the pore structure and texture of the DDIM-generated samples are highly consistent with the actual samples. Through these images, DDIM has demonstrated its remarkable ability to generate digital cores with realistic pore structures.

[0108] The present application also provides a three-dimensional digital core generation system for completing the above-mentioned three-dimensional digital core generation method, comprising: a data acquisition and preprocessing module, a geological feature input module, a model training and optimization module, a generation and post-processing module, and a system interaction and control module;

[0109] The data acquisition and preprocessing module is responsible for acquiring core images and preprocessing the image data, including denoising, standardization, and size adjustment, to adapt to subsequent generation tasks;

[0110] The geological feature input module is used to provide geological condition data through a dedicated interface and use these conditions as input to assist the model in maintaining the consistency of geological features when generating images;

[0111] The model training and optimization module is used to optimize the model through the training environment. It is responsible for using the self-attention mechanism of CSAUnet to improve the quality of digital core generation and gradually optimize the difference between the generated image and the actual core image;

[0112] The generation and post-processing module is used to generate digital core images by using the trained model, inputting noise and geological conditions, and gradually denoising the generated images using the DDIM diffusion model to generate three-dimensional digital core images that meet the target geological characteristics;

[0113] The system interaction and control module is used to provide a user interface and interactive functions, and supports system status monitoring and control of the model training process.

[0114] The above-mentioned 3D digital core generation method, firstly, obtains 3D digital data of real rock cores and preprocesses it to generate a core training dataset. Then, based on the generative model architecture of the DDIM diffusion model and CSAUnet, a digital core generation model is constructed. This model is trained using the training dataset, and the network parameters are gradually optimized to ultimately obtain the optimal model. During the generation process, noise and geological condition information are input to generate a target digital core image that matches the target geological characteristics. The generated target digital core image possesses the geological characteristics and visual effects of a real rock core, accurately simulating the porosity and bedding structure within the core. Furthermore, the self-attention mechanism of the DDIM and CSAUnet models enables more detailed restoration of geological features within the core, including porosity, pore distribution, and rock type. Generation accuracy is significantly improved, particularly in capturing complex bedding structures and subtle geological features. By introducing an enhanced MSE loss function that factors in noise level and geological feature weighting, the noise processing process is further optimized, reducing the interference of noise on the core image generation while improving the quality of detail restoration, particularly when simulating subtle pore structures and minor geological variations. Combining the advantages of the DDIM model and the CSAUnet model, the model can generate more accurate three-dimensional digital cores under complex geological conditions, especially in the simulation of complex geological features such as bedding structure and micropores, showing higher adaptability and generation accuracy.

[0115] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly indicate the number of the technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0116] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0117] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

Claims

1. A three-dimensional digital core generation method, characterized in that: The method includes: The noise data and geological conditions are input into the trained CSAUnet model for processing; the CSAUnet model includes a dual-channel convolution block, a downsampling module, an upsampling module, a dual attention module, and a convolution module; The two-channel convolution block extracts the features of the noisy data to obtain the initial feature map; The downsampling module performs three downsampling operations on the initial feature map, and then processes it using the dual attention module to obtain the first feature image block, the second feature image block, and the third feature image block. Each downsampling operation is processed by two first dual convolution blocks, and then auxiliary features such as time embedding, porosity information, and lithology category are introduced for integration. The upsampling module performs three upsampling operations on the third feature image block. Each upsampling operation is processed by two second double convolution blocks, and then auxiliary features such as time embedding, porosity information, and lithology category are introduced. After splicing and fusing with the first and second feature image blocks, they are processed by the dual attention module to obtain the fused features. The convolution module processes the fused features to generate an initial three-dimensional digital core image; The DDIM diffusion model is used to gradually denoise the noise in the initial 3D digital core image to generate the denoised target 3D digital core image. Acquire the original image of the real core and perform preprocessing to obtain a three-dimensional core data set; Based on the CSAUnet model and the DDIM diffusion model, a digital core generation model was constructed; The 3D core dataset and the corresponding geological conditions are input into the digital core generation model for training. A weighted loss function is introduced during the training process to obtain a trained digital core generation model. The weighted loss function is: Where, To generate pixel values in an image, is the pixel value in the real image, N is the total number of pixels in the image, is the weighting factor: Where, is the geological feature weight, is the noise weight, is the first hyperparameter, is the second hyperparameter.

2. The three-dimensional digital core generation method according to claim 1, characterized in that: The down-sampling module includes a first down-sampling block, a second down-sampling block and a third down-sampling block, and the up-sampling module includes a first up-sampling block, a second up-sampling block and a third up-sampling block.

3. The three-dimensional digital core generation method according to claim 2, characterized in that: The downsampling module performs three downsampling operations on the initial feature map, and then processes it using the dual attention module to obtain a first feature image block, a second feature image block, and a third feature image block, respectively, including: The first downsampling block downsamples the initial feature map. After two 3D convolutions and three activation functions in the two first biconvolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The dual attention module is used for processing to obtain the first feature image block with a size of 40×40×40 and 16 channels. The second downsampling block downsamples the first feature image block. After two 3D convolutions and three activation functions in the two first double convolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The block is then processed using a dual attention module to obtain a second feature image block with a size of 20×20×20 and 32 channels. The third downsampling block performs a downsampling operation on the two feature image blocks. After two 3D convolutions and three activation functions in the two first double convolution blocks, a 3D maximum pooling operation is performed. At the same time, auxiliary features of time embedding, porosity information and lithology category are introduced, and the dual attention module is used for processing to obtain a target feature image block with a size of 10×10×10 and 128 channels.

4. The three-dimensional digital core generation method according to claim 3, characterized in that: The upsampling module performs three upsampling operations on the third feature image block, and each upsampling operation is processed by two second double convolution blocks, and auxiliary features of time embedding, porosity information and lithology category are introduced, and after splicing and fusing with the first feature image block and the second feature image block, the dual attention module is used for processing to obtain the fusion feature, including: The first upsampling block performs an upsampling operation on the third feature image block. After two three-dimensional convolutions and a cubic activation function in the two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. The dual attention module is used for processing to obtain the first fused feature image block with a size of 20×20×20 and 32 channels. The second upsampling block performs an upsampling operation on the first fused feature image block. After two 3D convolutions and a cubic activation function in two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features such as time embedding, porosity information, and lithology category are introduced. After splicing and fusing with the second feature image block, it is processed using the dual attention module to obtain a second fused feature image block with a size of 40×40×40 and 16 channels. The third upsampling block performs an upsampling operation on the second fused feature image block. After two three-dimensional convolutions and three activation functions in the two second double convolution blocks, a linear interpolation operation is performed. At the same time, auxiliary features of time embedding, porosity information and lithology category are introduced. After splicing and fusing with the first feature image block, it is processed using the dual attention module to obtain a fused feature with a size of 80×80×80 and 8 channels.

5. The three-dimensional digital core generation method according to claim 4, characterized in that: The convolution module processes the fused features to generate the initial three-dimensional digital core image, including: The convolution module performs a convolution operation with a convolution kernel of 1×1×1 on the fused features with a size of 80×80×80 and a channel number of 8 to obtain an initial three-dimensional digital core image with a size of 80×80×80 and a channel number of 1.

6. The three-dimensional digital core generation method according to claim 5, characterized in that: The steps of gradually denoising the initial three-dimensional digital core image using the DDIM diffusion model to generate a denoised target three-dimensional digital core image include: The initial three-dimensional digital core image was sampled using the DDIM diffusion model; By iteratively applying the denoising equation, a real image is generated from the noise and optimized through conditional control to generate the target three-dimensional digital core image: in, is the sample at time t, Indicates category condition information, Represents the porosity condition information, is the parameter that controls the degree of noise diffusion at time t, is the parameter that controls the noise diffusion degree at time t-1, is the noise function, for The cumulative product of for The cumulative product of .

Images