A multi-component digital core reconstruction method, device, equipment and medium

Abstract

The present invention discloses a multi-component digital core reconstruction method, device, equipment and medium, and relates to the field of reservoir logging identification technology. The present invention introduces the Wasserstein distance to calculate the Wasserstein distance between the distribution of generated samples and real samples, minimizes the Wasserstein distance layer by layer during training, solves the non-convergence problem in the training process, and improves the training stability and the quality of generated samples. Prior knowledge decisions are embedded in the discriminator to regularize the mineral distribution, porosity and fracture characteristics, thereby improving the discriminator's ability to judge the authenticity of the generated samples and their compliance with domain characteristics, and improving the reconstruction accuracy of the multi-component structure. At the same time, a gradient penalty mechanism is introduced to improve the stability of model training and the distribution consistency of the generated samples. Finally, the trained network model is used to reconstruct the multi-component digital core, clearly showing the complex pore structure and multi-component structure of the reservoir.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir logging identification, and in particular to a multi-component digital core reconstruction method, device, equipment and medium. Background Art

[0002] Carbonate reservoirs occupy a vital position in the global energy sector, boasting abundant reserves of oil and natural gas. However, due to their complex pore structure and multi-component composition, traditional reservoir characterization methods face significant challenges. Well logging and microscopic imaging techniques, for example, struggle to characterize the microscopic, multi-component pore structure of carbonate rocks. Furthermore, techniques such as Qemscan (Quantitative Evaluation of Minerals by Scanning Electron Microscopy), which can identify the multi-component pore structure of microscopic carbonate cores, suffer from the disadvantage of high cost.

[0003] In recent years, artificial intelligence algorithms such as generative adversarial networks (GANs) have shown great potential in processing logging and physical property data, and have advantages in the structural characterization of carbonate porous media. At present, when generative adversarial networks are used to reconstruct the multi-component structure of carbonate reservoirs, the training process of generative adversarial networks relies on the discriminator's evaluation of the authenticity of the generated images, which requires a large amount of training data to support it. However, the structure of carbonate reservoirs is complex and changeable, and it is difficult to obtain more core images that show the structure of carbonate reservoirs. As a result, the reconstruction ratio is inaccurate when reconstructing the multi-component cores of carbonate reservoirs, making it difficult for the reconstructed multi-component digital cores to represent the complex pore structure and multi-component structure of carbonate reservoirs. Summary of the Invention

[0004] The embodiments of the present invention provide a multi-component digital core reconstruction method, device, equipment and medium, which can solve the problem in the prior art that the reconstructed multi-component digital core is difficult to represent the complex pore structure and multi-component structure of the carbonate reservoir.

[0005] An embodiment of the present invention provides a multi-component digital core reconstruction method, comprising the following steps:

[0006] Obtain multiple core images of small samples of carbonate rocks and the core multi-component structure corresponding to each core image;

[0007] The Wasserstein distance is introduced as the loss function of the generator of the generative adversarial network model GAN ​​to form an improved generator; prior knowledge decision and gradient penalty are embedded in the discriminator of the generative adversarial network model GAN ​​to form an improved discriminator; and an improved generative adversarial network model WGAN-GP is formed, which includes the improved generator and the improved discriminator;

[0008] An improved generative adversarial network model, WGAN-GP, is trained using the multi-component core structures corresponding to multiple core images. During training, the generator optimizes the network model parameters layer by layer by inputting a random noise vector and the corresponding domain knowledge vector. Simultaneously, the generator uses the Wasserstein distance as a loss function and adjusts the distribution of generated samples using feedback from the discriminator for adversarial training. The discriminator, based on embedded prior knowledge, imposes constraints on mineral distribution, porosity, and fracture connectivity to evaluate the degree of match between generated samples and real samples. Furthermore, during adversarial training of the generator and discriminator, the generator uses a gradient penalty mechanism to impose gradient constraints on the interpolation points between the real and generated samples, resulting in the trained generative adversarial network model, WGAN-GP.

[0009] The core multi-component structure corresponding to the core image to be reconstructed is input into the generator of the trained generative adversarial network model WGAN-GP, and the reconstructed multi-component digital core is output.

[0010] Preferably, the obtaining of multiple core images of carbonate rocks and the core multi-component structures corresponding to the core images includes:

[0011] Select core samples from carbonate cores and cut them into slices 5-10 mm thick using a slicer, ensuring that the slices contain the main pore and fracture features.

[0012] After the core sample slices were placed in a vacuum drying oven and dried for a preset time, the selected areas were scanned step by step to obtain high-resolution backscattered electron images (BSE) and multispectral energy spectra (EDS) of the samples. The QEMSCAN method was used to quantitatively evaluate the sample mineral components in the high-resolution backscattered electron images (BSE) and multispectral energy spectra (EDS) to obtain small sample core images of carbonate rocks and the multi-component structure of the core corresponding to the core images.

[0013] Preferably, the improvement of the generative adversarial network model GAN ​​includes:

[0014] The log-likelihood loss of the objective function in the generative adversarial network model GAN ​​is:

[0015] ;

[0016] Replace the log-likelihood loss with the Wasserstein distance, the equation of the Wasserstein distance is:

[0017] ;

[0018] in: and There are two possible distributions; There are two distributions and The set of joint distributions of , γ is ( , ) a possible joint distribution; x represents the real sample, y Indicates the generated sample, is a sample drawn from γ; Represents two samples x and y The distance between them; f represents the 1-Lipschitz function;

[0019] Prior knowledge decisions are embedded in the discriminator of the generative adversarial network model GAN ​​to obtain the loss error of multiple component proportions, where the loss function is expressed as:

[0020] ;

[0021] in: represents the sample prediction percentage of the proportion of the i-th component; Indicates the true percentage of the sample of the i-th component ratio; Indicates the number of rock component types included in the label; Represents the importance ratio of the original loss function in training the neural network; Indicates the importance ratio of prior knowledge in training neural networks;

[0022] In order to maintain the conditions of the 1-Lipschitz function, the weight clipping strategy in WGAN is replaced by gradient penalty to penalize the gradient of the discriminator output. The gradient penalty equation is:

[0023] ;

[0024] in: is from a uniform distribution Random variables sampled from ; is a hyperparameter.

[0025] Preferably, the training of the improved generative adversarial network model WGAN-GP includes:

[0026] During training, the generator takes as input a random noise vector z and a corresponding domain knowledge vector k, which are experimentally obtained mineral composition and pore distribution information. Based on this information, the generator optimizes the parameters of the network model layer by layer to generate realistic digital cores.

[0027] The discriminator receives real samples and generated samples. It imposes constraints on mineral distribution, porosity, and fracture connectivity through embedded prior knowledge decisions and outputs a scalar as an approximation of the Wasserstein distance to evaluate the difference between the generated samples and the real sample distribution. The generator uses the Wasserstein distance as a loss function and is updated based on the feedback from the discriminator to minimize the Wasserstein distance and adjust the distribution of the generated samples.

[0028] At the same time, the discriminator uses the gradient penalty mechanism to regularize the gradient of the interpolation points between the real samples and the generated samples to ensure that the discriminator meets the continuity condition of the 1-Lipschitz function.

[0029] Preferably, the output of the trained generative adversarial network model WGAN-GP includes:

[0030] After the training of the generative adversarial network model WGAN-GP is completed, the input of the model generator is a random noise vector z and a domain knowledge vector k; the random noise vector z is a high-dimensional vector sampled from a standard normal distribution, and the knowledge vector k is the mineral composition ratio extracted from the actual collected data;

[0031] After the random noise vector z and the domain knowledge vector k are concatenated, the generator maps them to a high-dimensional feature space through a fully connected layer and expands them into an initial low-resolution feature map. The generator then gradually upsamples the resolution of the feature map by performing multi-layer transposed convolution using a three-dimensional convolution kernel. Each layer of transposed convolution refines the features, gradually generating a multi-component digital core with mineral distribution, pore structure, and fracture connectivity.

[0032] An embodiment of the present invention further provides a multi-component digital core reconstruction device, comprising:

[0033] A data module is used to obtain multiple core images of small samples of carbonate rocks and the core multi-component structure corresponding to each core image;

[0034] A model building module for introducing Wasserstein distance as the loss function of the generator of the generative adversarial network model (GAN) to form an improved generator; embedding prior knowledge decision and gradient penalty in the discriminator of the generative adversarial network model (GAN) to form an improved discriminator; and forming an improved generative adversarial network model (WGAN-GP) including the improved generator and the improved discriminator;

[0035] The model training module is used to train an improved generative adversarial network model (WGAN-GP) using the multi-component core structures corresponding to multiple core images. During training, the generator optimizes the network model parameters layer by layer by inputting a random noise vector and the corresponding domain knowledge vector. Simultaneously, the generator uses the Wasserstein distance as a loss function and adjusts the distribution of generated samples based on feedback from the discriminator for adversarial training. The discriminator uses embedded prior knowledge to impose constraints on mineral distribution, porosity, and fracture connectivity to evaluate the degree of match between generated samples and real samples. Furthermore, during adversarial training of the generator and discriminator, the generator uses a gradient penalty mechanism to impose gradient constraints on interpolation points between real and generated samples to obtain the trained generative adversarial network model (WGAN-GP).

[0036] The reconstruction module is used to input the core multi-component structure corresponding to the core image to be reconstructed into the generator of the trained generative adversarial network model WGAN-GP, and output the reconstructed multi-component digital core.

[0037] An embodiment of the present invention further provides an electronic device, including a memory and a processor;

[0038] The memory is used to store computer programs;

[0039] The processor is configured to implement the steps of the multi-component digital core reconstruction method described above when executing the computer program stored in the memory.

[0040] An embodiment of the present invention further provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of the multi-component digital core reconstruction method described above.

[0041] The embodiments of the present invention provide a multi-component digital core reconstruction method, device, equipment, and medium. Compared with the prior art, the advantages thereof are as follows:

[0042] The present invention embeds prior knowledge decisions and gradient penalties in the discriminator of the generative adversarial network model (GAN) to form a network model WGAN-GP that improves the generative adversarial network model GAN. By embedding prior knowledge decisions in the discriminator, the present invention enables the model to consider geological constraints during training, that is, regularizes domain knowledge such as mineral distribution, porosity, and fracture characteristics, thereby improving the discriminator's ability to judge the authenticity of generated samples and their compliance with domain characteristics. This process only requires a small sample of core images to train the model WGAN-GP, thereby improving the accuracy of reconstructing the multi-component structure of carbonate reservoir cores.

[0043] In addition, the present invention introduces Wasserstein distance as the loss function of the generator in the generative adversarial network model GAN, takes Wasserstein distance as the optimization target, calculates the Wasserstein distance between the generated samples and the real sample distribution, and minimizes the Wasserstein distance layer by layer during training to solve the non-convergence problem in the model training process, thereby improving the training stability and the quality of the generated samples; at the same time, the present invention also introduces a gradient penalty mechanism to constrain the gradient of the interpolation points between the real samples and the generated samples to improve the stability of the model training and the distribution consistency of the generated samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the overall process of a multi-component digital core reconstruction method provided by an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of the neural network structure of a WGAN-GP model established by a multi-component digital core reconstruction method provided in an embodiment of the present invention;

[0046] Figure 3 Sample 1 of a multi-component digital core reconstruction method provided in an embodiment of the present invention uses the Qemscan method to collect a core image and a multi-component structure diagram of a small portion of carbonate rock;

[0047] Figure 4 Sample 2 of a multi-component digital core reconstruction method provided in an embodiment of the present invention uses the Qemscan method to collect a core image and a multi-component structure diagram of a small portion of carbonate rock;

[0048] Figure 5 Sample 3 of a multi-component digital core reconstruction method provided in an embodiment of the present invention uses the Qemscan method to collect a core image and a multi-component structure diagram of a small portion of carbonate rock;

[0049] Figure 6 A schematic diagram of a multi-component digital core reconstructed by a multi-component digital core reconstruction method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar modifications without violating the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0051] See also Figure 1 An embodiment of the present invention provides a multi-component digital core reconstruction method, specifically a multi-component digital core reconstruction method based on knowledge embedding optimization WGAN-GP. This method improves training stability and the quality of generated samples by introducing Wasserstein distance to improve the generative adversarial neural network (GAN), and combines knowledge embedding technology to achieve the embedding of prior knowledge of multi-component proportions and improve the reconstruction accuracy of multi-component structures. At the same time, considering the conditions of traditional methods for reconstructing multi-component structures of carbonate rocks, such as inaccurate multi-component proportions and high experimental costs, the method of the present invention is highly timely and low-cost, can reconstruct multi-component digital cores of carbonate rocks in batches, and is suitable for on-site measurement and analysis in oil fields.

[0052] During model training, the generator inputs a random noise vector z and the corresponding domain knowledge vector k, optimizing its network parameters layer by layer to generate realistic digital core samples. The generator uses the Wasserstein distance as a loss function and, through feedback from the discriminator, adjusts the distribution of the generated samples, gradually approximating the multi-component characteristics of real core samples. During the discriminator training process, an embedded prior knowledge decision module imposes constraints on domain characteristics such as mineral distribution, porosity, and fracture connectivity, ensuring that the discriminator can accurately assess the authenticity of the generated samples and their degree of match with real samples. The entire training process utilizes an alternating optimization approach. First, the generator is fixed and the discriminator is updated, enabling it to distinguish between real and generated samples while maintaining its 1-Lipschitz continuity through a gradient penalty mechanism. Subsequently, the discriminator is fixed and updated. By minimizing the discriminator's output, the generator gradually approximates the generated digital core samples in terms of multi-component distribution and geometric characteristics to real data.

[0053] After training, the trained generator model is used to generate new digital core samples. By inputting a new random noise vector z and domain knowledge k, the generator can efficiently produce high-resolution digital cores that conform to the specified mineral distribution and pore structure characteristics. These generated digital core samples are highly consistent with real cores in terms of mineral composition and can be directly used for reservoir analysis, seepage simulation, and pore network modeling, significantly improving the efficiency and accuracy of reservoir research.

[0054] The specific steps include:

[0055] Step 1: Select a representative sample from the carbonate core column and slice it into 5-10 mm thick slices using a microtome, ensuring that the slices capture the main pore and fracture features. The slices are then dried in a vacuum oven for 24 hours. Selected areas are then scanned stepwise to obtain high-resolution backscattered electron (BSE) images and energy-dispersive spectrometers (EDS) of the sample. The automated energy-dispersive spectrometer analysis function of the QEMSCAN method is used to quantitatively assess the sample's mineralogy. The QEMSCAN method was used to acquire core images and multi-component structures of a small section of the carbonate rock.

[0056] Step 2: Perform unique-hot encoding on each component label in each image sample so that the corresponding label is stored in vector form. Specifically:

[0057] First, the distribution information of all components, such as calcite, dolomite, quartz, and pore and fracture areas, is extracted from the image. In the mineral distribution map obtained using the QEMSCAN method, each pixel is assigned a label representing a mineral type or pore property. To facilitate the storage and processing of these discrete labels in a computer-friendly manner, a unique vector representation is created for each component using one-hot encoding. Specifically, assuming the sample contains N different components, each label is represented by an N-dimensional vector, where the position corresponding to the component is 1 and all other positions are 0. For example, [1, 0, …, 0] represents dolomite, and [0, 1, …, 0] represents quartz. In this way, the two-dimensional image data of the component labels is converted into a three-dimensional array structure, where each pixel is represented by a one-hot encoded vector, enabling efficient use in subsequent data integration, analysis, and simulation. This method can be further combined with pore characteristic data to construct a multi-scale digital core that includes mineral distribution and pore properties, providing standardized and easy-to-process input data for reservoir research.

[0058] Step 3: Divide the data into a training set and a test set. The training set is used for training data, and the test set is used to test the performance of the model. Use a random partitioning method to split the data according to a certain ratio (for example, 80% training set and 20% test set) to ensure that each subset has consistent component distribution and feature space.

[0059] Step 4: Initialize the WGAN-GP model, whose model structure is as follows Figure 2 As shown; the network structure of WGAN-GP (such as Figure 2 The process of generating multi-component digital cores through the WGAN-GP network model architecture is mainly based on the powerful learning ability of the generative adversarial network to accurately reconstruct and generate the multi-component characteristics of carbonate reservoirs (such as mineral distribution, pore structure, fracture connectivity, etc.) in the form of digital cores.

[0060] The entire model consists of two components: a generator and a discriminator, each with distinct tasks. Through adversarial optimization, they achieve high-quality digital core generation. The generator's primary function is to generate realistic digital cores from random noise z and a knowledge embedding vector k (e.g., experimentally acquired prior information such as mineral composition and pore distribution). The generator progressively maps the input into high-dimensional image data through multiple layers of fully connected networks and convolutional layers. Each layer refines the mineral distribution and pore characteristics, resulting in a multi-component digital core with the desired resolution at the output layer. The introduction of knowledge embeddings enables the generator to incorporate domain-specific features, resulting in digital cores that better reflect actual reservoir characteristics. The discriminator's task is to classify generated samples from real samples and output a Wasserstein distance to assess the difference between the generated and real sample distributions. By introducing a gradient penalty, the discriminator ensures consistency in the distribution of generated samples in terms of mineral composition and pore characteristics, while avoiding the mode collapse problem inherent in traditional GANs. In addition, a knowledge correction module is embedded in the discriminator to verify whether the generated samples meet the constraints of domain knowledge, thereby improving the quality and authenticity of the generated samples. The generation process first extracts real sample features and knowledge vectors from the preprocessed dataset and initializes the generator and discriminator. Then, by inputting random noise z and knowledge vector k into the generator, a digital core with preliminary mineral distribution and pore structure characteristics is generated. The discriminator receives the generated samples and real samples and evaluates the authenticity and distribution differences of the samples using the Wasserstein distance and the knowledge correction module. The generator continuously optimizes parameters based on the feedback from the discriminator, so that the generated samples gradually approach the distribution of the real samples.

[0061] The Wasserstein distance measures the difference between the generated sample distribution and the true sample distribution, overcoming the training instability caused by JS divergence in traditional GANs. Specifically, it refers to the minimum energy consumption along the optimal path. In WGAN-GP (Wasserstein GAN with Gradient Penalty), the Wasserstein distance is integrated into the generative adversarial network in the following way. The objective function of the traditional GAN ​​is based on the log-likelihood loss:

[0062] .

[0063] In WGAN-GP, this is replaced by the Wasserstein distance, which is formulated as:

[0064] .

[0065] in: and There are two possible distributions, the Wasserstein distance is defined by optimizing over all possible joint distributions whose marginal distributions are and , which quantifies the difference between two distributions. There are two distributions and The set of joint distributions of , γ is ( , ) a possible joint distribution; x represents the real sample, y Indicates the generated sample, is a sample drawn from γ; Represents two samples x and y The distance between them; f is the 1-Lipschitz function (implemented by weight constraint or gradient penalty).

[0066] In order to maintain the 1-Lipschitz condition, WGAN-GP replaces the weight clipping strategy in WGAN with a gradient penalty, which penalizes the gradient of the discriminator output, defined as:

[0067] .

[0068] in: is from a uniform distribution This interpolation method ensures that It is the "mixing" point between the real sample and the generated sample and is used to calculate the gradient of the discriminator output. is a hyperparameter that controls the weight of the gradient penalty.

[0069] The generator receives random noise z and a knowledge vector k, maps the input to a high-dimensional feature space using fully connected layers, and then gradually upsamples the features through multiple layers of transposed convolutions, generating digital rock cores with multi-scale mineral distribution and pore structure. Activation functions such as LeakyReLU further enhance the generator's nonlinear representation, making the output image closer to real rock core samples. The discriminator receives generated and real samples as input, extracts local and global features of the samples using convolutional layers, and outputs a scalar representing the Wasserstein distance between the generated and real sample distributions through fully connected layers.

[0070] Step 5: Calculate the normal loss function based on the training reconstructed image and the labeled image.

[0071] Step 6: Based on the knowledge embedding technology, calculate the loss error of the multi-component ratio. The specific loss function is shown in the following formula:

[0072] .

[0073] in: is the sample prediction percentage of the i-th component proportion, is the true percentage of the sample of the i-th component proportion, Indicates the number of rock component types contained in the label, represents the importance ratio of the original loss function in training the neural network, and Represents the importance of prior knowledge in training neural networks.

[0074] Step 7: Perform error propagation and feedback adjustment on the constructed WGAN-GP neural network, and finally update the weights according to the error.

[0075] During the training process for generating multi-component digital rock cores, WGAN-GP achieves high-quality sample generation through adversarial optimization between the generator and discriminator. Training begins with initializing model parameters, initializing weights for both the generator and discriminator, and configuring an optimizer (typically Adam) to ensure efficient model convergence. At each training iteration, the discriminator is first updated multiple times to ensure it accurately estimates the Wasserstein distance between generated and real samples. Specifically, the discriminator receives real and generated samples, extracts local and global features of the samples through a convolutional layer, and outputs a scalar approximation of the Wasserstein distance. Furthermore, a gradient penalty is introduced to regularize the gradients at interpolation points between real and generated samples, ensuring that the discriminator satisfies the 1-Lipschitz continuity condition, thereby improving training stability and sample distribution matching accuracy. After the discriminator is trained, the generator is updated based on the discriminator's feedback, aiming to minimize the Wasserstein distance and gradually bring the generated sample distribution closer to the real sample distribution. The input to the generator consists of a random noise vector z and a knowledge vector k. The noise is used to generate sample diversity, while the knowledge vector provides domain information to guide the generator in learning the key features of the samples. The generator gradually constructs a digital rock core with multi-component features through multiple layers of fully connected layers and transposed convolutions. The mineral distribution and pore structure of the output image gradually approach real data. The entire training process alternates between optimizing the generator and the discriminator. The generator challenges the discriminator by continuously improving the quality of generated samples, while the discriminator forces the generator to improve by continuously improving its ability to distinguish real from fake samples. During training, to ensure model convergence and generation quality, the number of optimization steps between the generator and discriminator is typically set asymmetric (for example, the discriminator is updated multiple times before each generator update). When the model converges, the generator can produce digital rock cores that are highly consistent with real rock samples in terms of multi-component features and spatial structure.

[0076] Step 8: Repeat steps 6 and 7 repeatedly until the required number of iterations is met and then output the model.

[0077] The trained WGAN-GP model uses the generator G to generate digital cores with multi-component features based on the input random noise z and knowledge vector k. The specific generation process is as follows: First, the generator input consists of two components: the random noise vector z and the domain knowledge vector k. The random noise z is a high-dimensional vector sampled from a standard normal distribution, which introduces diversity in the generated samples. The knowledge vector k is the mineral composition ratio extracted from actual collected data and guides the generator in generating digital core samples that meet specific reservoir conditions. The input vectors [z, k] are concatenated and processed layer by layer by the generator's neural network. Within the generator, the input is first mapped to a high-dimensional feature space through a fully connected layer, expanding it into an initial low-resolution feature map. The generator then performs multiple layers of transposed convolution using a three-dimensional convolution kernel to gradually upsample the resolution of the feature map. Each layer of transposed convolution refines the features, gradually generating a multi-component digital core with mineral distribution, pore structure, and fracture connectivity. During the generation process, activation functions (such as LeakyReLU and Tanh) are used to enhance nonlinear representation, making the generated samples more similar to real samples in terms of mineral distribution and pore geometry. The output of the generator is a high-resolution 3D image, where each pixel (voxel) represents the component at a local point in the digital core. To reflect multi-component characteristics, the generated core sample contains multiple channels, each corresponding to a mineral distribution or pore characteristic. For example, a voxel with coordinates (X, Y, Z) and a value of [1, 0, …, 0] indicates that the component generated at that location is dolomite.

[0078] The present invention introduces the Wasserstein distance as an optimization target in the generator of the generative adversarial network model GAN. By calculating the Wasserstein distance between the generated samples and the real sample distribution, and combining the input of random noise and prior knowledge vectors, the accurate generation of multi-component features of the generator is achieved.

[0079] The present invention embeds a priori knowledge decision module into the discriminator of the generative adversarial network model GAN. By regularizing the knowledge in the fields of mineral distribution, porosity and crack characteristics, the discriminator's ability to judge the authenticity of generated samples and their conformity to domain characteristics is improved.

[0080] The present invention adopts a gradient penalty mechanism in the training process of the generative adversarial network model WGAN-GP. By constraining the gradient of the interpolation points between real samples and generated samples, the discriminator is ensured to meet the 1-Lipschitz continuity condition, thereby improving the stability of model training and the distribution consistency of generated samples.

[0081] The generator of the generative adversarial network model WGAN-GP of the present invention gradually upsamples through multiple layers of fully connected layers and transposed convolutional layers, and combines the Tanh activation function to limit the output range, thereby achieving the generation of digital cores with high resolution and multi-component features.

[0082] The present invention introduces the concatenation vector of domain knowledge vector k and random noise z as the input of the generator, so that the digital core generated by the generator is more consistent with the actual reservoir conditions in terms of characteristics such as mineral distribution, pore structure and fracture connectivity, and meets the needs of reservoir analysis.

[0083] The present invention only requires a small amount of sample data collected by Qemscan technology to quickly and accurately reconstruct multi-component digital cores. It has high timeliness, low cost and good generalization ability, realizes the multi-component pore structure characterization of carbonate reservoirs, and can provide guiding suggestions for the exploration and development of carbonate oil reservoirs.

[0084] The present invention improves the generative adversarial neural network (GAN) by introducing the Wasserstein distance, thereby improving training stability and the quality of generated samples. It addresses the problem of inaccurate multi-component reconstruction ratios and combines knowledge embedding technology to embed prior knowledge of multi-component ratios, thereby improving the reconstruction accuracy of multi-component structures. Through the improved WGAN-GP framework, the complex pore structure and multi-component structure of carbonate reservoirs can be intelligently reconstructed. This method reduces time costs, improves processing efficiency and reconstruction accuracy, is suitable for on-site measurement and analysis in oil fields, and provides strong technical support for the development of conglomerate oil and gas reservoirs.

[0085] The generative adversarial neural network model of the present invention incorporates the Wasserstein distance, thereby achieving improved model training stability and generated sample quality. It is applicable to the generation of pore structure of carbonate reservoir core samples at various depths and has the characteristics of strong generalization. The entire evaluation process requires no complex calculations, no manual processing, and no huge computing resources, making it suitable for promotion to oil field sites.

[0086] Specific experiments:

[0087] (1) Taking a carbonate reservoir as an example, the core images and multi-component structures of a small part of carbonate rocks were collected by the QEMSCAN method, such as Figure 3Before the experiment began, core samples were cut into slices approximately 10 mm thick and vacuum-dried to remove moisture and impurities from the pores. Subsequently, each sample was scanned layer by layer using the QEMSCAN method, acquiring high-resolution 2D slice images and mineral composition distribution data. The left image shows a typical carbonate core image, clearly showing pores, fractures, and matrix structure. The right image shows the corresponding multi-component structure acquisition results, with different colors marking mineral components such as calcite, dolomite, and quartz, as well as pore areas. After preprocessing, these data were generated into a standardized sample matrix that not only captures the spatial characteristics of mineral distribution but also records the geometric parameters of the pore network, providing comprehensive input information for subsequent deep learning model training. To ensure the scientific nature of model training and validation, the collected core images and multi-component structure data were divided into training and test sets. Random sampling was used during this division process to ensure consistent data distribution in terms of mineral composition and pore structure. Finally, 80% of the samples were selected as the training set to optimize the model parameters of the generator and discriminator; the remaining 20% ​​of the samples were used as the test set to verify the effectiveness of the model in generating digital cores.

[0088] (2) Construct the WGAN-GP neural network model and initialize the parameters. The generator structure consists of multiple layers of fully connected layers and transposed convolutional layers. Its input is random noise z and knowledge vector k. It generates high-resolution digital cores by upsampling layer by layer. The discriminator structure consists of convolutional layers and fully connected layers. It is mainly used to evaluate the distribution difference between generated samples and real samples and output the Wasserstein distance value. The Xavier initialization method is used for model initialization to optimize the distribution of the weights of the generator and discriminator to ensure the stability of the parameters in the initial stage. In addition, to accelerate the training process, the Adam optimizer is selected and the learning rate is set to 10 −4 , and a gradient penalty term is introduced into the discriminator to satisfy the 1-Lipschitz continuity condition. The K vector is set to [0.02, 0.01, 0.26, 0.01, 0.26, 0.06, 0.06, 0.21, 0.25, 0.01, 0.01]. Each digit of this vector represents the component ratio of albite, quartz, calcite, illite, chlorite, pyrite, rutile, dolomite, anhydrite, other minerals, and pores. These ratios are averaged from core analysis experiments of a large number of core samples collected from adjacent wells.

[0089] (4) During the training of the WGAN-GP neural network model, the generator and discriminator are optimized alternately. In each iteration, the discriminator is first updated multiple times so that it can more accurately calculate the Wasserstein distance between the real sample and the generated sample. The generator generates realistic digital rock cores by minimizing the feedback loss of the discriminator. In each training iteration, the generator inputs random noise z and domain knowledge k, and the discriminator receives the generated samples and the real samples, and adjusts its parameters through the gradient penalty term to make the calculation of the Wasserstein distance more stable. During the training process, the loss curve of the model gradually converges, indicating that the output quality of the generator continues to improve, and the matching degree of the generated samples with the real samples in terms of mineral composition and pore structure is significantly improved.

[0090] (5) Save the parameters of the iterated WGAN-GP neural network model. After training is complete, save the final parameters of the generator and discriminator as a model file (HDF5 or PT format) and record key hyperparameters (such as learning rate, batch size, and training rounds) to ensure model reproducibility. In addition, the saved model contains the weight configuration and network structure of the generator and can be directly used for subsequent digital core generation.

[0091] (6) Read the saved WGAN-GP neural network model parameters and input the test set to verify the model effect. The generated digital core is as follows: Figure 4 During the validation process, the test data set was used to feed the saved generator model, generating digital core samples with multi-component characteristics. The generated digital cores not only closely matched the real samples in terms of mineral distribution, but also exhibited pore and fracture connectivity consistent with the physical properties of the actual reservoir. Figure 4 and Figure 5 The generated digital core sample is shown in the figure below. The left figure shows the overall geometric characteristics of the core, and the right figure further shows the distribution characteristics of different mineral components and pore areas through multi-component annotation. The digital core generated by the model can be directly applied to reservoir seepage simulation and pore network modeling, providing high-precision digital model support for the development of carbonate reservoirs; the final reconstructed multi-component digital core is as follows: Figure 6 As shown, it contains albite, quartz, calcite, illite, chlorite, pyrite, rutile, dolomite, anhydrite, other minerals, and pores.

[0092] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A multi-component digital core reconstruction method, characterized in that: The following steps are involved: Obtain multiple core images of small samples of carbonate rocks and the core multi-component structure corresponding to each core image; The Wasserstein distance is introduced as the loss function of the generator of the generative adversarial network model GAN ​​to form an improved generator; the prior knowledge decision and gradient penalty are embedded in the discriminator of the generative adversarial network model GAN ​​to form an improved discriminator; and forming an improved generative adversarial network model WGAN-GP including an improved generator and an improved discriminator; The improved generative adversarial network model WGAN-GP is trained using the core multi-component structure corresponding to multiple core images; During training, the generator optimizes the network model parameters layer by layer by inputting random noise vectors and corresponding domain knowledge vectors. Simultaneously, the generator uses Wasserstein distance as a loss function and adjusts the distribution of generated samples through feedback from the discriminator for adversarial training. The discriminator uses embedded prior knowledge to impose constraints on mineral distribution, porosity, and fracture connectivity, and evaluates the degree of match between generated samples and real samples. In the adversarial training of the generator and the discriminator, the generator uses a gradient penalty mechanism to impose gradient constraints on the interpolation points between the real samples and the generated samples to obtain the trained generative adversarial network model WGAN-GP; The core multi-component structure corresponding to the core image to be reconstructed is input into the generator of the trained generative adversarial network model WGAN-GP, and the reconstructed multi-component digital core is output; Improvements to the generative adversarial network model GAN ​​include: The log-likelihood loss of the objective function in the generative adversarial network model GAN ​​is: ； Replace the log-likelihood loss with the Wasserstein distance, the equation of the Wasserstein distance is: ； in: and There are two possible distributions; There are two distributions and The set of joint distributions of , γ is ( , ) a possible joint distribution; x represents the real sample, y Indicates the generated sample, is a sample drawn from γ; Represents two samples x and y The distance between them; f represents the 1-Lipschitz function; Prior knowledge decisions are embedded in the discriminator of the generative adversarial network model GAN ​​to obtain the loss error of multiple component proportions, where the loss function is expressed as: ； in: represents the sample prediction percentage of the proportion of the i-th component; Indicates the true percentage of the sample of the i-th component ratio; Indicates the number of rock component types included in the label; Represents the importance ratio of the original loss function in training the neural network; Indicates the importance ratio of prior knowledge in training neural networks; In order to maintain the conditions of the 1-Lipschitz function, the weight clipping strategy in WGAN is replaced by gradient penalty to penalize the gradient of the discriminator output. The gradient penalty equation is: ； in: is from a uniform distribution Random variables sampled from ; is a hyperparameter.

2. A multi-component digital core reconstruction method according to claim 1, characterized in that: The method of obtaining multiple core images of carbonate rocks and the core multi-component structures corresponding to the core images includes: Select core samples from carbonate cores and cut them into slices 5-10 mm thick using a slicer, ensuring that the slices contain the main pore and fracture features. After the core sample slices were placed in a vacuum drying oven and dried for a preset time, the selected areas were scanned step by step to obtain high-resolution backscattered electron images (BSE) and multispectral energy spectra (EDS) of the samples. The QEMSCAN method was used to quantitatively evaluate the sample mineral components in the high-resolution backscattered electron images (BSE) and multispectral energy spectra (EDS) to obtain small sample core images of carbonate rocks and the multi-component structure of the core corresponding to the core images.

3. A multi-component digital core reconstruction method according to claim 1, characterized in that: The training of the improved generative adversarial network model WGAN-GP includes: During training, the generator takes as input a random noise vector z and a corresponding domain knowledge vector k, which are experimentally obtained mineral composition and pore distribution information. Based on this information, the generator optimizes the parameters of the network model layer by layer to generate realistic digital cores. The discriminator receives real samples and generated samples. It imposes constraints on mineral distribution, porosity, and fracture connectivity through embedded prior knowledge decisions and outputs a scalar as an approximation of the Wasserstein distance to evaluate the difference between the generated samples and the real sample distribution. The generator uses the Wasserstein distance as a loss function and is updated based on the feedback from the discriminator to minimize the Wasserstein distance and adjust the distribution of the generated samples. At the same time, the discriminator uses the gradient penalty mechanism to regularize the gradient of the interpolation points between the real samples and the generated samples to ensure that the discriminator meets the continuity condition of the 1-Lipschitz function.

4. A multi-component digital core reconstruction method according to claim 1, characterized in that: The output of the trained generative adversarial network model WGAN-GP includes: After the training of the generative adversarial network model WGAN-GP is completed, the input of the model generator is a random noise vector z and a domain knowledge vector k; the random noise vector z is a high-dimensional vector sampled from a standard normal distribution, and the knowledge vector k is the mineral composition ratio extracted from the actual collected data; After the random noise vector z and the domain knowledge vector k are concatenated, the generator maps them to a high-dimensional feature space through a fully connected layer and expands them into an initial low-resolution feature map. The generator then gradually upsamples the resolution of the feature map by performing multi-layer transposed convolution using a three-dimensional convolution kernel. Each layer of transposed convolution refines the features, gradually generating a multi-component digital core with mineral distribution, pore structure, and fracture connectivity.

5. A multi-component digital core reconstruction device, characterized in that: include: A data module is used to obtain multiple core images of small samples of carbonate rocks and the core multi-component structure corresponding to each core image; A model building module for introducing Wasserstein distance as the loss function of the generator of the generative adversarial network model (GAN) to form an improved generator; embedding prior knowledge decision and gradient penalty in the discriminator of the generative adversarial network model (GAN) to form an improved discriminator; and forming an improved generative adversarial network model (WGAN-GP) including the improved generator and the improved discriminator; The model training module is used to train the improved generative adversarial network model WGAN-GP using the core multi-component structure corresponding to multiple core images; During training, the generator optimizes the network model parameters layer by layer by inputting random noise vectors and corresponding domain knowledge vectors. Simultaneously, the generator uses Wasserstein distance as a loss function and adjusts the distribution of generated samples through feedback from the discriminator for adversarial training. The discriminator uses embedded prior knowledge to impose constraints on mineral distribution, porosity, and fracture connectivity, and evaluates the degree of match between generated samples and real samples. In the adversarial training of the generator and the discriminator, the generator uses a gradient penalty mechanism to impose gradient constraints on the interpolation points between the real samples and the generated samples to obtain the trained generative adversarial network model WGAN-GP; The reconstruction module is used to input the core multi-component structure corresponding to the core image to be reconstructed into the generator of the trained generative adversarial network model WGAN-GP, and output the reconstructed multi-component digital core; Improvements to the generative adversarial network model GAN ​​include: The log-likelihood loss of the objective function in the generative adversarial network model GAN ​​is: ； Replace the log-likelihood loss with the Wasserstein distance, the equation of the Wasserstein distance is: ； in: and There are two possible distributions; There are two distributions and The set of joint distributions of , γ is ( , ) a possible joint distribution; x represents the real sample, y Indicates the generated sample, is a sample drawn from γ; Represents two samples x and y The distance between them; f represents the 1-Lipschitz function; Prior knowledge decisions are embedded in the discriminator of the generative adversarial network model GAN ​​to obtain the loss error of multiple component proportions, where the loss function is expressed as: ； in: represents the sample prediction percentage of the proportion of the i-th component; Indicates the true percentage of the sample of the i-th component ratio; Indicates the number of rock component types included in the label; Represents the importance ratio of the original loss function in training the neural network; Indicates the importance ratio of prior knowledge in training neural networks; In order to maintain the conditions of the 1-Lipschitz function, the weight clipping strategy in WGAN is replaced by gradient penalty to penalize the gradient of the discriminator output. The gradient penalty equation is: ； in: is from a uniform distribution Random variables sampled from ; is a hyperparameter.

6. An electronic device, characterized in that: include: memory and processor; The memory is used to store computer programs; The processor is configured to implement the steps of a multi-component digital core reconstruction method according to any one of claims 1 to 4 when executing the computer program stored in the memory.

7. A computer-readable storage medium, characterized in that Used to store a computer program, which, when executed by a processor, implements the steps of a multi-component digital core reconstruction method according to any one of claims 1 to 4.

Images