A fast multi-scale digital core reconstruction method based on surrogate model

Abstract

The present invention discloses a multi-scale digital core rapid reconstruction method based on an agent model, and belongs to the field of digital core rapid reconstruction. The present invention provides all-round support for constructing a multi-scale digital core by combining multiple steps with a model, covering the key links of data acquisition, feature extraction, model construction, and multi-scale information fusion, and provides a basis for reservoir physical property evaluation by accurately predicting permeability, combining microscopic images and macroscopic properties, and realizing effective connection of cross-scale information, and finally generating a digital core with a real microstructure, providing a high-fidelity data model for oil and gas exploration.

Description

Technical Field

[0001] The invention belongs to the technical field of rapid digital core reconstruction, and in particular relates to a multi-scale rapid digital core reconstruction method based on an agent model. Background Art

[0002] In recent years, digital core construction technology has developed rapidly in the study of rock physical properties of unconventional reservoirs, especially in the construction of multi-mineral and multi-scale digital cores. Multi-mineral digital cores can incorporate the microstructures of different mineral types and pores into digital models by combining micron-scale computed tomography (micron CT) and scanning electron microscopy (SEM) imaging, overcoming the limitation of traditional single mineral models that cannot accurately describe the complex pore structure of reservoirs, making the simulation of the physical properties of complex reservoirs such as tight sandstones more accurate.

[0003] At present, digital core construction is mainly divided into two categories: image fusion construction and pore network integrated construction. Image fusion modeling methods include mixed superposition, template matching and deep learning. By fusing image data of different resolutions, it can accurately reflect the three-dimensional structure inside the rock and perform numerical simulation of multiple physical fields. However, due to its computational efficiency limitations, it is difficult to handle mixed constructions with large scale differences. Relatively speaking, pore network integrated construction can realize the construction of multiple continuous scales, and the numerical simulation efficiency is relatively high, but it has certain limitations in describing mineral properties and complex physical properties.

[0004] Four existing technical solutions for multi-scale digital core construction are as follows: image fusion construction method; pore network integration construction method; generative adversarial neural network (GAN) modeling method; three-dimensional Voronoi volume modeling method.

[0005] The evaluation of existing technologies is summarized as follows: 1. Image fusion construction method: By fusing image data of different resolutions, such as micron CT and energy dispersive scanning electron microscopy (EDS), a digital core containing microstructure is generated. This method can accurately reflect the three-dimensional structure inside the rock and is suitable for numerical simulation of multi-physical fields.

[0006] Low computational efficiency: Since image data of different resolutions need to be processed, the amount of computation is large and time-consuming, especially when faced with large scale differences, it is difficult to perform effective fusion.

[0007] Difficulty in handling scale differences: When fusing data of different scales, the mismatch between field of view and resolution may lead to unsatisfactory fusion effects, thus affecting modeling accuracy.

[0008] 2. Pore network integrated construction method: The digital core model is constructed by explicit micropore network or adding additional throats and fracture systems. This method can realize the construction of multiple continuous scales and has high numerical simulation efficiency.

[0009] Limited to specific physical properties: This method is mainly used to describe physical properties such as seepage and conductivity. It is difficult to accurately simulate the complex mineral properties of rocks and their coupling with other physical processes.

[0010] Lack of description of complex mineral compositions: The details and diversity of minerals cannot be fully reflected, so there are limitations in simulating complex reservoirs containing multiple minerals.

[0011] 3. Generative Adversarial Neural Network (GAN) Construction Method: The introduction of deep learning technology, especially generative adversarial network (GAN) and its improved versions (such as AttentionGAN), can combine X-ray micro-CT (micro-CT) and scanning electron microscope (SEM) images to generate high-resolution multi-scale digital core images. This method can retain the original image spatial information and enhance the expression of image details and mineral information. It is a high-precision digital core numerical simulation solution.

[0012] Large data requirements: Deep learning methods require a large amount of high-quality image data for training, especially GAN and its improved versions, which require paired high-resolution and low-resolution image data, which may be difficult to obtain in practice.

[0013] High model complexity: It requires high computing resources and time to train the model. For complex digital core modeling with multiple scales and components, the implementation process may be too complicated.

[0014] 4. Three-dimensional Voronoi body construction method: Based on the Voronoi theory, a three-dimensional digital core model is established to characterize the spatial distribution of mineral composition, pores and cracks. This method can analyze the mechanical properties and damage evolution process of rocks from a microscopic perspective. The simulation results are highly consistent with laboratory experiments and are particularly suitable for mechanical behavior research.

[0015] The accuracy of simulated details is limited: Although Voronoi volume modeling can better describe the random microstructure of rocks, its modeling process simplifies the true morphology of mineral particles and may ignore some detailed features, especially in complex porous media.

[0016] Limited to mechanical behavior simulation: This method is particularly suitable for mechanical behavior research, but its ability to describe other physical properties of rocks such as seepage and electrical properties is relatively weak, and its application scope is relatively limited.

[0017] In summary, the existing digital core construction technology still has limitations in the application of multi-mineral, multi-scale fusion and its application in rock physics and mechanical behavior research, such as insufficient computing power, small representation space scale, and insufficient representation accuracy. A multi-scale digital core rapid reconstruction method based on an agent model combines deep learning, image fusion and multi-scale modeling technologies to establish a multi-scale digital core construction method. Its speed and accuracy are greatly improved compared with the existing methods, and the multi-scale representation is more accurate, providing more solid technical support for promoting the quantitative research of unconventional reservoirs and oil and gas field development. Summary of the invention

[0018] In view of the above technical problems existing in the prior art, the present invention proposes a multi-scale digital core rapid reconstruction method based on a proxy model, which has a reasonable design, overcomes the shortcomings of the prior art, and has good effects.

[0019] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a method for rapid reconstruction of multi-scale digital cores based on a proxy model, comprising the following steps: step 1: data acquisition and image stitching; step 2: image classification and texture feature extraction; step 3: hybrid proxy model construction; extracting depth features to achieve permeability prediction; step 4: constructing a Transformer model based on physics guidance; step 5: multi-scale digital core construction.

[0020] Preferably, in step 1, the core sample is imaged step by step by a scanning electron microscope to generate high-resolution images with overlapping edges, and these images are processed by an image stitching algorithm to generate a scanning electron microscope stitched image.

[0021] Preferably, in step 2, a Gaussian mixture model is used to perform regional feature classification on the stitched scanning electron microscope stitching image to extract the microstructural information of the core, and the image is segmented using the K-means clustering algorithm through the extracted texture features including contrast, homogeneity, energy, correlation and entropy, and the image is divided into a categories, each category representing a specific feature of the core.

[0022] Preferably, step 3 specifically includes the following steps: step 3.1: obtain a data set for training the proxy model, including images and their corresponding permeability labels; step 3.2: standardize, grayscale and resize the images, and adjust the image size to 256×256 pixels to adapt to the VGG network input requirements; step 3.3: use AVIZO software to calculate the permeability value of each microstructure sample to generate an "image-permeability" data set; step 3.4: extract the deep feature vector of the micro core image through the VGG convolutional neural network, and establish a nonlinear relationship between image features and permeability through the BP neural network to achieve permeability prediction.

[0023] Preferably, step 3.4 specifically includes the following steps: step 3.4.1: data preprocessing; retain the first 7 convolution and pooling layers for low-level and mid-level image feature extraction; obtain the high-dimensional feature vector output by the 7th layer as the complex feature expression of the micro core image; step 3.4.2: after completing the data preprocessing, the micro core image is input into the VGG network, and the high-dimensional feature vector is extracted through the convolution layer and the pooling layer. The extracted high-dimensional feature vector includes the geometric features of the core and the texture features of the image, which are used to train the BP neural network; step 3.4.3: the input layer of the BP neural network accepts the high-dimensional feature vector, and the output layer generates the permeability prediction value , each layer of the network implements the following nonlinear mapping: ;in, is the activation value of the Lth layer, The weight matrix of the Lth layer, is the bias vector, It is a nonlinear activation function. By back-propagation optimization of weights and biases, the network learns the mapping relationship between input features and output targets. Step 3.4.4: Complete the permeability prediction by learning the nonlinear mapping relationship between image features and permeability labels.

[0024] Preferably, step 3.4.4 specifically includes the following steps: step 3.4.4.1: using the feature vector output by the VGG network as the input layer node of the BP neural network, and the permeability label obtained by AVIZO calculation as the output target; step 3.4.4.2: network parameter adjustment; setting hyperparameters including the number of BP network layers, the number of nodes, and the learning rate, and using the mean square error as the loss function during the optimization process to minimize the error between the predicted permeability and the actual permeability; using the mean square error as the loss function in the training stage to measure the deviation between the predicted value and the true value, the formula is: ;in, For the The true penetration rate of samples is is the predicted value, N is the total number of samples; Step 3.4.4.3: Model training and optimization; The BP network parameters are iteratively updated through the back propagation algorithm, and the network weights are continuously adjusted on the training set until the optimal model is obtained; The optimization process uses the gradient descent algorithm, and the rule for updating parameters each time is as follows: ;in, is the learning rate, which determines the step size of parameter update. Through multiple iterations, the model gradually converges to the optimal parameters.

[0025] Preferably, in step 4, the physics-guided Transformer model uses the feature vector extracted by VGG as input, predicts the permeability through the BP neural network, adopts the mean square error as the loss function, and performs model training and optimization through back propagation and gradient descent algorithms to finally obtain the optimal prediction result; the physics-guided Transformer model framework includes an input module, a Transformer feature extraction module, a physical embedding module and a prediction module; wherein, the input module: is configured to input multimodal data, including core images and physical prior characteristics; the Transformer feature extraction module: uses the self-attention mechanism to extract the global features of the microscopic core image, and adds the multi-head attention mechanism to capture the features of different scales and directions; the physical embedding module: in Tr The middle layer of the ansformer model is used to introduce prior knowledge based on physical equations and constrain the feature extraction process; prediction module: use the fully connected layer to output the permeability prediction value, and add a physical consistency loss function to optimize the result; preferably, in step 4, the specific implementation steps of the physically guided Transformer model are as follows: step 4.1: data preparation; scanning electron microscope stitching images, uniformly adjust the images to 256×256 pixels, adapt to the Transformer input requirements, and bind each image to the corresponding physical properties to form a multimodal data pair; step 4.2: Transformer model construction; using the Transformer architecture, the image features are used as the main input, and the complete embedding input is generated by supplementing the physical property encoding; the specific contents are as follows: ① Embedding of image input X: ; Where X is the original input image data, It is the small block after the image is segmented; It is the embedded representation of the image patch, obtained after linear transformation and position encoding. It is a feature vector representing the image content and spatial position, and is the basic input for Transformer to process images. ② Physical property input: physical property input includes porosity, mineral composition and permeability label data. The physical property is mapped to the same feature dimension as the image embedding through the multi-layer perceptron to unify the data dimension. ③ Fusion of image and physical property: directly embed the image into and physical embedding To splice; ; Input: Image embedding and physical properties embedded ; Output: combined feature vector ; Through the input module, the model can simultaneously utilize the geometric features, texture features and physical property data of the core image, and organically integrate this information into a unified input representation. Among them, the geometric features include pore distribution, particle morphology and boundary connectivity, the texture features cover image attributes including contrast, homogeneity, energy and entropy, and the physical property data provides key parameters including porosity, mineral composition and permeability. After preprocessing and feature extraction, these multi-dimensional information are converted into high-dimensional vectors as the input of Transformer. The model integrates the spatial distribution relationship in geometric and texture characteristics and the regularity information in physical properties to achieve comprehensive modeling of the microstructure and macroscopic flow properties of the core.

[0026] Preferably, step 5 specifically includes the following steps: step 5.1: integrating microscopic features for preliminary construction of a multi-scale model through the scanning electron microscope stitching image obtained in step 1 and the classification information extracted in step 2; step 5.2: establishing a mapping relationship between microstructure and macroscopic physical properties through the depth features extracted in step 3 and the permeability prediction model; the prediction formula of the permeability K is constructed based on the input feature F and the BP network parameter θ, and the specific formula is: ; In the formula, The high-dimensional feature vector extracted by the VGG network is combined with texture features as input; As the activation function, nonlinear characteristics are introduced; are the weights and bias parameters in the BP network, which are obtained by back propagation optimization; Step 5.3: Combine steps 1, 2, and 4, and use the scanning electron microscope stitching image and physical property data and the physically guided Transformer model to realize the construction of multi-scale digital cores and capture the geometry, texture, and physical properties of the cores; through the multi-head attention mechanism of the Transformer model, image features from different scales are integrated to form a complete core feature expression.

[0027] The beneficial technical effects brought by the present invention are as follows: 1. By using the physics-guided Transformer model and the self-attention mechanism of deep learning, it is possible to capture the core image features while incorporating physical laws such as permeability, significantly improving the accuracy of permeability prediction, and in a physical sense it is more in line with the actual reservoir conditions.

[0028] 2. This solution can more accurately describe the relationship between microstructure and macroscopic properties by extracting image features at different scales and effectively mapping between different scales. The multi-scale digital core constructed can not only capture the detailed information of the core, but also simulate the physical properties of the core at different scales, effectively enhancing the authenticity and reliability of the digital core model.

[0029] 3. The physics-guided Transformer model introduces prior knowledge based on physical equations in the middle layer of the model through the physical embedding module, which improves the prediction accuracy. Traditional deep learning models are often difficult to explain their prediction results. This technical solution introduces physical laws to enable the model to capture image features while maintaining a high degree of consistency with physical features. The use of physical consistency loss functions further ensures that the model's prediction results conform to physical laws.

[0030] 4. By combining image segmentation results and macroscopic physical parameters such as permeability, the model can generate a core model containing microstructural details and macroscopic properties, and conduct subsequent verification and optimization. During the model training process, the comparison between the predicted permeability values ​​of the proxy model and the experimental values ​​can help optimize the network parameters and feature extraction process, thereby achieving continuous improvement of the model.

[0031] The present invention combines multiple steps with models to provide all-round support for building multi-scale digital cores, covering the key links of data acquisition, feature extraction, model construction and multi-scale information fusion. It provides a basis for reservoir physical property evaluation by accurately predicting permeability, combines microscopic images and macroscopic properties to achieve effective connection of cross-scale information, and finally generates digital cores with real microstructures, providing high-fidelity data models for oil and gas exploration. The use of proxy models reduces the amount of calculation for model construction and improves the calculation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A complete flowchart for data acquisition, image classification, and hybrid agent model construction; Figure 2 This is a diagram of the physics-guided Transformer model framework; Figure 3 Diagram of the physics-guided Transformer input and feature extraction process; Figure 4 Constructing a flow chart for multiscale digital cores. DETAILED DESCRIPTION

[0033] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Definition of key terms: (1) CT (Computed tomography) scanning: An X-ray source is used to emit X-rays through a sample, and a sensor is used to collect the attenuated X-ray signals, and finally the X-ray signals are converted into a three-dimensional CT image through computer signal processing.

[0034] (2) BP (Back Propagation) neural network: It is a multi-layer feedforward neural network that uses the back propagation algorithm for training and is suitable for implementing complex regression tasks. The BP neural network is responsible for learning the mapping relationship between the VGG feature vector and the permeability label.

[0035] (3) VGG (Visual Geometry Group, specific CNN architecture) network: The VGG-16 model contains 13 convolutional layers and 3 fully connected layers. The image feature extraction uses the first 7 layers of VGG-16 to obtain the intermediate feature map. Specifically, the first few layers of the VGG network extract low-level and mid-level features from the microscopic core image by stacking convolutional layers and pooling layers, thereby retaining the geometric and texture information contained in the image.

[0036] (4) Mean square error: It is a measure of the difference between the estimator and the estimated quantity and is an effective way to measure the average error.

[0037] (5) Attention Mechanism: It is a technology used in deep learning models, which aims to enable the model to "focus" on the most important part of the input data when processing the input data. This mechanism was originally proposed to solve sequence-to-sequence problems (such as translation tasks) in natural language processing (NLP), and was later widely used in various tasks such as image processing and text generation.

[0038] (6) SEM-MAPs: By stitching together high-resolution SEM scan images from multiple fields of view, large-scale, high-resolution SEM scan images can be obtained. SEM-MAPs are large-scale, high-resolution SEM images.

[0039] This technical solution combines microscopic core images and physical property data to accurately predict permeability and provide a basis for reservoir physical property evaluation. Its implementation process includes the following steps: 1. Data acquisition and image processing: Use a high-resolution scanning electron microscope (FE-SEM) to obtain core microscopic images, and generate high-resolution stitched images (SEM-MAPs) through an image stitching algorithm. Feature extraction and classification: Use a Gaussian mixture model (GMM) to extract image texture features, use the K-means algorithm for classification, and extract the microstructural information of the core (such as porosity and mineral composition).

[0040] 2. Model training and permeability prediction: Use VGG convolutional neural network to extract deep feature vectors, and use BP neural network to establish nonlinear relationship between image features and permeability to achieve permeability prediction. In order to improve model accuracy, the physics-guided Transformer model is combined with deep learning and physical laws to make accurate permeability predictions.

[0041] 3. Multi-scale digital core construction: Integrate core features (geometry, texture, physical properties) at different scales to build a unified digital core model, provide accurate permeability prediction, and provide a high-fidelity digital core data model for oil and gas exploration.

[0042] The details are as follows: 1. Data acquisition and image stitching: SEM is used to gradually image the core samples to generate high-resolution images with overlapping edges. These images are then stitched together into a large-viewing area, high-resolution SEM-MAPs using an image stitching algorithm. This high-resolution SEM-MAPs provides high-quality basic data for subsequent feature extraction and model training, and can truly reflect the microstructure of the core.

[0043] 2. Image classification and texture feature extraction: GMM is used to classify the regional features of SEM-MAPs and extract the microstructural information of the core. The image is segmented using the extracted texture features and the K-means clustering algorithm is used to divide the image into 5 categories. Each category represents a specific feature of the core (such as porosity, mineral composition, etc.), providing basic data for lithology analysis and reservoir evaluation.

[0044] 3. Hybrid proxy model construction and VGG convolutional neural network; The permeability value of each microstructure sample is calculated by AVIZO software, and a dataset of image and permeability label pairs is generated. The images are standardized, grayscaled, and resized (adjusted to 256×256 pixels) to adapt to the input requirements of the VGG convolutional neural network. Subsequently, the pre-trained VGG-16 network is used for feature extraction, and the first 7 convolutional and pooling layers are retained for the extraction of low-level and mid-level image features. By training the BP neural network, the nonlinear mapping relationship between image features and permeability labels is learned to complete the prediction of permeability.

[0045] The pre-trained VGG-16 network is used to adjust some of its structures: 1) the first 7 convolution and pooling layers are retained for low-level and mid-level image feature extraction; 2) the high-dimensional feature vector output by the 7th layer is obtained as the complex feature expression of the microscopic core image; 3) the extracted feature vector includes geometric and texture information, which is the input of the subsequent BP network; after completing the data preprocessing, the microscopic core image is input to the VGG network, and the high-dimensional feature vector is extracted through the convolution layer and the pooling layer. The features extracted in the VGG network not only cover the geometric features of the core such as porosity distribution, but also include the texture features of the image such as contrast and homogeneity. The extracted high-dimensional vector is used to train the BP neural network. Assume that the extracted feature vector is represented by a vector X=[X1, X2,…,Xn], whose dimension is n, corresponding to the multi-dimensional feature expression of the image.

[0046] The input layer of the BP neural network accepts these feature vectors, and the output layer generates the permeability prediction value , each layer of the network implements the following nonlinear mapping: ;in, is the activation value of the Lth layer, The weight matrix of the Lth layer, is the bias vector, It is a nonlinear activation function. By back-propagating and optimizing the weights and biases, the network learns the mapping relationship between input features and output targets. By learning the nonlinear mapping relationship between image features and permeability labels, permeability prediction is completed. The specific steps include: Input and training data: 1) Use the feature vector output by the VGG network as the input layer node of the BP neural network, and the permeability label (obtained by AVIZO calculation) as the output target; 2) Network parameter adjustment: Set the hyperparameters such as the number of BP network layers, the number of nodes, and the learning rate. During the optimization process, the mean square error (MSE) is used as the loss function to minimize the error between the predicted permeability and the actual permeability; In the training stage, the mean square error is used as the loss function to measure the deviation between the predicted value and the true value. The formula is: ;in is the true permeability of the ith sample, is the predicted value, N is the total number of samples, and by minimizing the loss function, the weight of the BP network , and bias is continuously updated to improve the prediction accuracy.

[0047] 3) Model training and optimization: The BP network parameters are iteratively updated through the back propagation algorithm, and the network weights are continuously adjusted on the training set until the optimal model is obtained; the optimization process uses the gradient descent algorithm, and the rule for updating parameters in each iteration is: ;in, is the learning rate, which determines the step size of parameter update. Through multiple iterations, the model gradually converges to the optimal parameters.

[0048] 4. Construction of the physics-guided Transformer model: In this step, the physics-guided Transformer model is used to replace the separated structure of VGG+BP. This model combines physical constraints and the self-attention mechanism of deep learning. While capturing the long-range dependency characteristics of the image, it incorporates the physical laws of permeability, improving the prediction accuracy and interpretability of the model. The Transformer model input includes core images and physical properties (such as porosity, mineral composition, etc.). The global features of the microscopic core images are extracted through the self-attention mechanism and the multi-head attention mechanism, and a physical embedding module is added to the model to ensure the physical consistency of the prediction results.

[0049] The physics-guided Transformer model framework is divided into the following modules: 1) Input module: input multimodal data, including core images and physical prior characteristics (such as porosity and mineral composition); 2) Transformer feature extraction module: use the self-attention mechanism to extract the global features of the microscopic core image, and add a multi-head attention mechanism to capture features of different scales and directions; 3) Physical embedding module: in the middle layer of the Transformer model, introduce prior knowledge based on physical equations to constrain the feature extraction process.

[0050] 5) Prediction module: Use the fully connected layer to output the permeability prediction value, and add a physical consistency loss function to optimize the result.

[0051] The specific implementation steps of the above Transformer model include: 1) Data preparation. SEM-MAPs stitched images (high-resolution electron microscope images) are uniformly adjusted to 256×256 pixels to adapt to the Transformer input requirements, and random rotation, cropping, flipping, noise addition and other methods are applied to enhance data diversity to form a core image dataset for training and testing. Porosity, mineral composition, and permeability values ​​(calculated or experimentally obtained) are collected, and each image is bound to the corresponding physical properties to form a multimodal data pair.

[0052] 2) Transformer model construction: Using the Transformer architecture, image features are used as the main input, and the complete embedding input is generated by encoding supplementary physical properties.

[0053] Embedding of image input X: ; where X is the original input image data (e.g. a 256×256 pixel two-dimensional core image), The small blocks after image segmentation; It is the embedded representation of the image patch, obtained after linear transformation and position encoding. It is a feature vector representing the image content and spatial position, and is the basic input for Transformer to process images. Position encoding.

[0054] 3) Physical property input. Physical properties include porosity, mineral composition, permeability labels and other data related to the physical properties of the core; The embedded representation of physical properties uses a multi-layer perceptron to map the physical properties to the same feature dimension as the image embedding (e.g. 512 dimensions). By unifying the data dimension, it is easier to process by the subsequent Transformer.

[0055] 4) Fusion of images and physical properties.

[0056] Embed images directly and physical embedding To splice; ; Input: Image embedding and physical properties embedded ; Output: combined feature vector , for subsequent processing by the Transformer model.

[0057] Through the above input modules, the model can simultaneously utilize the geometric features, texture features and physical property data of the core image, and organically integrate this information into a unified input representation. The geometric features include pore distribution, particle morphology and boundary connectivity, and the texture features cover image attributes such as contrast, homogeneity, energy and entropy. The physical property data provides key parameters such as porosity, mineral composition and permeability. After preprocessing and feature extraction, these multi-dimensional information are converted into high-dimensional vectors as the input of the Transformer. The model integrates the spatial distribution relationship in the geometric and texture characteristics and the regularity information in the physical characteristics to achieve comprehensive modeling of the core microstructure and macroscopic flow properties, thereby improving the accuracy in permeability prediction while maintaining consistency with actual physical laws, further enhancing the versatility and reliability of the model.

[0058] 5. Multi-scale digital core construction (combining steps 1, 2, and 3).

[0059] A preliminary multi-scale digital core model can be constructed by integrating the high-resolution SEM-MAPs obtained in step 1, the classification information extracted in step 2, and the depth features extracted in step 3. In this process, the model establishes the relationship between microstructure and macroscopic physical properties (such as permeability) through multi-dimensional feature extraction and cross-scale feature mapping, providing strong support for permeability prediction.

[0060] 6. Construction of multi-scale digital cores based on the physics-guided Transformer model.

[0061] 1) Combining steps 1, 2, and 3, the method for constructing a multi-scale digital core generates a comprehensive and accurate digital core model through multi-dimensional feature extraction and cross-scale characteristic mapping of microscopic images. The following is the specific method and its connection with the previous steps: through the high-resolution stitched image SEM-MAPs obtained in step 1 and the classification information extracted in step 2, the microscopic features are integrated for the preliminary construction of the multi-scale model, and the mapping relationship between the microstructure and the macroscopic physical property is established through the depth features extracted in step 3 and the permeability prediction model.

[0062] The prediction formula of permeability K is constructed based on the input feature F and BP network parameter θ. The specific formula is: ; In the formula, The high-dimensional feature vector extracted by the VGG network is combined with texture features as input; As the activation function, nonlinear characteristics are introduced; are the weight and bias parameters in the BP network, which are obtained by back propagation optimization.

[0063] 2) Combining steps 1, 2, and 4, using SEM-MAPs and physical property data, the physically guided Transformer model can realize the construction of multi-scale digital cores and capture the geometry, texture, and physical properties of the cores. Through the multi-head attention mechanism of the model, image features from different scales (such as porosity, grain morphology, boundary connectivity, etc.) are fused to form a complete core feature expression. The information of different scales is integrated into a unified input vector for subsequent permeability prediction and core property analysis; based on the input multimodal data (core images and physical properties), the permeability of the core is predicted, and the prediction results are optimized to ensure that the permeability prediction is consistent with the actual physical laws, and to improve the interpretability and accuracy of the model.

[0064] Combining steps 1, 2, and 4, a multi-scale digital core is constructed under the framework of the physics-guided Transformer model. The model uses the Transformer architecture to extract the geometry, texture, and physical properties of the core through the input of high-resolution SEM-MAPs and physical property data, and integrates information at different scales. Through the multi-head attention mechanism, the model can integrate the core image features at different scales, and optimize the permeability prediction through the physical consistency loss function to ensure that the model results conform to the actual physical laws. In this process, the microstructure and macroscopic flow characteristics of the core are comprehensively modeled, and the digital core finally generated can provide a high-fidelity data model for oil and gas exploration.

[0065] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A multi-scale digital core rapid reconstruction method based on a proxy model, characterized by: The steps include: Step 1: Data acquisition and image stitching; Step 2: Image classification and texture feature extraction; Step 3: Hybrid proxy model construction; extracting depth features to achieve permeability prediction; Step 4: Construction of Transformer model based on physics guidance; Step 5: Multi-scale digital core construction; Step 3 specifically includes the following steps: Step 3.1: Obtain the dataset for training the proxy model, including images and their corresponding permeability labels; Step 3.2: Standardize, grayscale and resize the image to 256×256 pixels to meet the VGG network input requirements; Step 3.3: Use AVIZO software to calculate the permeability value of each microstructure sample and generate an "image-permeability" data set; Step 3.4: The deep feature vector of the microscopic core image is extracted through the VGG convolutional neural network, and the nonlinear relationship between the image feature and the permeability is established through the BP neural network to achieve permeability prediction; In step 4, the physics-guided Transformer model uses the feature vector extracted by VGG as input, predicts the permeability through the BP neural network, adopts the mean square error as the loss function, and trains and optimizes the model through back propagation and gradient descent algorithms to finally obtain the optimal prediction result; The physics-guided Transformer model framework includes an input module, a Transformer feature extraction module, a physical embedding module, and a prediction module; An input module configured to input multimodal data including core images and physical prior properties; The Transformer feature extraction module is configured to extract global features of microscopic core images using the self-attention mechanism and add a multi-head attention mechanism to capture features of different scales and directions; The physical embedding module, in the middle layer of the Transformer model, is configured to introduce prior knowledge based on physical equations to constrain the feature extraction process; The prediction module is configured to output the permeability prediction value using a fully connected layer and add a physical consistency loss function to optimize the result.

2. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 1 is characterized by: In step 1, the core samples are imaged step by step by scanning electron microscopy to generate high-resolution images with edge overlap, and these images are processed by an image stitching algorithm to generate a scanning electron microscope stitched image.

3. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 1 is characterized by: In step 2, the Gaussian mixture model is used to classify the regional features of the stitched SEM images to extract the microstructural information of the core. The K-means clustering algorithm is used to segment the image through the extracted texture features including contrast, homogeneity, energy, correlation and entropy. The image is divided into a categories, each of which represents a specific feature of the core.

4. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 1 is characterized by: Step 3.4 specifically includes the following steps: Step 3.4.1: Data preprocessing; The first 7 convolution and pooling layers are retained for low-level and mid-level image feature extraction; Obtain the high-dimensional feature vector output by the 7th layer as the complex feature expression of the microscopic core image; Step 3.4.2: After completing data preprocessing, the microscopic core image is input into the VGG network, and the high-dimensional feature vector is extracted through the convolution layer and the pooling layer. The extracted high-dimensional feature vector includes the geometric features of the core and the texture features of the image, which are used to train the BP neural network; Step 3.4.3: The input layer of the BP neural network accepts high-dimensional feature vectors, and the output layer generates permeability prediction values Each layer of the network implements the following nonlinear mapping: a (l) =f(W (l) a (l-1) +b (l) ); Among them, a (1) is the activation value of the Lth layer, W (l) The weight matrix of the Lth layer, b (l) is the bias vector, f is the nonlinear activation function, and the network learns the mapping relationship between input features and output targets by back-propagating and optimizing the weights and biases; Step 3.4.4: Complete the permeability prediction by learning the nonlinear mapping relationship between image features and permeability labels.

5. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 4 is characterized by: Step 3.4.4 specifically includes the following steps: Step 3.4.4.1: Use the feature vector output by the VGG network as the input layer node of the BP neural network, and the permeability label obtained by AVIZO calculation as the output target; Step 3.4.4.2: Network parameter adjustment; Hyperparameters including the number of BP network layers, number of nodes, and learning rate were set. During the optimization process, mean square error was used as the loss function to minimize the error between the predicted permeability and the actual permeability. The mean square error is used as the loss function in the training phase to measure the deviation between the predicted value and the true value. The formula is: Among them, y i is the true permeability of the ith sample, is the predicted value, N is the total number of samples; Step 3.4.4.3: Model training and optimization; The BP network parameters are iteratively updated through the back propagation algorithm, and the network weights are continuously adjusted on the training set until the optimal model is obtained; The optimization process uses the gradient descent algorithm, and the rule for updating parameters in each iteration is: Among them, η is the learning rate, which determines the step size of parameter update. Through multiple iterations, the model gradually converges to the optimal parameters.

6. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 1 is characterized by: In step 4, the specific implementation steps of the physical guided Transformer model are as follows: Step 4.1: Data preparation; The scanning electron microscope stitched images and adjusted the image size to 256×256 pixels to adapt to the Transformer input requirements. Each image was bound to the corresponding physical properties to form a multimodal data pair. Step 4.2: Transformer model construction; Using the Transformer architecture, image features are used as the main input, and the complete embedding input is generated by encoding supplementary physical properties; the specific contents are as follows: Embedding of image input X: X emb =Linear(X patch )+PositionEmbedding; In the formula, X is the original input image data, X patch is the small block after image segmentation; X emb It is the embedded representation of the image patch, obtained after linear transformation and position encoding. It is a feature vector representing the image content and spatial position, and is the basic input for Transformer to process images. PositionEmbedding is the position encoding. Physical property input; The physical property input includes porosity, mineral composition and permeability label data. The physical properties are mapped to the same feature dimension as the image embedding through a multi-layer perceptron to unify the data dimension. Fusion of images and physical properties; Embed images directly into X emb and physical embedding P emb To splice; Z input =Concat(X emb ,P emb ); Input: image embedding X emb and physical properties embedded in P emb ; Output: Combined feature vector Z input ; Through the input module, the model can simultaneously utilize the geometric features, texture features and physical property data of the core image, and organically integrate this information into a unified input representation. Among them, the geometric features include pore distribution, particle morphology and boundary connectivity, the texture features cover image attributes including contrast, homogeneity, energy and entropy, and the physical property data provides key parameters including porosity, mineral composition and permeability. After preprocessing and feature extraction, these multi-dimensional information are converted into high-dimensional vectors as the input of Transformer. The model integrates the spatial distribution relationship in geometric and texture characteristics and the regularity information in physical properties to achieve comprehensive modeling of the microstructure and macroscopic flow properties of the core.

7. The multi-scale digital core rapid reconstruction method based on proxy model according to claim 1 is characterized by: Step 5 specifically includes the following steps: Step 5.1: Using the SEM stitching image obtained in step 1 and the classification information extracted in step 2, the microscopic features are integrated for the preliminary construction of the multi-scale model; Step 5.2: Establish the mapping relationship between microstructure and macroscopic physical properties through the depth features extracted in step 3 and the permeability prediction model; The prediction formula of permeability K is constructed based on the input feature F and BP network parameter θ. The specific formula is: K=f(F; θ)=W2·σ(W1·F+b1)+b2; In the formula, F is the high-dimensional feature vector extracted by the VGG network, combined with texture features as input; σ is the activation function, which introduces nonlinear characteristics; W1, W2, b1, b2 are the weight and bias parameters in the BP network, which are obtained by back propagation optimization; Step 5.3: Combine steps 1, 2, and 4, and use the SEM stitching image and physical property data and the physics-guided Transformer model to build a multi-scale digital core and capture the geometry, texture, and physical properties of the core. Through the multi-head attention mechanism of the Transformer model, image features from different scales are fused to form a complete core feature expression.

Images