Rock pore three-dimensional high-precision reconstruction method based on graph attention autoencoder

Abstract

This application belongs to the field of 3D geological modeling and deep learning, and specifically discloses a method for high-precision 3D reconstruction of rock pores based on graph attention autoencoders. The method includes: constructing a 3D geometric and topological graph of rock pores; applying random perturbations to the nodes and features in the constructed graph to enhance the graph data, resulting in an enhanced graph; encoding the enhanced graph using an encoder composed of multiple graph attention convolutional layers, dynamically assigning weights to different neighboring nodes through a self-attention mechanism and fusing neighboring node features to obtain a latent representation, and using a decoder to invert the latent representation to obtain the predicted 3D pore structure; and optimizing the encoder and decoder using a dedicated loss function adapted to mineral proportion differences, the dedicated loss function being a weighted combination of cross-entropy loss and loss based on mineral combination similarity. This application can effectively improve the robustness of 3D reconstruction of rock pore structures.

Description

Technical Field

[0001] This application belongs to the field of three-dimensional geological modeling and deep learning, and more specifically, relates to a method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoders. Background Technology

[0002] Complex geological evolution processes result in the heterogeneity and uncertainty of geological space. How to accurately and efficiently describe the invisible geological space in three dimensions is a major challenge in the field of geoscience. Three-dimensional geological models, with their highly intuitive visualization features, can effectively reveal the spatial distribution and topological relationships of geological objects, and describe geological structures, structural features, and attribute distribution.

[0003] In recent years, research on 3D digital rock reconstruction using deep learning technologies (such as CNNs and GANs) has gradually increased. Deep learning models can extract high-level features from existing rock data, capturing multi-scale, multi-attribute pore structure information. However, in the early stages of model training, a large number of high-quality microscopic rock structure images are required. For deep strata or rare rock samples, data acquisition costs are high, and data gaps may even occur. Real pore structure features are extremely complex, including various mineral composition types, connectivity, and other geological features. Current deep learning models mainly focus on the geometry of pores in digital rock structure reconstruction, without considering the global constraints of mineral composition, i.e., the spatial distribution and attribute ratio of different mineral components. This leads to inaccuracies in the mineral composition of the reconstruction results, resulting in poor physical consistency and geological plausibility of the model. Achieving high-precision, multi-attribute 3D digital rock reconstruction with limited data but complex features remains a problem that needs to be solved and overcome.

[0004] Accurately capturing spatial characteristics is a challenging task in processing geological and mineralization information. Current prediction models are mainly divided into three categories: single-pixel-based methods, image or pixel-based models, and model systems built upon graph structures. Single-pixel-based models typically ignore the spatial features of the information, while image (pixel-patch)-based models have stronger spatial feature learning capabilities than single-pixel models. However, the input image size is relatively fixed, and relying on a regular grid structure, they cannot fully consider the spatial anisotropy of geological information. In contrast, GNNs can directly model the complex nonlinear spatial dependencies between geological objects, adapt to graph structure data inputs of different sizes, and more accurately characterize the spatial topological relationships of geological bodies. However, the traditional equal-weight aggregation mechanism of graph convolution has limited ability to model complex heterogeneous structures and lacks robustness to sparse connections and missing local structures, thus limiting the reconstruction quality.

[0005] Therefore, improving the robustness of three-dimensional reconstruction of rock pore structure is an urgent problem to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a high-precision three-dimensional reconstruction method for rock pores based on graph attention autoencoders, which can effectively improve the robustness of three-dimensional reconstruction of rock pore structures.

[0007] To achieve the above objectives, in a first aspect, this application provides a method for high-precision three-dimensional reconstruction of rock pores based on a graph attention autoencoder, comprising the following steps: (1) Construct a three-dimensional geometric and topological relationship diagram of rock pores; (2) Apply random perturbations to the nodes and features in the constructed graph to achieve graph data augmentation and obtain the augmented graph; (3) The enhanced graph is encoded by an encoder composed of multiple graph attention convolutional layers. The weights of different neighboring nodes are dynamically assigned and the features of neighboring nodes are fused through a self-attention mechanism to obtain a latent representation. The latent representation is then inverted using a decoder to obtain the predicted three-dimensional pore structure. (4) The encoder and decoder are optimized by using a special loss function that adapts to the differences in mineral proportions. The special loss function is a weighted combination of cross-entropy loss and loss based on mineral combination similarity.

[0008] As a further preferred step, in step (1), a graph structure is created based on the voxelized three-dimensional rock data. The nodes of the graph represent each pixel in the three-dimensional voxel grid, and the edges connect adjacent pixels. The attribute value of each pixel is associated with the spatial coordinates of the pixel. Each pixel node is represented by the one-hot encoding of the node attribute value of the spatial coordinates and the profile condition data as the node feature. The attribute value is used as the target variable. The adjacency matrix is ​​constructed based on the spatially adjacent pixels.

[0009] As a further preferred embodiment, the edge is constructed as follows: for each node, the Euclidean distance between the node and all other nodes is calculated. If the distance between two nodes is less than a preset threshold, an edge is established between the two nodes. At the same time, each node is additionally connected by its K spatial nearest neighbors.

[0010] As a further preferred embodiment, in step (2), the random perturbation includes adding Gaussian noise to the spatial coordinate features of the node.

[0011] As a further preferred embodiment, in step (3), the encoder consists of multiple graph attention convolutional layers, each layer performing weighted fusion of neighboring node features through a self-attention mechanism; the graph attention network dynamically assigns differentiated weights to different neighboring nodes through a self-attention mechanism to achieve refined aggregation of node features, and each layer fuses neighboring information according to the attention weights between nodes, thereby incorporating the graph structure context into the node representation.

[0012] As a further preferred embodiment, the graph attention convolutional layer performs the following operation in each layer: Perform a linear transformation on the features of the current node and calculate the unnormalized attention score. , ,in This represents a vector concatenation operation. ∈R2Fl+1 is the first l The layer provides a learnable attention vector, with LeakyReLU as the activation function. For a shared weight matrix, and They are the first l Layer nodes i and nodes j Features; The unnormalized attention score is normalized into attention weights using the Softmax function. , ,in, It is a node i The neighboring nodes; Using normalized attention weights as coefficients, the transformed features of neighboring nodes are weighted and summed, then passed through a nonlinear activation function to obtain the node... i In the New features of layer +1 ; A multi-head attention mechanism is used to independently execute the above attention operation multiple times and concatenate the resulting K feature vectors: .

[0013] As a further preferred embodiment, in step S40, the total loss of the dedicated loss function is:

[0014] in, It is cross-entropy loss; It is a loss based on the similarity of mineral assemblages; and It is a weighting coefficient that controls the contribution of both factors to the total loss.

[0015] As a further preferred embodiment, the cross-entropy loss is calculated as follows:

[0016] in, It is the one-hot encoding of the actual tag; It is the class probability predicted by the model; K is the total number of samples; K is the total number of categories of the target variable.

[0017] As a further preferred embodiment, the calculation method for the loss based on mineral assemblage similarity is as follows:

[0018] Where N represents the total number of rock samples; the Similarity function is used to calculate the similarity of mineral composition proportions; For the true first i The proportion of each mineral component; For the predicted first i The proportion of each mineral component; The method for calculating the similarity of the mineral component proportions is as follows: .

[0019] Secondly, this application provides a high-precision three-dimensional reconstruction device for rock pores based on a graph attention autoencoder, including a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the steps of the method described in any one of the above.

[0020] The beneficial effects of this application are as follows: This application designs an encoder-decoder architecture based on graph attention network (GAT), which dynamically assigns weights to different neighbor nodes through a self-attention mechanism to achieve refined modeling of rock pore structure; to address the problems of data sparsity and noise, a graph data augmentation method is introduced to apply random perturbations to nodes and edges during training to improve model robustness; at the same time, by constructing a sufficiently large and structurally sound rock sample dataset and designing a dedicated loss function suitable for cases with multiple mineral components and unbalanced proportions, the generalization ability of the model in real-world scenarios is further enhanced. Attached Figure Description

[0021] Figure 1 This is a flowchart of the implementation method provided in this application; Figure 2 This is a flowchart illustrating the network structure and modeling process of the 3D digital rock reconstruction method based on graph autoencoders provided in this application. Figure 3 The method provided in this application is used to test and evaluate two sets of datasets; (a) for a binary discrete sandstone model, and (b) for a multivariate discrete sandstone model. Figure 4These are experimental cases and statistical characteristic comparisons of the three-dimensional binary attribute lithofacies model provided in this application; where (a), (b), and (c) are connectivity curves, and (d) is a variation function curve; Figure 5 This application presents a comparison of the experimental and statistical characteristics of a three-dimensional discrete-attribute lithofacies model; where (a), (b), and (c) are connectivity curves, and (d) is a variogram curve. Figure 6 This is the equipment structure diagram provided in this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] This application addresses the limitations of traditional deep learning, which relies on fixed receptive fields and struggles to comprehensively capture the global structural features of rock pores. It proposes a graph attention encoding strategy, a graph data augmentation strategy, and an encoder-decoder architecture. First, a 3D geometric and topological graph of rock pores is constructed, transforming the pore structure into a graph representation containing node features and heterogeneous connections. During the encoding stage, graph attention convolutional layers perform multi-round weighted aggregation of node features, dynamically assigning differentiated weights to different neighboring nodes through a self-attention mechanism, thus finely capturing the contextual relationships and heterogeneous connectivity characteristics of complex pore structures. To improve the model's robustness in noisy environments, a graph data augmentation strategy is introduced, applying random perturbations to nodes and their features to enhance the model's adaptability to structural changes and input interference. Finally, a decoder is used to achieve 3D reconstruction of the rock pore structure.

[0024] The method for high-precision 3D reconstruction of rock pore structure based on graph autoencoder and graph attention network provided in this application includes the following process: S1: Construct a three-dimensional geometric and topological diagram of rock pores; S2: Graph autoencoder captures the context and heterogeneous connectivity of complex porous structures; S3: Graph data augmentation strategies apply random perturbations to nodes and features; S4: The coded features are inverted through the decoder to realize the three-dimensional reconstruction of the rock pore structure.

[0025] Preferably, in S1, a graph structure is created based on the voxelized 3D rock data. The nodes of the graph represent each pixel in the 3D voxel grid, and the edges connect spatially adjacent pixels, associating the spatial coordinates of the pixels with attribute values. Each node feature consists of spatial coordinates (x, y, z) and one-hot encoding of profile condition data, with the target variable being the mineral composition attribute, and the adjacency matrix constructed based on spatial adjacency relationships.

[0026] Preferably, in S2, the encoder consists of multiple graph attention convolutional layers (GATConv), each layer performing weighted fusion of neighboring node features through a self-attention mechanism; GAT dynamically assigns differentiated weights to different neighboring nodes through a self-attention mechanism, realizing refined aggregation of node features and incorporating richer graph structure context into the node representation.

[0027] Preferably, in S3, the graph data augmentation strategy further enhances the model's adaptability to the complex structure and noisy environment of 3D digital rocks, including random perturbation of node positions and feature values, improving the model's robustness to structural changes and input interference, thereby enhancing the model's generalization performance in real complex scenarios while maintaining high inference efficiency.

[0028] Preferably, in S4, a dedicated loss function adapted to the differences in mineral proportions is used during the training process. This loss function consists of a mean squared error term and a structural consistency term. The two types of losses are balanced by weighting coefficients to simultaneously ensure the accuracy of attribute proportions and structural connectivity. Specifically, the total loss function is:

[0029] in, It is cross-entropy loss. It is a loss based on the similarity of mineral assemblages. and It is a weighting coefficient that controls the contribution of both factors to the total loss.

[0030] The cross-entropy loss function is commonly used in discrete classification tasks. It measures the difference between the probability distribution of a target variable across multiple discrete categories and the true value. The calculation method is as follows:

[0031] in, It is the one-hot encoding of the actual label. It is the class probability predicted by the model. This is the total number of samples. K is the total number of categories for the target variable.

[0032] To address the significant differences in mineral combinations and proportions, a loss function based on mineral combination similarity measures the degree of similarity between the predicted and actual mineral combinations and optimizes the model accordingly. This improves the simulation accuracy of the model and ensures that the predicted mineral combinations are as consistent as possible with the actual mineral combinations in terms of structure and proportion.

[0033]

[0034] Where N represents the total number of rock samples, and the Similarity function is used to calculate the similarity of mineral composition proportions. For the true first The proportion of each mineral component For the predicted first The proportion of each mineral component, since the mineral component proportion is a continuous value vector, can be calculated using the mean squared error (MSELoss) as the loss function. The calculation of mineral proportion similarity is as follows:

[0035] This similarity calculation maximizes the similarity between the predicted and actual mineral assemblages by calculating the squared difference between the predicted and actual proportions.

[0036] Preferably, the graph attention convolutional layer performs the following operations in each layer: performs a linear transformation on the current node features, calculates the similarity score between the current node features and the features of each neighboring node, normalizes it to attention weights through the Softmax function, and updates the current node features by weighted summation of the neighboring node features, thereby achieving multi-round information aggregation.

[0037] Based on the same inventive concept, such as Figure 6 As shown, this application also provides a high-precision three-dimensional reconstruction device for rock pore structure based on graph autoencoder and graph attention network, including: processor 402 and storage medium 403; the processor loads and executes instructions and data in the storage medium to implement the steps of any of the above methods.

[0038] In one embodiment, the technical solution for achieving the above objective can specifically be as follows: (Refer to...) Figure 1 This embodiment provides a method and device for high-precision three-dimensional reconstruction of rock pore structure based on graph autoencoder and graph attention network: S1: Construct a three-dimensional geometric and topological diagram of rock pores; Furthermore, step S1 specifically includes: S11: Data Acquisition and Preprocessing: First, high-resolution Micro-CT scan sequences of the target core are acquired. The obtained two-dimensional slices undergo image preprocessing, including grayscale normalization, filtering and denoising, and image segmentation (e.g., using Otsu thresholding or the watershed algorithm), binarizing the image into porous and solid skeleton phases. Subsequently, the processed two-dimensional slices are stacked to generate a three-dimensional binary voxel matrix V∈RH×W×D, where 1 represents pores and 0 represents solids.

[0039] S12: Definition of graph structure: such as Figure 2 As shown, the three-dimensional voxel space is mapped to an undirected graph G=(V,E,X). Where: Node set V: Each pore voxel is defined as a graph node. To control the graph size and highlight the main pore channels, the binary image can be skeletonized or merged using a supervoxel clustering method, merging spatially adjacent groups of pore voxels into a single node.

[0040] Edge set E: The construction of edges reflects the topological connectivity between pores. This embodiment employs a strategy combining K-nearest neighbors and a spatial distance threshold. For each node vi, its Euclidean distance to all other nodes is calculated. If the distance between two nodes is less than a preset threshold τdist, an edge is established between them. Simultaneously, to ensure graph connectivity and capture a broader context, each node is additionally connected by its K spatial nearest neighbors. This strategy ensures that the graph reflects both the tight connectivity of local pores and the establishment of topological relationships spanning longer distances, thus better modeling the infiltration path.

[0041] Node feature matrix X: Each node vi is accompanied by a feature vector xi, which describes the geometry and properties of the pore represented by that node. For example... Figure 2 As shown, node features may include: Spatial coordinates: The normalized coordinates (x, y, z) of a node in three-dimensional space.

[0042] Mineral composition attributes: One-hot encoding of node attribute values ​​in profile condition data.

[0043] S2: Graph autoencoder captures the context and heterogeneous connectivity of complex porous structures; S21: The goal of the encoder is to compress high-dimensional node features and complex graph structure information into a low-dimensional, semantically rich latent space representation. This embodiment abandons the fixed, degree-based weight allocation method in traditional GCNs and adopts a Graph Attention Network (GAT).

[0044] S22: As Figure 3As shown, the encoder consists of L stacked graph attention convolutional layers (GATConv). The operation of each layer can be broken down in detail as follows: For the l-th layer, the input is the set of node features from the previous layer. The output is a new set of node features. For node i and its neighboring nodes (including nodes) The calculation process for the attention coefficient of a self-loop (i.e., a self-loop) is as follows: Linear transformation and attention mechanism: First, through a shared weight matrix A linear transformation is performed on the features of each node. Then, an attention mechanism is used. To calculate node pairs Unnormalized attention scores between :

[0045] in, This represents a vector concatenation operation, where a(l)∈R2Fl+1 is the learnable attention vector of this layer, and LeakyReLU is the activation function.

[0046] Attention weight normalization: To make attention coefficients easier to compare across different nodes, a softmax function is used to normalize the node weights. The attention scores of all neighbors are normalized to obtain the final attention weights. :

[0047] Feature weighted aggregation: Normalized attention weights are used as coefficients to perform a weighted sum of the transformed features of neighboring nodes, followed by a non-linear activation function to obtain the node... In the New features of layer +1:

[0048] To stabilize the learning process and improve expressive ability, this embodiment employs a multi-head attention mechanism at each layer. That is, the above attention operation is executed independently K times, and the resulting K feature vectors are concatenated:

[0049] After multiple layers of GATConv propagation and aggregation, the final representation z of each node is... i Each node incorporates differentiated weighted information from its multi-hop neighbors, enabling a fine-grained characterization of the local role and global context of nodes within complex porous networks. The final features of all nodes constitute the latent representation Z of the graph.

[0050] S3: Graph data augmentation strategies apply random perturbations to nodes and features; S31: In practical applications, CT scan data often contains noise, and the training samples may be limited. To enhance the model's generalization ability and robustness to incomplete / noisy data, this embodiment introduces a graph data augmentation strategy during the model training phase. This strategy mainly adds small amounts of Gaussian noise to the spatial coordinate features of nodes. This increases the diversity of the input data, makes the model less sensitive to small geometric changes, and improves the stability of the reconstruction results.

[0051] S4: The coded features are inverted through the decoder to realize the three-dimensional reconstruction of the rock pore structure.

[0052] S41: As shown in Figure 3, Dataset 1 is a 400×400×400 resolution binary discrete lithofacies model, used to evaluate the applicability of the method in basic 3D digital rock reconstruction and when including a small amount of profile data. Dataset 2 is a 1200×1200×1200 resolution multivariate discrete lithofacies model, used to test the simulation capability of the method under complex conditions such as increased attribute categories and significant differences in mineral composition ratios. For each set of training images, the experiment sets a fixed sliding window of 64×64×64 (unit: pixel), and segments the original image along the horizontal and vertical directions with a step size of 1 pixel to construct a standardized dataset. The size of each segmented 3D lithofacies sub-model is 64×64×64.

[0053] S42: During training, a dedicated loss function adapted to the differences in mineral proportions is used. This loss function includes a mean squared error term and a structural consistency term. The two types of losses are balanced through weighting coefficients to simultaneously ensure the accuracy of attribute proportions and structural connectivity. In deep learning architectures, the cross-entropy loss function is commonly used for discrete classification tasks. It measures the difference between the probability distribution of a target variable across multiple discrete categories and its true value. The calculation method is as follows:

[0054] in, It is the one-hot encoding of the actual label. It is the class probability predicted by the model. This is the total number of samples. K is the total number of categories for the target variable.

[0055] However, this function alone cannot optimize the matching degree of real physical constraints, including the composition ratio of minerals and their spatial distribution. To simultaneously optimize the classification task and mineral combination similarity, the two losses are combined with a weighted average:

[0056] in, It is cross-entropy loss. It is a loss based on the similarity of mineral assemblages. and It is a weighting coefficient that controls the contribution of both factors to the total loss.

[0057] To address the significant differences in mineral combinations and proportions, a loss function based on mineral combination similarity measures the degree of similarity between the predicted and actual mineral combinations and optimizes the model accordingly. This improves the simulation accuracy of the model and ensures that the predicted mineral combinations are as consistent as possible with the actual mineral combinations in terms of structure and proportion.

[0058]

[0059] Where N represents the total number of rock samples, and the Similarity function is used to calculate the similarity of mineral composition proportions. For the true first The proportion of each mineral component For the predicted first The proportion of each mineral component, since the mineral component proportion is a continuous value vector, can be calculated using the mean squared error (MSELoss) as the loss function. The calculation of mineral proportion similarity is as follows:

[0060] This similarity calculation maximizes the similarity between the predicted and actual mineral assemblages by calculating the squared difference between the predicted and actual proportions.

[0061] S43: Figure 4 The paper presents four reference models and their corresponding reconstruction results, as well as four conditional profile models and their corresponding reconstruction results. The comparison clearly shows that the reconstructed models can reproduce the attribute characteristics of the reference models relatively completely, thus strongly verifying the applicability and effectiveness of this method in this reconstruction task. To further evaluate the pattern matching degree between the reconstruction results and the reference models, and to achieve a quantitative assessment of the model reconstruction quality, this experiment employs various statistical analysis methods. Connectivity analysis is a method for quantitatively analyzing the spatial connectivity between different units in a geological body or spatial unit. The variogram characterizes the covariance structure or spatial correlation of a spatial variable at different locations. Connectivity curves for 20 reference models and their reconstruction results are plotted respectively. Figure 4 (a), (b), (c)) and the variogram curve ( Figure 4 In the middle (d), the variogram curve of the reconstructed model (blue) is highly consistent with that of the reference model (red) within the main range of variation, indicating that the method can reproduce the spatial variation characteristics of the reference model well, while also maintaining good connectivity characteristics.

[0062] S44: Figure 5 The paper showcases four reference models and the overall reconstruction results, along with the profile condition data and their corresponding reconstructed profiles. By comparing the visual features of the reconstructed results with those of the reference models, it can be observed that the spatial distribution of various attributes on the reconstructed profiles and the condition profiles is basically consistent. The connectivity features of the lithofacies structure were successfully learned and reconstructed by the proposed method, demonstrating that this method can effectively reconstruct complex three-dimensional geological structural features. The connectivity function curves (a, b, c) and variogram curve (d) of the reconstruction results are presented. The connectivity function curves show that the connectivity trends of the reconstructed results in the X and Y directions are basically consistent with the reference models, indicating that the proposed method has a good effect on reconstructing the spatial connectivity of geological models. The connectivity function curves illustrate the effectiveness and reliability of the proposed method in the reconstruction of three-dimensional geological models. The variogram curves show that the curve trends are relatively consistent, the differences between the reconstructed model and the reference model are relatively small, and the spatial variability of the reconstructed results is close to the variability of the reference model, indicating that the geological structure reconstructed by this method has a high degree of consistency with the reference geological features. The variogram curve proves the high accuracy of the proposed method in the geological model reconstruction process, enabling it to generate reconstruction results that match the actual geological conditions.

[0063] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A rock pore three-dimensional high-precision reconstruction method based on a graph attention autoencoder, characterized in that, Includes the following steps: (1) Construct a three-dimensional geometric and topological relationship diagram of rock pores; (2) Apply random perturbations to the nodes and features in the constructed graph to achieve graph data augmentation and obtain the augmented graph; (3) The enhanced graph is encoded by an encoder composed of multiple graph attention convolutional layers. The weights of different neighboring nodes are dynamically assigned and the features of neighboring nodes are fused through a self-attention mechanism to obtain a latent representation. The latent representation is then inverted using a decoder to obtain the predicted three-dimensional pore structure. (4) The encoder and decoder are optimized by using a special loss function that adapts to the differences in mineral proportions. The special loss function is a weighted combination of cross-entropy loss and loss based on mineral combination similarity.

2. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 1, characterized in that, In step (1), a graph structure is created based on the voxelized 3D rock data. The nodes of the graph represent each pixel in the 3D voxel grid, and the edges connect adjacent pixels. The attribute value of each pixel is associated with the spatial coordinates of the pixel. Each pixel node is represented by the one-hot encoding of the node attribute value of the spatial coordinates and the profile condition data as the node feature. The attribute value is used as the target variable. The adjacency matrix is ​​constructed based on the spatially adjacent pixels.

3. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 2, characterized in that, The edges are constructed as follows: for each node, the Euclidean distance between the node and all other nodes is calculated. If the distance between two nodes is less than a preset threshold, an edge is established between the two nodes. At the same time, each node is additionally connected by its K spatial nearest neighbors.

4. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 1, characterized in that, In step (2), the random perturbation includes adding Gaussian noise to the spatial coordinate features of the nodes.

5. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 1, characterized in that, In step (3), the encoder consists of multiple graph attention convolutional layers. Each layer performs weighted fusion of the features of neighboring nodes through a self-attention mechanism. The graph attention network dynamically assigns differentiated weights to different neighboring nodes through a self-attention mechanism to achieve refined aggregation of node features. Each layer fuses neighborhood information according to the attention weights between nodes, thereby incorporating the graph structure context into the node representation.

6. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 5, characterized in that, The graph attention convolutional layer performs the following operations in each layer: Perform a linear transformation on the features of the current node and calculate the unnormalized attention score. , ,in This represents a vector concatenation operation. ∈R2Fl+1 is the first l The layer provides a learnable attention vector, with LeakyReLU as the activation function. For a shared weight matrix, and They are the first l Layer nodes i and nodes j Features; pass The Softmax function normalizes unnormalized attention scores into attention weights. , ,in, It is a node i The neighboring nodes; Using normalized attention weights as coefficients, the transformed features of neighboring nodes are weighted and summed, then passed through a nonlinear activation function to obtain the node... i In the New features of layer +1 ; A multi-head attention mechanism is used to independently execute the above attention operation multiple times and concatenate the resulting K feature vectors: .

7. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 1, characterized in that, In step (4), the total loss of the specific loss function is: in, It is cross-entropy loss; It is a loss based on the similarity of mineral assemblages; and It is a weighting coefficient that controls the contribution of both factors to the total loss.

8. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 7, characterized in that, The cross-entropy loss is calculated as follows: in, It is the one-hot encoding of the actual tag; It is the class probability predicted by the model; K is the total number of samples; K is the total number of categories of the target variable.

9. The method for high-precision three-dimensional reconstruction of rock pores based on graph attention autoencoder as described in claim 7, characterized in that, The method for calculating the loss based on mineral assemblage similarity is as follows: Where N represents the total number of rock samples; the Similarity function is used to calculate the similarity of mineral composition proportions; For the true first i The proportion of each mineral component; For the predicted first i The proportion of each mineral component; The similarity of the mineral composition proportions is calculated as follows: 。 10. A high-precision three-dimensional reconstruction device for rock pores based on a graph attention autoencoder, characterized in that, It includes a processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the steps of the method according to any one of claims 1 to 9.

Images