A three-dimensional geological modeling method based on a DGCR-GNN architecture

Abstract

The application provides a three-dimensional geological modeling method based on a DGCR-GNN architecture, and relates to the field of three-dimensional geological modeling. The method comprises the following steps: constructing a known semantic feature point connected graph and a full node connected graph through geological sampling data and a three-dimensional regular grid of a target research area; based on a single-layer graph attention mechanism, adaptively weighting and transferring node features in the known semantic feature point connected graph to generate high-dimensional geological prior features; injecting the high-dimensional geological prior features into the full node connected graph, and performing deep topological feature extraction through a plurality of cascaded message passing residual convolution layers; for unmarked encrypted grid nodes in the full node connected graph, adopting a semi-supervised classification strategy to output geological category probability distribution of each node, and completing reconstruction of a three-dimensional geological model. The application realizes high-precision reconstruction from sparse geological sampling data to dense three-dimensional geological bodies, and improves modeling precision and automation level.

Description

Technical Field

[0001] This application relates to the field of three-dimensional geological modeling, and in particular to a three-dimensional geological modeling method based on the DGCR-GNN architecture. Background Technology

[0002] 3D geological modeling is mainly divided into explicit and implicit modeling. Traditional explicit modeling relies on manual interaction and expert knowledge, which is cumbersome, subjective, and inefficient, making it difficult to adapt to the automated modeling needs of complex geological scenes. Compared with explicit modeling, implicit modeling has advantages such as reproducibility, automation, flexibility, easy updating, less human bias, and the ability to integrate multi-source information. In recent years, machine learning and deep learning have driven the automation and intelligent transformation of 3D geological modeling. As a cutting-edge branch for processing non-Euclidean structural data, GNN has broken through the limitations of traditional meshing methods. Through node message passing mechanisms, it can accurately depict the complex spatial relationships of boreholes, rock strata interfaces, and faults.

[0003] Graph Neural Networks (GNNs) are deep learning models for graph-structured data. They enable feature interaction between nodes and edges through information diffusion and relaxation mechanisms, and rely on random walks for parameter learning, making them a key technology for processing topologically structured data. Kipf and Welling proposed a semi-supervised classification method based on graph convolutional networks, which can effectively encode graph structure and node features. GNN convolutions are divided into spectral and spatial types: the spectral method requires eigenvalue decomposition of the graph Laplacian operator, while the spatial method is based on the local neighborhood of nodes and uses a message-passing mechanism to aggregate features, making it more intuitive and easier to implement. Experiments have shown that it performs better in 3D modeling. Hillier et al. applied GNNs to 3D geological modeling, constructing an expression framework and integrating interpolation constraints, which can simultaneously handle continuous and discrete geological attributes, providing an important reference for subsequent research.

[0004] However, existing GNN-based 3D geological modeling still has significant shortcomings: First, large-scale three-dimensional geological discrete data requires multi-layer network training. Too many layers can easily lead to homogenization and oversmoothing of node features, reducing model accuracy. Secondly, real geological discrete data are unevenly distributed, and the spatial characteristics of different regions are significantly different. Conventional GNNs only aggregate the features of adjacent nodes and cannot take into account the overall features of the geological body. Finally, the model failed to achieve both sample distribution difference correction and deep network topology feature co-optimization. Complex construction regions are prone to fitting distortion and topological disorder, affecting model accuracy and practical application value.

[0005] A three-stage end-to-end modeling paradigm of feature encoding, message passing, and feature decoding is adopted. Through the construction of a heterogeneous computing space coupled with dual graphs, attention-weighted feature passing, multi-level message passing residual network, and encrypted point semi-supervised classification, high-precision reconstruction from sparse geological sampling data to dense three-dimensional geological bodies is achieved. Summary of the Invention

[0006] The purpose of this invention is to address the technical challenges of large spatial feature differences among nodes with the same semantic label, feature oversmoothing, and gradient vanishing caused by multi-layer network iteration during the transformation of sparse geological sampling data into dense 3D geological bodies, and to provide a 3D geological modeling method based on the DGCR-GNN architecture.

[0007] The above-mentioned objective of this application is achieved through the following technical solution: Step S1: Construct a connected graph of known semantic feature points and a fully connected graph using geological sampling data and a three-dimensional regular grid of the target study area; Step S2: Based on the single-layer graph attention mechanism, the node features of the known semantic feature points in the connected graph are adaptively weighted and passed to generate high-dimensional geological prior features; Step S3: Inject high-dimensional geological prior features into the fully connected graph and extract deep topological features through multiple cascaded message-passing residual convolutional layers; Step S4: For the unlabeled encrypted grid nodes in the fully connected graph, a semi-supervised classification strategy is adopted. Based on the node features extracted from the deep topological features, the probability distribution of the geological category of each node is output to complete the reconstruction of the three-dimensional geological model.

[0008] Optionally, step S1 includes: The known semantic feature point connected graph is based on the discretized nodes of geological sampling data, and the edge relationships between nodes are constructed through K-nearest neighbor or Delaunay triangulation algorithms; The fully connected graph is constructed based on a three-dimensional regular grid, and its edge relationships represent the spatial adjacency relationships between grid nodes. It is known that a connected graph with known semantic feature points and a fully connected graph share a node feature space containing spatial coordinates and semantic encoding; Geological sampling data includes: borehole resampling data or discretized geological profile sampling data; The edge relationships of the fully connected node graph specifically adopt a six-adjacency relationship, that is, each internal node is connected to its six adjacent nodes in the three-dimensional orthogonal grid direction.

[0009] Optionally, step S2 includes: After linearly upscaling the known semantic feature point connected graph to a high-dimensional latent space through the input layer, it is then input into a single-layer graph attention layer constructed based on the known semantic feature point connected graph. The single-layer graph attention layer adopts a single-head attention strategy, which calculates the attention coefficients between nodes through learnable parameters, assigns adaptive weights to different neighbor nodes, and dynamically expands the feature reception range of the central node.

[0010] Optionally, step S3 includes: Each message-passing residual convolutional layer sequentially performs neighborhood message aggregation, residual fusion of current node features and aggregated features, and linear transformation of feature dimensions; The neighborhood message aggregation adopts the mean aggregation strategy to gather feature information within the neighborhood of a node. The residual fusion adds the aggregated neighborhood features to the current features of the node element by element through skip connections; The linear transformation of the feature dimension maps the fused features to the target dimension through a learnable linear layer.

[0011] Optionally, step S4 includes: The semi-supervised classification strategy includes: High-dimensional geological prior features are used as monitoring signals, and label information is diffused to neighboring encrypted nodes through a multi-level message passing process. During the diffusion process, the spatial adjacency relationship of grid nodes is used to constrain the continuous transmission of geological features, and the pre-set edge connections in fault areas are disconnected or their weights are reduced to zero.

[0012] Optionally, step S4 may also include: The probability distribution of geological categories of each node is normalized using the Softmax function to obtain the probability value of each encrypted grid point belonging to different lithologies or geological categories, which is used for the visualization of three-dimensional geological bodies.

[0013] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a three-dimensional geological modeling method based on the DGCR-GNN architecture.

[0014] A computer-readable storage medium storing instructions that, when executed, perform a three-dimensional geological modeling method based on a DGCR-GNN architecture.

[0015] The beneficial effects of the technical solution provided in this application are: 1. A dual-graph coupled heterogeneous computing space design is proposed to solve the problem of feature transfer of sparse geological data: For the first time, a heterogeneous computing space is constructed that coordinates a connected graph of known semantic feature points and a connected graph of all nodes. Through differentiated topology and shared feature space, the effective association between sparse geological sampling data and dense grid nodes is realized, which solves the model learning difficulty caused by large differences in spatial features of nodes with the same semantic label, and provides a new data organization form for high-precision feature transfer.

[0016] 2. Design a single-layer attention-weighted feature transfer mechanism to balance feature fidelity and parameter efficiency: Unlike the traditional multi-layer GAT stacking strategy, a single-layer graph attention layer is designed in advance. This not only utilizes the adaptive weight advantage of the attention mechanism to dynamically capture the spatial correlation of geological bodies and avoid the blurring of geological boundaries, but also controls the amount of parameters and prevents feature homogenization through the single-layer structure. At the same time, it accelerates the label propagation efficiency and alleviates the label scarcity bottleneck in semi-supervised learning.

[0017] 3. A multi-level message-passing residual convolution mechanism is proposed to alleviate gradient vanishing and feature oversmoothing: The innovative design of a three-stage residual convolution layer consisting of message aggregation, residual fusion, and linear transformation preserves high-frequency information of the original geological features through skip residual connections. Combined with multi-level stacking, multi-scale geological information aggregation is achieved, which effectively alleviates the gradient vanishing and feature oversmoothing problems in deep network training and ensures that detailed information such as thin geological bodies and complex structures is not lost.

[0018] 4. Integrating geological constraints and semi-supervised classification to achieve high-precision geological body reconstruction: Geological spatial physical constraints and topological relationships are integrated into the semi-supervised classification process. The continuity of geological features is ensured through grid adjacency relationships, and feature aliasing is avoided by adjusting the edge relationships at faults. This achieves end-to-end accurate reconstruction from sparse sampling points to dense 3D geological bodies, taking into account both model accuracy and geological realism. Attached Figure Description

[0019] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a diagram of the DGCR-GNN network architecture in the embodiments of this application; Figure 2 This is a flowchart illustrating the weighted geological feature transfer process that incorporates an attention mechanism in the embodiments of this application. Figure 3 This is a diagram of the multi-level message passing residual network architecture in an embodiment of this application; Figure 4 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0020] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0021] The embodiments of this application provide a three-dimensional geological modeling method based on the DGCR-GNN architecture.

[0022] Please refer to Figure 1 , Figure 1 This is a diagram of the DGCR-GNN network architecture of a three-dimensional geological modeling method based on the DGCR-GNN architecture in an embodiment of this application, including: Step S1: Construct a connected graph of known semantic feature points and a fully connected graph using geological sampling data and a three-dimensional regular grid of the target study area; Step S2: Based on the single-layer graph attention mechanism, the node features of the known semantic feature points in the connected graph are adaptively weighted and passed to generate high-dimensional geological prior features; Step S3: Inject high-dimensional geological prior features into the fully connected graph and extract deep topological features through multiple cascaded message-passing residual convolutional layers; Step S4: For the unlabeled encrypted grid nodes in the fully connected graph, a semi-supervised classification strategy is adopted. Based on the node features extracted from the deep topological features, the probability distribution of the geological category of each node is output to complete the reconstruction of the three-dimensional geological model.

[0023] As one embodiment, this invention effectively solves the core technical challenges in the transformation of sparse geological sampling data into dense 3D geological bodies through the collaborative design of the DGCR-GNN architecture. It demonstrates excellent performance in comparison and ablation experiments on actual datasets: it significantly alleviates the problems of feature oversmoothing and gradient vanishing caused by multi-layer networks, and improves the discriminative power of node features; it effectively addresses the problems of large differences in spatial features of nodes with the same semantic label and uneven distribution of geological data, accurately capturing local details and global correlations of geological bodies; it avoids fitting distortion and topological disorder in complex structural regions, and greatly improves the classification accuracy, geological body continuity, and boundary sharpness of 3D geological modeling.

[0024] Meanwhile, the single-layer attention design reduces the number of model parameters and improves training efficiency and generalization ability; the integration of semi-supervised classification process and geological constraints enables the model to achieve reliable modeling even in scenarios with scarce labels, without the need for a large amount of manual annotation; the end-to-end modeling paradigm eliminates tedious manual post-processing steps, improves the automation and intelligence level of 3D geological modeling, and provides reliable technical support for accurate cognition of underground geological bodies.

[0025] This application employs the aforementioned technical solution, utilizing a three-stage end-to-end modeling paradigm of feature encoding, message passing, and feature decoding. Through dual-graph coupled heterogeneous computational space construction, attention-weighted feature passing, multi-level message-passing residual networks, and encrypted point semi-supervised classification, it achieves high-precision reconstruction from sparse geological sampling data to dense 3D geological bodies. The technical solution of this application can be applied to regional geological surveys, oil and gas exploration, solid mineral exploration, geothermal resource exploration, urban geological surveys, digital mine construction, engineering geological exploration, and water resource assessment. It can directly construct high-precision, topologically sound, and accuracy-compliant 3D geological models based on relatively sparse geological borehole data.

[0026] Step S1 includes: The known semantic feature point connected graph is based on the discretized nodes of geological sampling data, and the edge relationships between nodes are constructed through K-nearest neighbor or Delaunay triangulation algorithms; The fully connected graph is constructed based on a three-dimensional regular grid, and its edge relationships represent the spatial adjacency relationships between grid nodes. It is known that a connected graph with known semantic feature points and a fully connected graph share a node feature space containing spatial coordinates and semantic encoding; Geological sampling data includes: borehole resampling data or discretized geological profile sampling data; The edge relationships of the fully connected node graph specifically adopt a six-adjacency relationship, that is, each internal node is connected to its six adjacent nodes in the three-dimensional orthogonal grid direction.

[0027] As one implementation, this module establishes the main framework and data flow of the DGCR-GNN network, constructing two heterogeneous computational spaces: a known semantic feature point connected graph and a full node connected graph. The known semantic feature point connected graph uses gridded data from borehole resampling and profile discretization as nodes, constructing edge relationships through K-nearest neighbors or Delaunay triangulation to characterize the overall topological adjacency relationships of the geological sample data. The full node graph is constructed based on a three-dimensional regular grid, with nodes covering the entire study area, and edge relationships representing the spatial six-adjacency relationships of grid points. Both graphs share the same node feature space (integrating spatial coordinates and semantic encoding), but have different graph topologies and sampling densities, laying the foundation for the transfer of sparse features to dense spaces.

[0028] Step S2 includes: After linearly upscaling the known semantic feature point connected graph to a high-dimensional latent space through the input layer, it is then input into a single-layer graph attention layer constructed based on the known semantic feature point connected graph. The single-layer graph attention layer adopts a single-head attention strategy, which calculates the attention coefficients between nodes through learnable parameters, assigns adaptive weights to different neighbor nodes, and dynamically expands the feature reception range of the central node.

[0029] As one implementation, a cross-graph attention mechanism is introduced to achieve weighted transfer of geological body features from the sparse sampling point space to the dense prediction space. The input structured geological data is linearly up-dimensional after being processed by the input layer, and then enters a single-layer graph attention layer based on a connected graph of known semantic feature points. A single-head attention strategy is adopted, and adaptive weight allocation is achieved through learnable parameters and attention coefficients. This dynamically expands the node receptive field, allowing the model to prioritize nodes with strong geological relevance, avoiding the blurring of geological boundaries caused by fixed weight aggregation. At the same time, it reduces the number of model parameters, preserves the original feature discriminativeness, and provides high-quality prior features for subsequent dense prediction.

[0030] Step S3 includes: Each message-passing residual convolutional layer sequentially performs neighborhood message aggregation, residual fusion of current node features and aggregated features, and linear transformation of feature dimensions; The neighborhood message aggregation adopts the mean aggregation strategy to gather feature information within the neighborhood of a node. The residual fusion adds the aggregated neighborhood features to the current features of the node element by element through skip connections; The linear transformation of the feature dimension maps the fused features to the target dimension through a learnable linear layer.

[0031] As one embodiment, a deep topology is constructed by cascading multiple message-passing residual convolutional layers. Each convolutional layer sequentially executes a three-stage process: message aggregation, residual fusion, and linear transformation. The message aggregation stage employs a mean aggregation strategy to gather information from the six neighborhoods of a node, ensuring a smooth transition in the geological feature space. The residual fusion stage introduces skip residual connections, adding the aggregated neighborhood features element-wise to the current node's features, mitigating gradient vanishing and feature homogenization problems. The linear transformation stage uses learnable linear layers to calibrate feature dimensions and perform semantic mapping, gradually extracting hierarchical geological representations to ensure that the deep network retains high-frequency information of the original geological attributes.

[0032] Step S4 includes: The semi-supervised classification strategy includes: High-dimensional geological prior features are used as monitoring signals, and label information is diffused to neighboring encrypted nodes through a multi-level message passing process. During the diffusion process, the spatial adjacency relationship of grid nodes is used to constrain the continuous transmission of geological features, and the pre-set edge connections in fault areas are disconnected or their weights are reduced to zero.

[0033] As one example, a semi-supervised classification strategy is adopted to achieve accurate geological semantic annotation for a large number of unlabeled densified grid points in a fully connected graph. First, initialization is completed through label propagation. Preprocessed features of known semantic sampling points are injected into the fully connected graph as supervisory signals, and label information is disseminated to neighboring densified points through multi-level message passing. The six-adjacency relationship of grid nodes ensures continuous transmission of geological features, and edge relationships at fault locations are adjusted to avoid feature aliasing. Finally, Softmax normalization is used to output the probability distribution of lithology categories for each node, achieving reliable classification from sparse sampling points to dense densified points, providing data support for the detailed visualization of 3D geological bodies.

[0034] Step S4 also includes: The probability distribution of geological categories of each node is normalized using the Softmax function to obtain the probability value of each encrypted grid point belonging to different lithologies or geological categories, which is used for the visualization of three-dimensional geological bodies.

[0035] In one embodiment, Figure 2 This is a flowchart illustrating the weighted geological feature transfer process that incorporates an attention mechanism in the embodiments of this application. Figure 3 This is a diagram of the multi-level message passing residual network architecture in an embodiment of this application.

[0036] In another embodiment, comparative and ablation experiments were conducted based on actual datasets to quantitatively verify the modeling performance of the DGCR-GNN architecture from three core dimensions: classification accuracy, geological body continuity, and boundary sharpness. This ensures that the method can effectively address the shortcomings of existing technologies and meet the needs of actual 3D geological modeling.

[0037] In one exemplary embodiment, a specific implementation method for reconstructing dense 3D geological bodies from sparse geological sampling data is provided based on a dual-graph coupled residual graph neural network (DGCR-GNN) architecture. The following describes the data processing flow step-by-step according to the inference stage: (1) Input preprocessing We acquired actual geological sampling borehole data from a mining area and resampled the boreholes. All data were standardized and a unified coordinate system (Gaussian projection coordinate system) was adopted to ensure consistency in lithological classification and geological boundaries, meeting the standardization requirements for 3D geological modeling data. Spatial coordinates (x, y, z) and lithological semantic labels of the boreholes were extracted to construct structured geological data, and Z-score normalization was used to standardize the spatial coordinates. The standardized structured data was divided into training and validation sets for model training and performance verification.

[0038] (2) Construction of a dual-graph coupled heterogeneous computation space Two heterogeneous computational spaces, a known semantic feature point connected graph and a full node connected graph, are constructed to establish the main framework and data flow of the DGCR-GNN network. The known semantic feature point connected graph uses gridded data from borehole resampling as nodes, and edge relationships are constructed using the Delaunay triangulation algorithm to characterize the overall topological adjacency relationships of the geological sample data. The full node graph is constructed based on a three-dimensional regular grid, covering the entire mining area, and edge relationships represent the spatial six-adjacency relationships of grid points. The two graphs share the same node feature space, but their topological structures and sampling densities differ significantly, laying the foundation for the transfer of sparse features to dense spaces.

[0039] (3) Attention-weighted transfer of geological features The preprocessed structured geological data is input into the input layer, and the initial feature dimensions are mapped to a 128-dimensional high-dimensional latent space through linear dimensionality upscaling. The upscaled features are then input into a single-layer graph attention layer based on a connected graph of known semantic feature points. A single-head attention strategy is adopted, and the attention coefficients are calculated through a single-layer feedforward network. The hidden layer uses the Leaky ReLU activation function. Through adaptive weight allocation, the node receptive field is dynamically expanded, allowing the model to prioritize nodes with strong geological relevance, avoiding the blurring of geological boundaries caused by fixed weight aggregation. At the same time, the number of model parameters is reduced, the original feature discriminability is preserved, and the 128-dimensional high-dimensional features of the sampling points are output, providing high-quality prior features for subsequent full-node dense prediction.

[0040] (4) Construction of multi-level message passing residual network A deep topology is constructed using four cascaded message-passing residual convolutional layers. Each convolutional layer sequentially executes a three-stage process: message aggregation, residual fusion, and linear transformation. The message aggregation stage employs a mean aggregation strategy to gather information from the six neighbors of a node, ensuring a smooth spatial transition of geological features and adapting to sedimentary stratigraphic regions with gradual lithological changes. The residual fusion stage introduces skip residual connections, adding the aggregated neighborhood features element-wise to the current node's features, maintaining the feature dimension and mitigating gradient vanishing and feature homogenization issues. The linear transformation stage uses learnable linear layers to perform dimensional calibration and semantic mapping on the fused features, gradually extracting hierarchical geological representations from local lithology to regional tectonics. The entire multi-level message-passing process uses the Leaky ReLU nonlinear activation function to ensure that the deep network retains high-frequency information of the original geological attributes, avoiding the loss of thin-layer stratigraphic information.

[0041] (5) Encryption point semi-supervised classification process To address the large number of unlabeled densified grid points in the fully connected graph, a semi-supervised classification strategy is employed to achieve accurate geological semantic annotation. Initialization is achieved through label propagation. Known semantic sampling point features, preprocessed by a single-layer graph attention layer, are injected into the fully connected graph as supervisory signals. These nodes serve as core "feature sources," and label information is propagated to neighboring densified points via a 4-level residual message passing mechanism. The six-adjacency relationship of grid nodes forces geological features to propagate along a spatially continuous direction. Cross-fault connections are broken at nodes on both sides of a fault (weights are set to 0) to prevent feature aliasing caused by local mean aggregation of the residual network at the fault, thus preserving the geological discontinuity of lithology on both sides of the fault. After 4-level residual block iterations, each densified grid point completes multi-scale contextual information aggregation. The final layer outputs the probability distribution of each lithology category obtained through Softmax normalization, achieving reliable semi-supervised classification from sparse sampling points to dense densified points, establishing a high-quality data foundation for the detailed visualization of 3D geological bodies.

[0042] (6) Training strategies The model is trained end-to-end using the PyTorch framework and the PyTorch Geometric library, with a maximum of 5000 training epochs. To ensure sufficient model convergence while avoiding redundant training, early stopping is triggered when the validation set accuracy does not increase after 120 consecutive training epochs. The learning rate parameter is set to 0.0001, the weight decay is set to 0.0001, the optimizer is set to Adam, the activation function is Leaky ReLU, the cross-entropy loss is obtained by the criterion function, and the hidden layer dimension is set to 128.

[0043] This application also discloses an electronic device. (See reference...) Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0044] The communication bus 502 is used to enable communication between these components.

[0045] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0046] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0047] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the aforementioned three-dimensional geological modeling method based on the DGCR-GNN architecture.

[0048] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0049] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A three-dimensional geological modeling method based on the DGCR-GNN architecture, characterized in that, The method includes the following steps: Step S1: Construct a connected graph of known semantic feature points and a fully connected graph using geological sampling data and a three-dimensional regular grid of the target study area; Step S2: Based on the single-layer graph attention mechanism, the node features of the known semantic feature points in the connected graph are adaptively weighted and passed to generate high-dimensional geological prior features; Step S3: Inject high-dimensional geological prior features into the fully connected graph and extract deep topological features through multiple cascaded message-passing residual convolutional layers; Step S4: For the unlabeled encrypted grid nodes in the fully connected graph, a semi-supervised classification strategy is adopted. Based on the node features extracted from the deep topological features, the probability distribution of the geological category of each node is output to complete the reconstruction of the three-dimensional geological model.

2. The three-dimensional geological modeling method based on the DGCR-GNN architecture as described in claim 1, characterized in that, Step S1 includes: The known semantic feature point connected graph is based on the discretized nodes of geological sampling data, and the edge relationships between nodes are constructed through K-nearest neighbor or Delaunay triangulation algorithms; The fully connected graph is constructed based on a three-dimensional regular grid, and its edge relationships represent the spatial adjacency relationships between grid nodes. It is known that a connected graph with known semantic feature points and a fully connected graph share a node feature space containing spatial coordinates and semantic encoding; Geological sampling data includes: borehole resampling data or discretized geological profile sampling data; The edge relationships of the fully connected node graph specifically adopt a six-adjacency relationship, that is, each internal node is connected to its six adjacent nodes in the three-dimensional orthogonal grid direction.

3. The three-dimensional geological modeling method based on the DGCR-GNN architecture as described in claim 1, characterized in that, Step S2 includes: After linearly upscaling the known semantic feature point connected graph to a high-dimensional latent space through the input layer, it is then input into a single-layer graph attention layer constructed based on the known semantic feature point connected graph. The single-layer graph attention layer adopts a single-head attention strategy, which calculates the attention coefficients between nodes through learnable parameters, assigns adaptive weights to different neighbor nodes, and dynamically expands the feature reception range of the central node.

4. The three-dimensional geological modeling method based on the DGCR-GNN architecture as described in claim 1, characterized in that, Step S3 includes: Each message-passing residual convolutional layer sequentially performs neighborhood message aggregation, residual fusion of current node features and aggregated features, and linear transformation of feature dimensions; The neighborhood message aggregation adopts the mean aggregation strategy to gather feature information within the neighborhood of a node. The residual fusion adds the aggregated neighborhood features to the current features of the node element by element through skip connections; The linear transformation of the feature dimension maps the fused features to the target dimension through a learnable linear layer.

5. The three-dimensional geological modeling method based on the DGCR-GNN architecture as described in claim 1, characterized in that, Step S4 includes: The semi-supervised classification strategy includes: High-dimensional geological prior features are used as monitoring signals, and label information is diffused to neighboring encrypted nodes through a multi-level message passing process. During the diffusion process, the spatial adjacency relationship of grid nodes is used to constrain the continuous transmission of geological features, and the pre-set edge connections in fault areas are disconnected or their weights are reduced to zero.

6. The three-dimensional geological modeling method based on the DGCR-GNN architecture as described in claim 1, characterized in that, Step S4 also includes: The probability distribution of geological categories of each node is normalized using the Softmax function to obtain the probability value of each encrypted grid point belonging to different lithologies or geological categories, which is used for the visualization of three-dimensional geological bodies.

7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform the three-dimensional geological modeling method based on the DGCR-GNN architecture as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the three-dimensional geological modeling method based on the DGCR-GNN architecture as described in any one of claims 1-6.

Images