Graph convolution three-dimensional geological modeling method fusing geological cause constraint

Abstract

The invention belongs to the technical field of geological modeling, and particularly discloses a geological cause constraint fused graph convolution three-dimensional geological modeling method, which comprises the following steps of: obtaining and performing spatial sampling on original geological data to obtain a geological attribute distribution data set; constructing an underground physical field model based on the geological attribute distribution data set, and performing adaptive space division on the research area by adopting an octree structure to obtain multi-scale geological units as leaf nodes; mapping each leaf node into a graph node, and taking the space coordinates, density, water content and compression modulus of the leaf nodes as node features; establishing a spatial adjacency relation between graph nodes through Delaunay triangulation, and introducing a gravity model to quantify cause influence based on volume and distance between the nodes to construct a weighted graph structure; and performing lithology classification prediction on the weighted graph structure by using a graph convolutional neural network to obtain a three-dimensional geologic model. The method can effectively improve the prediction precision of the model in a complex geological environment.

Description

Technical Field

[0001] This application belongs to the field of geological modeling technology, and more specifically, relates to a graph convolution three-dimensional geological modeling method that integrates geological genetic constraints. Background Technology

[0002] Because underground space is inherently a complex, three-dimensional, continuous system with interconnected genesis, three-dimensional geological models can visualize abstract geological structures, vividly and intuitively showcasing the spatial distribution, geometric forms, and interrelationships of geological units such as strata, faults, and rock masses. This is of great significance for a deeper understanding of regional geological evolution, guiding mineral resource exploration, ensuring the safety of urban underground space development and utilization, and reducing the risks of engineering construction and geological disasters.

[0003] With the development of artificial intelligence technology, 3D geological modeling has gradually shifted towards data-driven implicit modeling methods. The introduction of machine learning technology enables models to learn lithological distribution patterns from limited borehole samples, achieving intelligent prediction of unknown areas. However, current mainstream machine learning-based modeling methods typically use regular voxels or discrete points as basic units, relying solely on their spatial coordinates, density, magnetic susceptibility, and other local features for training and prediction, making it difficult to fully reflect the spatial continuity and genetic correlation of geological bodies. When faced with scenarios of complex geological structures and sparse borehole distribution, such methods often lack the ability to adaptively express multi-scale structural features and cannot effectively characterize the interaction relationships between different geological units determined by differences in volume, distance, and physical properties. Furthermore, traditional neighborhood modeling mechanisms are limited by a fixed perception range, neglecting the influence between non-adjacent but genetically related geological bodies, resulting in deficiencies in model structural integrity and geological rationality.

[0004] Therefore, there is an urgent need to propose a new modeling method that can break through the limitations of regular grids and quantify the causal influence between geological units in order to improve the prediction accuracy of models in complex geological environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a graph convolution 3D geological modeling method that integrates geological genetic constraints, which can effectively improve the prediction accuracy of models in complex geological environments.

[0006] To achieve the above objectives, in a first aspect, this application provides a graph convolution 3D geological modeling method that integrates geological genetic constraints, comprising the following steps: S10, Obtain raw geological data, construct grid cells for the study area, and perform spatial sampling on the raw geological data to obtain a geological attribute distribution dataset; S20. Based on the geological attribute distribution dataset, a subsurface physical field model is constructed, and an octree structure is used to adaptively divide the study area into spatial parts to obtain multi-scale geological units as leaf nodes. S30, map each leaf node to a graph node, and combine the spatial coordinates, density, water content and compressibility modulus of the leaf node as node features; S40, establish the spatial adjacency relationship between the graph nodes through Delaunay triangulation, and introduce a gravity model to quantify the causal influence between nodes based on volume and distance, and construct a weighted graph structure; S50, using a graph convolutional neural network to perform lithology classification and prediction on the weighted graph structure, a three-dimensional geological model is obtained.

[0007] As a further preferred option, step S10 specifically includes: S11, Determine the modeling range boundary based on the study area, and establish a bounding box based on the modeling range boundary; S12, Divide the bounding box into regular meshes according to a preset precision to obtain a three-dimensional mesh simulation model; S13, Extract lithological information from lithological stratification borehole data, wherein the lithological stratification borehole data originates from the original geological data; S14 maps the lithological information in the borehole to the corresponding three-dimensional regular mesh model voxels to generate a geological attribute distribution dataset.

[0008] As a further preferred option, step S20 specifically includes: S21, Extract physical attribute data from the geological attribute distribution dataset; S22, Spatial interpolation processing is performed on the physical attribute data to generate a physical distribution field model of the study area; S23, Divide the boundaries and minimum unit size according to the unit boundary mapping relationship of the study area; S24, recursively subdivide according to the set conditions until the modeling accuracy requirements are met.

[0009] As a further preferred option, step S23 specifically includes: S231, construct a physical lithology mapping box plot of all boreholes and real features in the entire area, and obtain the division threshold between each type and physical attribute through the box plot; S232, based on the boundary values ​​of the box plot, the avoidance rules for dividing geological units are obtained.

[0010] As a further preferred option, step S30 specifically involves: S31, the divided geological unit blocks are used as leaf node units of the entire octree structure; S32, calculate the mean values ​​of spatial coordinates, density, water content and compressibility modulus of the geological units contained in each leaf node as node features.

[0011] As a further preferred option, step S40 specifically includes: S41, map the leaf nodes in the octree structure to graph nodes; S42, obtain the spatial coordinates, density, water content and compressibility modulus of the graph nodes as node features; S43, establish spatial adjacency relationships between graph nodes through Delaunay triangulation; S44 introduces a gravity model to quantify the causal influence between nodes based on volume and distance, and calculates the geological influence weight; S45 uses the geological influence weight as the edge weight to construct a weighted graph structure.

[0012] As a further preferred embodiment, in step S50, the graph convolutional neural network includes an input layer, a hidden layer, and an output layer; The input layer receives a set of graph nodes mapped from the leaf nodes of an octagonal tree. The feature vector of each node contains its physical properties such as spatial coordinates, density, water content, and compressibility modulus, forming a six-dimensional node feature vector. The hidden layer consists of two stacked graph convolutional layers. Each layer executes a message passing mechanism to aggregate the feature information of neighboring nodes and update the representation of the current node. The output layer uses the Softmax function to classify the final node features and outputs the probability distribution of the lithology of each node, thereby achieving a refined prediction of the lithology distribution in underground space.

[0013] As a further preferred embodiment, step S50 specifically includes: S51, based on the map model of the study area, the geological category of each predicted node is obtained through graph convolution operation; S52 maps the classification prediction value of each graph node to the actual geological unit category to obtain a three-dimensional geological model.

[0014] Secondly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the graph convolution three-dimensional geological modeling method with fused geological genetic constraints as described above.

[0015] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the graph convolution 3D geological modeling method incorporating geological genetic constraints as described above. The graph convolution 3D geological modeling method, storage medium, and device provided in this application, which integrates geological genesis constraints, have the following effects: (1) Breaking through the limitations of traditional regular grid modeling: By using an octree structure combined with Delaunay triangulation and gravity model to calculate edge weights, a weighted graph is constructed, making the method applicable to 3D geological space modeling under complex geological conditions that are non-uniform and discontinuous; (2) Improving modeling speed: By intelligently dividing the original geological units into multiple small blocks, the geological features within each processing unit are more consistent. This method can reduce the amount of data and the number of nodes that need to be processed, reduce the computational burden in graph convolution operations, and smaller blocks mean that the adjacency matrix and node feature matrix have lower dimensions, simplifying the calculation process; (3) Enhancing the geological and physical significance: Introducing a "gravity" model based on volume and distance to calculate the edge weights between different geological units not only enhances the geological and physical significance of the graph structure, but also helps to more accurately reflect the mutual influence relationship between geological units; (4) High-precision lithology classification prediction: Using multi-layer graph convolution operations to achieve layer-by-layer aggregation and nonlinear transformation of spatial features, lithology classification prediction is performed on multi-scale geological units generated by octree division. This method can effectively capture the spatial distribution characteristics of geological data, thereby improving the accuracy of lithological classification. Attached Figure Description

[0016] Figure 1 This is a flowchart of the graph convolution 3D geological modeling method that integrates geological genesis constraints provided in this application; Figure 2 This is a flowchart of the graph convolution 3D geological modeling method that integrates geological genetic constraints, provided in the embodiments of this application; Figure 3 This is an architecture diagram of the graph convolution 3D geological modeling system that integrates geological genesis constraints provided in this application; Figure 4 This is a structural diagram of the electronic device provided in this application. Detailed Implementation

[0017] 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.

[0018] like Figure 1 As shown, this application provides a graph convolution 3D geological modeling method that integrates geological genetic constraints, including steps S10 to S50, which are detailed below: Step S10 involves acquiring raw geological data, constructing a grid cell for the study area, and spatially sampling the raw geological data to obtain a dataset of geological attribute distributions. This step provides a structured dataset with a unified spatial benchmark for subsequent 3D geological modeling and graph neural network analysis, ensuring data quality and spatial consistency.

[0019] Specifically, step S10 can be as follows: S11, determine the modeling range boundary according to the study area, and establish a bounding box according to the modeling range boundary; S12, divide the bounding box into regular grids according to the preset accuracy to obtain a three-dimensional grid simulation model; S13, extract lithological information from the lithological stratification borehole data; S14, map the lithological information in the borehole to the corresponding three-dimensional regular grid model voxels to generate a geological attribute distribution dataset.

[0020] Step S20 involves constructing a subsurface physical field model based on a geological attribute distribution dataset, and using an octree structure to adaptively partition the study area, obtaining multi-scale geological units as leaf nodes. This step enables high-precision 3D modeling of complex geological bodies, generating a set of geological units that combines geometric multi-scale characteristics with lithological consistency, thus enhancing the model's adaptive representation capability for non-uniform geological conditions.

[0021] Specifically, step S20 can be: S21, extracting physical attribute data from the geological attribute distribution dataset; S22, performing spatial interpolation on the physical attribute data to generate a physical distribution field model of the study area; S23, dividing the boundaries and minimum unit size according to the unit boundary mapping relationship of the study area; S24, recursively subdividing according to the set conditions until the modeling accuracy requirements are met.

[0022] Specifically, step S23 may include: S231, constructing a physical lithology mapping box plot of all boreholes and real features in the entire area, and obtaining the division threshold between each type and physical attribute through the box plot; S232, obtaining the avoidance rules when dividing geological units based on the boundary values ​​of the box plot.

[0023] Step S30 involves mapping each leaf node to a graph node and combining the spatial coordinates, density, water content, and compression modulus of the leaf nodes as node features. This step transforms the spatial partitioning results into structured data suitable for graph neural network processing, providing compact and representative node feature inputs for subsequent graph convolution operations.

[0024] Specifically, step S30 can be: S31, taking the divided geological unit blocks as leaf node units of the entire octree structure; S32, calculating the average values ​​of the spatial coordinates, density, water content and compressibility modulus of the geological units contained in each leaf node as node features.

[0025] Step S40 establishes spatial adjacency relationships between graph nodes through Delaunay triangulation and introduces a gravity model to quantify the causal influence between nodes based on volume and distance, constructing a weighted graph structure. This step overcomes the limitations of traditional regular grid modeling, enhances the geophysical significance of the graph structure, and effectively reflects the mutual influence relationships between geological units.

[0026] Specifically, step S40 can be as follows: S41, mapping the leaf nodes in the octree structure to graph nodes; S42, obtaining the spatial coordinates, density, water content, and compressibility modulus of the graph nodes as node features; S43, establishing the spatial adjacency relationship between graph nodes through Delaunay triangulation; S44, introducing a gravity model to quantify the causal influence between nodes based on volume and distance, and calculating the geological influence weight; S45, using the geological influence weight as the graph edge weight to construct a weighted graph structure.

[0027] Step S50: Use graph convolutional neural network to perform lithology classification prediction on the weighted graph structure to obtain a three-dimensional geological model.

[0028] Specifically, step S50 can be: S51, obtaining the geological category of each predicted node through graph convolution operation based on the map model of the study area; S52, mapping the classification prediction value of each graph node to the real geological unit category to obtain a three-dimensional geological model.

[0029] The graph convolutional neural network comprises an input layer, hidden layers, and an output layer. The input layer receives a set of graph nodes mapped from the leaf nodes of an octagonal tree. Each node's feature vector contains its spatial coordinates (X, Y, Z), density, water content, and compressibility modulus, forming a six-dimensional node feature vector. The hidden layer consists of two stacked graph convolutional layers, each executing a message-passing mechanism to aggregate the feature information of neighboring nodes and update the representation of the current node. The output layer uses the Softmax function to classify the final node features, outputting the probability distribution of the lithology to which each node belongs, thus achieving a refined prediction of the lithology distribution in underground space. This application achieves layer-by-layer aggregation and nonlinear transformation of spatial features through multi-layer graph convolutional operations, capturing the spatial distribution characteristics of geological data, thereby completing high-precision lithology classification and prediction.

[0030] The graph convolution 3D geological modeling method with integrated geological genesis constraints provided in this application has the following effects: (1) Breaking through the limitations of traditional regular grid modeling: By using an octree structure combined with Delaunay triangulation and gravity model to calculate edge weights, a weighted graph is constructed, making the method applicable to 3D geological space modeling under complex geological conditions that are non-uniform and discontinuous; (2) Improving modeling speed: By intelligently dividing the original geological units into multiple small blocks, the geological features within each processing unit are more consistent. This method can reduce the amount of data and number of nodes that need to be processed, reduce the computational burden in graph convolution operations, and smaller blocks mean that the adjacency matrix and node feature matrix have lower dimensions, simplifying the calculation process; (3) Enhancing the geological and physical significance: Introducing a "gravity" model based on volume and distance to calculate the edge weights between different geological units not only enhances the geological and physical significance of the graph structure, but also helps to more accurately reflect the mutual influence relationship between geological units; (4) High-precision lithology classification prediction: Using multi-layer graph convolution operations to achieve layer-by-layer aggregation and nonlinear transformation of spatial features, lithology classification prediction is performed on multi-scale geological units generated by octree division. This method can effectively capture the spatial distribution characteristics of geological data, thereby improving the accuracy of lithological classification.

[0031] In one embodiment, the technical solution for achieving the above objective can be specifically as follows: Figure 2 As shown, this embodiment provides a graph convolution 3D geological modeling method that integrates geological genetic constraints, including the following steps: S1: Obtain raw geological data, construct grid cells for the study area, and spatially sample the raw geological data to obtain a dataset of geological attribute distribution. Specifically, in step S1, to support subsequent 3D geological modeling and graph neural network analysis, this embodiment first preprocesses the original geological exploration data to construct a structured dataset with a unified spatial benchmark. This process includes steps such as defining the modeling domain, spatial grid generation, borehole data parsing, and attribute mapping.

[0032] Specifically, step S1 mainly includes: S11: Determine the modeling range boundary based on the study area, and establish a bounding box based on the modeling range boundary; Spatially overlaying the borehole distribution area with the DEM coverage area in the study area, the horizontal union of the two is used to determine the horizontal boundary of the 3D model. In the vertical direction, the lower boundary of the modeling depth is taken as the elevation value extending downward from the maximum depth of all boreholes, while the upper boundary is consistent with the surface elevation described by the DEM.

[0033] Based on the modeling scope boundary, an axis-aligned cuboid spatial container, i.e., a 3D bounding box, is constructed. Its bottom and top faces correspond to the preset minimum and maximum surface elevations, respectively, while its sides are parallel to the coordinate axes and completely encompass the spatial distribution area of ​​all borehole trajectories and target geological bodies. This bounding box serves as the benchmark framework for subsequent spatial discretization processing.

[0034] S12: Divide the bounding box into regular meshes according to a preset precision to obtain a three-dimensional mesh model; Specifically, spatial resolution parameters, including horizontal and vertical resolution, are determined. Then, within the bounding box, nodes are arranged at equal intervals along the X, Y, and Z coordinate axes at a preset resolution to generate a regular three-dimensional node array. Based on this node array, an S-grid type three-dimensional mesh unit is constructed. Each unit has a unique spatial index and corresponding geometric center coordinates, forming a discretized spatial framework covering the entire study area.

[0035] S13: Extracting lithological information from lithological stratification borehole data; All lithological borehole data within the study area were acquired. Each record includes borehole coordinates, borehole trajectory, and lithological descriptions by depth. Structured analysis was performed on each stratigraphic interval to determine the top and bottom depths of each layer. Lithological labels were standardized and unified to a pre-defined classification system.

[0036] S14: Map the lithological information in the borehole to the corresponding three-dimensional regular mesh model voxels to generate a geological attribute distribution dataset.

[0037] Specifically, the data expansion from discrete observation points to a continuous spatial model involves the following mapping operations: resampling the intervals between adjacent stratigraphic boundaries in the same borehole, inserting several intermediate sampling points to ensure that the stratum has representative points across its entire thickness; assigning a corresponding lithology category label to each sampling point to form a discrete set of attributes; projecting these attributed sampling points onto a three-dimensional regular grid model, and assigning a lithology label to any grid cell if its spatial extent intersects the borehole path and contains at least one attributed sampling point; all assigned grid cells together constitute the initial geological attribute distribution dataset.

[0038] S2: Based on the borehole dataset, a subsurface physical field model is constructed, and an octree structure is used to adaptively divide the study area into spatial parts to obtain multi-scale geological units as leaf nodes. To achieve high-precision 3D modeling of complex geological bodies, a continuous 3D field model reflecting the physical properties of the underground medium is constructed based on measured borehole data. An adaptive spatial partitioning mechanism is then introduced to generate a set of geological units that combine geometric multi-scale characteristics with lithological consistency.

[0039] The specific steps in S2 mainly include: S21: Extract physical property data from physical measurement borehole data; All available physical measurement borehole data within the study area were acquired. Each borehole record includes spatial location information (x, y, z) and lithological descriptions and physical parameters at the corresponding depth. Quantitative physical property data closely related to the lithology were extracted from these data.

[0040] S22: Perform spatial interpolation on the physical attribute data to generate a physical distribution field model for the study area; To construct a continuous, smooth, and geostatistically plausible physical property distribution field in three-dimensional space, this embodiment employs the Ordinary Kriging interpolation method to extend discrete physical borehole measurement data to a voxel grid covering the entire study area, generating a high-resolution three-dimensional physical property distribution model.

[0041] Specifically, in performing Kriging interpolation: First, a regular three-dimensional voxel mesh is constructed based on the spatial extent of the study area and the preset modeling resolution. Each voxel is uniquely identified by its spatial coordinates (x, y, z), serving as the target location for interpolation. Second, all available physical borehole data within the study area are integrated. Each record contains spatial coordinates and corresponding measured values ​​of physical properties, such as density, water content, or compressibility modulus. These borehole points constitute the known sample point set for ordinary Kriging interpolation. The above interpolation process is repeated across all voxel units within the study area, ultimately generating a three-dimensional physical property distribution field model covering the entire region. Each voxel is assigned the density, water content, or compressibility modulus value obtained through ordinary Kriging interpolation, forming a structurally continuous and spatially self-consistent physical field data volume.

[0042] S23: Define the boundaries and minimum unit size based on the boundary mapping relationship of the study area units; To achieve refined modeling of geological bodies, this embodiment proposes an adaptive partitioning strategy based on statistical analysis. Its core lies in establishing a mapping relationship between physical properties and actual stratigraphic categories using borehole data, and accordingly setting threshold conditions for octree recursive subdivision, while avoiding the impact of ambiguous boundary regions on modeling accuracy. Specifically, step S23 includes: S231: Construct a physical lithology mapping box plot of all boreholes and real features in the entire region, and obtain the threshold for the division between each type and physical attribute through the box plot.

[0043] Specifically, all physical measurement borehole data within the study area were acquired, and physical property parameters closely related to lithology, including but not limited to density, water content, and compression modulus, were extracted. For each known stratigraphic unit (such as sandstone, silt, and clay), statistical analysis was performed on its corresponding physical property values, and box plots were generated. In the box plots, the lower quartile (Q1) and upper quartile (Q3) were taken as the typical value range of the corresponding physical properties for that stratigraphic unit, i.e., the mapping range. This range reflects the fluctuation range of physical properties of this type of stratum under normal geological conditions, and has good representativeness and stability.

[0044] Specifically, this embodiment explicitly excludes outliers and upper / lower boundary values ​​(such as outliers other than the minimum and maximum values) in the box plot when setting the partitioning threshold, and only uses the interval defined by Q1 and Q3 as the effective threshold range. This ensures that the physical attribute threshold used for octree partitioning does not include boundary values ​​in the category transition zone or under measurement noise interference, thereby avoiding misclassification caused by ambiguous samples.

[0045] S232: Obtain the avoidance rules when dividing geological units based on the boundary values ​​of the box plot.

[0046] Specifically, the method for confirming the avoidance rule in step S32 is as follows: In the actual partitioning process, if the physical attribute value of a spatial unit falls into the boundary area between two adjacent stratigraphic mapping intervals (i.e., close to the Q1 or Q3 boundary), the unit may have lithological transition or data uncertainty characteristics. To improve the geological rationality and classification robustness of the model, this embodiment introduces an avoidance rule: when the physical attribute statistical value of the unit to be partitioned is less than a preset tolerance threshold from the boundary of any stratigraphic mapping interval, the unit is not directly classified based on a single threshold, but rather further subdivision is triggered until the attributes within the subunit tend to be consistent, or its attribute values ​​completely fall within the interior of a stratigraphic mapping interval (i.e., a safe area far from the boundary). This avoidance mechanism effectively prevents the misclassification of units located in lithological transition zones or data noise areas, ensuring the input quality of subsequent graphical model construction.

[0047] S24: Recursively subdivide according to the set conditions until the modeling accuracy requirements are met; In the octree construction process, this embodiment abandons the traditional uniform or simple adaptive partitioning method that only uses the number of voxels or geometric dimensions as criteria, and instead adopts the intelligent subdivision strategy based on the physical attribute mapping interval and boundary avoidance rules. Specifically, for the spatial region corresponding to the current octree node, the physical attribute statistical characteristics of all boreholes or interpolation points within it are calculated; if the attribute value distribution of this region exceeds the effective range of a single stratum mapping interval, or falls into the boundary adjacent area that needs to be avoided, then the node is recursively subdivided into eight equal parts; otherwise, the node is determined to be a leaf node, and the subdivision stops. This process continues to iterate until all leaf nodes meet the following conditions: (1) Its internal physical attribute values ​​fall entirely within the mapping range of a certain stratigraphic category; (2) And it is far from the upper and lower boundaries of the interval, which meets the preset avoidance tolerance requirements.

[0048] The octree structure constructed using the above method not only achieves multi-scale adaptive partitioning in space, but also ensures that the geological unit represented by each leaf node has a clear and stable lithological attribution in semantics, providing a high-quality and low-noise input foundation for the subsequent construction of graph models and graph convolution prediction.

[0049] S3: Map each leaf node to a graph node, and combine its spatial coordinates, density, water content and compressibility modulus as node features; To transform the spatial partitioning results into structured data suitable for graph neural network processing, this embodiment abstracts the multi-scale geological units (i.e., leaf nodes) obtained from the octree partitioning into nodes in the graph model, and extracts feature vectors based on their internal physical properties and geometric information as the basic input for subsequent graph convolution operations.

[0050] Specifically, step S3 includes the following sub-steps: S31: The divided geological unit blocks are used as leaf node units of an octree structure.

[0051] After completing the adaptive octree spatial partitioning described in step S2, the study area is divided into several non-overlapping three-dimensional blocks with well-defined spatial boundaries. Each block corresponds to a leaf node in the octree, representing a geological unit with internal consistency in physical properties.

[0052] S32: Calculate the mean values ​​of spatial coordinates, density, water content and compressibility modulus of the geological units contained in each leaf node as node features; For each leaf node, attribute information of all voxel units within it is extracted and statistically aggregated to generate compact and representative node features. Specifically, this includes: Spatial coordinate features: The geometric center coordinates of the spatial region covered by the leaf node are taken as its spatial location representation; Physical property features: The arithmetic mean of the structural strength, water content, and density representative data of all voxels within the leaf node are calculated respectively. The values ​​of the above six dimensions are combined into a six-dimensional feature vector: as the initial feature representation of the graph node.

[0053] S4: Spatial adjacency relationships are established through Delaunay triangulation, and a gravity model is introduced to quantify the causal influence between nodes based on volume and distance, thus constructing a weighted graph structure; The specific steps in S4 mainly include: S41: Map the leaf nodes in the octree structure to graph nodes; Each leaf node of the octree represents a specific spatial region and contains information about the geological properties within that region. S42: Obtain the spatial coordinates and physical property values ​​of the graph nodes as node features; For each graph node transformed from an octagonal leaf node, extract its corresponding spatial coordinates (x, y, z) and parameters reflecting geological characteristics, such as density, water content, and compressibility modulus, as node features.

[0054] S43: Establish spatial adjacency relationships between graph nodes through Delaunay triangulation; Delaunay triangulation is used to determine the connections between nodes in the graph. The distribution of the triangulation network is related to the spatial distribution after partitioning; the triangulation network is sparse in areas with little change, and dense in areas with more change, thus capturing more information.

[0055] S44: Introduce a gravity model to quantify the causal influence between nodes based on volume and distance, and calculate the geological influence weight; Based on the law of gravity in physics, a model is defined to quantify the interactions between nodes. For any two connected nodes i and j, the geological influence weight W is calculated.

[0056] S45: Use the geological influence weight as the edge weight to construct a weighted graph structure.

[0057] Specifically, node features and spatial adjacency relationships are used as features of the graph model.

[0058] S5: Use a graph convolutional neural network to perform lithology classification and prediction on the weighted graph structure to obtain a three-dimensional geological model.

[0059] The network takes the multidimensional features of graph nodes (including spatial coordinates, density, water content, compression modulus, etc.) as input, aggregates neighborhood information through multi-layer graph convolution operations, and finally outputs the lithology category probability distribution of each node, thereby generating a three-dimensional geological model with geophysical significance and spatial consistency.

[0060] Specifically, step S5 includes the following sub-steps: S51: Based on the map model of the study area, the geological category of each predicted node is obtained through graph convolution operation. The graph model constructed in step S4 is input into a pre-trained graph convolutional neural network. This network contains several graph convolutional layers. Each layer aggregates and transforms the features of a node and its first- or multi-order neighboring nodes through a learnable weight matrix, gradually enhancing the semantic representation ability of the node. After multiple propagations, each graph node outputs a normalized category probability vector, the dimension of which corresponds to a preset set of lithology categories. The category with the highest probability is selected as the final lithology prediction result for that node, forming a structured node classification label set.

[0061] S52: Map the classification prediction result of each graph node to the actual geological unit category to generate a three-dimensional geological model.

[0062] It should be noted that the output of the graph convolutional neural network is the classification label on the discrete graph nodes, while the ultimate goal is to construct a continuous, spatially aligned 3D geological model. To this end, this embodiment designs a semantic backfilling mechanism from the graph domain to the spatial domain, specifically including the following sub-steps: S521: Map the stratigraphic category predicted for each graph node to the geological unit represented by its corresponding octagonal leaf node.

[0063] Specifically, each graph node is initially mapped from a leaf node in an octree structure during the modeling phase. This leaf node spatially corresponds to a multi-scale geological unit with defined geometric boundaries and volume. After obtaining the lithological classification label (e.g., sandstone, clay, gravel) of the graph node, this label is assigned to its original corresponding octree leaf node, thus achieving semantic backfilling from the graph domain to the spatial domain. This mapping process ensures the consistency between the classification results and the original geological spatial division, avoiding geological information distortion caused by node abstraction.

[0064] S522: Based on the mapped geological unit categories, assign values ​​to the voxel mesh within the study area to generate a complete three-dimensional geological model.

[0065] In this embodiment, geological units are represented in voxels at the spatial resolution level. The spatial region covered by each octagonal leaf node can be further subdivided into several regular or irregular voxel units. These voxels serve as the basic spatial units for 3D modeling and are used to carry the final stratigraphic category attributes.

[0066] Specifically, after completing step S521, the system traverses all spatial location points (or voxel center points) within the study area and assigns them corresponding stratigraphic category labels based on their respective octagonal leaf nodes. For any voxel located within a leaf node, its stratigraphic attribute value is the lithological category mapped by that leaf node. In this way, the entire study area is filled with voxels one by one, forming a three-dimensional geological model with a unified coordinate system, complete spatial coverage, and clear lithological semantics.

[0067] In some embodiments, the above-described graph convolutional 3D geological modeling method incorporating geological genetic constraints can be implemented through the following methods: Please refer to Figure 3 This embodiment provides a graph convolution 3D geological modeling system that integrates geological genetic constraints, specifically including the following modules: Module 1, the borehole dataset preprocessing module, receives and cleans raw lithological stratification borehole data, standardizing the format of information such as borehole coordinates, depth, and lithological description. Simultaneously, it transforms unstructured stratigraphic records into structured lithology-depth sampling point sequences and generates attributed spatial observation point sets, providing high-quality input data for subsequent spatial modeling.

[0068] Format standardization: Convert borehole data from different sources to the same geographic coordinate system (such as WGS84) and ensure that the units are consistent.

[0069] Structured Transformation: Unstructured stratigraphic information records are transformed into a structured lithology-depth sampling point sequence. For each borehole, attribute boundary points are extracted in depth order, and interpolation is performed at 0.5m intervals between adjacent strata to increase vertical resolution. The final result is a set of spatial sampling data with attribute values ​​(x, y, z, ml, yx, md), where x, y, and z represent the spatial coordinates of the sampling point, ml, yx, and md represent the compressive modulus of structural strength, the liquid limit of water content, and the natural density, respectively.

[0070] Module 2 Spatial Dataset Partitioning Module: Based on preset spatial resolution parameters, a bounding box is established for the study area, and regular grids are generated in the X, Y, and Z directions to produce a three-dimensional regular grid model (such as S-grid). Each grid cell has a unique spatial index and geometric center coordinates, forming a discretized spatial framework to carry geological attribute information.

[0071] Specifically, the bounding box is constructed as follows: The study area's extent is determined based on the borehole distribution data depth and the DEM data range. The horizontal extent is determined by the union of the two data points in the horizontal direction, while the vertical extent is determined by the upper limit of the DEM and the lower limit of the borehole data. The resolutions in the horizontal and vertical directions are set. Since the depth in a real modeling scenario is much smaller than the horizontal scale of the study area, the spatial resolution in the x and y axes is set to 50m, and the z-axis resolution is set to 2m. This results in a spatially distributed mesh element at this resolution.

[0072] Furthermore, spatial interpolation was performed using the ordinary kriging method to obtain the spatial distribution of physical properties throughout the study area.

[0073] Further, the entire study area is divided into grid cells: statistical analysis is performed on its corresponding physical properties (such as density), box plots are drawn, and the lower quartile (Q1) and upper quartile (Q3) are extracted. The statistical values ​​of the physical properties of all voxel cells contained in the divided octagonal leaf nodes must strictly lie within this mapping interval and must not be equal to the boundary values. Let a certain leaf node... If the value of the internal physical property is z, then it satisfies the division termination condition if and only if there exists a stratigraphic category c such that:

[0074] in The threshold is defined. Simultaneously, there are no boundary values ​​(P1, P2...) within the region, where p1 and p2 are the quarter points of the stratigraphic category in the box plot and the boundary portion, respectively.

[0075] Module 3 Spatial Feature Extraction Module: This module uses tree nodes as nodes in the subsequent graph model and calculates the physical properties of each grid cell to obtain feature vectors. It further combines Delaunay triangulation to establish topological connections between nodes and introduces a gravity model to quantify the volume and distance influence between nodes, constructing a weighted graph structure where nodes represent grid cells and edge weights reflect the intensity of geological interactions.

[0076] Node features consist of spatial coordinates and attribute features. Specifically, for each tree node v, its feature vector fv is a multidimensional vector composed of the mean values ​​of the following geological attributes: Spatial coordinates: The coordinates of the geometric center of the spatial region covered by the tree node. , () as a location representation.

[0077] Physical properties: Density: Calculates the arithmetic mean of the density of all voxels within this tree node. .

[0078] Water content: Calculate the arithmetic mean of the water content of all voxels within this tree node. .

[0079] Compression modulus: Calculates the arithmetic mean of the compression modulus of all voxels within this tree node. .

[0080] Therefore, for each tree node v, its feature vector fv can be represented as a six-dimensional vector:

[0081] Module 4, Graph Neural Network Prediction Module: Utilizing the constructed weighted graph as input, this module employs a Graph Convolutional Neural Network (GNN) model for spatial propagation and feature learning. Through multi-layer graph convolution operations, it captures the non-local correlations and complex dependencies of geological attributes in space, enabling intelligent prediction and inference of lithological categories in unobserved areas.

[0082] The specific graph network construction method used here is Delaunay triangulation: Delaunay triangulation is performed on the center point of the tree node unit to obtain the connection relationship between the nodes.

[0083] Further elimination of adjacency between nodes is calculated using a gravity model: for any two connected nodes i and j, their edge weights... It can be calculated using the following formula:

[0084] in, These are the volumes corresponding to nodes i and j, respectively. It is the Euclidean distance between the two nodes.

[0085] The weighted graph structure is further input into the GNN model, and multi-layer graph convolution operations are performed: A multi-layer graph convolution model is used, and during the operation of each layer, the features of neighboring nodes are aggregated and the representation of the current node is updated. Classification prediction: The last layer typically uses the Softmax function to output the probability distribution of the lithology to which each node belongs.

[0086] Module 5, Spatial Mapping Modeling Module: This module maps the prediction results output by the graph neural network back to a 3D mesh model, completing continuous spatial interpolation and assignment of lithological properties to generate a high-precision, high-fidelity 3D geological model. This model can intuitively display the spatial distribution patterns of lithology and supports applications such as geological modeling, engineering assessment, and risk prediction.

[0087] Prediction result mapping: The classification prediction result of each graph node is back-mapped to the geological unit represented by its corresponding octagonal leaf node. Attribute assignment: Traverse all S-grid units and assign corresponding lithological category labels according to their respective octagonal leaf nodes to construct a geological model.

[0088] like Figure 4 As shown, this application also proposes an electronic device including but not limited to the following components: at least one processor 121, at least one communication interface 123, memory 124, and at least one bus 122 for implementing communication between the components.

[0089] This electronic device, through the collaborative operation of its internal components, supports the execution of a graph convolutional geological modeling method that incorporates geological genetic constraints. Specifically, the communication bus 122 enables interconnection and data exchange between the processor 121, the communication interface 123, and the memory 124. The communication bus 122 can adopt various standard designs, such as PCI or EISA, and can be configured as an address bus, data bus, or control bus according to actual needs. For the sake of simplicity, the diagram uses only a single line to represent the communication path, but this does not mean that the system contains only a single type of bus.

[0090] The processor 121, as the core computing unit, can be a central processing unit (CPU), a network processor (NP), or a combination of both. Furthermore, it may integrate application-specific integrated circuits (ASICs) or programmable logic devices (PLDs), including complex programmable logic devices (CPLDs) and field-programmable gate arrays (FPGAs), to enhance specific task processing capabilities.

[0091] Memory 124 is responsible for storing the operating system, applications, and data. It can consist of high-speed random access memory (RAM) or non-volatile memory such as flash memory, hard disk drive (HDD), or solid-state drive (SSD). In some configurations, memory 124 can also be part of a remote storage device connected to the main system via a network.

[0092] Communication interface 123 provides data interaction capabilities with the external environment, and may include a display screen, keyboard, wired or wireless communication interface, etc. This enables the device to receive input data, display results, and exchange information with other devices.

[0093] The key to this embodiment is that the application stored in memory 124 can execute a graph convolution geological modeling process that incorporates geological genetic constraints under the call of processor 121, thereby achieving accurate simulation of geological structures.

[0094] Furthermore, this embodiment also relates to a graph convolution geological modeling device that incorporates geological genesis constraints. This device is not limited to platforms such as desktop computers, laptops, or cloud servers. The device includes a processor, memory, and other necessary hardware components, designed to run the aforementioned geological modeling program and complete the entire process from raw geological data analysis to the generation of a three-dimensional geological model.

[0095] 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 graph convolution 3D geological modeling method incorporating geological genetic constraints, characterized in that, Includes the following steps: S10, Obtain raw geological data, construct grid cells for the study area, and perform spatial sampling on the raw geological data to obtain a geological attribute distribution dataset; S20. Based on the geological attribute distribution dataset, a subsurface physical field model is constructed, and an octree structure is used to adaptively divide the study area into spatial parts to obtain multi-scale geological units as leaf nodes. S30, map each leaf node to a graph node, and combine the spatial coordinates, density, water content and compressibility modulus of the leaf node as node features; S40, establish the spatial adjacency relationship between the graph nodes through Delaunay triangulation, and introduce a gravity model to quantify the causal influence between nodes based on volume and distance, and construct a weighted graph structure; S50, using a graph convolutional neural network to perform lithology classification and prediction on the weighted graph structure, a three-dimensional geological model is obtained.

2. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 1, characterized in that, Step S10 is as follows: S11, Determine the modeling range boundary based on the study area, and establish a bounding box based on the modeling range boundary; S12, Divide the bounding box into regular meshes according to a preset precision to obtain a three-dimensional mesh simulation model; S13, Extract lithological information from lithological stratification borehole data, wherein the lithological stratification borehole data originates from the original geological data; S14 maps the lithological information in the borehole to the corresponding three-dimensional regular mesh model voxels to generate a geological attribute distribution dataset.

3. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 1, characterized in that, Step S20 is as follows: S21, Extract physical attribute data from the geological attribute distribution dataset; S22, Spatial interpolation processing is performed on the physical attribute data to generate a physical distribution field model of the study area; S23, Divide the boundaries and minimum unit size according to the unit boundary mapping relationship of the study area; S24, recursively subdivide according to the set conditions until the modeling accuracy requirements are met.

4. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 3, characterized in that, Step S23 is as follows: S231, construct a physical lithology mapping box plot of all boreholes and real features in the entire area, and obtain the division threshold between each type and physical attribute through the box plot; S232, based on the boundary values ​​of the box plot, the avoidance rules for dividing geological units are obtained.

5. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 1, characterized in that, Step S30 is as follows: S31, the divided geological unit blocks are used as leaf node units of the entire octree structure; S32, calculate the mean values ​​of spatial coordinates, density, water content and compressibility modulus of the geological units contained in each leaf node as node features.

6. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 1, characterized in that, Step S40 is as follows: S41, map the leaf nodes in the octree structure to graph nodes; S42, obtain the spatial coordinates, density, water content and compressibility modulus of the graph nodes as node features; S43, establish spatial adjacency relationships between graph nodes through Delaunay triangulation; S44 introduces a gravity model to quantify the causal influence between nodes based on volume and distance, and calculates the geological influence weight; S45 uses the geological influence weight as the edge weight to construct a weighted graph structure.

7. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 6, characterized in that, In step S50, the graph convolutional neural network includes an input layer, a hidden layer, and an output layer; The input layer receives a set of graph nodes mapped from the leaf nodes of an octagonal tree. The feature vector of each node contains its physical properties such as spatial coordinates, density, water content, and compressibility modulus, forming a six-dimensional node feature vector. The hidden layer consists of two stacked graph convolutional layers. Each layer executes a message passing mechanism to aggregate the feature information of neighboring nodes and update the representation of the current node. The output layer uses the Softmax function to classify the final node features and outputs the probability distribution of the lithology of each node, thereby achieving a refined prediction of the lithology distribution in underground space.

8. The graph convolution 3D geological modeling method with integrated geological genetic constraints as described in claim 1, characterized in that, Step S50 is as follows: S51, based on the map model of the study area, the geological category of each predicted node is obtained through graph convolution operation; S52 maps the classification prediction value of each graph node to the actual geological unit category to obtain a three-dimensional geological model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the graph convolution three-dimensional geological modeling method that incorporates geological genetic constraints as described in any one of claims 1 to 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the graph convolution three-dimensional geological modeling method that incorporates geological genetic constraints as described in any one of claims 1 to 8.

Images