A three-dimensional mineralization prediction method based on 3DGCN-CNN model

Abstract

The present invention relates to the intersection of computer and geology, and in particular to a three-dimensional mineralization prediction method based on a 3DGCN-CNN model, comprising: step S1: constructing a three-dimensional geological database; step S2: constructing a three-dimensional geological model and a three-dimensional ore body model; step S3: constructing a three-dimensional mineralization prediction information set; step S4: constructing a 3DGCN-CNN model; step S5: constructing a 3DCNN data set and 3DGCN graph structure data; step S6: inputting a training set into the model for training; and step S7: inputting data of a region to be predicted into the trained model for prediction to obtain a prediction result. The present invention effectively integrates spatial and local features through the 3DGCN-CNN model, eliminates the need for quantitative analysis of the three-dimensional geological model, and improves the prediction capability of mineralization potential.

Description

Technical Field

[0001] The present invention relates to the intersection of computer science and geology, and specifically to a three-dimensional mineralization prediction method based on a 3DGCN-CNN model. Background Art

[0002] With the acceleration of global industrialization and the surge in demand for mineral resources driven by emerging technologies, shallow, easily exploitable mineral resources are becoming increasingly depleted. Exploration of deep, concealed ore bodies has become an inevitable choice to ensure a sustainable supply of resources. Deep deposits are significantly more difficult to explore than shallow deposits due to their great depth, complex geological structure, and weak mineralization that is difficult to directly observe. Prospecting and prediction for these deposits are extremely challenging. Therefore, developing efficient and accurate three-dimensional mineralization prediction methods is of strategic importance for reducing the risks of deep prospecting and improving resource development efficiency.

[0003] Currently, 3D mineralization prediction and its methods and technologies have become an important means of predicting deep and concealed mineral deposits, and have achieved significant results in practical prospecting applications. 3D mineralization prediction is bound to become a key direction for the future development of mineral exploration techniques. Key areas of 3D mineralization prediction include the mining of 3D prediction elements, uncertainty analysis and optimization research for 3D mineralization prediction, 3D deep learning models and methods for 3D mineralization prediction, and multi-scale 3D mineralization prediction methods. Addressing these key issues will further improve the efficiency and level of deep mineralization prediction and exploration, and facilitate breakthroughs in deep prospecting.

[0004] After searching, it was found that Chinese patent literature disclosed a three-dimensional prospecting prediction method, system and storage medium (application number: CN202411693382.0). The method steps are as follows: collecting multi-source geological data in the target mining area and establishing a geological database; standardizing the data format of the geological database to obtain preprocessed data; constructing a three-dimensional geological model based on the preprocessed data; quantitatively analyzing the three-dimensional geological model to extract favorable mineralization information; using a three-dimensional mineralization process numerical simulation method to numerically simulate the mineralization process of the target mining area and extract numerical simulation variable information; fusing the favorable mineralization information with the numerical simulation variable information to construct a three-dimensional comprehensive prospecting prediction model; predicting the prospecting target area based on the three-dimensional comprehensive prospecting prediction model to obtain prospecting prediction results; it quantitatively analyzes the three-dimensional geological model to extract favorable mineralization information, but this artificial secondary processing method has the risk of feature expression distortion when dealing with complex structure-lithology combinations, which may cause inaccurate prediction results; Chinese patent literature also discloses a three-dimensional quantitative prediction method, equipment and product for a mineral cluster scale area (application number: CN202410286484.4). The method steps include obtaining geological survey data of different scales; establishing a three-dimensional geological model of the ore cluster scale based on the geological survey data of different scales and the distribution relationship and mineralization law of the ore deposits in the ore cluster; the three-dimensional geological model is used for visualization; quantitatively extracting favorable mineralization information based on the three-dimensional geological model and mineralization law of the ore cluster scale; establishing a three-dimensional quantitative prediction model based on the CNN network using the favorable mineralization information; and using the three-dimensional quantitative prediction model to perform three-dimensional quantitative prediction of the ore cluster scale region; the three-dimensional quantitative prediction model is established based on the CNN network, which can learn the spatial position relationship between data, and in the three-dimensional mineralization prediction task, the three-dimensional spatial information contains the mineralization position relationship, so the CNN network has unique advantages in the three-dimensional mineralization prediction task; but CNN focuses more on the extraction of local features and relies on a regular grid structure for convolution operations. The feature extraction efficiency of irregular or sparse data is low, which may lead to the loss of key information. Moreover, the fixed-size convolution kernel used in traditional CNN is difficult to capture geological phenomena of different scales at the same time, resulting in a lack of prediction ability of the model. Summary of the Invention

[0005] The purpose of the present invention is to provide a three-dimensional mineralization prediction method based on the 3DGCN-CNN model, which does not require quantitative analysis of the three-dimensional geological model, and directly uses the favorable mineralization structures of the three-dimensional geological model as Boolean prediction variables. It can realize the coordinated extraction of geological local features and geological spatial features, and combines multi-scale modules and spatial attention mechanisms in CNN to improve the reliability of three-dimensional mineralization prediction results.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a three-dimensional mineralization prediction method based on a 3DGCN-CNN model, comprising the following steps:

[0007] Step S1: Collect geological data of the study area and construct a three-dimensional geological database;

[0008] Step S2: Using geological data, construct a three-dimensional geological model of the study area and a three-dimensional ore body model of known ore deposits;

[0009] Step S3: Each favorable metallogenic structure of the 3D geological model is directly used as a Boolean prediction variable, and is used together with the numerical simulation variables to construct a 3D metallogenic prediction information set;

[0010] Step S4: constructing a 3DGCN-CNN model; including a 3DCNN module and a 3DGCN module, which are used to extract local features and spatial features respectively, and the two features are spliced ​​and fused and classified through a fully connected layer; the 3DCNN module is divided into a geological body processing part and a numerical simulation variable processing part, the geological body processing part adopts a ResNet18 architecture, and the numerical simulation variable processing part adopts a LeNet-5 architecture, and the geological body processing part and the numerical simulation variable processing part both combine a multi-scale module and a spatial attention mechanism, and the multi-scale module is integrated into the Inception module; the 3DGCN module models the spatial relationship of the geological body, and the input is the constructed graph structure data. The information of the first-order and second-order neighborhood is captured respectively through two layers of graph convolution layers, and the nonlinear transformation of the spatial features is realized through Batch Normalization and ReLU activation function;

[0011] Step S5: Determine a positive sample data set and a negative sample data set, each subset in the data set is a unit block, and the positive and negative sample unit blocks are divided into a training set and a test set in a ratio of 8:2, and construct a 3DCNN data set and a 3DGCN graph structure data; when constructing the graph structure data for inputting the 3DGCN module, first extract the Boolean geological body attributes of each unit block, construct a node feature matrix, then construct an edge set of the graph, and use the normalized value of the Euclidean distance to represent the edge weight, respectively construct the feature matrix, edge set, edge weight and label of the training set and test set, and use the PyTorchGeometric Data data structure to encapsulate the graph structure data; when constructing the 3DCNN data set, take each positive sample unit block and negative sample unit block as the center, expand at least 4 blocks in each direction of the three-dimensional space to construct a large cubic unit, and form a three-dimensional cubic block containing the central unit and its neighborhood as the input sample;

[0012] Step S6: Input the training set into the model for training; after each round of training, use the test set to adjust the model weights until the model converges;

[0013] Step S7: Input the data of the area to be predicted into the trained model and output the prediction result through the fully connected layer.

[0014] Preferably, in step S1, the geological data include a comprehensive interpretation profile of geological and geophysical data, a regional geological map, digital elevation data, a mineral deposit profile, drilling data and physical property parameters of geological bodies in the study area.

[0015] Preferably, in step S2, the three-dimensional geological model is optimized using three-dimensional geophysical forward modeling and full lithologic inversion methods according to the physical property parameter values ​​of each geological body.

[0016] Preferably, in step S3, the three-dimensional block model is used as a carrier to discretize the large block model of the study area into unit blocks of equal size, and each unit block is assigned a Boolean prediction variable and a numerical simulation variable, and the variables assigned to all unit blocks together constitute a three-dimensional mineralization prediction information set; each favorable mineralization structure is a favorable mineralization rock body, a favorable mineralization stratum and a favorable mineralization fault; the Boolean prediction variables include favorable mineralization rock body, favorable mineralization stratum and favorable mineralization fault, and the numerical simulation variables include volume strain increment, shear strain increment, volume strain rate, shear strain rate, pore pressure, maximum principal stress and temperature.

[0017] Preferably, in step S3, the three-dimensional geological model is directly learned without performing spatial analysis on the three-dimensional geological model to convert it into a continuous prediction variable.

[0018] Preferably, in step S5, the positive sample data set is composed of the main mineralized unit blocks of the known ore deposits in the study area; the negative sample data set construction process is: with the drilling point as the center, a buffer zone is applied to it to indicate the average exploration spacing, and for drill holes without ore, the unit blocks within the range are used as negative sample unit blocks; for drill holes with ore, the non-ore body unit blocks around the known ore body revealed within the range are used as negative sample unit blocks.

[0019] Preferably, in step S6, after each round of training is completed, the model is performance tested using the test set, the predictive ability of the model is evaluated by calculating the cross entropy loss function value on the test set, the loss function values ​​of the model on the training set and the test set are recorded and visualized for analyzing the convergence of the training process, and the hyperparameter settings of the model are dynamically adjusted according to the changes in the loss function value of the test set, including learning rate decay and training rounds, and the trained model is obtained after multiple rounds of iterations.

[0020] Preferably, in step S7, the prediction result is assigned to the three-dimensional block model as a mineralization probability value, and the model prediction effect is judged by screening high mineralization probability blocks and visualizing them, and comparing them with known ore deposits.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention performs deep three-dimensional mineralization prediction by integrating the 3DGCN and 3DCNN models, combining the processing of spatial and local features of deep geological bodies. The 3DGCN module focuses on capturing the spatial relationship between geological bodies, while the 3DCNN module focuses on learning the geometric morphology, spatial distribution and other local features of geological bodies and the relationship between numerical simulation variables. The spatial and local features are effectively integrated through the 3DGCN-CNN hybrid model.

[0023] 2. The present invention extracts information directly from the three-dimensional geological model by fusing the 3DGCN and 3DCNN models, thereby eliminating the steps of quantitative analysis of the three-dimensional geological model and extracting favorable mineralization information in other three-dimensional mineralization prediction methods, and avoiding the feature expression distortion problem caused by this processing process.

[0024] 3. Based on the characteristics of three-dimensional mineralization prediction, the present invention introduces a spatial attention mechanism and an Inception module into the 3DCNN module. By adding convolutional layers of different scales into the Inception module, the network can capture features within different receptive fields while reducing computational costs, and enhance attention to key areas, thereby improving the ability to predict mineralization potential.

[0025] 4. After constructing the initial three-dimensional geological model, the present invention uses three-dimensional geophysical forward modeling and full lithologic inversion to optimize the model based on the physical parameter values ​​of the regional geological body. When selecting positive and negative samples, a buffer zone strategy centered on the borehole is adopted, and the non-ore bodies around the known ore bodies revealed by the borehole are used as negative sample unit blocks, so that the prediction results are more consistent with geological facts. At the same time, when constructing the data set for the 3DCNN module, the selected unit block is used as the central unit and then expanded outward in the three coordinate axis dimensions, enriching the model's understanding and utilization of the spatial information around the center point and enhancing the model's predictive ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the three-dimensional mineralization prediction method of the present invention;

[0027] Figure 2 This is a structural diagram of the three-dimensional mineralization prediction method of the present invention;

[0028] Figure 3 A three-dimensional geological model diagram of the study area in an embodiment of the present invention;

[0029] Figure 4A comparison diagram of the measured magnetic anomaly in the study area, the magnetic anomaly in the initial three-dimensional geological model, and the magnetic anomaly in the optimized three-dimensional geological model in the embodiment of the present invention;

[0030] Figure 5 This is a structural diagram of the 3DGCN-CNN hybrid model of the present invention;

[0031] Figure 6 This is a model structure diagram of the numerical simulation variable processing part in the 3DCNN module of the present invention;

[0032] Figure 7 This is a model structure diagram of the geological body processing part in the 3DCNN module of the present invention;

[0033] Figure 8 This is the improved three-dimensional Inception module diagram of the present invention;

[0034] Figure 9 Schematic diagram of the three-dimensional spatial attention mechanism of the present invention;

[0035] Figure 10 A schematic diagram of selecting negative sample unit blocks based on drilling and ore body data according to the present invention;

[0036] Figure 11 This is the process of expanding and building a single input sample for the 3DCNN module of the present invention;

[0037] Figure 12 This is the 3DGCN-CNN model training curve diagram of the present invention;

[0038] Figure 13 This is a comparison chart of ROC curves of 3DGCN-CNN, 3D CNN method, and RF method in an embodiment of the present invention;

[0039] Figure 14 The figure shows the target area drilling histogram and typical geological phenomena in the embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] This method is based on a 3D geological model constructed from a comprehensive interpretation of geological and geophysical data, including profiles, regional geological maps, digital elevation data, ore deposit profiles, drill hole data, and physical parameters of the geological bodies in the study area. This 3D geological model captures the deep geological structure and numerically simulates the mineralization process based on it. This model identifies a series of numerical simulation variables with ore-indicating significance, and, based on deep learning technology, constructs an intelligent 3D mineralization prediction model that directly learns from the 3D geological model.

[0042] like Figure 1 and Figure 2 As shown, a three-dimensional mineralization prediction method based on the 3DGCN-CNN model includes the following steps:

[0043] Step S1: Collect geological data of the study area and construct a three-dimensional geological database;

[0044] Collect comprehensive interpretation profiles of geological and geophysical data covering a certain area, regional geological maps, digital elevation data, ore deposit profiles, drill hole data and physical parameters of geological bodies in the study area.

[0045] Step S2: Using geological data, construct a three-dimensional geological model of the study area and a three-dimensional ore body model of known ore deposits;

[0046] Using the collected geological data of a certain area, a 3D geological model of the study area was constructed using a 3D implicit modeling method. Figure 3 , which is a three-dimensional geological model of the study area in the embodiment of the present invention; based on the physical parameter values ​​of each geological body, the morphological characteristics of the intrusive rock body are optimized using three-dimensional geophysical forward modeling and full lithologic inversion methods. The optimized three-dimensional geological model makes its rock mass and stratum morphological distribution more consistent with the measured data and more in line with geological facts; the ore body model of the known ore deposits in the study area is constructed using a three-dimensional display modeling method; please refer to Figure 4 , which is a comparison diagram of the measured magnetic anomaly in the study area, the initial three-dimensional geological model magnetic anomaly, and the optimized three-dimensional geological model magnetic anomaly in the embodiment of the present invention.

[0047] Step S3: directly using each favorable metallogenic structure of the three-dimensional geological model as a Boolean prediction variable, and constructing a three-dimensional metallogenic prediction information set with the Boolean prediction variable and the numerical simulation variable obtained by the three-dimensional metallogenic process numerical simulation calculation;

[0048] Using a 3D block model as a carrier, the large block model of the study area was discretized into unit blocks of 100m×100m×50m in size. Each unit block was assigned 3D mineralization prediction attributes, namely, the geological structure and numerical simulation variables that are favorable for mineralization in the region. The geological structure that is favorable for mineralization is represented by a Boolean prediction variable of 0 or 1, where 0 indicates that the unit block is not the geological body, and 1 indicates that the unit block is the geological body. Each unit block includes Boolean prediction variables and numerical simulation variables. Among the Boolean prediction variables, the mineralization-favorable rock bodies, mineralization-favorable strata, and mineralization-favorable faults are selected for 3D mineralization prediction. Among the numerical simulation variables, seven continuous numerical simulation variables with ore-indicating significance, namely, volume strain increment, shear strain increment, volume strain rate, shear strain rate, pore pressure, maximum principal stress, and temperature, are selected for 3D mineralization prediction. The variables assigned to all unit blocks together constitute the 3D mineralization prediction information set.

[0049] Furthermore, Boolean predictor variables are directly derived by learning the 3D geological model, eliminating the need for spatial analysis of the 3D geological model to extract continuous predictor variables. This eliminates the need for quantitative analysis of the 3D geological model to extract favorable mineralization information, a step often encountered in other 3D mineralization prediction methods, and avoids the distortion of feature representation that can arise from this process.

[0050] Step S4: constructing a 3DGCN-CNN hybrid model; the 3DGCN-CNN hybrid model includes a 3DCNN module and a 3DGCN module; the local features extracted by the 3DCNN module and the spatial features extracted by the 3DGCN module are fused through a splicing operation and classified in a fully connected layer;

[0051] 3DCNN is a three-dimensional convolutional neural network, a type of feedforward neural network with a deep structure that includes convolution calculations. It is one of the representative algorithms of deep learning and is mainly used to process grid data. It is good at extracting local spatial information. 3DGCN is a three-dimensional graph convolutional neural network, which is mainly used to process graph structure data. Graph structure data is irregular grid structure data. 3DGCN is good at capturing complex relationships in graph structures. Based on this, the 3DGCN-CNN hybrid model constructed by the present invention includes a 3DCNN module and a 3DGCN module. The 3DGCN module is used to learn the spatial correlation between geological bodies, and the 3DCNN module is used to learn local features such as the geometric form and spatial distribution of geological bodies, as well as the relationship between numerical simulation variables. The 3DGCN-CNN hybrid model integrates the local and spatial features of geological bodies, and improves the mineralization prediction ability of complex geological bodies such as deep mineralization. Please refer to Figure 5 , which is the structural diagram of the 3DGCN-CNN hybrid model of the present invention, the specific model construction method is as follows:

[0052] (1) Building a 3DGCN model

[0053] The 3DGCN module models the spatial relationship of geological bodies. The input is the constructed graph structure data. The first layer of GCN convolution convolves the input feature matrix through the GCNConv operation to capture the information of the first-order neighborhood of each node (that is, the directly adjacent nodes). The output of this layer is normalized by Batch Normalization to ensure the balance of features, and the ReLU activation function is applied for nonlinear transformation to obtain the feature representation of each node; the second layer of GCN convolution captures the second-order neighborhood information of each node, that is, in addition to the directly adjacent nodes, it also includes the neighbors of its neighbors, so that it can capture deeper spatial dependency information. Similar to the first layer, the output is processed by Batch Normalization and ReLU activation function to obtain the node features after the second layer of graph convolution. The final output feature dimension is 32, representing the spatial characteristics of the geological body.

[0054] (2) Building a 3DCNN model

[0055] The 3DCNN module is divided into geological body processing part and numerical simulation variable processing part, please refer to Figure 6 , is a model structure diagram of the numerical simulation variable processing part in the 3DCNN module of the present invention, Figure 7 This is a structural diagram of the geological body processing model in the 3DCNN module of the present invention.

[0056] The geological body processing part adopts the ResNet18 architecture. The kernel size of the first convolution layer is 3×3×3 and the stride is 1. After the first convolution layer, the improved Inception module is used. It converts the traditional Inception module into a three-dimensional structure and introduces a multi-scale module. Please refer to Figure 8 , is the improved 3D Inception module diagram of the present invention. First, a 1×1×1 convolution layer is used to reduce the number of channels and aggregate information. Then, 3×3×3 and 5×5×5 convolution layers and a 3×3×3 pooling layer are used to extract feature information of different scales. A ReLU activation function is immediately applied after each convolution layer to introduce nonlinearity. Finally, these features are superimposed and output. The 1×1×1 convolution layer can actually be regarded as a pixel-by-pixel fully connected layer. In this way, the number of channels is reduced while keeping the size of the feature map unchanged, thereby reducing the computational cost. This 3D Inception module design can more effectively capture and utilize the spatial features in 3D sample data. A spatial attention mechanism is added after the Inception module to enhance the focus on key areas. Please refer to Figure 9, which is a diagram of the three-dimensional spatial attention mechanism of the present invention. The maximum pooling layer is used for dimensionality reduction, the pooling window size is 2×2×2, and the dimension of the final extracted local features is 64. The numerical simulation variable processing part adopts the LeNet-5 architecture, which is also combined with the improved Inception module and spatial attention mechanism. The final local feature dimension extracted is 256. By introducing the spatial attention mechanism and Inception module into the 3DCNN module and adding convolutional layers of different scales to the Inception module, the network can capture features within different receptive fields while reducing computational costs, and enhance the focus on key areas, thereby improving the ability to predict mineralization potential.

[0057] (3) Fusion of local features and spatial features

[0058] The local features and spatial features obtained by 3DCNN and 3DGCN are fused through a splicing operation. The fused feature dimension is 352 (spatial feature dimension 32 + local feature dimension 64 + local feature dimension 256), and the final prediction output is generated through a fully connected layer.

[0059] Step S5: Determine the positive sample data set and the negative sample data set. Each subset in the data set is a unit block. The positive sample unit block and the negative sample unit block are divided into a training set and a test set in a ratio of 8:2. The 3DCNN data set and the 3DGCN graph structure data are constructed.

[0060] The positive sample dataset is composed of the main mineralized unit blocks of the known ore deposits in the study area; the negative sample dataset is constructed by applying a buffer zone to the drill hole point to indicate the average exploration spacing. For drill holes that do not reveal ore, the unit blocks within the range are used as negative sample unit blocks; for drill holes that reveal ore, the non-ore body unit blocks around the known ore bodies revealed within the range are used as negative sample unit blocks. Figure 10 , which is a schematic diagram of selecting negative sample unit blocks based on drilling and ore body data in the present invention.

[0061] Based on the gridded three-dimensional geological model, the study selected the Chating, Qiaomaishan and Changshan deposits, which have been widely explored, as data-driven end-elements, and used the unit blocks contained in the three-dimensional ore body models of these three deposits as positive sample unit blocks. The positive sample data set contains a total of 734 positive sample unit blocks. During the exploration process, the drill holes in the area control the morphology of the entire ore body. The mineralization properties of the unit blocks within a certain range of the drill holes can be considered to be known. The unit blocks within 300 meters around the drill holes in the study area are considered to be the areas explored by the drill holes. Therefore, for the negative sample data set, a 300-meter buffer zone is applied to the drill hole point as the center to indicate the average exploration spacing. For drill holes without ore, the unit blocks within the range are used as negative sample unit blocks. For drill holes with ore, the non-ore body units around the known ore bodies revealed within the range are used as negative sample unit blocks. The negative sample data set contains a total of 8875 negative sample unit blocks. In order to maintain the balance of positive and negative samples, 800 negative sample unit blocks were randomly selected from the negative sample data set to participate in model training. The positive and negative sample unit blocks were divided into training set and test set in a ratio of 8:2. The training set contained 587 positive sample unit blocks and 640 negative sample unit blocks, and the test set contained 147 positive sample unit blocks and 160 negative sample unit blocks. Each positive and negative sample unit block contained 3 Boolean predictor variable attributes and 7 numerical simulation variable attributes. The Boolean variable attributes are represented by 0 or 1; the numerical simulation variables are continuous values.

[0062] When constructing the graph structure data for input into the 3DGCN module, the Boolean predictor variable attributes of each positive and negative sample unit block are first extracted, and the node feature matrix is ​​constructed with each positive and negative sample unit block as a node. Then, the edge set of the graph is constructed, and the normalized value of the Euclidean distance is used to represent the edge weight. The constructed edge represents the spatial relationship between geological bodies. The feature matrix, edge set, edge weight and label of the training set and test set are constructed respectively. The Geometric Data structure encapsulates graph-structured data. When constructing a 3DCNN dataset, since the formation of ore bodies is often affected by the surrounding environment, including geological structure, lithology, mineralization type, and other factors, it is necessary to consider and utilize the predicted information of the space surrounding the positive and negative sample blocks, so that the model can fully learn the spatial correlation information between the known mineralization and the surrounding geological bodies. Therefore, when constructing the dataset required by the 3DCNN module, the selected positive and negative sample blocks are used as the central unit, and then the length of the n blocks is expanded outward in the three coordinate axis dimensions to construct a large cubic unit, where n ≥ 4, so as to construct a cubic block with a shape of (2n+1)×(2n+1)×(2n+1). Please refer to Figure 11, which is the single input sample expansion and construction process of the 3DCNN module of the present invention. Each cube block contains Boolean prediction variables and numerical simulation variable attributes that are beneficial to mineralization, enriching the model's understanding and utilization of spatial information around the center point. Among them, the Boolean prediction variable attributes that are beneficial to mineralization enter the geological body processing part, and the numerical simulation variable attributes enter the numerical simulation variable processing part.

[0063] Step S6: Input the training set into the 3DGCN-CNN hybrid model for model training; after each round of training, use the test set data to adjust the model weights until the model converges;

[0064] The training data was fed into the 3DGCN-CNN hybrid model. The 3DGCN module extracted geological spatial features based on graph-structured data through graph convolution operations. The 3DCNN module trained the training data, extracting local features such as the geological geometry and spatial distribution, as well as the relationships between numerical simulation variables. The model was trained using the Adam optimizer with a batch size of 16 and an initial learning rate of 0.001 that decreased linearly with the number of training rounds. After each round of training, the model weights were adjusted using the test data. After 20 rounds of iteration, the trained model achieved an overall accuracy of 95.94%. (See [1] for details.) Figure 12 , which is the 3DGCN-CNN model training curve diagram of the present invention.

[0065] In order to analyze the effectiveness of the 3DGCN-CNN hybrid model in performing 3D mineralization prediction without performing quantitative analysis on the 3D geological model to extract Boolean prediction variables, a comparative analysis was conducted between 3DGCN-CNN and traditional 3D CNN and RF algorithms. The 3D CNN and RF algorithms used the mineralization favorable information extracted by quantitative analysis of the 3D geological model and the same numerical simulation variables as those of 3DGCN-CNN as prediction attributes. For further comparison, the results of the 3D CNN and RF algorithms using prediction variables without quantitative analysis were also calculated. Based on the three-dimensional mineralization prediction analysis of 3DGCN-CNN, three-dimensional CNN, RF, three-dimensional CNN without quantitative analysis, and RF without quantitative analysis, by comparing the consistency of high mineralization probability areas with known deposits and the performance indicators of the models in these five cases, it can be seen that 3DGCN-CNN can predict the known Shuangjing and Chashan deposits near the Changshan deposit, while the three-dimensional CNN and RF methods only perform well in the vicinity of the Shuangjing and Chashan deposits, and the fit with the known deposit locations is poor. At the same time, drilling projects in the small-scale rock pillar area south of the Chating deposit have shown that the mineralization capacity of this area is weak. The prediction results of the 3DGCN-CNN method are fewer high mineralization probability units in this area than those of the three-dimensional CNN and RF algorithms, which is more consistent with the drilling facts. Secondly, the three-dimensional CNN without quantitative analysis and the RF algorithm without quantitative analysis are almost completely unable to predict known deposits other than data-driven end members. See Figure 13 , which is a comparison chart of the ROC curves of the 3DGCN-CNN, 3D CNN method, RF method, 3D CNN without quantitative analysis, and RF algorithm without quantitative analysis in the embodiment of the present invention. It can be concluded that the ROC-AUC values ​​of 3DGCN-CNN, 3D CNN, RF, 3D CNN without quantitative analysis, and RF without quantitative analysis are 0.982, 0.968, 0.943, 0.959, and 0.931, respectively. The analysis results show that for previous 3D mineralization prediction methods based on models such as 3D CNN and RF, the step of pre-quantification analysis of the 3D geological model is effective and necessary. The performance of the 3DGCN-CNN model without pre-3D spatial analysis is better than that of the 3D CNN and RF models with pre-3D spatial analysis. This shows that by fusing 3DGCN and 3DCNN to directly model the spatial relationship between geological data, greater stability can be achieved in capturing complex geological features.

[0066] Step S7: Input the data of the area to be predicted into the pre-trained 3DGCN-CNN model for prediction, and obtain the prediction result through the fully connected layer.

[0067] The study area dataset is input into the pre-trained 3DGCN-CNN model for prediction. During the prediction process, the 3DGCN module processes the input graph structure data and extracts geological spatial features through graph convolution operations. The 3DCNN module processes the input data through convolutional layers and fully connected layers to extract local features such as the geometric morphology and spatial distribution of geological bodies, as well as the relationships between numerical simulation variables. The feature results obtained by the two modules are fused to form a comprehensive feature, and finally the prediction result is obtained through the final classifier.

[0068] See also Figure 14 The following is a drill hole histogram and typical geological phenomena for the target area in the embodiment of the present invention. Drilling verification was conducted on the predicted target area with the highest mineralization probability, located in the northeastern part of the Chating deposit, on the northwest wing of the Jiulianshan-Liqiao anticline. The drill hole, located at the northwest end of the target area, demonstrated good predictive capability using the 3DGCN-CNN model-based three-dimensional mineralization prediction method. The drilling results were consistent with the predictions: the drill hole revealed a large-scale quartz diorite porphyry body, revealing alteration characteristics of skarn and porphyry deposits, and lead, zinc, and copper (gold) mineralization. Pyrite sericite is well-developed within the porphyry, with sulfide stockwork developed in some areas, including vein-like and disseminated chalcopyrite. Skarns are well-developed in the contact area between the rock body and the stratum, and numerous pyrite sulfide clusters containing copper, lead, zinc, and gold are visible.

[0069] The present invention directly extracts information from the geological model by fusing the 3DGCN and 3DCNN models, thereby eliminating the need for quantitative analysis of the 3D geological model in other 3D mineralization prediction methods and avoiding the problem of feature expression distortion caused by the processing process. It effectively integrates spatial features and local features, and introduces a spatial attention mechanism and an improved Inception module in the 3DCNN module, thereby improving the prediction ability of deep mineral exploration.

[0070] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0071] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional mineralization prediction method based on the 3DGCN-CNN model, characterized in that: The following steps are involved: Step S1: Collect geological data of the study area and construct a three-dimensional geological database; Step S2: Using geological data, construct a three-dimensional geological model of the study area and a three-dimensional ore body model of known ore deposits; Step S3: Each favorable metallogenic structure of the 3D geological model is directly used as a Boolean prediction variable, and is used together with the numerical simulation variables to construct a 3D metallogenic prediction information set; Step S4: constructing a 3DGCN-CNN model; including a 3DCNN module and a 3DGCN module, which are used to extract local features and spatial features respectively, and the two features are spliced ​​and fused and classified through a fully connected layer; the 3DCNN module is divided into a geological body processing part and a numerical simulation variable processing part, the geological body processing part adopts a ResNet18 architecture, and the numerical simulation variable processing part adopts a LeNet-5 architecture, and the geological body processing part and the numerical simulation variable processing part both combine a multi-scale module and a spatial attention mechanism, and the multi-scale module is integrated into the Inception module; the 3DGCN module models the spatial relationship of the geological body, and the input is the constructed graph structure data. The information of the first-order and second-order neighborhood is captured respectively through two layers of graph convolution layers, and the nonlinear transformation of the spatial features is realized through Batch Normalization and ReLU activation function; Step S5: Determine a positive sample data set and a negative sample data set, each subset in the data set is a unit block, and the positive and negative sample unit blocks are divided into a training set and a test set in a ratio of 8:2, and construct a 3DCNN data set and a 3DGCN graph structure data; when constructing the graph structure data for inputting the 3DGCN module, first extract the Boolean geological body attributes of each unit block, construct a node feature matrix, then construct an edge set of the graph, and use the normalized value of the Euclidean distance to represent the edge weight, respectively construct the feature matrix, edge set, edge weight and label of the training set and test set, and use the PyTorch Geometric Data data structure to encapsulate the graph structure data; when constructing the 3DCNN data set, take each positive sample unit block and negative sample unit block as the center, expand at least 4 blocks in each direction of the three-dimensional space to construct a large cubic unit, and form a three-dimensional cubic block containing the central unit and its neighborhood as the input sample; Step S6: Input the training set into the model for training; after each round of training, use the test set to adjust the model weights until the model converges; Step S7: Input the data of the area to be predicted into the trained model and output the prediction result through the fully connected layer.

2. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S1, the geological data includes a comprehensive interpretation profile of geological and geophysical data, a regional geological map, digital elevation data, a ore deposit profile, drilling data and physical property parameters of geological bodies in the study area.

3. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S2, the three-dimensional geological model is optimized using three-dimensional geophysical forward modeling and full lithologic inversion methods according to the physical property parameter values ​​of each geological body.

4. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S3, the large block model of the study area is discretized into unit blocks of equal size using the three-dimensional block model as a carrier, and each unit block is assigned a Boolean prediction variable and a numerical simulation variable. The variables assigned to all unit blocks together constitute a three-dimensional mineralization prediction information set; each favorable mineralization structure is a favorable mineralization rock body, a favorable mineralization stratum and a favorable mineralization fault; the Boolean prediction variables include the favorable mineralization rock body, the favorable mineralization stratum and the favorable mineralization fault, and the numerical simulation variables include the volume strain increment, the shear strain increment, the volume strain rate, the shear strain rate, the pore pressure, the maximum principal stress and the temperature.

5. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S3, the three-dimensional geological model is directly learned without performing spatial analysis on the three-dimensional geological model to convert it into a continuous prediction variable.

6. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S5, the positive sample data set is composed of the main mineralized unit blocks of the known ore deposits in the study area; the negative sample data set is constructed by applying a buffer zone to the drill point to indicate the average exploration spacing. For drill holes without ore deposits, the unit blocks within the range are used as negative sample unit blocks; for drill holes with ore deposits, the non-ore body unit blocks around the known ore bodies revealed within the range are used as negative sample unit blocks.

7. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S6, after each round of training is completed, the model is performance tested using the test set, and the prediction ability of the model is evaluated by calculating the cross entropy loss function value on the test set. The loss function values ​​of the model on the training set and the test set are recorded and visualized for analyzing the convergence of the training process, and the hyperparameter settings of the model, including learning rate decay and training rounds, are dynamically adjusted according to the changes in the loss function value of the test set. After multiple rounds of iteration, the trained model is obtained.

8. The three-dimensional mineralization prediction method based on the 3DGCN-CNN model according to claim 1, characterized in that: In step S7, the prediction results are assigned to the three-dimensional block model as the mineralization probability value, and the model prediction effect is evaluated by screening and visualizing blocks with high mineralization probability and comparing them with known mineral deposits.

Images