Polymetallic ore deposit prediction method based on three-dimensional geological modeling

Abstract

The application discloses a polymetallic ore mineralization prediction method based on three-dimensional geological modeling and belongs to the technical field of geological exploration. In view of the problems that the prior art is difficult to automatically extract deep mineralization characteristics and ignores the polymetallic paragenetic relationship, leading to low prediction accuracy, a multi-source heterogeneous three-dimensional voxel data model is constructed; a three-dimensional convolutional neural network is adopted to automatically extract the spatial form features of geological bodies; a multi-task learning mechanism is introduced, a shared feature extraction layer and independent prediction branches for different metal elements are set; and an attention mechanism is introduced in the independent prediction branches to capture the paragenetic and associated relationship between different metal elements; and finally, the collaborative positioning result and the resource quantity estimation result of the deep polymetallic ore body are output. The application effectively utilizes the polymetallic paragenetic law and significantly improves the prediction accuracy and reliability of the deep concealed ore body.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to a method for predicting the mineralization of polymetallic deposits based on three-dimensional geological modeling. Background Technology

[0002] Mineral resources are a vital material foundation of the national economy. With the increasing depletion of shallow resources, prospecting is gradually shifting towards deeper deposits and concealed ore bodies. Three-dimensional geological modeling technology, capable of visually displaying underground geological structures, has become an important tool for predicting deep mineralization.

[0003] Existing mineralization prediction methods are mostly based on geostatistics or traditional machine learning algorithms. Typically, these methods utilize borehole data, geophysical and geochemical data to construct three-dimensional models, and then perform feature extraction and mineralization regularity analysis through human-computer interaction. For example, the weight of evidence method or logistic regression method is used to delineate prospective mineralization areas.

[0004] However, the aforementioned existing technologies have certain limitations. On the one hand, traditional methods mainly rely on expert experience for manual interpretation, making it difficult to automatically and objectively extract the nonlinear spatial characteristics of deep geological bodies from massive amounts of multi-source heterogeneous data, resulting in insufficient ability to identify deep concealed ore bodies. On the other hand, existing technologies often make independent predictions for single metal elements, ignoring the complex symbiotic and associated relationships between multiple metal elements, and failing to effectively utilize the spatial correlation information between elements, thus limiting the accuracy and reliability of multi-metal ore collaborative prediction. Summary of the Invention

[0005] The main technical problem solved by this invention is that existing three-dimensional geological modeling techniques are unable to automatically extract deep mineralization features, and that the prediction accuracy is insufficient due to the neglect of the symbiotic relationship between polymetallic elements.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting polymetallic mineralization based on three-dimensional geological modeling, comprising the following steps: Acquire multi-source geological data, including borehole data, geophysical data, and geochemical data; The multi-source geological data is preprocessed and transformed into a three-dimensional voxel grid to construct a multi-source heterogeneous three-dimensional voxel data model. A three-dimensional convolutional neural network model is constructed, and the three-dimensional voxel data model is convolved using a three-dimensional convolutional kernel to automatically extract the spatial morphological features of the geological body. A multi-task learning mechanism is introduced, in which a shared feature extraction layer and independent prediction branches for different metal elements are set in the three-dimensional convolutional neural network; An attention mechanism is introduced into the independent prediction branch to capture the symbiotic and associated relationships between different metal elements; Based on the spatial morphological characteristics and the symbiotic and associated relationships, the collaborative positioning results and resource estimation results of deep polymetallic ore bodies are output.

[0007] Furthermore, the construction of the multi-source heterogeneous three-dimensional voxel data model includes: performing spatial interpolation processing on borehole data, performing inversion and standardization processing on geophysical and geochemical data, and mapping the processed data to a preset three-dimensional voxel grid space to adapt to the spatial structure of complex geological bodies.

[0008] Furthermore, the three-dimensional convolutional neural network includes multiple convolutional layers and pooling layers, and the three-dimensional convolutional kernel is used to slide in three-dimensional space to capture the nonlinear mineralization patterns of geological bodies.

[0009] Furthermore, the multi-task learning mechanism optimizes network parameters by minimizing the joint loss function of different metal element prediction tasks, enabling the shared feature extraction layer to learn common geological and mineralization features.

[0010] Furthermore, the attention mechanism is a spatial attention module, which is used to weight the spatial feature maps corresponding to different metal elements to enhance the regional features rich in mineralization information.

[0011] Furthermore, the attention mechanism is a channel attention module, which is used to model the correlation between different feature channels and automatically enhance feature channels that are sensitive to mineralization.

[0012] Furthermore, the resolution of the three-dimensional voxel grid is adaptively adjusted according to the exploration depth, wherein the voxel resolution in the shallow region is higher than that in the deep region.

[0013] Furthermore, the output of the collaborative positioning results and resource estimation results of deep polymetallic ore bodies includes: outputting the probability values ​​of each voxel location belonging to different metal mineralization types, and delineating the prospective mineralization areas by setting thresholds.

[0014] Furthermore, the three-dimensional convolutional neural network introduces nonlinearity using the ReLU activation function and reduces the feature dimension using a max pooling layer.

[0015] Furthermore, the preprocessing also includes spatial location calibration and lithological coding of borehole data, inversion processing of geophysical data to convert profile data into three-dimensional attribute volumes, and background removal and standardization of geochemical data.

[0016] The present invention has the following beneficial effects:

[0017] (1) Because the three-dimensional convolutional neural network is used to process the three-dimensional voxel grid, it can automatically mine the spatial features of the deep hidden ore body and capture the nonlinear mineralization law, thereby reducing the dependence on human experience and improving the objectivity of feature extraction.

[0018] (2) By introducing a multi-task learning architecture and attention mechanism, the symbiotic and associated laws among multi-metal elements are utilized to establish spatial correlations among different metal elements, thereby effectively improving the accuracy of multi-metal ore collaborative prediction.

[0019] (3) Because the data model is constructed by using voxelization, it can effectively adapt to the spatial structure of complex geological bodies and enhance the ability to process unstructured data.

[0020] (4) By adaptively adjusting the voxel resolution, the computational efficiency of deep large-scale prediction can be effectively improved while ensuring prediction accuracy.

[0021] (5) By introducing spatial attention or channel attention mechanisms, it is possible to adaptively enhance mineralization favorable areas or mineralization sensitive feature channels, thereby further improving the reliability of prediction. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0023] Example 1:

[0024] This embodiment provides a method for predicting polymetallic mineralization based on three-dimensional geological modeling, which specifically includes the following steps: Step 1: Obtain multi-source geological data.

[0025] The multi-source geological data includes borehole data, geophysical data, and geochemical data. Specifically, borehole columnar sections, core sampling analysis data (grade data), and geophysical data such as high-precision magnetic and gravity measurements, as well as soil geochemical measurement data, are collected within the study area. This embodiment uses a porphyry copper-molybdenum deposit as an example, selecting a 5km × 5km × 2km area in the mining zone, and collecting Cu, Mo, and Au grade data from 120 boreholes, as well as magnetic three-dimensional inversion data covering the area.

[0026] Step S: Preprocess the multi-source geological data to convert it into a three-dimensional voxel grid and construct a multi-source heterogeneous three-dimensional voxel data model.

[0027] First, the borehole data underwent spatial location calibration and lithological coding. Kriging interpolation was then used to perform 3D spatial interpolation on sparsely sampled points. Geophysical data was inverted, transforming the profile data into a 3D attribute volume. Geochemical data underwent background removal and standardization. Subsequently, the boundary range and voxel resolution of the 3D model were determined, for example, setting the voxel size to 25m × 25m × 25m. The processed attribute data was mapped to the corresponding voxel grid, forming a 3D voxel matrix containing multiple attribute information. Each voxel cell in this matrix contains data from multiple channels, including lithological coding, magnetic susceptibility values, and various elemental grades.

[0028] Step 3: Construct a three-dimensional convolutional neural network and use a three-dimensional convolutional kernel to perform convolution processing on the three-dimensional voxel data model to automatically extract the spatial morphological features of the geological body.

[0029] The three-dimensional convolutional neural network includes multiple convolutional layers and pooling layers, with the input being the three-dimensional voxel matrix constructed in step 2. In this embodiment, the network contains 5 convolutional layers with a kernel size of 3×3×3, a stride of 1, and a "same" padding method to preserve spatial dimensionality information. Nonlinearity is introduced through the ReLU activation function, and max-pooling layers (pooling window size of 2×2×2, stride of 2) are used to reduce feature dimensionality. The three-dimensional convolutional kernel slides in three-dimensional space, enabling it to capture the nonlinear mineralization patterns of geological bodies, such as the morphology of rock contact zones and the distribution of fault zones.

[0030] Step 4: Introduce a multi-task learning mechanism by setting up a shared feature extraction layer and independent prediction branches for different metal elements (such as copper, gold, and molybdenum) in the three-dimensional convolutional neural network.

[0031] The shared feature extraction layer is used to extract general geological mineralization features, such as fault structures and rock mass morphology; the independent prediction branch is used to predict the mineralization probability for specific metal elements. The multi-task learning mechanism optimizes network parameters by minimizing the joint loss function of different metal element prediction tasks, enabling the shared feature extraction layer to learn general geological mineralization features. In this embodiment, the joint loss function is the sum of the cross-entropy losses of each branch: , where N is the number of metal element types and Li is the prediction loss for the i-th element.

[0032] Step 5: Introduce an attention mechanism module into the independent prediction branch to capture the symbiotic and associated relationships between different metal elements.

[0033] This embodiment employs a spatial attention mechanism. Specifically, for a shared feature map, the spatial attention module generates a weight map with the same spatial size as the feature map through convolution operations. The values ​​of the weight map represent the importance of different spatial locations for mineralization prediction. Then, the weight map is multiplied element-wise with the original feature map to weight different spatial regions, strengthening the features of regions rich in mineralization information and suppressing background or non-mineralized areas. In this way, the model can automatically learn spatial co-occurrence patterns such as "molybdenum is often associated with high-grade copper areas."

[0034] Step 6: Based on the spatial morphological characteristics and the symbiotic and associated relationships, output the collaborative positioning results and resource estimation results of the deep polymetallic ore body.

[0035] The network outputs the probability values ​​of each voxel location belonging to different metal mineralization types, and delineates potential mineralization areas by setting a threshold (e.g., 0.7). The output results can be displayed in a three-dimensional visualization, intuitively presenting the spatial distribution and enrichment areas of polymetallic ore bodies such as copper, molybdenum, and gold.

[0036] To verify the effectiveness of this invention, an experiment was conducted using a porphyry copper-molybdenum deposit as an example. Parameter settings: voxel size was set to 25m × 25m × 25m, training set to validation set ratio was 8:2, initial learning rate was set to 0.001, and batch size was set to 16. Comparative experiments: Comparative Example 1 (traditional three-dimensional evidence weight method) and Comparative Example 2 (single-task 3D CNN model, without multi-task learning and attention mechanisms) were set. The results are shown in Table 1.

[0037] Table 1. Comparison of prediction accuracy of different methods on the validation set Comparative Example 1 (Three-Dimensional Evidence Rights Law) 72.5% 68.4% 0.70 Comparative Example 2 (Single Task 3D CNN) 81.2% 79.6% 0.80 Method of the present invention 89.3% 87.1% 0.88 As can be seen from Table 1, the method provided by this invention is significantly better than traditional methods and single-task models in terms of the accuracy of polymetallic ore body prediction, proving the effectiveness of multi-task learning and attention mechanisms in capturing element symbiotic relationships.

[0038] Example 2:

[0039] This embodiment is basically the same as embodiment 1, except for the type of attention mechanism in step 5.

[0040] In this embodiment, the attention mechanism module employs a channel attention mechanism. The channel attention mechanism focuses on the correlation between different feature channels (such as different geophysical or geochemical properties). Specifically, it involves global average pooling of the shared feature map to obtain a feature vector with one channel dimension. Then, it learns the weights of each channel through two fully connected layers (with dimensionality reduction and expansion operations in between). Finally, the weights are multiplied by the corresponding channels in the original feature map. In this way, the model can automatically enhance feature channels sensitive to mineralization (such as channels corresponding to areas with abnormally high copper values) and suppress irrelevant or interfering feature channels. This embodiment is particularly suitable for scenarios with high noise in geochemical data. Experiments show that its prediction accuracy is comparable to that of Embodiment 1, and its robustness to outliers is superior.

[0041] Example 3: This embodiment is basically the same as embodiment 1, except that the resolution of the three-dimensional voxel mesh is set in step 2.

[0042] In this embodiment, the resolution of the three-dimensional voxel grid is adaptively adjusted according to the exploration depth. Specifically, for shallow areas (0-500m), the voxel size is set to 10m×10m×10m to finely depict the morphology of shallow ore bodies; for deep areas (below 500m), the voxel size is set to 50m×50m×50m to reduce computation and expand the prediction range for deep areas. In practical implementation, before constructing the voxel model, multiple voxel grid layers with different resolutions can be generated according to the depth range, and then features can be extracted separately. Finally, the multi-scale features are fused and input into the subsequent network. This adaptive multi-resolution voxelization strategy can effectively improve data processing efficiency while ensuring prediction accuracy, and is especially suitable for large-area deep exploration projects.

[0043] Example 4: This embodiment is basically the same as Embodiment 1, except for a more detailed explanation of the ReLU activation function, max pooling, and preprocessing.

[0044] ReLU activation function and max pooling: In the 3D convolutional neural network of this embodiment, each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer. The ReLU activation function is expressed as f(x) = max(0,x), which can effectively alleviate the gradient vanishing problem and accelerate network convergence. The pooling layer uses max pooling with a pooling window size of 2×2×2 and a stride of 2, which reduces the spatial dimension of the feature map and the number of parameters while preserving significant features.

[0045] Preprocessing details: This embodiment provides a more detailed definition of the preprocessing steps. Specifically, it includes: For borehole data: First, spatial coordinate calibration is performed (converted to a unified projected coordinate system), then lithology is coded (e.g., granite is coded as 1, sandstone as 2, etc.), and finally, ordinary kriging is used to perform three-dimensional grid interpolation on the grade data.

[0046] For geophysical data: perform three-dimensional inversion on magnetic, gravity and other profile data to obtain physical property parameters (such as magnetic susceptibility and density) corresponding to each voxel location in the subsurface.

[0047] For geochemical data: First, outliers obviously caused by surface pollution are removed. Then, logarithmic transformation is performed to make its distribution close to a normal distribution. Finally, Z-score standardization is used to unify the dimensions of each element.

[0048] The above preprocessing steps ensure that multi-source data can be effectively fused in the same voxel space, providing high-quality input for subsequent neural networks.

[0049] Example 5: This embodiment provides a specific network training parameter configuration to support the aforementioned joint loss function optimization.

[0050] In this embodiment, the three-dimensional convolutional neural network has a total of 12 layers (5 convolutional layers, 5 pooling layers, and 2 fully connected layers). The shared feature extraction layer consists of the first 4 convolutional layers and the corresponding pooling layers. The independent prediction branch consists of the 5th convolutional layer and two fully connected layers. The number of output nodes in each branch is equal to the number of predicted categories for the metal element (e.g., 3 categories: rich ore, poor ore, no ore). The joint loss function is a weighted sum of the loss functions of each branch. The weights wi can be dynamically adjusted based on the number of samples for each element (e.g., using inverse frequency weighting). The optimizer used is Adam, with an initial learning rate of 0.001, which decays to 0.9 times the original rate every 10 epochs. Early stopping is employed during training to prevent overfitting. After the above training, the model achieves the excellent performance shown in Table 1 of Example 1 on the validation set.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting polymetallic mineralization based on three-dimensional geological modeling, characterized in that, include: Acquire multi-source geological data, including borehole data, geophysical data, and geochemical data; The multi-source geological data is preprocessed and transformed into a three-dimensional voxel grid to construct a multi-source heterogeneous three-dimensional voxel data model. A three-dimensional convolutional neural network model is constructed, and the three-dimensional voxel data model is convolved using a three-dimensional convolutional kernel to automatically extract the spatial morphological features of the geological body. A multi-task learning mechanism is introduced, in which a shared feature extraction layer and independent prediction branches for different metal elements are set in the three-dimensional convolutional neural network; An attention mechanism is introduced into the independent prediction branch to capture the symbiotic and associated relationships between different metal elements; Based on the spatial morphological characteristics and the symbiotic and associated relationships, the collaborative positioning results and resource estimation results of deep polymetallic ore bodies are output.

2. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The construction of the multi-source heterogeneous three-dimensional voxel data model includes: performing spatial interpolation processing on borehole data, performing inversion and standardization processing on geophysical and geochemical data, and mapping the processed data to a preset three-dimensional voxel grid space.

3. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The three-dimensional convolutional neural network includes multiple convolutional layers and pooling layers, and the three-dimensional convolutional kernel is used to slide in three-dimensional space to capture the nonlinear mineralization patterns of geological bodies.

4. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The multi-task learning mechanism optimizes network parameters by minimizing the joint loss function of different metal element prediction tasks, enabling the shared feature extraction layer to learn common geological and mineralization features.

5. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The attention mechanism is a spatial attention module, which is used to weight the spatial feature maps corresponding to different metal elements and enhance the regional features rich in mineralization information.

6. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The attention mechanism is a channel attention module, which is used to model the correlation between different feature channels and automatically enhance feature channels that are sensitive to mineralization.

7. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The resolution of the three-dimensional voxel grid is adaptively adjusted according to the exploration depth, with the voxel resolution in shallow regions being higher than that in deep regions.

8. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The output of the collaborative positioning results and resource estimation results of deep polymetallic ore bodies includes: outputting the probability values ​​of each voxel location belonging to different metal mineralization types, and delineating the prospective mineralization areas by setting thresholds.

9. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The three-dimensional convolutional neural network introduces nonlinearity using the ReLU activation function and reduces feature dimensionality using max pooling layers.

10. The polymetallic mineralization prediction method based on three-dimensional geological modeling according to claim 1, characterized in that, The preprocessing also includes spatial location calibration and lithological coding of borehole data, inversion processing of geophysical data to convert profile data into three-dimensional attribute volumes, and background removal and standardization of geochemical data.