Geochemical halo ore-prospecting method based on graph neural network

Abstract

The application discloses a geochemical halo ore prospecting method based on a graph neural network and relates to the field of mineral prediction, which comprises the following steps: acquiring interpolation grid data of element combinations in geological sample data; taking pixel points with effective data as nodes of an undirected graph to construct a global spatial adjacency matrix of the undirected graph; determining a primary halo zoning sequence according to a geological mineralization rule, extracting element abundance values corresponding to the nodes, and constructing an initial node feature matrix; based on the initial node feature matrix, calculating correlation coefficients of two elements in the element combinations to construct a symmetric correlation matrix and normalizing the symmetric correlation matrix to obtain a local primary halo feature collaborative tensor; constructing a graph convolution network, calculating features of the graph convolution network based on the initial node feature matrix, the global spatial adjacency matrix and the local primary halo feature collaborative tensor; and based on the trained graph convolution network, predicting minerals in a study area. The application solves the problem that a mineral prediction method based on geochemical halos lacks geophysical constraints.

Description

Technical Field

[0001] This invention relates to the field of mineral prediction technology, and in particular to a geochemical halo prospecting method based on graph neural networks. Background Technology

[0002] Mineral potential prediction aims to delineate metallogenic potential areas and determine priorities for mineral exploration by analyzing and integrating relevant geological exploration data (such as geochemical, geophysical, and remote sensing data). This plays a crucial guiding role in reducing exploration risks and improving the success rate of mineral discovery. Given the current surge in demand for strategic mineral resources and the increasing depletion of easily identifiable shallow deposits, the quantitative prediction and assessment of deep and concealed deposits in overly covered areas has become a core task in the field of geological exploration.

[0003] In recent years, data-driven algorithms based on artificial intelligence have sparked a new wave of quantitative assessment of mineral resources. At its core, machine learning methods (such as support vector machines and random forests) can automatically analyze and mine the complex nonlinear relationships between massive amounts of exploration data and known mineral deposits. With the improvement of computing power, deep learning methods (such as convolutional neural networks (CNNs) and graph neural networks (GCNs) have shown great potential in the field of mineral potential prediction due to their superior high-dimensional feature extraction and complex spatial pattern recognition capabilities, effectively overcoming the limitations of traditional machine learning that relies on manually designed features. However, existing deep learning models that consider spatial structure have a fundamental underlying architectural flaw when applied to geochemical anomaly identification and mineral prediction: the "implicit learning" of the spatial symbiotic relationships of geochemical elements is treated as a black box, failing to effectively integrate and explicitly express the a priori mineralization laws of the "geochemical primary halo."

[0004] In real hydrothermal mineralization systems, the migration and precipitation of primary ore-forming elements and associated indicator elements are controlled by specific physicochemical conditions, forming geochemical primary halos with strict spatial zoning. Taking typical low-to-medium temperature hydrothermal disseminated deposits (such as Carlin-type gold deposits) as an example, their primary ore-forming temperature is usually 140–250℃, and the ore-forming fluids are rich in H2S, CO2, and CH4. Their mineralization system dictates that only elements that are sulfide-loving, low-temperature, easily migrating, and co-precipitate with sulfides can form stable geochemical anomalies. Long-term geological research and exploration practice have shown that Au, As, Sb, and Hg are the core indicator element combinations for this type of gold deposit, forming a complete zoning structure in space: Hg (long-range halo) – Sb (front halo) – As (near-ore halo) – Au (ore body). This symbiotic combination of multiple elements and its dynamic spatial evolution (i.e., spatial non-stationarity) is the core basis for geochemical prospecting of mineral deposits. However, in existing deep neural network architectures, different geological and geochemical elements (such as Au, As, Sb, and Hg) are often simply stacked as independent "feature channels" input into the network. The network model relies entirely on the weight matrix to "implicitly" fit the relationships between elements in a black box. This purely mathematical approach isolates the objectively existing mineralization synergies between elements, causing the network to be unable to perceive the leading-edge halo features that "predict a deep, concealed main ore body due to abnormally high levels of associated indicator elements (such as As and Sb)" when extracting features. Under limited exploration sample conditions, this fitting, lacking constraints from geophysical mechanisms, is easily interfered with by background noise, resulting in predicted prospecting targets that often violate geological common sense and lack reliable geological interpretability. Summary of the Invention

[0005] The purpose of this invention is to address the lack of geophysical constraints in existing mineral prediction methods based on geochemical primary halos, and to propose a geochemical halo-based mineral exploration method using graph neural networks, comprising the following steps: S1. Determine the indicator element combination of minerals and obtain the interpolated raster data of element combinations in geological sample data; S2. Using the pixels with valid data as nodes of the undirected graph, construct the sparse adjacency matrix of the undirected graph to obtain the global adjacency matrix. Based on the geological mineralization law, the primary halo zoning sequence of mineral indicator element combinations is determined, and the element abundance values ​​corresponding to each node are extracted according to the zoning sequence order to construct the initial node feature matrix. S3. Based on the initial node feature matrix, calculate the correlation coefficient between pairs of elements in the element combination, extract the absolute value of the effective correlation coefficient, construct a symmetric correlation matrix and normalize it to obtain the local native halo feature co-tensor. S4. Construct a graph convolutional network. Calculate the features of the graph convolutional network based on the initial node feature matrix, the global spatial adjacency matrix, and the local native halo feature co-current tensor. S5. Based on the trained graph convolutional network, mineral resources are predicted in the study area.

[0006] Furthermore, the sparse adjacency matrix of the undirected graph is constructed using the eight-neighbor rule, and self-connections are introduced.

[0007] Furthermore, a moving window of size n×n is set, and for any node, the corresponding band sequence within the window is extracted, and the correlation coefficient between pairs of elements in the combination of band sequence elements is calculated.

[0008] Furthermore, The correlation coefficient between any two elements in a combination of elements is calculated using the following formula:

[0009] in, This represents the Pearson correlation coefficient between elements u and v within the local window of node i; n is the number of effective pixels within the window. and These represent the abundance values ​​of elements u and v at the m-th pixel within the window, respectively. and These are the average abundance values ​​of elements u and v within the window, respectively; The symmetric correlation matrix is ​​normalized using the following formula:

[0010] in, This represents the value in the j-th row and k-th column of the local native halo feature co-tensor of node i after normalization; represents the absolute correlation coefficient value in the j-th row and k-th column of the symmetric correlation matrix of node i before normalization; c is the column index; m indicates the number of elements in the element combination; To prevent constants with a denominator of zero; After traversing all nodes, the local native halo feature co-current tensor is generated by splicing them together.

[0011] Furthermore, the formula for calculating the features of the first layer of the graph convolutional network is as follows:

[0012] in, This is the hidden layer feature matrix output by the first layer graph convolution; It is a non-linear activation function; This is the initial feature tensor after adding dimensions; This is a batch matrix multiplication operator; For dimensionality reduction operators; This is the learnable weight matrix of the first layer of the network; This is the global adjacency matrix; The features of a graph convolutional network are calculated using local feature tensors; The formulas for calculating the features of layers 2 through L of a graph convolutional network are as follows:

[0013] in, and The first and the The hidden layer feature matrix output by the layer graph convolution; For the first The layer learnable weight matrix, where L is the number of graph convolutional layers; The formula for calculating the output layer of a graph convolutional network is:

[0014] Where Y is the class prediction probability matrix for all nodes, and softmax is the softmax function.

[0015] Furthermore, the loss function during training of the graph convolutional network is:

[0016] in, The loss value for training the graph convolutional network; is the total number of labeled samples participating in the calculation; i is the sample index; c is the category index, where 1 represents mineralization and 2 represents non-mineralization; One-hot encoding of the true label of node i belonging to category c; The predicted probability that node i belongs to category c in the network output; The L2 regularization coefficient; This represents all trainable weight parameters in the network.

[0017] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described geochemical halo prospecting method based on graph neural networks.

[0018] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described geochemical halo prospecting method based on graph neural networks.

[0019] The present invention also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described geochemical halo prospecting method based on graph neural networks.

[0020] The beneficial effects of the technical solution provided by this invention are: This invention constructs an undirected graph using interpolated raster data of element combinations as nodes. Based on the spatial zoning evolution law of multi-element elements, it constructs a node feature matrix according to the element abundance values ​​corresponding to each node, and obtains a global spatial adjacency matrix. It uses the absolute value of the correlation coefficient between pairs of elements as a measure of the cooperative strength between elements, and obtains a local primary halo feature cooperative tensor. Based on the node feature matrix, the global spatial adjacency matrix, and the local feature tensor, it calculates the features of the graph convolutional network. The constrained graph convolutional network performs calculations according to geological mineralization laws, transforming the local symbiotic relationships between elements into local spatial correlation constraints that can be computed by the neural network. This invention overcomes the problem of existing mineral prediction methods based on geochemical primary halos lacking geophysical constraints, improving the geoscientific rationality and hit rate of hidden mineral prediction. Attached Figure Description

[0021] Figure 1 This is a flowchart of a geochemical halo prospecting method based on graph neural networks, according to an example of the present invention. Figure 2 This is a map showing the actual distribution of Au (gold), As (arsenic), Sb (antimony), and Hg (mercury) in the study area; Figure 3 This is a probability distribution map of ore deposit prediction corresponding to the traditional semi-supervised model (Standard GCN) and the method of this invention; Figure 4 This is a comparison chart of the ROC curves of the method of this invention and the traditional semi-supervised model; Figure 5 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] A flowchart of a geochemical halo prospecting method based on graph neural networks, as an example of this invention, is shown below. Figure 1 Specifically, it includes: S1. Based on the deposit types and regional metallogenic theories of the study area, indicator element combinations closely related to mineralization are screened. Geological samples (rocks, soils, stream sediments, etc.) are collected, and multi-element content data are obtained through laboratory analysis. Data cleaning and outlier processing are then performed. Based on this, continuous spatial interpolation raster data are generated for each indicator element using methods such as Kriging interpolation or inverse distance weighting. In this invention, four original geochemical interpolation raster data for Au (gold), As (arsenic), Sb (antimony), and Hg (mercury) of the study area are obtained, with a raster size of 250×403. Figure 2As shown. Invalid values ​​were filtered out, and pixels with valid data were converted into nodes of an undirected graph, resulting in a total of N = 66939 valid nodes.

[0024] S2. Map the pixels with valid observations in the raster to nodes in the undirected graph to be constructed. Each node corresponds to a spatial location, and its geographic coordinates or raster row and column numbers are used to determine the spatial adjacency relationship between nodes. Based on the spatial proximity relationship between nodes, construct the edges and sparse adjacency matrix of the undirected graph to obtain the global spatial adjacency matrix.

[0025] Specifically, based on the undirected graph structure, the spatial connectivity between nodes is defined using the K-Nearest Neighbors (KNN) or Delaunay triangulation method, generating a sparse adjacency matrix. Self-connections are introduced into this matrix to obtain a global spatial adjacency matrix. This matrix reflects the spatial proximity and connectivity of each sampling point within the study area.

[0026] In this invention, based on the physical spatial coordinates of the aforementioned 66,939 nodes, a sparse adjacency matrix A of the undirected graph is constructed using the eight-neighborhood rule. To stabilize the graph information transmission mechanism, self-connections are introduced into the matrix, and symmetric normalization is applied to obtain the global spatial adjacency matrix. The calculation formula is as follows:

[0027] Where I is the identity matrix, D is the degree matrix, and its diagonal elements are... .

[0028] Based on geological mineralization regularities and the primary halo zoning patterns of typical mineral deposits, the axial / vertical primary halo zoning sequence of indicator element assemblages is determined. Following the order of this zoning sequence, element abundance values ​​are extracted from the interpolation raster corresponding to each node to form an initial node feature matrix. In this invention, the primary halo zoning sequence of mineral indicator element assemblages is arranged according to the channel order [Au, As, Sb, Hg], and the element abundance values ​​corresponding to each node are extracted to construct the initial node feature matrix. , where 66939 represents the number of valid nodes, and 4 represents the four element channels Au, As, Sb, and Hg.

[0029] S3. Based on the initial node feature matrix obtained in S2, calculate the local correlation coefficients of each pair of elements in the element combination, and construct the local native halo feature co-current tensor.

[0030] Based on the same initial node feature matrix, the correlation coefficients between pairs of elements in the indicator element combination are calculated. The absolute value of the correlation coefficient is taken as the measure of the cooperative strength between elements, and a symmetric correlation matrix is ​​constructed. The correlation matrix is ​​normalized and extended into a local native halo feature cooperative tensor to characterize the internal cooperative variation features of element combinations at different spatial locations.

[0031] In this invention, the size is set as A moving window (covering 49 pixels) is used. For any node i, the corresponding sequence within the window is extracted, and the local Pearson correlation coefficients of the six pairs of elements (Au-As, Au-Sb, Au-Hg, As-Sb, As-Hg, and Sb-Hg) are calculated sequentially using the following formula. :

[0032] in, This represents the Pearson correlation coefficient between elements u and v within the local window of node i; n is the number of effective pixels within the window (in this embodiment). ); and These represent the abundance values ​​of elements u and v at the m-th pixel within the window, respectively. and These are the average abundance values ​​of elements u and v within this window, respectively.

[0033] Extract the absolute value of the effective correlation coefficient and initialize a new one for each node. Symmetric correlation matrix Its diagonal elements are set to .

[0034] To prevent gradient explosion during deep network training, row normalization is performed along the column direction of the matrix using the following formula:

[0035] in, This represents the value in the j-th row and k-th column of the local constraint tensor of node i after normalization; represents the absolute correlation coefficient value in the j-th row and k-th column of the symmetric correlation matrix of node i before normalization; c is the column index, and the value of c ranges from 1 to 4, representing the four element channels Au, As, Sb, and Hg respectively. To prevent extremely small constants with a denominator of zero, the value is taken as... After traversing all 66,939 nodes, the final local native halo feature co-tensor is generated by concatenating the data. .

[0036] S4. Construct a graph convolutional network by calculating the features of the graph convolutional network based on the initial node feature matrix, the global spatial adjacency matrix, and the local feature tensor.

[0037] To further adapt to subsequent matrix operations on the local native halo feature co-coordinated tensor, the tensor dimension of the initial node feature matrix is ​​expanded. Specifically, let the number of effective nodes be N and the number of indicator elements be M, the initial node feature matrix is ​​represented as follows: Each row corresponds to the M abundance values ​​of a node. The initial node feature matrix is ​​expanded by adding one dimension along the feature column direction to obtain the expanded initial feature tensor. This dimension expansion operation does not change the number of nodes or the number of element channels. Instead, it represents the element abundance vector of each node as a column vector, enabling batch matrix multiplication with the corresponding local native halo feature co-matrix. In this invention, the initial feature tensor after dimension expansion...

[0038] Construct a graph convolutional network. Then, expand the feature tensor by... Global adjacency matrix and local feature tensor Substitute the following formula to calculate the first layer of convolution:

[0039] in, This is the hidden layer feature matrix output by the first layer graph convolution; It is a non-linear activation function; This is the initial feature tensor after adding dimensions; This is a batch matrix multiplication operator; For dimensionality reduction operators; This is the learnable weight matrix for the first layer of the network.

[0040] The physical meaning of this operation is as follows: before aggregating spatial neighbors, the Au feature of each node is forced to undergo geological law coordination and weighted correction based on its local affinity with As, Sb, and Hg. The global spatial adjacency matrix is ​​used to propagate spatial neighborhood information; the local native halo feature coordination tensor is used to fuse the intrinsic correlation of element combinations. By aggregating the information of nodes and their neighbors layer by layer through graph convolutional layers, multi-scale feature extraction from local to global is achieved, and finally, a high-order embedding representation of each node is output to characterize its mineralization potential.

[0041] First layer graph convolution output Subsequently, the subsequent graph convolutional layers no longer directly operate on the original four element channels (Au, As, Sb, Hg). Instead, they use the hidden layer feature matrix output from the previous layer as input to continue spatial neighborhood aggregation based on the global spatial adjacency matrix. For the ... The subsequent graph convolutional layer is calculated using the following formula:

[0042] in, For the first The hidden layer feature matrix output by the layer graph convolution; For the first Layer learnable weight matrix.

[0043] The calculation formula for the final output layer is:

[0044] Where Y is the class prediction probability matrix for all nodes, and softmax is the softmax function.

[0045] S5. Supervised training of the graph convolutional network is performed using known mineral deposits (points) and negative samples to optimize model parameters. After training, all nodes in the study area are input into the trained model to output mineral predictions.

[0046] During the model training phase, 70 known gold deposits in the exploration area were extracted as positive samples (label 1), and 70 areas with extremely low background were randomly selected as negative samples (label 0) to construct a semi-supervised mask. The negative log-likelihood loss is calculated using the following formula:

[0047] in, This represents the total loss value during network training. is the total number of labeled samples participating in the calculation; i is the sample index; c is the category index (1 represents mineralization, 2 represents non-mineralization); One-hot encoding of the true label of node i belonging to category c; The predicted probability that node i belongs to category c in the network output; The L2 regularization coefficient has a value of [value missing]. ; This represents all trainable weight parameters in the network. The Adam algorithm is used to optimize the parameters, and the training is performed iteratively 1000 times.

[0048] After convergence, the mineralization probabilities of its 66,939 output nodes are extracted. Combining the coordinate system and affine transformation parameters of the original geochemical raster, the one-dimensional probability array is mapped back to its original form. The system generates a two-dimensional GeoTIFF prediction map. It also outputs the actual latitude and longitude or geodetic coordinates of high-probability target areas, forming a target area optimization data table.

[0049] This invention uses the PyTorch deep learning framework to conduct mineralization prediction tests in the study area and compares and evaluates them with traditional semi-supervised models that do not explicitly embed native halo features. To visually demonstrate the prediction results, Figure 3 The paper presents the probability distribution diagrams of ore deposit predictions for the traditional semi-supervised model (Standard GCN) and the method of this invention. Figure 3 The left side shows the probability distribution of ore deposit predictions corresponding to the traditional semi-supervised model, while the right side shows the probability distribution of ore deposit predictions corresponding to the method of this invention. To further quantify and evaluate the model's ability to distinguish between ore deposit and non-ore deposit samples, Figure 4A comparison of ROC curves for the two methods is presented, where the AUC of the method of this invention is 0.9971, and the AUC of the traditional method is 0.9571.

[0050] from Figure 3 As can be seen, when using the traditional Standard GCN, because the model treats Au, As, Sb, and Hg as independent channels for implicit feature fusion, it cannot explicitly perceive the geological "leading halo" correlation features. This results in insufficient spatial continuity and geological orientation of high-probability anomaly areas in the prediction results, and some areas are prone to noise-driven discrete anomaly responses. Especially in Figure 3 The concealed ore area indicated by the black box shows no significant Au anomaly, but a clear Hg anomaly (see...). Figure 2 Traditional Standard GCN struggles to adequately identify deep mineralization information reflected by leading-edge halo indicators such as Hg in weak Au response regions, resulting in relatively weak mineralization prediction responses in these areas. In contrast, when employing the algorithm of this invention, the explicit constraint of the graph convolutional layer's feature generation on the strong positive correlation between As and Sb in the leading-edge halo features of this region allows the network to automatically enhance the mineralization information supported by indicators such as As, Sb, and Hg in the weak Au region during the local batch matrix calculation stage. This leads to stronger consistency between predicted high-value areas and multi-element primary halo combination anomalies, demonstrating that explicitly embedding primary halo constraints helps improve the model's ability to identify concealed ore bodies. Figure 4 As can be seen, the ROC curve of the method of this invention is closer to the upper left corner, demonstrating a stronger ability to distinguish between positive and negative samples. Specifically, the AUC value of the traditional method is 0.9571, while the AUC value of the method of this invention is improved to 0.9971. This result indicates that after explicitly embedding the synergistic relationship of native halo elements, the model can more fully utilize the local correlation structure among Au, As, Sb, and Hg, maintaining a higher deposit identification ability under low false alarm conditions, thereby improving the stability and reliability of the prediction results.

[0051] The experimental results above demonstrate that this invention not only effectively overcomes the overfitting problem under limited sample conditions but also possesses strong geological interpretability. By explicitly injecting the spatial combination law of multi-element mineralization, it significantly improves the adaptability and accuracy of deep concealed mineral prediction in complex geological backgrounds.

[0052] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the above-described geochemical halo prospecting method based on graph neural networks.

[0053] Please see Figure 5In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0054] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned geochemical halo prospecting method based on graph neural networks.

[0055] In one exemplary embodiment, a computer program product is proposed, comprising a computer program / instructions that, when executed by a processor, implement the steps of the above-described graph neural network-based geochemical halo prospecting method.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A geochemical halo prospecting method based on graph neural networks, characterized in that, Includes the following steps: S1. Determine the indicator element combination of minerals and obtain the interpolated raster data of element combinations in geological sample data; S2. Using the pixels with valid data as nodes of the undirected graph, construct the sparse adjacency matrix of the undirected graph to obtain the global adjacency matrix. Based on the geological mineralization law, the primary halo zoning sequence of mineral indicator element combinations is determined, and the element abundance values ​​corresponding to each node are extracted according to the zoning sequence order to construct the initial node feature matrix. S3. Based on the initial node feature matrix, calculate the correlation coefficient between pairs of elements in the element combination, extract the absolute value of the effective correlation coefficient, construct a symmetric correlation matrix and normalize it to obtain the local native halo feature co-tensor. S4. Construct a graph convolutional network. Calculate the features of the graph convolutional network based on the initial node feature matrix, the global spatial adjacency matrix, and the local native halo feature co-current tensor. S5. Based on the trained graph convolutional network, mineral resources are predicted in the study area.

2. The geochemical halo prospecting method based on graph neural networks according to claim 1, characterized in that, We construct a sparse adjacency matrix for an undirected graph using the eight-neighbor rule and introduce self-connections.

3. The geochemical halo prospecting method based on graph neural networks according to claim 1, characterized in that, Set a moving window of size n×n. For any node, extract the corresponding zone sequence within the window and calculate the correlation coefficient between pairs of elements in the combination of zone sequence elements.

4. The geochemical halo prospecting method based on graph neural networks according to claim 3, characterized in that, The correlation coefficient between any two elements in a combination of elements is calculated using the following formula: in, This represents the Pearson correlation coefficient between elements u and v within the local window of node i; n is the number of effective pixels within the window. and These represent the abundance values ​​of elements u and v at the m-th pixel within the window, respectively. and These are the average abundance values ​​of elements u and v within the window, respectively; The symmetric correlation matrix is ​​normalized using the following formula: in, This represents the value in the j-th row and k-th column of the local native halo feature co-tensor of node i after normalization; represents the absolute correlation coefficient value in the j-th row and k-th column of the symmetric correlation matrix of node i before normalization; c is the column index; m indicates the number of elements in the element combination; To prevent constants with a denominator of zero; After traversing all nodes, the local native halo feature co-current tensor is generated by splicing them together.

5. The geochemical halo prospecting method based on graph neural networks according to claim 1, characterized in that, The formula for calculating the features of the first layer of a graph convolutional network is as follows: in, This is the hidden layer feature matrix output by the first layer graph convolution; It is a non-linear activation function; This is the initial feature tensor after adding dimensions; This is a batch matrix multiplication operator; For dimensionality reduction operators; This is the learnable weight matrix of the first layer of the network; This is the global adjacency matrix; The features of a graph convolutional network are calculated using local feature tensors; The formulas for calculating the features of layers 2 through L of a graph convolutional network are as follows: in, and The first and the The hidden layer feature matrix output by the layer graph convolution; For the first The layer learnable weight matrix, where L is the number of graph convolutional layers; The formula for calculating the output layer of a graph convolutional network is: Where Y is the class prediction probability matrix for all nodes, and softmax is the softmax function.

6. The geochemical halo prospecting method based on graph neural networks according to claim 1, characterized in that, The loss function for training a graph convolutional network is: in, The loss value for training the graph convolutional network; is the total number of labeled samples participating in the calculation; i is the sample index; c is the category index, where 1 represents mineralization and 2 represents non-mineralization; One-hot encoding of the true label of node i belonging to category c; The predicted probability that node i belongs to category c in the network output; The L2 regularization coefficient; This represents all trainable weight parameters in the network.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, The device includes a processor and a memory, the processor being interconnected with the memory, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to perform the method as described in any one of claims 1 to 6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the method described in any one of claims 1 to 6.

Images