A method for inferring the migration patterns of chemical elements in a geochemical field

Abstract

This invention relates to the field of geological modeling and spatial analysis technology, specifically disclosing a method for inferring the migration patterns of chemical elements in a geochemical field. Employing a data-driven approach, it directly learns migration patterns from geochemical data. Compared to traditional numerical simulation methods that rely on precise hydrogeological parameters, this method reduces the dependence on the accuracy of physical models and parameters. Furthermore, this invention models the spatial topological relationships between sampling points using a GAT network, enabling the capture of complex spatial structural relationships, dynamic adjustment of the influence of different neighboring nodes, and clear identification of branching and converging features of migration paths. This effectively characterizes the nonlinear dynamic processes in the migration of chemical elements in a geochemical field.

Description

Technical Field

[0001] This application relates to the field of geological modeling and spatial analysis technology, and specifically discloses a method for inferring the migration pattern of chemical elements in a geochemical field. Background Technology

[0002] Inferring the spatial distribution characteristics of geochemical fields and their material migration pathways is a core scientific issue in fields such as mineral resource exploration, environmental pollutant tracing, groundwater cycle research, and geological disaster early warning. Accurately characterizing the migration patterns of chemical elements in underground media, such as rocks, soil, and water bodies, within the study area is of vital importance for understanding surface chemical anomalies and hydrogeochemical cycles.

[0003] Currently, the mainstream methods for inferring chemical element migration patterns in geochemical fields mainly come from the following three technical fields: (1) In the field of geostatistics, based on spatial autocorrelation theory, spatially discrete sampling points are regarded as regional variables. The spatial correlation structure of the study area is described by the variogram, and continuous and smooth geochemical field distribution maps are generated by spatial interpolation algorithms, such as kriging and its variants. The pattern inference of this type of method is often implicit in the gradient changes or anomalous patterns of chemical element content, and it requires manual interpretation.

[0004] (2) In the field of physical or chemical numerical simulation, based on fluid dynamics, solute transport theory and chemical reaction equations, a mathematical model of material migration in the underground medium within the study area is constructed and solved using the finite element method or finite volume method. The main steps of this type of method are to establish a set of partial differential equations describing the groundwater flow field, solute transport and chemical reaction. Under given initial conditions, boundary conditions and parameter fields such as permeability, porosity, dispersion coefficient, reaction rate, etc., the migration and evolution of materials over time and space is simulated through numerical calculation, and the distribution of chemical element content and possible transport trajectories are output: streamlines or particle tracking. This type of method requires detailed geological structure, hydrogeological parameters and boundary conditions as input.

[0005] (3) In the field of empirical geological analysis and path tracing, it relies on the knowledge and experience of geological experts. It requires geological experts to comprehensively study the geological background, structural features, lithological combinations, alteration zoning, known mineralization information, etc. of the research area, and combine the spatial distribution of geochemical anomalies, such as anomaly trends and concentration centers. They should use geological theories and empirical rules to subjectively judge and delineate the material migration pattern. Usually, it is necessary to draw manually or use simple spatial analysis tools, such as flow direction analysis, to complete the inference.

[0006] The aforementioned existing technical solutions all have significant shortcomings. They struggle to effectively integrate multi-source heterogeneous data, lack the ability to capture complex spatial correlations and nonlinear dynamics, and lack explicit model exploration capabilities. Specifically, field distribution maps generated based on geostatistics mainly reflect the spatial distribution of chemical elements, failing to clearly and quantitatively reveal the specific trajectories of material migration, source-sink relationships, and the topological structure of channels, such as branching and convergence. Model inference heavily relies on subjective human interpretation, easily leading to ambiguity. Numerical simulation methods heavily depend on precise geological and hydrogeological conceptual models and spatial parameters, such as permeability and dispersion coefficients, which are difficult to obtain accurately. Model construction is complex, boundary conditions are difficult to set, the calculation process is usually extremely time-consuming, and the simulation results are extremely sensitive to uncertainties in input parameters and model structure, exhibiting poor robustness in data-sparse areas or complex geological conditions. Empirical geological analysis, on the other hand, is highly subjective, lacks repeatability and universality, has a weak ability to handle complex nonlinearities, and struggles to effectively handle massive, high-dimensional geochemical data and the complex spatial patterns and nonlinear relationships they contain.

[0007] This invention provides a method for inferring the migration patterns of chemical elements in a geochemical field, in order to solve the above-mentioned problems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for inferring the migration patterns of chemical elements in a geochemical field. A graph model is constructed, and the chemical element content of all nodes is predicted separately. The edge probabilities are obtained by a self-supervised learning method. After joint optimization, a reliable migration pattern probability distribution is obtained. The chemical element migration pattern probability map is generated directly and explicitly, clearly showing the topological structure and migration intensity of the path.

[0009] To achieve the above objectives, the basic solution of the present invention provides a method for inferring the migration patterns of chemical elements in a geochemical field, comprising the following steps: Step A1: Obtain the three-dimensional geological model of the study area, establish the attribute field of the three-dimensional geological model and discretize it into a three-dimensional network model, and construct the initial graph structure with the grid as nodes and the connection between nodes as edges; Step A2: Determine the node features and edge features, and after preprocessing, integrate them to obtain the node feature vector and edge feature vector respectively, and then concatenate them to obtain the sample data; Step A3: Based on the multi-head attention mechanism with added edge features and residual structure, construct an improved GAT network as the spatial propagation layer. Input sample data and execute the following processes in sequence: concatenate the previous layer embeddings of the node and its neighbor nodes and the edges between them, calculate and normalize the attention weights, aggregate neighbor embedding messages, fuse multi-head attention, adjust residuals, and output the final node embedding representation shared by multiple tasks. Step A4: Using the final representation of the node embedding as input, the chemical element content is regressed through a deep regression network combined with the Softplus activation function to output the predicted value of the chemical element content of the node. At the same time, by fusing the final representation of the two nodes connected by the edge with the edge's own features, the edge-level binary classifier composed of a linear layer and a Sigmoid activation layer outputs the edge connection migration probability. Step A5: Construct a multivariate loss function that incorporates physical constraints and geological knowledge, and iteratively train steps A3 to A4 until the multivariate loss function converges to obtain the final node chemical element content and edge connection migration probability.

[0010] Furthermore, in step A2, the node features include chemical element content data, spatial coordinates, lithological characteristics, and uncertainty indicators of lithological classification, while the edge features include the distance between nodes, fault probability, and chemical element content gradient.

[0011] Furthermore, the lithological characteristics include lithological type, whether it is a fault, and whether it is an ore body. The uncertainty indicators for lithological classification include information entropy, first lithological probability difference, and second lithological probability difference.

[0012] Furthermore, in step A3, the multi-head attention mechanism incorporating edge features is performed as follows: Step B1: Concatenate the features of the node, its neighboring nodes, and the edge between them, and use GATv2 to move the attention weights outside the nonlinear result for calculation to obtain the initial attention weights. Step B2: Normalize the initial attention weights of all adjacent edges of the node using the softmax function to obtain normalized attention coefficients; Step B3: Utilize the normalized attention coefficient to perform message aggregation on the previous layer embedding of neighboring nodes, and update the current layer embedding of the node; Step B4: Use K independent fusion attention mechanisms to complete the aggregation process from step B1 to step B3 respectively, to obtain k sets of node embedding results. After splicing in the intermediate layer, obtain the intermediate layer output, and then aggregate the K sets of node embedding results through the mean operation in the last layer to obtain the final node embedding after the multi-head attention mechanism with added edge features.

[0013] Furthermore, in step A3, the execution of the residual structure includes calculating the attention coefficient residual, calculating the node embedding residual, taking the final node embedding as input, and outputting the final representation of the node embedding shared by multiple tasks.

[0014] Furthermore, the formula for calculating the attention coefficient residual is as follows: ; In the formula, This represents the attention coefficient after adding the residual coefficient. Indicates hyperparameters, , This represents the attention coefficient of the current layer. This represents the attention coefficient of the previous layer; The formula for calculating the node embedding residual is as follows: ; In the formula, This represents the final representation of the embedding of node i. This represents a trainable weight matrix used to align the dimensions of the input and output. .

[0015] Furthermore, when hour, , It is a diagonal matrix.

[0016] Furthermore, in step A4, the formula for calculating the predicted chemical element content is: , In the formula, This represents the predicted chemical element content of node i. , This represents the weight matrix of the regression layer. , This represents the bias vector. .

[0017] Furthermore, in step A4, the formula for calculating the edge migration probability is: ; In the formula, This represents the weight matrix of the linear layer. This represents the bias vector.

[0018] Furthermore, in step A5, the multivariate loss function includes weighted regression loss, edge loss, geological constraint loss, and regularization loss.

[0019] The principle and effect of this solution are as follows:

[0020] 1. Compared with the prior art, the present invention constructs a graph model, predicts the chemical element content of all nodes separately and obtains the edge probability by a self-supervised learning method, and obtains a reliable migration mode probability distribution after joint optimization. It directly and explicitly generates a chemical element migration mode probability map, which clearly shows the topology and migration intensity of the path.

[0021] 2. Compared with existing technologies, this invention adopts a data-driven approach, directly learning migration patterns from geochemical data. Compared with traditional numerical simulation methods that rely on precise hydrogeological parameters, this invention reduces the dependence on physical models and parameter accuracy. This invention models the spatial topological relationships between sampling points using a GAT network, enabling it to capture complex spatial structural relationships, dynamically adjust the influence of different neighboring nodes, clearly identify the branching and converging characteristics of migration paths, and effectively characterize the nonlinear dynamic processes in the migration of chemical elements in geochemical fields.

[0022] 3. Compared with existing technologies, this invention also supports multi-source data fusion, which can flexibly integrate multi-source information such as geological structure and remote sensing, improving the geological rationality of model inference, and is especially suitable for complex tectonic areas such as fault zones. After acquiring sample data, a migration pattern probability map can be automatically generated end-to-end, with side probability weights quantifying migration intensity, intuitively displaying the channels and intensity of material migration, facilitating verification and adjustment by geological experts, significantly reducing the cost of manual intervention, and being more efficient and repeatable than empirical geological analysis methods. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart of a method for inferring the migration pattern of chemical elements in a geochemical field, as proposed in an embodiment of this application, is shown.

[0025] Figure 2 The illustration shows a schematic diagram of various losses during the iteration process in a practical application according to the embodiments of this application;

[0026] Figure 3 A schematic diagram showing the copper content distribution in a practical application according to the embodiments of this application is shown;

[0027] Figure 4 A schematic diagram of the edge probability distribution proposed in the practical application of the embodiments of this application is shown;

[0028] Figure 5 This illustration shows a schematic diagram of the relationship between the inferred path and the distribution of faults and ore bodies in a practical application proposed in the embodiments of this application. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] A method for inferring the migration patterns of chemical elements in a geochemical field, implementing, for example... Figure 1 As shown, the process includes the following steps: Step A1: Obtain the three-dimensional geological model of the study area, establish the attribute field of the three-dimensional geological model and discretize it into a three-dimensional network model, and construct the initial graph structure with the grid as nodes and the connection between nodes as edges.

[0031] Specifically, in step A1, after obtaining the three-dimensional geological model of the study area, based on the geological data such as the geological structure, chemical element content, environmental parameters, and lithological distribution of the study area, multiple attribute fields of the three-dimensional geological model are established by interpolation, including but not limited to the lithological field, chemical element content field, and fault distribution field.

[0032] Next, each attribute field is discretized according to a pre-defined grid size to obtain a three-dimensional mesh model. In this embodiment, the grid is used as nodes, each node corresponds to a 3×3 neighborhood, and each node connects to 8 adjacent nodes, with connecting lines as edges to construct a graph structure. In particular, faults are the main channels for element migration, so an enhanced connection strategy is applied to the fault nodes obtained from the discretized fault distribution sites, and full connectivity is implemented for fault nodes within the same fault.

[0033] Step A2: Determine the node features and edge features, and after preprocessing, integrate them to obtain the node feature vector and edge feature vector respectively, and then concatenate them to obtain the sample data.

[0034] Specifically, node characteristics include chemical element content data, spatial coordinates, lithological characteristics, and uncertainty indicators for lithological classification. Among these, lithological characteristics include lithological type, whether it is a fault, and whether it is an ore body; uncertainty indicators for lithological classification include information entropy, first lithological probability difference, and second lithological probability difference.

[0035] Edge features include the distance between nodes, fault probability, and chemical element content gradient. Edge features are used to reflect the factors that affect the migration of chemical elements.

[0036] After obtaining the node features and edge features, the following preprocessing is performed: logarithmic standardization is applied to the chemical element content data in all node features, that is, the logarithm of each chemical element content data is taken and then standardized to eliminate differences in data units.

[0037] For continuous data in node features, max-min normalization is used to shrink the data to the range [0,1].

[0038] For lithological characteristics, such as lithology type, whether it is a fault, and whether it is an ore body, they are converted into binary vectors through unique thermal encoding.

[0039] For edge features, all data are continuous, and the data is scaled to the range [0,1] using max-min normalization.

[0040] The preprocessed node and edge features are then processed using a feature projection mechanism. Specifically, this involves mapping the node features to a high-dimensional semantic space using a node feature encoder, as shown below: ; In the formula, This represents the initial embedding of node i. Representing nodes respectively Uncertainty indices regarding chemical element content data, spatial coordinates, lithological characteristics, and lithological classification. The weight matrix represents the feature mapping of nodes. This represents the bias vector for the node feature mapping.

[0041] The edge feature encoder processes edge features synchronously: ; In the formula, Let represent the initial characteristics of the edge connecting node i and node j. Indicates the node spacing. Indicates the probability of a fault. Represents the concentration gradient. The weight matrix represents the edge feature mapping. The bias vector represents the edge feature mapping.

[0042] Step A3: Based on the multi-head attention mechanism with added edge features and residual structure, construct an improved GAT network as the spatial propagation layer. Input sample data and execute the following processes in sequence: concatenate the previous layer embeddings of the node and its neighboring nodes and the edges between them, calculate and normalize the attention weights, aggregate neighbor embedding messages, fuse multi-head attention, adjust residuals, and output the final node embedding representation shared by multiple tasks.

[0043] Specifically, in the GAT network, each node's features are updated by a weighted average of the features of its neighbors. The weights are determined by an attention mechanism to dynamically adjust the influence of different neighboring nodes during information propagation. In the propagation of chemical element content data, the properties of the propagation trajectory, such as distance, chemical element content difference, and connection type, are crucial to propagation. Therefore, a multi-head attention mechanism is constructed that incorporates edge features according to the following steps.

[0044] Step B1: Concatenate the features of the node, its neighboring nodes, and the upper-level embeddings of the edges between them. Use GATv2 to move the attention weights outside the non-linear calculation to obtain the initial attention weights, as shown in the following expression: ; In the formula, The initial attention weights of node i and its neighbor node j in the current layer l are represented by ||, where || denotes the feature concatenation operation and the attention mechanism. It is a single-layer feedforward neural network. This represents the weight matrix of the current layer l, used to perform linear transformations on the node / edge embeddings. Indicates the parent level of node i: Layer embedding, This indicates the parent layer embedding of node j. This represents the upper-level embedding of the edge between node i and its neighbor node j.

[0045] Among them, GATv2 is an improved variant of the graph attention network GAT. Its core is to achieve a dynamic attention mechanism by adjusting the order of operations, which is specifically used for feature interaction and weight learning between nodes and their neighbors in graph structure data.

[0046] Step B2: Normalize the initial attention weights of all adjacent edges of the node using the softmax function to obtain normalized attention coefficients.

[0047] Specifically, the initial attention weights of all adjacent edges of node i are normalized using the softmax function as follows: ; In the formula, This represents the normalized attention coefficient of node i and its neighbor node j at the current layer. Represents the set of all neighboring nodes of node i. This represents an exponential function.

[0048] Step B3: Utilize the normalized attention coefficient to perform message aggregation on the previous layer embedding of neighboring nodes, and update the current layer embedding of the node.

[0049] Specifically, the current embedding of node i is obtained by performing message aggregation on the previous-layer embedding of neighboring nodes using the normalized attention coefficients as follows: ; In the formula, This indicates the current layer embedding of node i after the aggregated message. Represents the weight matrix. Represents a non-linear activation function. Specifically, the LeakyRelu function has the following expression: ; In the formula, It is the negative slope of the activation function, that is, at the input value When the value is negative, it represents the slope to be multiplied. It is usually set to a small positive number to control the slope of the linear portion in the negative region. The default value is 0.01.

[0050] Step B4: Use K independent fusion attention mechanisms to complete the aggregation process from step B1 to step B3 respectively, to obtain k sets of node embedding results. After splicing in the intermediate layer, obtain the intermediate layer output, and then aggregate the K sets of node embedding results through the mean operation in the last layer to obtain the final node embedding after the multi-head attention mechanism with added edge features.

[0051] Specifically, based on K independent fusion attention mechanisms, node aggregation is performed on each layer of the improved GAT network according to the following formula to obtain k sets of node embedding results: ; In the middle layer of the network, the embedding results of k groups of nodes are concatenated to serve as the output of the current layer, as calculated below: ; Finally, in the last layer of the improved GAT network, the mean operation is used instead of the concatenation operation to aggregate the results of K groups of multi-head attention mechanisms, resulting in the multi-task shared node embedding after multi-head attention mechanism processing with added edge features. The calculation formula is as follows: , In this embodiment, the execution of the residual structure includes calculating the attention coefficient residual, calculating the node embedding residual, and finally obtaining the final representation of the node embedding shared by multiple tasks.

[0052] The calculation of the attention coefficient residual is as follows: the attention coefficient of the current layer is weighted and fused with the attention coefficient of the previous layer, and the calculation formula is as follows: ; In the formula, This represents the attention coefficient after adding the residual coefficient. Indicates hyperparameters, , This represents the attention coefficient of the current layer. This represents the attention coefficient of the previous layer.

[0053] The formula for calculating the node embedding residual is as follows: ; In the formula, This represents the final representation of the embedding of node i. This represents a trainable weight matrix used to align the dimensions of the input and output. In particular, when hour, , It is a diagonal matrix.

[0054] By using residual structures, the attention coefficients of the current layer and the previous layer are weighted and fused, residual connections are added to message aggregation, and the input and output dimensions are adapted through trainable weight matrices, which alleviates the problems of oversmoothing and gradient numerical values ​​and enhances the robustness of the model.

[0055] Step A4: Construct the task layer, which contains two tasks.

[0056] The task of predicting the chemical element content of nodes takes the final representation of node embedding as input, and uses a deep regression network combined with the Softplus activation function to realize the regression of chemical element content and output the predicted value of the chemical element content of nodes.

[0057] The edge connection migration probability prediction task integrates the final representations of the two nodes connected by the edge with the edge's own features, and outputs the edge connection migration probability through an edge-level binary classifier composed of a linear layer and a sigmoid activation layer.

[0058] Specifically, in the process of predicting the chemical element content of a node, the final representation of node i is input into a deep regression network. After linear layer transformation, it is processed using the Softplus piecewise linear activation function, and the predicted chemical element content of node i is calculated using the following formula: ; In the formula, This represents the predicted chemical element content of node i. , This represents the weight matrix of the regression layer. , This represents the bias vector. , This indicates the subscript for regression tasks, distinguishing them from classification tasks.

[0059] In the process of performing edge connection migration probability prediction, the connection edges are first merged. The two nodes ultimately represent and and edge characteristics The fused features are then input into an edge-level binary classifier consisting of a linear layer and a sigmoid activation layer, and the edge connection transfer probability is output as follows: ; In the formula, This represents the weight matrix of the linear layer. This represents the bias vector.

[0060] Step A5: Construct a multivariate loss function that incorporates physical constraints and geological knowledge, and iteratively train steps A3 to A4 until the multivariate loss function converges to obtain the final node chemical element content and edge connection migration probability.

[0061] In this embodiment, a multivariate loss function is constructed as shown in the following formula: ; in, This represents the total loss. Task 1 represents the task of predicting the chemical element content of a node, and the loss it contains corresponds to the loss of the task of predicting the chemical element content of a node. Task 2 represents the task of predicting the probability of edge connection migration, and the loss it contains corresponds to the loss of the task of predicting the probability of edge connection migration. , , , These are weighting coefficients. , , , The sum of is 1.

[0062] in, The weighted regression loss is expressed as follows: ; In the formula, N is the total number of nodes in the known data. and These are the predicted and actual values ​​of the chemical element content at each node.

[0063] in, The edge loss is expressed as follows: ; In the formula, y represents the true label and p represents the predicted probability. When the true label y is 1, if the predicted probability p is close to 1, the loss is close to 0. Conversely, if p is close to 0, the loss will be close to infinity. Similarly, when y is 0, if p is close to 0, the loss is close to 0. If p is close to 1, the loss will become very large.

[0064] in, The geological constraint loss is expressed as follows: ; In the formula, For edge fault coding, Predicted edge probabilities Let be the total number of edges. It is a rectifier linear unit. hour If the value is positive, the model will lower the predicted values ​​for these regions. ; hour No loss will occur, ensuring that only when This means that a penalty is incurred when the fault is not encountered.

[0065] in, The regularization loss is expressed as follows: ; In the formula, It is the weight matrix of the linear transformation in the GAT convolutional layer. It is the chemical element in the i-th row and j-th column of the weight matrix. This represents the Frobenius norm of the matrix, which is the square root of the sum of the squares of all chemical elements.

[0066] This application employs a data-driven approach, directly learning migration patterns from geochemical data. Compared to traditional numerical simulation methods that rely on precise hydrogeological parameters, this reduces the dependence on physical models and parameter accuracy. By modeling the spatial topological relationships between sampling points using a GAT network, this application can capture complex spatial structural relationships, dynamically adjust the influence of different neighboring nodes, and clearly identify the branching and converging characteristics of migration paths. This effectively characterizes the nonlinear dynamic processes in the migration of chemical elements in geochemical fields.

[0067] This application also supports multi-source data fusion, which can flexibly integrate information from multiple sources such as geological structures and remote sensing, improving the geological rationality of model inferences, and is especially suitable for complex tectonic areas such as fault zones. After acquiring sample data, migration pattern probability maps can be automatically generated end-to-end, with side probability weights quantifying migration intensity, intuitively displaying the channels and intensity of material migration, facilitating verification and adjustment by geological experts, significantly reducing the cost of manual intervention, and being more efficient and repeatable than empirical geological analysis methods.

[0068] The following is a description of the practical application verification of the geochemical field chemical element migration model inference method provided in this application in a certain mineral deposit area. This mineral deposit area has a complex geological structure with multiple fault zones. The experimental data comes from a vertical profile taken along the y-axis of the established area, with a grid size of 10m × 10m and a total of 44,070 grids. The data includes grid center coordinates, lithological data, ore body and fault data, content data of five chemical elements (sulfur, zinc, silver, gold, and copper), and uncertainty index data for lithological classification.

[0069] During graph construction, each grid point serves as a node in the graph. Node features include processed chemical element content data, spatial coordinates, lithological characteristics, and uncertainty indicators for lithological classification. The initial edges are established using a multi-criteria strategy: each node selects several nodes closest to it to establish initial edges; pairs of points located within the same fault layer receive additional geologically related edges; and the final edge weights are dynamically adjusted during training via a graph attention network.

[0070] A grid search method is used to search for the hyperparameters affecting the model and find the optimal combination of hyperparameters. This application employs a two-layer bi-headed graph attention network (GAT) architecture, with each layer having a hidden dimension of 128. The optimizer is Adam, the learning rate is set to 0.005, the dropout rate is set to 0.2, and the training is conducted for 200 epochs. The loss function is as follows. Figure 2 As shown. After training, the predicted chemical element: copper content and side probability distribution are as follows. Figure 3 , Figure 4 As shown in the figure, the method for inferring the migration pattern of chemical elements in a geochemical field provided in this application can effectively predict the content of chemical elements by regression. Simultaneously, by utilizing self-supervised learning to generate migration correlation probabilities between nodes reflected by edge connection probabilities, the method simulates the migration pattern of copper chemical elements. The distribution of chemical element content and migration probability maps correspond to surface anomalies, verifying the effectiveness of the chemical element migration pattern simulation.

[0071] And such as Figure 4 As shown, in deep space, the predicted migration of chemical elements is controlled by regional faults. The areas with extremely high migration probabilities—the deep red areas—highly coincide with the spatial distribution of the main fault zones, revealing the core mechanism of faults as the dominant pathway for chemical element migration. Overlaying the chemical element migration probability distribution map with the spatial locations of ore bodies and faults, as shown... Figure 5 As shown, this application can also help analyze the migration characteristics of copper chemical elements. For example, in fault areas, the areas with extremely high migration probabilities predicted by the model are highly consistent with the spatial distribution of the main fault zone, revealing that the fault provides a rapid channel for Cu migration. At the same time, there are significant red high-connection-probability areas around the ore body, which is consistent with the chemical element diffusion model. Surface anomaly I is directly connected to the underlying ore body through a high-probability edge, with a clear channel. Cu chemical elements may have been transported to the surface by the fault, resulting in a surface anomaly. Its two peaks correspond to the surface depressions, possibly because the depressions are prone to water accumulation or slow water flow, causing mineral accumulation. At the same time, the accumulated water evaporates, and the minerals may be concentrated on the surface, forming the anomalous peaks. Surface anomaly II has a high anomaly intensity and a large anomaly range. Its highest peak on the left is due to the exposed ore body in this area. The high-probability migration trend indicates that chemical elements in both deep ore body areas are accumulating here, but the channels are relatively dispersed and there are no obvious faults, indicating that its formation may include other factors, such as groundwater migration, microbial activity, and other multiple sources superimposed.

[0072] According to the results shown in the figure, the migration of Cu chemical elements is significantly controlled by faults, exhibiting a concentration of edge-connection probabilities around the fault zone. A logically closed "migration channel" exists between surface anomaly I and the ore body, presumably a surface response to direct migration of primary Cu. Both known deep ore bodies show a tendency to converge towards surface anomaly II, but their distribution is relatively dispersed, possibly due to superimposed formation or indirect migration. This application effectively reveals the spatial coupling relationship between the ore body, fault, and surface, confirming the model's reasonable ability to characterize the vertical migration process of chemical elements and providing a basis for determining the deep source of surface chemical element anomalies. This demonstrates that the GATv2 model can quantitatively generate the probability of chemical element migration networks jointly controlled by tectonics and lithology, providing a scientific tool for deep mineral prediction that combines process mechanism and spatial location capabilities.

[0073] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any indirect modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for inferring the migration pattern of chemical elements in a geochemical field, characterized in that, Includes the following steps: Step A1: Obtain the three-dimensional geological model of the study area, establish the attribute field of the three-dimensional geological model and discretize it into a three-dimensional network model, and construct the initial graph structure with the grid as nodes and the connection between nodes as edges; Step A2: Determine the node features and edge features, and after preprocessing, integrate them to obtain the node feature vector and edge feature vector respectively, and then concatenate them to obtain the sample data; Step A3: Based on the multi-head attention mechanism with added edge features and residual structure, construct an improved GAT network as the spatial propagation layer. Input sample data and execute the following processes in sequence: concatenate the previous layer embeddings of the node and its neighbor nodes and the edges between them, calculate and normalize the attention weights, aggregate neighbor embedding messages, fuse multi-head attention, adjust residuals, and output the final node embedding representation shared by multiple tasks. Step A4: Using the final representation of the node embedding as input, the chemical element content is regressed through a deep regression network combined with the Softplus activation function to output the predicted value of the chemical element content of the node. At the same time, by fusing the final representation of the two nodes connected by the edge with the edge's own features, the edge-level binary classifier composed of a linear layer and a Sigmoid activation layer outputs the edge connection migration probability. Step A5: Construct a multivariate loss function that incorporates physical constraints and geological knowledge, and iteratively train steps A3 to A4 until the multivariate loss function converges to obtain the final node chemical element content and edge connection migration probability.

2. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 1, characterized in that, In step A2, node features include chemical element content data, spatial coordinates, lithological characteristics, and uncertainty indicators of lithological classification, while edge features include distance between nodes, fault probability, and chemical element content gradient.

3. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 2, characterized in that, The lithological characteristics include lithological type, whether it is a fault, and whether it is an ore body. The uncertainty indicators for lithological classification include information entropy, first lithological probability difference, and second lithological probability difference.

4. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 1, characterized in that, In step A3, the multi-head attention mechanism incorporating edge features is performed as follows: Step B1: Concatenate the features of the node, its neighboring nodes, and the edge between them, and use GATv2 to move the attention weights outside the nonlinear result for calculation to obtain the initial attention weights. Step B2: Normalize the initial attention weights of all adjacent edges of the node using the softmax function to obtain normalized attention coefficients; Step B3: Utilize the normalized attention coefficient to perform message aggregation on the previous layer embedding of neighboring nodes, and update the current layer embedding of the node; Step B4: Use K independent fusion attention mechanisms to complete the aggregation process from step B1 to step B3 respectively, to obtain k sets of node embedding results. After splicing in the intermediate layer, obtain the intermediate layer output, and then aggregate the K sets of node embedding results through the mean operation in the last layer to obtain the final node embedding after the multi-head attention mechanism with added edge features.

5. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 4, characterized in that, In step A3, the execution of the residual structure includes calculating the attention coefficient residual, calculating the node embedding residual, taking the final node embedding as input, and outputting the final representation of the node embedding shared by multiple tasks.

6. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 5, characterized in that, The formula for calculating the attention coefficient residual is as follows: ； In the formula, This represents the attention coefficient after adding the residual coefficient. Indicates hyperparameters, , This represents the attention coefficient of the current layer. This represents the attention coefficient of the previous layer; The formula for calculating the node embedding residual is as follows: ； In the formula, This represents the final representation of the embedding of node i. This represents a trainable weight matrix used to align the dimensions of the input and output. .

7. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 6, characterized in that, when hour, , It is a diagonal matrix.

8. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 1, characterized in that, In step A4, the formula for calculating the predicted chemical element content is: ； In the formula, This represents the predicted chemical element content of node i. , This represents the weight matrix of the regression layer. , This represents the bias vector. .

9. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 8, characterized in that, In step A4, the formula for calculating the edge migration probability is: ； In the formula, This represents the weight matrix of the linear layer. This represents the bias vector.

10. The method for inferring the migration pattern of chemical elements in a geochemical field according to claim 1, characterized in that, In step A5, the multivariate loss function includes weighted regression loss, edge loss, geological constraint loss, and regularization loss.

Images