A deep ore-forming prediction three-dimensional implicit modeling method based on multi-source geophysical and geochemical data fusion

Abstract

This invention discloses a three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data, belonging to the interdisciplinary field of geological exploration and artificial intelligence. Addressing the shortcomings of existing methods, such as rasterization destroying data topology and the lack of geological process constraints in the implicit field, this invention constructs a cross-modal geological coupling graph containing causal edges of geological processes, and uses graph neural networks for relation-aware encoding; furthermore, it constructs an implicit neural field with geological semantic constraints, aligning the direction of the mineralization potential gradient with the direction of geological action mechanism transmission; and dynamically evolves and optimizes the graph structure during training, achieving adaptive adjustment of edge weights and topology pruning growth. This method realizes end-to-end generation driven by geological processes from discrete heterogeneous sampling to continuous three-dimensional implicit mineralization fields, significantly improving the geological interpretability and predictive stability of deep mineralization prediction.

Description

Technical Field

[0001] This invention relates to the field of geological exploration and mineral resource prediction technology, and in particular to a three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data. Background Technology

[0002] Deep mineralization prediction is a core component of strategic mineral resource exploration. As exploration depth increases, direct surface observation information decreases dramatically, necessitating the fusion of heterogeneous information from multiple sources, such as geophysical network data and geochemical sampling data, to assess the mineralization potential of deep, unexplored areas. Existing 3D implicit modeling techniques mainly include traditional geostatistical methods based on radial basis functions, kriging, or inverse distance weighting, as well as deep learning methods based on multilayer perceptrons or implicit neural representations. Regarding multi-source data fusion, current techniques typically interpolate geophysical and geochemical fields separately to a regular 3D grid and then perform post-fusion through weighted overlay or machine learning models, or concatenate multi-source data into a feature vector and directly input it into a neural network for end-to-end training.

[0003] However, the aforementioned existing technologies have the following inherent defects: traditional rasterization preprocessing or simple coordinate stitching methods destroy the original topological relationship between geophysical survey networks and geochemical sampling points in space, and fail to explicitly model geological a priori constraints such as ore-controlling structures, lithological ore-controlling structures, and elemental zoning, resulting in systematic smoothing distortion during deep inference; existing implicit neural fields only use spatial coordinates as input and rely entirely on the network's own generalization ability to fit mineralization potential, and their deep extrapolation lacks geological process mechanism constraints, and the gradient direction of the implicit field is not subject to geological a priori constraints, resulting in strong ambiguity in deep prediction results and a disconnect from regional mineralization patterns; in addition, existing methods generally use static predefined data relationship graphs or simple spatial adjacency graphs, which cannot adaptively correct the graph structure during training based on the self-consistency between the implicit field output and the geological action mechanism, resulting in insufficient stability of deep prediction under small sample sparse sampling conditions.

[0004] Therefore, there is an urgent need for a modeling method that can generate continuous three-dimensional implicit mineralization fields by dynamic graph evolution driven by geological process mechanisms and joint optimization of implicit fields, while preserving the original topological relationships and geological semantics of multi-source heterogeneous data. Summary of the Invention

[0005] This invention aims to at least solve the technical problems existing in the prior art, and in particular, proposes an innovative solution.

[0006] To achieve the above-mentioned objectives of this invention, this invention provides a three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data, comprising the following steps: S1. Acquire multi-source heterogeneous geological data in the study area and perform coordinate system and dimensional standardization on the multi-source heterogeneous geological data. S2, based on the processed multi-source heterogeneous geological data, constructs a cross-modal geological coupling map; S3, using graph neural networks to perform relation-aware encoding on cross-modal geological coupling maps to obtain multimodal node embedding representations; S4, based on multimodal node embedding representation, constructs an implicit neural field; S5 trains the implicit neural field and dynamically optimizes the cross-modal geological coupling map during the training process; S6, based on the trained implicit neural field, performs deep mineralization prediction.

[0007] Optionally, in step S1, the multi-source heterogeneous geological data includes geophysical network data, geochemical sampling data, and geological survey data.

[0008] Optionally, in step S2, the cross-modal geological coupling map includes geophysical field nodes, geochemical sampling nodes, and geological constraint nodes. The geophysical field nodes and geochemical sampling nodes are connected by geological process causal edges, which are constructed based on the metallogenic geological mechanism and carry directional geological semantics.

[0009] Optionally, the causal edges of geological processes include tectonic ore-controlling conduction edges, hydrothermal conduit conduction edges, and lithological ore-controlling differentiation edges; The tectonic ore-controlling conduction edge connects the tectonic fault node with the geochemical sampling node controlled by the fault, with the direction pointing from the tectonic fault to the element-enriched area; The conductive edge of the hydrothermal channel connects the geophysical low-resistivity anomaly node with the corresponding geochemical alteration element enrichment node, and the direction is consistent with the hydrothermal migration path. The lithological ore-controlling differentiation boundary connects the nodes on both sides of the lithological interface, with the direction pointing from the geophysical property abrupt change interface to the geochemical background differentiation zone.

[0010] Optionally, in step S3, the relation-aware encoding performs directed message passing along the directional geological semantics of the causal edge of the geological process, so that the embedded representation of the target node aggregates the feature information of its geological genesis upstream nodes.

[0011] Optionally, in step S4, the implicit neural field responds to the spatial coordinate query and outputs the mineralization potential value, and the gradient of the implicit neural field is constrained by the directional consistency of the causal edges of the geological process, so that the trend of mineralization potential change is consistent with the spatial transmission direction of the mineralization geological mechanism.

[0012] Optionally, directional consistency constraints include the following steps: S41, For the source node and target node connected by the causal edge of the geological process, calculate the gradient vector of the implicit neural field at the spatial coordinates of the source node; S42, calculate the cosine similarity between the gradient vector and the direction vector of the causal edge of the geological process; S43 uses cosine similarity as a regularization term in the loss function and backpropagates to optimize the implicit neural field parameters.

[0013] Optionally, in step S5, the dynamic evolution optimization uses the consistency between the gradient direction of the implicit neural field and the direction vector of the causal edge of the geological process as feedback basis to perform adaptive adjustment of edge weights and adaptive update of topology on the cross-modal geological coupling graph; the adaptive adjustment of edge weights adjusts the weights of the corresponding causal edges of the geological process according to the consistency; the adaptive update of topology includes pruning low-consistency causal edges of the geological process and generating new causal edges of the geological process.

[0014] Optionally, adaptive edge weight adjustment and adaptive topology update include the following sub-steps: S51, calculate the cosine similarity between the gradient vector of the implicit neural field at the two ends of the causal edge of the geological process and the direction vector of the edge; S52, adjust the edge weights of causal edges of geological processes based on cosine similarity. When the cosine similarity is lower than a preset threshold, decrease the corresponding weight; when the cosine similarity is higher than the preset threshold, increase the corresponding weight. S53, when the weight of a causal edge in a geological process remains below the pruning threshold for more than a preset number of training rounds, the edge is removed from the cross-modal geological coupling graph; S54. When the spatial distance between a pair of nodes in the cross-modal geological coupling graph is less than the preset growth radius, and the gradient direction consistency of the implicit neural field at the pair of nodes is higher than the growth threshold, and there is no geological process causal edge between the pair of nodes, a new geological process causal edge is generated between the pair of nodes, and the direction vector of the new geological process causal edge is consistent with the gradient direction at the node pair.

[0015] Optionally, dynamic evolution optimization and implicit neural field network parameter optimization are performed alternately, including the following steps: S55, in the first training phase, fix the topology of the cross-modal geological coupling map and optimize the implicit neural field parameters until convergence; S56, in the second training phase, the implicit neural field parameters are fixed, and the edge weights and topological structure of the causal edges of the geological process are optimized according to the degree of consistency. S57, repeat the first training phase and the second training phase until the overall consistency between the gradient direction of the implicit neural field and the direction vector of the causal edge of the geological process satisfies the preset convergence condition.

[0016] The beneficial effects of this invention are as follows: 1. By constructing causal edges of geological processes carrying directional geological semantics to connect geophysical field nodes and geochemical sampling nodes, the cross-modal geological coupling map itself contains mineralization dynamics semantics such as tectonic mineralization control, hydrothermal channels, and lithological differentiation, avoiding information loss and geological semantic fragmentation caused by traditional rasterization preprocessing.

[0017] 2. By utilizing graph neural networks to perform directed message passing along directional geological semantics, the node embedding representation can simultaneously integrate local sampling features and global geological genetic information, significantly enhancing the information compensation capability of non-uniform sampling areas and deep blind areas.

[0018] 3. This invention introduces a dynamic evolution optimization mechanism during the implicit neural field training process. It uses the consistency between the gradient direction of the implicit neural field and the causal edge direction vector of the geological process as the feedback basis, adaptively adjusts the edge weights, and performs topology pruning and growth. This allows the graph structure and implicit field parameters to be mutually calibrated in the alternating optimization process, breaking through the bottleneck that static predefined graph structures cannot adapt to the uncertainty of deep inference.

[0019] 4. By embedding the directional consistency constraint of the causal edge of the geological process into the implicit neural field loss function, and supplementing it with geological boundary jump loss and mineralization physical process constraint loss, the implicit neural field generates discontinuous jumps that conform to geological significance at lithological or tectonic interfaces. Moreover, the trend of deep mineralization potential changes is consistent with the spatial transmission direction of mineralization geological mechanisms such as hydrothermal migration, which effectively suppresses the problem of geological boundary ambiguity caused by the global smooth transition of the traditional implicit field.

[0020] In summary, this invention achieves end-to-end generation driven by geological process mechanisms, from discrete heterogeneous geological sampling to continuous three-dimensional implicit mineralization fields, significantly improving the geological interpretability and predictive stability of deep mineralization prediction results.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Fig. 1 This is a schematic diagram of the overall process of the three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data of the present invention. Fig. 2 This is a schematic diagram illustrating the principle of constructing cross-modal geological coupling maps according to the present invention; Fig. 3This is a schematic diagram of the effect of three-dimensional implicit field slice and target area delineation for deep mineralization prediction in a certain research area according to the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] Example 1, Please see Figs. 1-3 This embodiment takes a copper polymetallic ore cluster as the study area, and the exploration depth target is a range of two kilometers below the surface, and describes in detail the specific execution process of the method of the present invention.

[0025] S1. Acquisition of Multi-Source Heterogeneous Geological Data: Acquisition of multi-source heterogeneous geological data within the study area, specifically including three categories: geophysical network data, geochemical sampling data, and geological survey data. The study area refers to the target geological exploration area for deep mineralization prediction and 3D implicit modeling. Its spatial extent is determined by the exploration task and typically includes known mineralized outcrops, the geophysical network coverage area, and the geochemical sampling area. Specifically, the geophysical network data includes 1:50,000 high-precision gravity network data (500-meter spacing between measuring points), recording the plane coordinates and Bouguer gravity anomalies of each measuring point; 1:25,000 surface magnetic data (250-meter spacing between measuring points), recording the total magnetic field strength anomaly; and magnetotelluric sounding profile data (1,000-meter spacing between points), recording resistivity inversion results. The geochemical sampling data includes 1:50,000 soil geochemical measurement data (500-meter spacing between sampling points), analyzing the content of twelve elements including copper, gold, silver, and molybdenum; and borehole core sample data, from twelve boreholes, with samples collected continuously at lengths of one to two meters. The geological survey data includes digitized boundaries of 1:50,000 geological maps, traces of known fault structures, and vector data of lithological contact zones.

[0026] The aforementioned multi-source heterogeneous geological data underwent coordinate system and dimensional standardization. All spatial data were transformed to three-dimensional Cartesian coordinates, denoted as... Regarding the content of geochemical elements After performing a logarithmic transformation Standardization, the standardization formula is ,in and These represent the mean and standard deviation of the logarithmic content of the element in all samples from the study area, respectively. For geophysical anomalies, range standardization is used, as shown in the formula: This maps the data to the zero-to-one range. After the above processing, all types of data have a unified numerical scale and can be directly used for subsequent graph node feature construction.

[0027] S2, Construction of cross-modal geological coupling map: Based on the multi-source heterogeneous geological data processed by S1, a cross-modal geological coupling map is constructed. .in For a set of nodes, Let be the set of edges. A collection of node types It is a set of edge types.

[0028] Node set It includes three types: the first type is object detection points. Each gravity, magnetic, or electrical measurement point constitutes a node, and the node eigenvector is denoted as . ,in To standardize gravity anomalies, To standardize magnetic anomalies, The first category is standardized resistivity; the second category is geochemical sample points. Each soil or core sample constitutes a node, and the node feature vector is denoted as . The third category is geological constraint points. The nodes were generated by sampling at 200-meter intervals along the geological boundary, and the node feature vectors were denoted as... ,in This is the unique thermal encoding vector for lithology categories. This is the Euclidean distance from the point to the nearest fracture trace.

[0029] edge set There are three types. The first type is spatial adjacency edges. For each type of node, a spherical domain with a variable radius in three-dimensional space is constructed, with the radius... Taking a radius of one kilometer, if the Euclidean distance between two nodes is less than this radius, an undirected edge is established to characterize the continuity of spatial sampling. The second type is geological causal edges. Based on the geological mechanisms of mineralization and carrying directional geological semantics, it specifically includes three subtypes: tectonic ore-controlling conduction edges, hydrothermal conduit conduction edges, and lithological ore-controlling differentiation edges. The structural ore-controlling transmission edge refers to a directed edge constructed based on the geological mechanism of structural ore-forming processes. Its source end is a structural fault node, and its target end is a geochemical sampling node controlled by that fault. The direction is from the structural fault towards the element-enriched region. Specifically, a structural fault node is a geologically constrained node generated by sampling at predetermined intervals along a known fault trace. Its node feature vector includes the point's three-dimensional spatial coordinates, lithological category unique thermal code, and the distance from the point to the nearest fault trace, which is zero or close to zero. A geochemical sampling node controlled by that fault is a geochemical sampling node located spatially within a predetermined influence bandwidth on both sides of the fault trace, and whose standardized content of the target ore-forming element is greater than the logarithmic mean of the element's content in the study area plus one standard deviation. This indicates that the fault has a controlling effect on element enrichment and ore formation. An element-enriched region refers to the spatial range in the geochemical sampling data where the content of the target ore-forming element, after logarithmic transformation and Z-score standardization, is greater than zero; that is, the spatial region where the element content is higher than the logarithmic mean of the study area. The condition for establishing the structural ore-controlling transmission edge is: for structural fault nodes on a known fault trace... Geochemical sampling nodes within 500 meters of the fault zone ,like Standardized copper content If the value is greater than the mean of the study area plus one standard deviation, then a system based on... point to Directed edge, direction vector It characterizes the geological genesis of the enrichment of elements controlled by fractures.

[0030] The hydrothermal channel conduction edge refers to a directed edge constructed based on the mineralization geological mechanism of hydrothermal channels. Its source end is a geophysical low-resistivity anomaly node, and its target end is a geochemical alteration element enrichment node. The direction is consistent with the hydrothermal migration path. Among them, the geophysical low-resistivity anomaly node refers to a geophysical detection point node whose resistivity value is lower than the lower quartile of the resistivity of all geophysical detection points in the study area, characterizing the resistivity reduction anomaly caused by hydrothermal alteration or metal sulfide enrichment; the geochemical alteration element enrichment node refers to a geochemical sampling node whose content of hydrothermal alteration indicator elements is higher than the preset anomaly lower limit. The alteration indicator elements include at least one of molybdenum, arsenic, and antimony, and their content, after logarithmic transformation and Z-score standardization, is greater than the mean of the study area plus one standard deviation; the hydrothermal migration path refers to the trajectory direction of hydrothermal fluids moving from the deep source area to the shallow precipitation enrichment area inferred from the regional mineralization geological rules and geophysical field characteristics. In this invention, it is simplified to the spatial vector direction from the geophysical low-resistivity anomaly node to the corresponding geochemical alteration element enrichment node. The conditions for establishing a conduction edge in a hydrothermal channel are: for geophysical low-resistivity anomaly nodes within a geophysical low-resistivity anomaly region. If its resistivity value It is below the lower quartile of the study area, and contains geochemical alteration element enrichment nodes within its 800-meter neighborhood. If the molybdenum content is higher than the abnormal lower limit, then a system based on... point to The directed edges have the same direction vector as the hydrothermal migration path.

[0031] Lithological ore-controlling differentiation edges refer to directed edges constructed based on the lithological ore-controlling geological mechanisms, connecting nodes on both sides of the lithological interface, with the direction pointing from the geophysical property abrupt change interface to the geochemical background differentiation zone. Nodes on both sides of the lithological interface refer to pairs of nodes located on either side of the same lithological contact zone but belonging to different lithological units, including geologically constrained nodes and / or geochemical sampling nodes, whose spatial location is determined by the digitized boundary of the geological map and sampling coordinates. Geophysical property abrupt change interfaces refer to the interface locations where geophysical field attribute values ​​change significantly over a short distance, identified by the gradient modulus of geophysical detection point attribute values ​​exceeding a preset threshold; physical properties include at least one of density, magnetic susceptibility, or resistivity. Geochemical background differentiation zones refer to areas where the background content of geochemical elements on both sides of the lithological interface exhibits statistically significant differences, identified by the target element content of sampling points on both sides. The significance of the difference is determined by a test, with a p-value less than 0.05 considered significant. The conditions for establishing the lithologically controlled ore-differentiating boundary are: for the nodes on both sides of the lithological interface... and If there is a geochemical sampling node on each side of the interface and If the difference in copper background content on both sides is significant after t-test, then a directed edge is established from the side node of the geophysical property abrupt change interface to the side node of the geochemical background differentiation zone.

[0032] The third category is physical response-related edges. This is used to connect nodes that are statistically correlated with geophysical responses and geochemical anomalies. Specifically, it is achieved by calculating the Pearson correlation coefficient between geophysical detection points and geochemical sample points within a 500-meter neighborhood. If the absolute value of the correlation coefficient is greater than 0.6, then an undirected edge is established.

[0033] Through the above construction process, the cross-modal geological coupling map G simultaneously contains spatial sampling topology, causal transmission of mineralization geological processes, and physical-chemical coupling relationships, providing a structured geological semantic foundation for subsequent coding.

[0034] S3, Multimodal Node Embedding Encoding: Relationship-aware encoding of a cross-modal geological coupling graph G is performed using a heterogeneous graph attention network. Relationship-aware encoding refers to the process of differentially modeling and extracting features of the geological semantic relationships carried by different types of nodes and edges in the cross-modal geological coupling graph using a graph neural network; directed message passing refers to strictly performing neighbor information aggregation along the preset direction of the causal edges of the geological process in the message passing framework of the graph neural network, so that the embedding representation of the target node only aggregates the feature information of its upstream nodes in geological genesis; embedding representation refers to the low-dimensional dense real-valued vector obtained after encoding by the graph neural network, which is used to represent the deep geological semantic features of the node.

[0035] Graph neural networks refer to deep learning network models that take graph structure data as input and learn deep representations of nodes and edges through message passing mechanisms between nodes; in this embodiment, a heterogeneous graph attention network is specifically used as the implementation architecture of the graph neural network.

[0036] The target node refers to the node that receives neighbor information and updates its own embedded representation during the message passing process in the graph neural network. It corresponds to the directed edge terminal node that serves as the information convergence end in the causal edge of the geological process.

[0037] The upstream node of geological genesis refers to the node located at the source end and transmitting geological genesis information to the target node in the direction of transmission of the metallogenic geological mechanism characterized by the causal edge of the geological process; it corresponds to the tectonic fault node in the tectonic ore-controlling transmission edge, the geophysical low resistivity anomaly node in the hydrothermal channel transmission edge, and the node located on the side of the geophysical property abrupt change interface in the lithological ore-controlling differentiation edge.

[0038] Feature information refers to the geological attribute data and mineralization semantic information contained in the original feature vector of a node and its embedded representation after being encoded by a graph neural network, including geophysical anomalies, geochemical element content, lithological category coding, and spatial coordinate information.

[0039] For a target node v, the message passing formula from its neighbor u along the edge type τ is: ; in Let u be the embedding representation of the l-th layer node. These are the learnable weight matrix and bias vector corresponding to edge type τ, respectively. Attention coefficients. Calculated through a type-aware attention mechanism: ; ; in edge type The corresponding learnable query projection matrix and key projection matrix, For the attention head dimension. Target node. In the The layer embedding update formula is: ; in For the activation function, a modified linear unit is used here. Specifically, for upstream and downstream nodes connected by geological causal edges, message passing is strictly performed along the directed edge direction, enabling the target node's embedding representation to aggregate the feature information of its upstream geological nodes, thereby encoding mineralization semantics such as tectonic mineralization control and hydrothermal migration into the node embedding. =3-layer message passing to obtain the deep embedded representation of each node , where the embedded dimension =128.

[0040] S4, Construction of Implicit Fields with Geological Semantic Constraints: Constructing Implicit Neural Fields with Geological Semantic Constraints Based on Deep Embedded Representations This makes it responsive to queries of any spatial coordinates. Output mineralization potential value .

[0041] The implicit neural field comprises three functional modules. The first is the local geological feature extraction module, which retrieves query points in the cross-modal geological coupling map G based on spatial proximity. The set of neighboring nodes , where the search radius Take 800 meters. Adaptively aggregate deep embedding representations of neighborhood nodes through a graph attention mechanism, local graph features. The calculation formula is: ; Attention weight , The distance attenuation parameter is set to 500 meters, and ATT is a learnable attention scoring function implemented through a two-layer fully connected network. The second module is a global geological background aggregation module, which obtains the overall metallogenic geological background characteristics of the study area through global average pooling, using the following formula: The third is an implicit mapping network, constructed using a four-layer fully connected multilayer perceptron, with each layer containing 256 neurons. The first three layers use a sinusoidal activation function. To capture high-frequency geological boundary details, where frequency parameters The last layer uses The activation function outputs a normalized mineralization potential value. The feature vector input to the implicit mapping network is... The concatenated vector has a dimension of 128+128+3=259.

[0042] In this invention, the loss function refers to a mathematical function used during the training of the implicit neural network to measure the difference between the model's predicted output and the desired target, and to guide the iterative optimization of network parameters. Specifically, the loss function is based on the deviation between the observed mineral potential values ​​of known sampling points and the predicted output values ​​of the implicit neural network, superimposed with geological prior constraints. The gradient of the loss function with respect to the learnable parameters of each layer of the network is calculated using the backpropagation algorithm, and the parameters are updated along the gradient descent direction to minimize the loss function value, thereby making the output of the implicit neural network approximate the true mineral potential distribution. Backpropagation refers to the algorithmic process of calculating the gradient layer by layer from the output layer of the loss function to the input layer based on the chain rule of differentiation, and passing the gradient to each learnable parameter to perform update optimization.

[0043] The training loss of the implicit neural field consists of four components. The first component is the data fitting loss. To constrain the consistency between the implicit field output values ​​at known geochemical sample points and the observed mineralization indicators, the formula is as follows: ,in This is the standardized mineralization indicator value for sample point v, where the combination of copper content and associated elements molybdenum and silver is used as the differentiation index.

[0044] A source node is the starting node of a directed edge in a causal edge of a geological process. It corresponds to an upstream node of geological genesis and is used to transmit geological genesis information to the target node.

[0045] The trend of mineralization potential refers to the increase or decrease of the mineralization potential value output by the implicit neural field in spatial coordinates, and is characterized by the partial derivative of the implicit neural field with respect to spatial coordinates, i.e., the gradient vector.

[0046] The spatial transmission direction refers to the direction of the physical action of the mineralization geological mechanism in three-dimensional space, including the extension direction of tectonic ore-controlling action along the fault zone, the flow direction of hydrothermal migration from deep to shallow, and the direction of lithological ore-controlling differentiation across lithological interfaces; in this invention, it is quantitatively characterized by the direction vector of the causal edge of the geological process.

[0047] The second term is the gradient direction consistency loss. The gradient used to realize the implicit neural field is constrained by the consistency of the causal edge direction of the geological process. For the causal edge of the geological process... Calculate the implicit neural field at the source node gradient vector at spatial coordinates and the direction vector of that edge. Calculate the cosine similarity between the two. The formula for gradient direction consistency loss is: ; This loss term optimizes the implicit neural field parameter θ through backpropagation, ensuring that the trend of mineralization potential variation aligns with the spatial transmission direction of the mineralization geological mechanism. The third term is the geological boundary jump loss. The formula used to constrain the difference in output values ​​of the implicit neural field on both sides of a known geological boundary is: ; in This is the set of node pairs on both sides of a known lithological interface or tectonic boundary. The boundary difference threshold is set to 0.3. This loss causes the mineralization potential to produce geologically significant discontinuous jumps at geological interfaces rather than a globally smooth transition. The fourth term is the loss constrained by the mineralization physical processes. This is achieved by embedding the residuals of the convection and diffusion equations for ore-forming hydrothermal migration into a loss function. The convection and diffusion equations are in the following form: ; in For elemental concentration, The diffusion coefficient is... For hydrothermal flow velocity vector, These are source and sink terms. Under the steady-state assumption, they simplify to... Adding a learnable linear mapping layer to the output of the implicit mapping network will transform the mineralization potential field. Mapped to elemental concentration field ( , (As learnable parameters), physical loss is applied to : ; in For a set of three-dimensional spatial sampling points, geophysical field nodes provide flow velocity. With pressure boundary conditions, geochemical sample nodes provide source and sink terms. Constraints. The total loss function is: ; in, , These are learnable parameters that are automatically optimized through training, and the weight coefficients are... =0.5, =0.1, =0.01.

[0048] S5, Dynamic Evolution Optimization: During the training of the implicit neural field, the edge weights and topological structure of the causal edges of the geological process are dynamically evolved and optimized, with the consistency between the gradient direction of the implicit neural field and the causal edge direction vector of the geological process as the feedback basis.

[0049] Dynamic evolution optimization and implicit neural field parameter updates are performed alternately. In the first training phase, the topology of the cross-modal geological coupling graph G is fixed, i.e., the edge set E and edge weights are fixed, and only the implicit neural field parameters are optimized. ,use Optimizer, learning rate Training to the loss function Convergence or reaching a preset number of rounds. Convergence refers to the fact that, during the iterative optimization process, the change in the corresponding loss function value of the implicit neural field parameters is less than a preset threshold, or the error between the network output and the target value reaches a stable state. In this invention, convergence of the implicit neural field parameters means that the relative change rate of the total loss function value is less than one-thousandth for several consecutive rounds.

[0050] In the second training phase, the implicit neural field parameter θ is fixed, and the edge weights and topological structure of the causal edges of the geological process are optimized according to the degree of directional consistency.

[0051] Two-end nodes refer to the two nodes connected by the same causal edge of a geological process, corresponding to the source node and the target node of the edge, respectively.

[0052] The adaptive adjustment formula for edge weights is: ; in, This means cropping the values ​​to the closed interval [0,1]. For the first Edge during round iteration The weight, The learning rate for the graph is set to 0.01. For the current implicit neural field at the edge The cosine similarity between the gradient vector at the source node and the direction vector of that edge. When the cosine similarity is below a preset threshold... When the value is 0.3, the weight of the corresponding causal edge of the geological process is reduced; when it is higher than the threshold, the weight of the edge is reduced. When the value is 0.7, the corresponding weight is increased.

[0053] The topology pruning rule is: when the weight of a causal edge of a geological process... Continuously below the pruning threshold =0.1 Exceeds the preset training rounds After 10 rounds, the edge is removed from the cross-modal geological coupling graph, indicating that the geological causal assumption is inconsistent with the current implicit field inference.

[0054] The topology growth rule is: when traversing cross-modal geological coupling maps, the spatial distance is less than the preset growth radius. =600 meters of node pairs If there is no causal edge between the nodes, and the gradient direction consistency of the implicit neural field at the node pair is higher than the growth threshold, =0.8, that is Then, a new causal edge for the geological process is generated between the node pairs. The direction vector of the new edge is consistent with the gradient direction at the node pair, and the initial weight is set to... 0.5.

[0055] Repeat the first and second training phases until the overall consistency between the gradient direction of the implicit neural field and the direction vector of the causal edge of the geological process satisfies the preset convergence condition, i.e. ,in Set the value to 0.75. Overall consistency satisfies the preset convergence condition, which means that the mean value of the gradient direction and the cosine similarity of the direction vector of all geological process causal edges is greater than or equal to the preset threshold.

[0056] S6, Deep Mineralization Prediction: Based on the implicit neural field optimized by dynamic evolution of S5, coordinates are traversed and queried in the three-dimensional space of the study area at a resolution of 50m×50m×20m. Calculate the mineralization potential value at each query point. This process generates a continuous three-dimensional implicit mineralization field. Isosurfaces with mineralization potential values ​​greater than 0.7 are extracted to delineate favorable target areas for deep mineralization. Based on known geological patterns and geophysical anomalies, geological interpretations and borehole deployment recommendations are provided for the target areas.

[0057] Fig. 3 The figure shows a three-dimensional rendering of deep mineralization prediction obtained after applying this embodiment in a certain study area. Horizontal slices at different depths show that the high-potential mineralization area is ellipsoidally distributed near a depth of -2,500 meters, consistent with the extension direction of known fault structures. The spatial location of the high-potential isopleths matches the mineralization depth of subsequent verification boreholes, verifying the effectiveness and geological interpretability of the method of this invention in deep mineralization prediction.

[0058] Example 2, The main difference between this embodiment and Embodiment 1 lies in the specific implementation of the constraint loss in the ore-forming physical process in S4. For hydrothermal ore-forming systems, Darcy's law is coupled with convection and diffusion equations to describe the hydrothermal transport and element precipitation processes. Velocity field Given by Darcy's Law ,in For penetration rate, For fluid viscosity, For pressure, For density, Let gravitational acceleration be the velocity. Substituting the Darcy velocity into the convection and diffusion equations, we obtain the residuals of the coupled equations: ; ; in For pressure source items, The element source and sink terms are defined by boundary constraints provided by the geophysical field nodes and the geochemical sample nodes, respectively. The physical loss term is: ; The remaining steps are the same as in Example 1. This example is applicable to hydrothermal copper-gold polymetallic deposits, ensuring that the implicit neural field not only fits the observation data but also satisfies the laws of mass and momentum conservation in terms of physical mechanisms, significantly improving the physical consistency of deep extrapolation.

[0059] Example 3, This embodiment details the numerical implementation of the alternating optimization strategy in S5. Let the implicit neural field parameters be... The set of causal edge weights for geological processes is The topological adjacency matrix is The joint optimization objective is defined as: ; Because of simultaneous optimization of continuous parameters Discrete topology This is a mixed-integer optimization problem, which is difficult to solve directly; therefore, an alternating optimization strategy is adopted. Specifically, in the... In the next outer iteration, inner optimization is performed first. The Adam optimizer was trained for fifty epochs; then the frequency was fixed. }, Execute graph structure optimization ), among which Gradient descent is used for updating. Discrete updates are performed using the aforementioned pruning-growth rule. The outer iteration terminates when the rate of change of the total loss between two adjacent iterations is less than one-thousandth or the graph structure no longer changes. This alternating strategy ensures the numerical stability of graph evolution and implicit field training, avoiding gradient conflict issues in joint optimization.

[0060] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data, characterized in that, Includes the following steps: Acquire multi-source heterogeneous geological data within the study area, and perform coordinate system and dimensional standardization processing on the multi-source heterogeneous geological data; Based on the processed multi-source heterogeneous geological data, a cross-modal geological coupling map is constructed. A graph neural network is used to perform relation-aware encoding on the cross-modal geological coupling map to obtain a multimodal node embedding representation; Based on the multimodal node embedding representation, an implicit neural field is constructed; The implicit neural field is trained, and the cross-modal geological coupling map is dynamically evolved and optimized during the training process; S6. Based on the trained implicit neural field, perform deep mineralization prediction.

2. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 1, characterized in that, The multi-source heterogeneous geological data includes geophysical network data, geochemical sampling data, and geological survey data.

3. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 1, characterized in that, The cross-modal geological coupling map includes geophysical field nodes, geochemical sampling nodes, and geological constraint nodes. The geophysical field nodes and the geochemical sampling nodes are connected by geological process causal edges, which are constructed based on the metallogenic geological mechanism and carry directional geological semantics.

4. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 3, characterized in that, The causal edges of the geological processes include structural ore-controlling conduction edges, hydrothermal channel conduction edges, and lithological ore-controlling differentiation edges. The tectonic ore-controlling conduction edge connects the tectonic fault node and the geochemical sampling node controlled by the fault, with the direction pointing from the tectonic fault to the element-enriched area; The conductive edge of the hydrothermal channel connects the geophysical low-resistivity anomaly node with the corresponding geochemical alteration element enrichment node, and its direction is consistent with the hydrothermal migration path. The lithological ore-controlling differentiation edge connects the nodes on both sides of the lithological interface, with the direction pointing from the geophysical property abrupt change interface to the geochemical background differentiation zone.

5. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 3, characterized in that, The relation-aware encoding performs directed message passing along the directional geological semantics of the causal edge of the geological process, enabling the embedded representation of the target node to aggregate the feature information of its upstream geological genesis nodes.

6. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 3, characterized in that, The implicit neural field responds to the spatial coordinate query and outputs the mineralization potential value, and the gradient of the implicit neural field is constrained by the directional consistency of the causal edges of the geological process, so that the trend of mineralization potential change is consistent with the spatial transmission direction of the mineralization geological mechanism.

7. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 6, characterized in that, The direction consistency constraint includes the following steps: S41, For the source node and target node connected by the causal edge of the geological process, calculate the gradient vector of the implicit neural field at the spatial coordinates of the source node; S42, Calculate the cosine similarity between the gradient vector and the direction vector of the causal edge of the geological process; S43, the cosine similarity is used as a regularization term in the loss function, and the implicit neural field parameters are optimized by backpropagation.

8. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 3, characterized in that, The dynamic evolution optimization uses the consistency between the gradient direction of the implicit neural field and the direction vector of the geological process causal edge as feedback to perform adaptive adjustment of edge weights and adaptive update of topology on the cross-modal geological coupling graph. The adaptive adjustment of edge weights adjusts the weights of the corresponding geological process causal edges according to the consistency. The adaptive update of topology includes pruning low-consistency geological process causal edges and generating new geological process causal edges.

9. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 8, characterized in that, The adaptive adjustment of edge weights and the adaptive update of the topology include the following sub-steps: S51, Calculate the cosine similarity between the gradient vector of the implicit neural field at the two ends of the causal edge of the geological process and the direction vector of the edge; S52, adjust the edge weight of the causal edge of the geological process according to the cosine similarity. When the cosine similarity is lower than a preset threshold, reduce the corresponding weight. When the cosine similarity is higher than the preset threshold, increase the corresponding weight. S53, when the weight of the causal edge of the geological process is continuously lower than the pruning threshold for more than a preset training round, the edge is removed from the cross-modal geological coupling graph; S54, when the spatial distance between a pair of nodes in the cross-modal geological coupling graph is less than the preset growth radius, and the gradient direction consistency of the implicit neural field at the pair of nodes is higher than the growth threshold, and there is no geological process causal edge between the pair of nodes, a new geological process causal edge is generated between the pair of nodes, and the direction vector of the new geological process causal edge is consistent with the gradient direction at the node pair.

10. The three-dimensional implicit modeling method for deep mineralization prediction based on the fusion of multi-source geophysical and geochemical data as described in claim 8, characterized in that, The dynamic evolution optimization and the network parameter optimization of the implicit neural field are performed alternately, including the following steps: S55, in the first training phase, the topology of the cross-modal geological coupling map is fixed, and the implicit neural field parameters are optimized until convergence; S56, in the second training phase, the implicit neural field parameters are fixed, and the edge weights and topological structure of the geological process causal edges are optimized according to the degree of consistency. S57, repeat the first training phase and the second training phase until the overall consistency between the gradient direction of the implicit neural field and the direction vector of the causal edge of the geological process satisfies the preset convergence condition.

Images