A mineralization prediction method, system, device and medium based on multi-modal geological data and a knowledge graph

Abstract

A mineralization prediction method, system, device and medium based on multi-modal geological data and knowledge graph are related to the technical field of mineral exploration. Multi-modal geological data and geological information of a target area are obtained; multi-scale feature extraction and fusion are performed on the multi-modal geological data to obtain a preliminary fusion vector; a knowledge graph related to the target area is constructed and coded based on the geological information to generate a knowledge embedding vector; the knowledge embedding vector and the preliminary fusion vector are input into a knowledge reasoning model for processing to obtain a mineralization probability value and a prediction uncertainty value of each spatial unit in the target area; when the mineralization probability value of the target spatial unit is greater than or equal to a preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to a preset uncertainty threshold, the target spatial unit in the target area is determined to be delineated as a mineral target area. By implementing the method, the problem of inaccurate prediction results caused by focusing on a single scale is solved.

Description

Technical Field

[0001] This application relates to the field of mineral exploration technology, specifically to a mineralization prediction method, system, equipment, and medium based on multimodal geological data and knowledge graphs. Background Technology

[0002] As mineral exploration enters the era of "big data" and "deep prospecting," exploration targets are becoming increasingly concealed and complex. The development of technologies such as remote sensing, geophysics, geochemistry, and geological surveys has generated massive amounts of heterogeneous, multimodal geological data. How to utilize artificial intelligence technology to efficiently and accurately predict mineralization and delineate target areas from this vast amount of multi-source geological data has become a research hotspot and an important development direction in the field of mineral exploration.

[0003] Currently, existing mineralization prediction technologies mainly employ data-driven machine learning or deep learning algorithms. For example, convolutional neural networks are used to extract features and identify anomalies from remote sensing images or geophysical and geochemical grid images, or geological, geophysical, and geochemical data from different sources are directly stitched together or simply stacked at the feature level and then input into a classification model for joint inference, thereby outputting the mineralization probability of the target area.

[0004] However, existing technologies for fusing multi-source data are too simplistic, neglecting the causal hierarchy from low-level geological phenomena to high-level metallogenic models. This results in insufficient utilization of the fused data, and the output of the component models lacks effective constraints from geoscientific knowledge, leading to inaccurate prediction results. Summary of the Invention

[0005] This application provides a mineralization prediction method, system, equipment, and medium based on multimodal geological data and knowledge graphs. This method solves the problems of insufficient information utilization and inaccurate prediction results caused by existing methods that only focus on a single scale.

[0006] Firstly, this application provides a mineralization prediction method based on multimodal geological data and knowledge graphs. The method includes: acquiring multimodal geological data, geological exploration reports, and ore deposit survey information for a target area; extracting multi-scale features from the multimodal geological data to obtain multiple feature vectors, where each feature vector corresponds to a type of scale feature in the multi-scale features; processing and fusing the multiple feature vectors based on a causal hierarchy to obtain a preliminary fused vector, where the causal hierarchy is the relationship from low-level geological phenomena to high-level mineralization models; constructing a knowledge graph related to the target area based on the geological exploration reports and ore deposit survey information, and using a graph neural network to process the knowledge graph. Encode and generate knowledge embedding vectors; input the knowledge embedding vectors and preliminary fusion vectors into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit in the target area; determine the preset probability threshold and preset uncertainty threshold based on the geological state of the target area, and compare the mineralization probability value of each spatial unit with the preset probability threshold; when the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, determine that the target spatial unit in the target area is delineated as the mineral target area, and the target spatial unit is any one of the spatial units.

[0007] By adopting the above technical solutions, multimodal geological data, geological exploration reports, and mineral deposit survey information of the target area are acquired, making full use of different types of geological information. When performing multi-scale feature extraction on the multimodal geological data, multiple feature vectors are obtained, solving the problem of insufficient information utilization caused by existing methods focusing only on a single scale. Based on the causal hierarchical relationship, multiple feature vectors are processed and fused to obtain preliminary fused features. A knowledge graph is constructed and a graph neural network is used to generate knowledge embedding vectors, which structure and deeply encode the geological exploration reports and mineral deposit survey information, enabling historical exploration experience and metallogenic theories to effectively guide the current prediction task. When the knowledge embedding vectors and preliminary fused vectors are input into the knowledge reasoning model for processing, the mineralization probability value and prediction uncertainty value of each spatial unit are obtained. Based on the geological state of the target area, the preset probability threshold and preset uncertainty threshold are dynamically determined, and the mineralization target area is delineated by combining a dual-threshold judgment mechanism, realizing site-specific adaptive decision-making, effectively reducing the risk of missed and misjudgment, and significantly improving the accuracy of mineralization prediction.

[0008] Optionally, multi-scale feature extraction is performed on the multimodal geological data to obtain multiple feature vectors. Specifically, this includes: acquiring remote sensing image data, geophysical and geochemical data, and geological vector map data from the multimodal geological data; performing multi-level feature extraction on the remote sensing image data to obtain small-scale, mesoscale, and large-scale remote sensing feature vectors, respectively; using wavelet transform to perform multi-band signal decomposition on the geophysical and geochemical data to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors, respectively; and performing multi-level semantic parsing on the geological vector map data based on geological layers to obtain... We obtain small-scale, medium-scale, and large-scale geological feature vectors; we concatenate the large-scale remote sensing feature vector, large-scale geophysical and geochemical feature vector, and large-scale geological feature vector to obtain a large-scale evidence vector; we concatenate the medium-scale remote sensing feature vector, medium-scale geophysical and geochemical feature vector, and medium-scale geological feature vector to obtain a medium-scale evidence vector; we concatenate the small-scale remote sensing feature vector, small-scale geophysical and geochemical feature vector, and small-scale geological feature vector to obtain a small-scale evidence vector; and we output the small-scale evidence vector, medium-scale evidence vector, and large-scale evidence vector as feature vectors.

[0009] By employing the above technical solutions, remote sensing image data, geophysical and geochemical exploration data, and geological vector map data are obtained from multimodal geological data. When performing multi-layer feature extraction on the remote sensing image data, feature maps of different depths are extracted using convolutional neural networks. This allows small-scale remote sensing feature vectors to capture local texture and spectral anomaly details, medium-scale remote sensing feature vectors to characterize geological body outlines and linear structural combinations, and large-scale remote sensing feature vectors to reflect the regional macro-geological background. When using wavelet transform to perform multi-band signal decomposition on the geophysical and geochemical exploration data, by separating high-frequency, medium-frequency, and low-frequency components, the small-scale geophysical and geochemical exploration feature vectors focus on local anomaly peaks, while the medium-scale geophysical and geochemical exploration feature vectors... Large-scale geophysical and geochemical feature vectors are used to characterize anomalous halos or gradient zones, and reveal regional background field characteristics, effectively solving the contradiction that traditional methods cannot simultaneously take into account local anomalies and regional background. When performing multi-level semantic analysis on geological vector map data based on geological hierarchy, semantic information is extracted layer by layer from the stratigraphy and lithology of spatial units to regional tectonic units, ensuring the hierarchical expression of geological knowledge. Remote sensing feature vectors, geophysical and geochemical feature vectors, and geological feature vectors of the same scale are spliced ​​to form evidence vectors of corresponding scales, realizing the effective integration of different data sources at the same cognitive level. The output small-scale evidence vectors, medium-scale evidence vectors, and large-scale evidence vectors are used as feature vectors.

[0010] Optionally, multiple feature vectors are processed and fused based on causal hierarchical relationships to obtain a preliminary fused vector. Specifically, this includes: calculating the similarity between the large-scale evidence vector and each background pattern in a pre-defined mineralization background pattern library; selecting the target background pattern corresponding to the maximum value from multiple similarity results; and using the confidence vector corresponding to the target background pattern as the regional background diagnosis vector. Furthermore, matrix multiplication is performed based on a pre-defined geological inference matrix and the regional background diagnosis vector to generate a mesoscale attention weight template. The pre-defined geological inference matrix is ​​an n×m dimensional matrix, where n represents the n rows of the pre-defined geological inference matrix corresponding to n mesoscale features, and m represents... The m columns of the preset geological inference matrix correspond to m regional mineralization models. The mesoscale evidence vector is multiplied element-wise with the mesoscale attention weight template to obtain the ore-controlling structure judgment vector. Based on the preset local indicator matrix and the ore-controlling structure judgment vector, matrix multiplication is performed to obtain the small-scale weight template. The preset local indicator matrix is ​​a p×q dimensional matrix, where p represents the p rows of the preset local indicator matrix corresponding to p small-scale mineral exploration marker features, and q represents the q columns of the preset local indicator matrix corresponding to q mesoscale ore-controlling structure types. The small-scale evidence vector is multiplied element-wise with the small-scale weight template to obtain the preliminary fusion vector.

[0011] By employing the above technical solutions, the similarity between the large-scale evidence vector and various background patterns in the preset metallogenic background pattern library is calculated. The target background pattern corresponding to the maximum similarity is selected, and the confidence vector is extracted as the regional background diagnostic vector, accurately locating the macroscopic metallogenic geological background type of the target area. When generating the mesoscale attention weight template by matrix multiplication based on the preset geological inference matrix and the regional background diagnostic vector, the interference of mesoscale features unrelated to the current metallogenic background is effectively suppressed when the mesoscale evidence vector and the mesoscale attention weight template are multiplied element-wise to obtain the ore-controlling structure judgment vector. When calculating the small-scale weight template based on the preset local indicator matrix and the ore-controlling structure judgment vector, the guiding role of specific ore-controlling structure types on each small-scale prospecting indicator feature is characterized by the matrix element values. The preliminary fusion vector is obtained by multiplying the small-scale evidence vector and the small-scale weight template element-wise, ensuring that the weight allocation of small-scale prospecting indicator features conforms to the upper-level geological logic, avoiding the problems of weight imbalance and information redundancy of different scale features in the traditional simple splicing method, and realizing the orderly integration and efficient utilization of multi-scale geological information under the causal hierarchy framework.

[0012] Optionally, a knowledge graph related to the target area is constructed based on geological exploration reports and mineral deposit survey information. A graph neural network is then used to encode the knowledge graph to generate knowledge embedding vectors. Specifically, this includes: extracting geological knowledge triples from the geological exploration reports and mineral deposit survey information, storing these triples in a graph database to construct a global knowledge graph related to the target area; dividing the target area into multiple spatial units, identifying the core geological entities contained within each spatial unit, and performing a correlation retrieval in the global knowledge graph centered on these core geological entities to extract local knowledge subgraphs corresponding to the target spatial units. The current spatial unit is any grid unit among the multiple spatial units. A multi-round iterative aggregation encoding is performed on each node in the local knowledge subgraph based on a graph neural network to obtain the node embedding vectors of each node in the local knowledge subgraph; graph pooling is performed on the node embedding vectors of all nodes in the local knowledge subgraph to obtain the initial knowledge embedding vector corresponding to the target spatial unit; and the initial knowledge embedding vectors corresponding to all spatial units are summarized to obtain the knowledge embedding vector.

[0013] By adopting the above technical solution, geological knowledge triples are extracted from geological exploration reports and mineral deposit survey information and stored in a graph database to construct a global knowledge graph. After dividing the target area into multiple spatial units, core geological entities are identified for the current spatial unit, and local knowledge subgraphs are extracted through association retrieval in the global knowledge graph. This achieves accurate mapping from global knowledge to local prediction tasks. When obtaining node embedding vectors by performing multi-round iterative aggregation encoding on each node in the local knowledge subgraph based on graph neural networks, the message passing and feature aggregation mechanism between nodes ensures that the representation of each geological entity not only includes its own attribute information but also integrates the association information of neighboring nodes. Graph pooling is performed on the node embedding vectors of all nodes in the local knowledge subgraph to generate initial knowledge embedding vectors. The initial knowledge embedding vectors corresponding to all spatial units are summarized to obtain the knowledge embedding vector, providing personalized expert knowledge guidance for each spatial unit and enhancing the model's generalization ability and prediction reliability under conditions of scarce samples.

[0014] Optionally, the knowledge reasoning model includes an evidence questioning module, an evidence construction module, and an evidence evaluation module. The knowledge embedding vector and the preliminary fusion vector are input into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit within the target area. Specifically, this includes: deconstructing the preliminary fusion vector into a multi-level data evidence set; using the evidence questioning module, guided by the knowledge embedding vector, performing correlation questioning on the data evidence set to obtain a reasoned evidence vector; concatenating the reasoned evidence vector with the knowledge embedding vector using the evidence construction module to generate a comprehensive evidence vector; inputting the comprehensive evidence vector into the evidence network of the evidence evaluation module, mapping it through the output layer of the evidence network and a non-negative activation function to output the first evidence strength supporting the mineralization hypothesis and the second evidence strength supporting the non-mineralization hypothesis; calculating 1 plus the sum of the first and second evidence strengths to obtain the evidence sum, and dividing 1 by the evidence sum to obtain the prediction uncertainty value; calculating the sum of 1 and the first evidence strength to obtain the numerator evidence value; calculating the sum of the first and second evidence strengths with the preset values ​​corresponding to the number of categories in the classification problem to obtain the denominator evidence value; and dividing the numerator evidence value by the denominator evidence value to obtain the mineralization probability value.

[0015] By adopting the above technical solution, the initial fusion vector is deconstructed into a multi-level data evidence set. The evidence questioning module uses the knowledge embedding vector as a guide to perform correlation questioning on the data evidence set to obtain the inferred evidence vector. This achieves effective correction and semantic enhancement of data-driven features by expert knowledge, effectively solving the problem of lack of domain knowledge guidance in pure data-driven methods. When the evidence construction module concatenates the inferred evidence vector with the knowledge embedding vector to generate a comprehensive evidence vector, it retains the data evidence information after knowledge questioning and directly incorporates structured expert knowledge representation. When the comprehensive evidence vector is input into the evidence network of the evidence evaluation module and the first and second evidence strengths are output through the output layer and non-negative activation function, the confidence level of evidence supporting the mineralization hypothesis and the non-mineralization hypothesis is quantified, respectively. By calculating the evidence sum and dividing it by 1, the prediction uncertainty value is obtained, which effectively quantifies the reliability of the model prediction result. Dividing the numerator evidence value by the denominator evidence value yields the mineralization probability value. Based on subjective logic theory, a reasonable conversion from evidence strength to probability judgment is achieved, avoiding the defect of traditional methods that only output a single probability value and ignore the prediction confidence.

[0016] Optionally, the evidence questioning module, guided by the knowledge embedding vector, performs correlation questioning on the data evidence set to obtain the inferred evidence vector. Specifically, this includes: projecting the knowledge embedding vector into the first linear transformation layer of the evidence questioning module to generate a query vector; projecting each sub-feature vector in the data evidence set into the second and third linear transformation layers of the evidence questioning module to generate key vectors and value vectors respectively; calculating the inner product of the query vector and each key vector, and normalizing it using a normalized exponential function to obtain the corresponding attention weights; and using the attention weights to perform a weighted summation of all value vectors to generate the inferred evidence vector.

[0017] By employing the aforementioned technical solution, when the knowledge-embedded vector is projected into the first linear transformation layer of the evidence inquiry module to generate a query vector, expert knowledge is transformed into a query representation form that can interact with data evidence. This allows the mineralization patterns and causal logic contained in the knowledge graph to actively retrieve relevant information from the data evidence. When each sub-feature vector in the data evidence set is input into the second and third linear transformation layers for projection to generate corresponding key and value vectors, and the inner product of the query vector and each key vector is calculated and normalized using a normalized exponential function to obtain the attention weight, the knowledge-embedded vector is applied to different sub-features in the data evidence set. The quantitative assessment of vector correlation enables data features highly correlated with mineralization knowledge to receive higher weight allocations. When generating the post-inference evidence vector by weighted summation of all value vectors using attention weights, dynamic weighted fusion selectively aggregates the multi-level feature information scattered in the data evidence set according to the importance of knowledge guidance. This ensures that the post-inference evidence vector can fully absorb effective information consistent with expert knowledge while filtering out noise interference. It not only realizes the interactive fusion of knowledge and data, but also provides flexible adaptability through learnable projection transformation and dynamic weight allocation, effectively solving the limitation of fixed weight fusion methods in adapting to different geological scenarios.

[0018] Optionally, a preset probability threshold and a preset uncertainty threshold are determined based on the geological conditions of the target area. Specifically, this includes: acquiring the ore-controlling element characteristics of each spatial unit within the target area, including fault structural density, known mineral deposit distribution density, and stratigraphic mineralization favorability; calculating the geological mineralization favorability index of each spatial unit by weighted summation based on the fault structural density, known mineral deposit distribution density, and stratigraphic mineralization favorability; calculating the preset probability threshold corresponding to each spatial unit based on the geological mineralization favorability index; and calculating the preset uncertainty threshold corresponding to each spatial unit based on the geological mineralization favorability index.

[0019] By adopting the above technical solution, the ore-controlling element characteristics of each spatial unit within the target area are obtained. The density of fault structural lines, the distribution density of known mineral occurrences, and the mineralization favorability of stratigraphy are selected. Based on these three ore-controlling element characteristics, a weighted summation is performed to calculate the geological mineralization favorability index. Multi-dimensional geological information is comprehensively quantified into a unified regional mineralization potential evaluation index. Based on the geological mineralization favorability index, a preset probability threshold corresponding to each spatial unit is calculated, effectively solving the problem that the traditional fixed threshold method is difficult to adapt to different geological scenarios. Based on the geological mineralization favorability index, a preset uncertainty threshold is calculated to ensure that the uncertainty tolerance is relaxed in areas with complex geological conditions and unclear mineralization background, while the confidence requirement is increased in favorable areas with clear geological conditions. This dual-threshold adaptive adjustment method fully considers the geological differences of different spatial units, achieves a dynamic balance between accurate delineation and risk control, significantly reduces the risk of missed and misjudgments, and improves the scientificity and practicality of target area delineation.

[0020] The second aspect of this application provides a mineralization prediction system based on multimodal geological data and knowledge graphs. The system includes an acquisition unit, an extraction and fusion unit, a construction unit, and a prediction unit. The acquisition unit acquires multimodal geological data, geological exploration reports, and mineral deposit survey information for the target area. The extraction and fusion unit performs multi-scale feature extraction on the multimodal geological data to obtain multiple feature vectors, where each feature vector corresponds to a type of scale feature in the multi-scale features. Based on a causal hierarchy, the multiple feature vectors are processed and fused to obtain a preliminary fused vector, where the causal hierarchy represents the relationship from low-level geological phenomena to high-level mineralization models. The construction unit constructs knowledge graphs related to the target area based on the geological exploration reports and mineral deposit survey information. The system generates a knowledge graph and encodes it using a graph neural network to generate knowledge embedding vectors. A prediction unit processes the knowledge embedding vectors and preliminary fusion vectors into a knowledge reasoning model to obtain the mineralization probability value and prediction uncertainty value of each spatial unit within the target area. Based on the geological conditions of the target area, a preset probability threshold and a preset uncertainty threshold are determined, and the mineralization probability value of each spatial unit is compared with the preset probability threshold. When the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, the target spatial unit within the target area is designated as a mineralization target area. The target spatial unit can be any one of the various spatial units.

[0021] In a third aspect, this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory, causing the electronic device to perform any of the methods described above in this application.

[0022] In a fourth aspect, this application provides a computer-readable storage medium storing instructions that, when executed, perform any of the methods described above in this application.

[0023] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Acquire multimodal geological data, geological exploration reports, and mineral deposit survey information of the target area, making full use of different types of geological information; when extracting multi-scale features from multimodal geological data, multiple feature vectors are obtained, solving the problem of insufficient information utilization caused by existing methods focusing only on a single scale; multiple feature vectors are processed and fused based on causal hierarchical relationships to obtain preliminary fused features; a knowledge graph is constructed and knowledge embedding vectors are generated using graph neural networks to structurally express and deeply encode geological exploration reports and mineral deposit survey information, enabling historical exploration experience and metallogenic theories to effectively guide current prediction tasks; when the knowledge embedding vectors and preliminary fused vectors are input into the knowledge reasoning model for processing, the mineralization probability value and prediction uncertainty value of each spatial unit are obtained; based on the geological state of the target area, preset probability thresholds and preset uncertainty thresholds are dynamically determined, and the mineralization target area is delineated by combining a dual-threshold judgment mechanism, realizing site-specific adaptive decision-making, effectively reducing the risk of missed and misjudgments, and significantly improving the accuracy of mineralization prediction. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a mineralization prediction method based on multimodal geological data and knowledge graphs provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a mineralization prediction system based on multimodal geological data and knowledge graph provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

[0025] Explanation of reference numerals in the attached figures: 201, acquisition unit; 202, extraction and fusion unit; 203, construction unit; 204, prediction unit; 300, electronic device; 301, processor; 302, memory; 303, user interface; 304, network interface; 305, communication bus. Detailed Implementation

[0026] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0028] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0029] Therefore, how to improve the accuracy of mineralization prediction by addressing the insufficient utilization of geological information in existing technologies is a pressing issue. This application provides a mineralization prediction method based on multimodal geological data and knowledge graphs, applied in a server. The server in this application can be a platform providing mineralization prediction services to geological exploration units, mining companies, and research institutions. Figure 1 This is a flowchart illustrating a mineralization prediction method based on multimodal geological data and knowledge graphs provided in an embodiment of this application. (Refer to...) Figure 1 The method includes the following steps S101-S107.

[0030] S101: Acquire multimodal geological data, geological exploration reports, and mineral deposit survey information for the target area.

[0031] In S101 above, the target area refers to a specific geographic spatial range for in-depth metallogenic prediction, delineated based on regional geological background surveys, preliminary mineral potential assessments, or existing exploration work instructions. This area can be a prospective metallogenic area within a provincial or municipal administrative region, an extension of a known mineralized area, or an anomalous concentrated distribution area delineated based on remote sensing alteration information. The spatial scale is typically between tens and thousands of square kilometers, and it must possess a certain metallogenic geological background and preliminary prospecting clues. After determining the target area, the multi-source data interface submodule in the data preprocessing and fusion module acquires multimodal geological data, geological exploration reports, and ore deposit survey information for the target area. The reason for acquiring these three types of data simultaneously is that metallogenesis is a complex geological process, and a single type of data cannot fully characterize its multidimensional features and genetic mechanisms. Multimodal geological data can provide complementary observational information from different physical and chemical field perspectives, while geological exploration reports and ore deposit survey information embody the field experts' understanding of metallogenic regularities and a summary of historical exploration results.

[0032] After identifying the target area, the multi-source data interface submodule within the data preprocessing and fusion module acquires multimodal geological data, geological exploration reports, and ore deposit survey information for the target area. The reason for acquiring these three types of data simultaneously is that mineralization is a complex geological process, and a single type of data cannot fully characterize its multidimensional features and genetic mechanisms. Multimodal geological data can provide complementary observational information from different physical and chemical fields, while geological exploration reports and ore deposit survey information embody the field experts' understanding of mineralization regularities and summaries of historical exploration results. In the specific implementation process, remote sensing image data of the target area is retrieved from the databases of national or local geological survey institutions, including multispectral satellite imagery such as the Landsat series, Gaofen series, or Sentinel-2 data. This remote sensing image data reflects surface lithology, structural linearity, and alteration information. Spectral characteristics can identify the spatial distribution patterns of wall-rock alteration zones and tectonic fracture zones related to mineralization. Simultaneously, geophysical and geochemical data are acquired. Geophysical data includes observations of geophysical fields such as gravity anomalies, magnetic anomalies, induced polarization, or audio-frequency magnetotelluric sounding. Geochemical data includes soil geochemical measurements, stream sediment measurements, or rock geochemical analysis results. This geophysical and geochemical data can reveal the physical characteristics and elemental enrichment patterns of concealed ore bodies or mineralization anomalies, providing crucial evidence for identifying blind and deep ore deposits. Further geological vector map data is collected, including regional geological maps, structural outline maps, and mineral geological maps in vector format. This data records basic geological information such as stratigraphic distribution, intrusive relationships of rock masses, fault structure distribution, and the location of known mineral occurrences, providing a spatial reference framework for metallogenic background analysis and identification of ore-controlling elements. Multimodal geological data also includes geochemical point source data and text report data. The geochemical point source data requires the processing of elemental content analysis results from each sampling point within the target area. These discrete point data reflect the spatial distribution patterns of chemical elements and the enrichment anomalies of ore-forming elements.

[0033] Furthermore, through literature retrieval systems and geological archive management platforms, relevant geological exploration reports and mineral deposit survey information for the target area are obtained. Geological exploration reports are typically in PDF or Word format and contain unstructured or semi-structured content such as drilling columnar sections, laboratory analysis results, geological profiles, and prospecting conclusions. Mineral deposit survey information covers structured knowledge summarized by experts, including the type of discovered mineral deposits, metallogenic epochs, ore-controlling factors, and ore grades. This textual data contains rich knowledge of metallogenic regularities and practical experience. During data acquisition, it is necessary to ensure that the cloud cover rate of remote sensing image data is below a set threshold to guarantee image quality, that the sampling density and measurement accuracy of geophysical and geochemical data meet the prediction scale requirements, that the scale of geological vector map data matches the target area, and that the timeliness and authority of geological exploration reports and mineral deposit survey information are verified.

[0034] S102: Multi-scale feature extraction is performed on multimodal geological data to obtain multiple feature vectors. Each feature vector corresponds to one type of scale feature in the multi-scale features.

[0035] In S102 above, after acquiring multimodal geological data of the target area, it is necessary to systematically extract multi-scale features from the multimodal geological data to construct a hierarchical characterization of mineralization evidence. The multi-scale feature extraction strategy is adopted because mineralization itself is a geological process that spans multiple spatial scales, from microscopic mineral alteration and local mineralization enrichment to mesoscopic ore-controlling structural combinations and macroscopic regional metallogenic background. Geological phenomena at different scales have different levels of indicative significance for mineralization prediction. Single-scale features are difficult to fully characterize such multi-level mineralization control factors, while multi-scale feature extraction can systematically capture the spatial distribution patterns and hierarchical relationships of mineralization information from details to the whole.

[0036] Furthermore, multi-scale feature extraction was performed on the multimodal geological data to obtain multiple feature vectors. Specifically, this included: acquiring remote sensing image data, geophysical and geochemical data, and geological vector map data from the multimodal geological data; performing multi-level feature extraction on the remote sensing image data to obtain small-scale, mesoscale, and large-scale remote sensing feature vectors, respectively; using wavelet transform to perform multi-band signal decomposition on the geophysical and geochemical data to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors, respectively; and performing multi-level semantic parsing on the geological vector map data based on geological levels to obtain... We obtain small-scale, medium-scale, and large-scale geological feature vectors; we concatenate the large-scale remote sensing feature vector, large-scale geophysical and geochemical feature vector, and large-scale geological feature vector to obtain a large-scale evidence vector; we concatenate the medium-scale remote sensing feature vector, medium-scale geophysical and geochemical feature vector, and medium-scale geological feature vector to obtain a medium-scale evidence vector; we concatenate the small-scale remote sensing feature vector, small-scale geophysical and geochemical feature vector, and small-scale geological feature vector to obtain a small-scale evidence vector; and we output the small-scale evidence vector, medium-scale evidence vector, and large-scale evidence vector as feature vectors.

[0037] Specifically, three types of basic data—remote sensing image data, geophysical and geochemical exploration data, and geological vector map data—are separated from multimodal geological data for independent multi-scale feature extraction. For multi-layer feature extraction of remote sensing image data, a pre-trained convolutional neural network, such as ResNet-50 or VGG-16, is used as the feature extractor. The multispectral remote sensing image of the target area is input into the convolutional neural network. Due to the hierarchical feature learning capability of the convolutional neural network, the shallow convolutional layers mainly capture local details such as edges and textures, the middle convolutional layers gradually integrate to form the ability to recognize the outline and linear structure of geological bodies, and the deep convolutional layers extract highly abstract semantic features and regional background information. After the second or third convolutional layer of the convolutional neural network, a feature extraction branch is connected. The feature map of this layer is reduced to a one-dimensional vector by global average pooling to obtain a small-scale remote sensing feature vector. The dimension of this small-scale remote sensing feature vector is, for example, 256 dimensions. The value of each dimension reflects the local texture intensity of the input image under the response of a specific convolutional kernel. For example, when there is a spectral anomaly caused by kaolinite alteration in a certain spatial unit, the corresponding feature vector will show a significantly high value in the dimension related to hydroxyl absorption features.

[0038] After the fourth or fifth convolutional block in the intermediate layer of the convolutional neural network, global average pooling is used to extract mesoscale remote sensing feature vectors, with a dimension of, for example, 512. These feature vectors can characterize the spatial distribution patterns of ring-shaped image structures or linear tectonic zones. For example, when a complete ring-shaped alteration halo is identified, the mesoscale remote sensing feature vector will generate an activation response in the dimension representing circular or elliptical geometric features. Before the last fully connected layer of the convolutional neural network or through global pooling of deep feature maps, large-scale remote sensing feature vectors are extracted, with a dimension of, for example, 1024. This vector comprehensively encodes the geological background of the entire spatial unit, such as whether it is located within a volcanic rock zone, the overall level of regional alteration intensity, and macroscopic information such as spatial similarity to known mineralized areas.

[0039] Mesoscale remote sensing feature vectors are further extracted from the mid-level feature maps of the convolutional neural network. These mid-level feature maps correspond to the outputs of the third to fourth convolutional layers of the network, with a receptive field extending to hundreds of pixels. They can integrate local texture information to form an abstract expression of the outline, shape regularity, and linear structural combination patterns of geological bodies. Mesoscale remote sensing feature vectors are mainly used to characterize medium-scale mineralization control elements such as the boundary morphology of rock masses, the extension of fault zones, and the intersection relationships of multiple sets of structures. Large-scale remote sensing feature vectors are extracted from the deep feature maps or global average pooling layers of the convolutional neural network. The deep feature maps correspond to the outputs of the fifth convolutional layer or deeper layers of the network, with a receptive field covering the entire input image or its main part. They can integrate global information to form a high-level semantic representation of the overall geological background, main structural framework, and lithofacies zoning patterns of the region. Large-scale remote sensing feature vectors are mainly used to characterize factors that play a macroscopic control role in mineralization, such as the location of tectonic rocks, the distribution of regional magmatic rock belts, and the dominant tectonic system.

[0040] For multi-band signal decomposition of geophysical and geochemical data, the data is resampled from the original point source format or irregular grid to the same regular grid as the remote sensing image data through Kriging interpolation or inverse distance weighted interpolation methods to form standardized two-dimensional field data. For example, geochemical element content data is interpolated to a 500-meter resolution raster map, and magnetic anomaly data is interpolated to a magnetic anomaly field of the same resolution. Wavelet transform is applied to each geophysical and geochemical data layer for multi-scale decomposition. Daubechies wavelet or Haar wavelet is selected as the wavelet basis function. The original signal is separated into components of different frequency bands through multi-level wavelet decomposition. Three-level wavelet decomposition is performed on the geochemical data. The first level of decomposition obtains high-frequency detail coefficients and low-frequency approximation coefficients. The high-frequency detail coefficients reflect the local spikes and abrupt changes in geochemical anomalies. After reconstructing the high-frequency coefficients into a spatial distribution map, the statistical features of the high-frequency field within each spatial cell, such as the maximum value, mean, and standard deviation, are extracted and combined to form a small-scale geophysical and geochemical feature vector. For example, a vector containing the maximum value of high-frequency anomalies of copper, the mean value of high-frequency anomalies of gold, etc., totaling 6 dimensions. This small-scale geophysical and geochemical feature vector can accurately locate the locally enriched mineralization center or anomaly point. For example, when there is a peak of copper content of 1000 ppm in a certain 500-meter grid cell, the dimension of the maximum value of high-frequency anomalies of copper in the small-scale geophysical and geochemical feature vector will record the intensity of the anomaly.

[0041] The low-frequency approximation coefficients are further decomposed into a second-level wavelet decomposition. The resulting mid-frequency components reflect the anomalous gradient bands or halos at a medium spatial scale. After reconstructing the mid-frequency components, the mid-frequency field mean, gradient direction, and gradient amplitude of each spatial unit are extracted to form a mid-scale geophysical and geochemical feature vector. For example, a vector containing the mid-frequency anomalous mean of mercury and the gradient angle indicating the direction of ore-forming fluid activity, etc., totaling 8 dimensions, can characterize the migration path and alteration zoning characteristics of the ore-forming hydrothermal system. The low-frequency components obtained by continuing the third-level wavelet decomposition represent the regional geophysical or geochemical background field. The mean and coefficient of variation of this low-frequency field within the spatial unit are extracted to form a large-scale geophysical and geochemical feature vector. For example, a vector containing the low-frequency background intensity of the magnetic field and the regional anomaly trend value of the gravity field, etc., is a total of 4 dimensions. This large-scale geophysical and geochemical feature vector reflects the overall characteristics of the regional geophysical field in which the spatial unit is located, such as whether it is located in a region with high regional magnetic anomaly or within a gravity gradient zone, providing a basis for judging whether the region has a large-scale mineralization geophysical background.

[0042] For multi-level semantic parsing of geological vector map data, the vector data such as geological layers, structural layers and mineral layers are first imported into the spatial analysis module. Through spatial overlay and attribute query operations, the geological elements are extracted hierarchically for each spatial unit to be predicted. At the small-scale level, a spatial intersection query is performed on each spatial unit to extract the stratigraphic lithological attributes of the location where the unit's center point falls. For example, if the lithological code obtained by the query is "J2a", representing Middle Jurassic andesite, the lithological code is converted into a numerical feature through a preset lithological coding table. For example, andesite is mapped to lithological category code 15. At the same time, the distance between the spatial unit and the nearest fault line and the fault's attitude attributes are extracted. For example, the nearest fault distance is 200 meters, the fault strike is 310 degrees northeast, and the dip angle is 75 degrees. These attribute values ​​are combined to form a small-scale geological feature vector. For example, a vector containing lithological code, distance to the nearest fault, fault strike angle, and fault dip angle, etc., totaling 4 dimensions. This small-scale geological feature vector accurately describes the direct geological environment of the spatial unit, such as whether the unit is located on favorable volcanic lithology and whether it is adjacent to favorable ore-guiding structures. At the mesoscale level, the focus shifts from solely on point attributes to extracting stratigraphic information of the spatial unit's stratigraphic system and secondary structural attribution information of adjacent faults. For example, by querying a geological map database, it can be determined that the spatial unit belongs to the "Middle Jurassic" stratigraphic system, which is labeled as "the middle segment of the XX-level fault zone" in mineralization analysis. The stratigraphic system is coded, for example, as stratigraphic level 3, and the secondary structural zone is coded as structural attribution code 7. These are combined to form a mesoscale geological feature vector, such as a 3-dimensional vector containing stratigraphic system stratigraphic level, secondary structural zone attribution code, and total fault length within the unit. This mesoscale geological feature vector can characterize the spatial unit's position within the regional stratigraphic framework and structural system, such as whether the unit is located in a favorable position within a known ore-controlling fault zone.

[0043] At a large scale, information about the tectonic unit in which the spatial unit is located and the regional tectonic zone attributes of the adjacent faults are extracted. For example, by querying the tectonic outline map, it can be found that the spatial unit is located within the tectonic unit of the "western margin of the Yangtze Plate" and the nearby main fault belongs to the regional compressional-shear tectonic zone of the "Longmenshan Fault Zone". The tectonic unit is encoded, for example, as tectonic domain code 2, and the regional tectonic zone is encoded as tectonic system code 5, forming a large-scale geological feature vector. For example, a 2-dimensional vector containing tectonic domain code, regional tectonic zone system code, etc. This large-scale geological feature vector reveals the regional tectonic background and large-scale ore-controlling environment of the spatial unit, such as whether the region belongs to the plate margin active zone or the regional metallogenic zone.

[0044] After completing the multi-scale feature extraction of the three types of data, the next stage is the hierarchical splicing and fusion of feature vectors. The purpose of this stage is to integrate different modal features at the same scale level to form a comprehensive multimodal evidence vector. First, the large-scale remote sensing feature vector, the large-scale geophysical and geochemical feature vector, and the large-scale geological feature vector are spliced ​​according to their feature dimensions. For example, the large-scale remote sensing feature vector is 1024-dimensional, the large-scale geophysical and geochemical feature vector is 4-dimensional, and the large-scale geological feature vector is 2-dimensional. The three are concatenated through a vector splicing function to form a 1030-dimensional large-scale evidence vector. This vector comprehensively encodes the regional background information of the spatial unit in the three dimensions of remote sensing, geophysical and geochemical exploration, and geology. It can comprehensively reflect whether the unit has macroscopic metallogenic geological conditions, such as whether it is located in a favorable tectonic environment, whether it is located in a regional geophysical anomaly zone, and whether it has macroscopic remote sensing alteration features related to metallogenesis. Similarly, by concatenating the 512-dimensional mesoscale remote sensing feature vector, the 8-dimensional mesoscale geophysical and geochemical feature vector, and the 3-dimensional mesoscale geological feature vector, a 523-dimensional mesoscale evidence vector is obtained. This vector characterizes the geological body combination features, anomalous gradient zone distribution, and tectonic unit affiliation of the spatial unit at the mesoscale, enabling the determination of whether the unit is located within a favorable segment of a mineralization alteration zone or within the influence range of a secondary ore-controlling structure. Concatenating the 256-dimensional small-scale remote sensing feature vector, the 6-dimensional small-scale geophysical and geochemical feature vector, and the 4-dimensional small-scale geological feature vector yields a 266-dimensional small-scale evidence vector. This vector finely describes the local anomaly details, direct surrounding rock properties, and occurrence of adjacent ore-controlling elements of the spatial unit, enabling the identification of microscopic indicators of mineralization alteration within the unit. Finally, the small-scale, mesoscale, and large-scale evidence vectors are output as three independent feature vectors, which together constitute a multi-scale, multi-modal holographic feature representation of the spatial unit.

[0045] S103: Based on the causal hierarchy, multiple feature vectors are processed and fused to obtain a preliminary fused vector. The causal hierarchy is the relationship from low-level geological phenomena to high-level mineralization models.

[0046] In S103 above, after completing the multi-scale feature extraction, feature vector processing and fusion based on causal hierarchical relationship are performed to simulate the geologist's "from macro to micro, top to bottom" mineralization analysis logic. Because mineralization has obvious hierarchical control characteristics, that is, the regional mineralization background determines the possible mineralization type and ore-controlling structure style, and the ore-controlling structure further constrains the expression of local mineral exploration indicators. If the feature vectors of the three scales are simply spliced ​​or averaged and fused, this causal progression relationship will be ignored, leading to logical contradictions between features of different scales.

[0047] Multiple feature vectors are processed and fused based on causal hierarchical relationships to obtain a preliminary fused vector. Specifically, this includes: calculating the similarity between the large-scale evidence vector and various background patterns in a pre-set mineralization background pattern library; selecting the target background pattern corresponding to the maximum value from multiple similarity results; and using the confidence vector corresponding to the target background pattern as the regional background diagnosis vector. Furthermore, matrix multiplication is performed based on a pre-set geological inference matrix and the regional background diagnosis vector to generate a mesoscale attention weight template. The pre-set geological inference matrix is ​​an n×m dimensional matrix, where n represents the n rows of the pre-set geological inference matrix corresponding to n mesoscale features, and m represents the pre-... Let m columns of the geological inference matrix correspond to m regional mineralization models; multiply the mesoscale evidence vector with the mesoscale attention weight template element-wise to obtain the ore-controlling structure judgment vector; perform matrix multiplication based on the preset local indicator matrix and the ore-controlling structure judgment vector to obtain the small-scale weight template. The preset local indicator matrix is ​​a p×q dimensional matrix, where p represents the p rows of the preset local indicator matrix corresponding to p small-scale mineral exploration marker features, and q represents the q columns of the preset local indicator matrix corresponding to q mesoscale ore-controlling structure types; multiply the small-scale evidence vector with the small-scale weight template element-wise to obtain the preliminary fusion vector.

[0048] Specifically, the highest-level large-scale evidence vector is processed first. The large-scale evidence vector contains regional background information of the spatial unit. The large-scale evidence vector is then matched with a preset metallogenic background pattern library to diagnose the metallogenic background type to which the unit most likely belongs. The pre-defined metallogenic background model library is a knowledge base constructed through a systematic summary of historical metallogenic data and geological literature. It stores standard feature templates for various typical regional metallogenic models. For example, the metallogenic background model library contains a total of m background models, such as "island arc environment porphyry copper deposit model", "collision orogenic belt orogenic gold deposit model", and "cratonic rift type lead-zinc deposit model". Each background model corresponds to a standard feature vector. The dimension of the standard feature vector is the same as that of the large-scale evidence vector, which is 1030 dimensions. The values ​​of each dimension of the standard feature vector represent the typical range or expected value of various geological elements under the metallogenic background model. For example, in the standard feature vector of the "island arc environment porphyry copper deposit model", the value of the tectonic domain coding dimension is 2, representing an active continental margin; the value of the regional magnetic anomaly background intensity dimension is a high magnetic anomaly interval; and the value of the large-scale remote sensing alteration index dimension is a strong hydroxyl alteration response interval. In practice, the large-scale evidence vector of the current spatial unit is compared with the m standard feature vectors in the mineralization background pattern library to calculate their similarity. Cosine similarity is used as the metric, and the formula is the inner product of the two vectors divided by the product of their respective norms. For example, let X be the large-scale evidence vector of the current spatial unit and T be the standard feature vector of the i-th background pattern. i Then the similarity S i Calculated as: S i=X×T i / (||X||2×||T i ||2), by traversing m background patterns, m similarity values ​​S1 to Sm are calculated, and the maximum value Smax and its corresponding index k are selected. The background pattern indicated by the index k is the target background pattern. For example, if the calculation result shows that the similarity between the current spatial unit and the "island arc environment porphyry copper deposit pattern" reaches 0.87, which is the highest value, then it is determined that the spatial unit is most likely in the porphyry copper deposit mineralization background.

[0049] After determining the target background pattern, the confidence vector corresponding to that pattern is obtained as the regional background diagnostic vector. The confidence vector is an m-dimensional vector pre-stored in the metallogenic background pattern library and associated with each background pattern. Each dimension of this vector represents the confidence level allocation of the current diagnostic result to the m background patterns. For example, when the target background pattern is the k-th pattern, "island arc environment porphyry copper deposit pattern", the corresponding confidence vector has the highest value in the k-th dimension, such as 0.8, and has the second highest value in the dimensions of other patterns with similar geological environments, such as the "post-collision extensional environment porphyry molybdenum deposit pattern", which has a dimension of 0. .15, which is zero in completely incompatible model dimensions, such as the "sedimentary basin stratabound lead-zinc deposit model" dimension being 0. This soft confidence assignment rather than a hard single classification design can more realistically reflect the uncertainty and ambiguity of geological background diagnosis. This m-dimensional confidence vector is extracted and denoted as the regional background diagnosis vector B. For example, in a model library containing 10 background models, B is a 10-dimensional vector with a numerical distribution of [0,0.8,0,0.15,0,0,0.05,0,0,0]. This vector quantifies the probability of the current spatial unit being in various metallogenic backgrounds.

[0050] The second level, causal guidance, involves adjusting the weights of features in the mesoscale evidence vector based on the regional background diagnostic vector. This process is achieved through a pre-set geological inference matrix, a knowledge matrix compiled by geological experts based on metallogenic theories and exploration experience. This matrix has n×m dimensions, where n represents the number of feature dimensions of the mesoscale evidence vector (e.g., 523 dimensions), and m represents the number of regional metallogenic models (e.g., 10 types). The element value W in the i-th row and j-th column of the matrix is... ij This indicates the importance weight of the i-th mesoscale geological feature within the context of the j-th regional metallogenic model. For example, in the column "Island Arc Environment Porphyry Copper Deposit Model", the weight of mesoscale features closely related to porphyry mineralization, such as "ring-shaped image structure features", is set to 0.9, and the weight of "mid-frequency copper anomaly gradient zone features" is 0.85. Meanwhile, the weight of features less related to porphyry mineralization, such as "layered sulfide anomaly features", is only 0.2. This weighting configuration reflects the differentiated requirements of different metallogenic backgrounds for ore-controlling elements in geological principles.

[0051] The pre-defined geological inference matrix (n×m dimensional) and the m-dimensional regional background diagnostic vector B are multiplied using a matrix-vector multiplication operation. Mathematically, this is expressed as the inner product of each row of the matrix and vector B. For example, the result of the calculation for the i-th row is the sum of the product of the m elements of that row with the corresponding m elements of B, resulting in a scalar value 'a'. i After traversing n rows, an n-dimensional vector A is obtained. This vector A is the mesoscale attention weight template. For example, when n is 523 and m is 10, the A obtained by matrix multiplication is a 523-dimensional vector. Among them, the feature dimensions that are highly correlated with the current background pattern receive higher weight values. For example, the weight of the ring image structure feature dimension is 0.76, while the weight of the irrelevant feature dimension is close to zero. The role of this weight template is to tell the system which mesoscale geological features should be focused on and which irrelevant features should be ignored under the current diagnosed mineralization background conditions.

[0052] After obtaining the mesoscale attention weight template, the mesoscale attention weight template is multiplied element-wise with the mesoscale evidence vector, that is, the corresponding dimensions are multiplied one by one. For example, if the mesoscale evidence vector is 523-dimensional and denoted as vector C, and the mesoscale attention weight template is 523-dimensional and denoted as vector A, the element-wise multiplication yields a new 523-dimensional vector D, where Di = Ci × Ai. This operation realizes dynamic weighted modulation of the mesoscale evidence vector, which enhances the mesoscale features that match the current mineralization background and suppresses the features that do not match. For example, the original value of the ring image structure feature dimension in the original mesoscale evidence vector is 0.65, which becomes 0.494 after modulation with a weight of 0.76, while the original value of the layered sulfide anomaly feature dimension is 0.70, which is attenuated to 0.14 after modulation with a weight of 0.2. Although the latter original value is higher, its importance is correctly reduced after background-guided modulation. The modulated vector D is named the ore-controlling structure judgment vector. This vector comprehensively reflects the intensity distribution of the ore-controlling structure features of the spatial unit under the constraints of a specific mineralization background.

[0053] Finally, we enter the third level of causal guidance, which requires adjusting the weights of each feature in the small-scale evidence vector based on the ore-controlling structure judgment vector. This process is achieved through a pre-set local indicator matrix. This pre-set local indicator matrix is ​​also a knowledge matrix constructed based on geological prospecting experience, with dimensions p×q. Here, p represents the feature dimension of the small-scale evidence vector, for example, 266 dimensions, and q represents the number of types of mesoscale ore-controlling structures. This number can be determined through cluster analysis of the mesoscale evidence vector or predefined by experts. For example, defining 12 typical ore-controlling structure types includes "Hercyan magma segregation type copper-nickel deposit or nickel-as-associated element annular alteration zone structure," "hydrothermal vein type fracture-filling structure," and "contact metasomatic skarn zone structure," etc., then q equals 12, and the pre-set local indicator matrix is ​​266×12 dimensional. The element value V in the i-th row and j-th column of the matrix...ij This represents the importance weight of the i-th small-scale prospecting indicator feature under the j-th type of ore-controlling structure. For example, in the column "Haixiian magmatic segregation type copper-nickel deposit or nickel-associated ring alteration zone structure", the weight value of the small-scale feature "kaolinite alteration spectral anomaly peak" is set to 0.95, the weight value of "maximum value of high-frequency anomaly of copper element" is set to 0.90, while the weight value of the small-scale feature "layered lead-zinc mineral reflectance characteristics" is only 0.1. This weight setting reflects the differentiated indicative significance of different ore-controlling structures for prospecting indicators. However, at this time, the dimension of the ore-controlling structure judgment vector D is 523-dimensional instead of q-dimensional 12-dimensional. Therefore, it is necessary to convert D into an expression of the degree of belonging to each ore-controlling structure type. The specific conversion method is to pre-establish a mapping matrix from the 523-dimensional mesoscale feature space to the 12-dimensional ore-controlling structure type space. This mapping matrix is ​​12×523-dimensional and can be obtained by training after experts annotate the ore-controlling structure types of typical samples or generated by using a rule-based feature combination discrimination method. The 523-dimensional ore-controlling structure judgment vector D is then mapped to the 12-dimensional ore-controlling structure type space. Multiplying the 2×523-dimensional mapping matrices yields a 12-dimensional ore-controlling structure type attribution vector E. The values ​​of each dimension of vector E represent the degree to which the current spatial unit belongs to various ore-controlling structure types. For example, if E is a 12-dimensional vector [0.75,0.15,0.05,0,0,0,0.05,0,0,0,0,0], it means that 75% of the features of this unit conform to "Hercynian magmatic segregation type copper-nickel deposit or annular alteration zone structure with nickel as an associated element" and 15% conform to "hydrothermal vein type fracture filling structure".

[0054] After obtaining the ore-controlling structure type classification vector E, a matrix-vector multiplication operation is performed between the preset local indicator matrix p×q and the q-dimensional vector E. The calculation process is similar to the previous level. Each row of the matrix is ​​multiplied by the vector E. For example, the calculation result of the i-th row is the sum of the q elements of the row multiplied by the corresponding q elements of E, resulting in a scalar value fi. After traversing p rows, a p-dimensional vector F is obtained. This vector F is the small-scale weight template. For example, when p is 266 and q is 12, the F obtained after matrix multiplication is a 266-dimensional vector. Among them, the small-scale prospecting features that are highly related to the current ore-controlling structure type receive higher weight values. For example, the weight of the dimension "kaolinite alteration spectral anomaly peak" is 0.72, while the weight of the unrelated dimension "layered lead-zinc mineral reflectance features" is only 0.08. This small-scale weight template clearly indicates which local prospecting features should be identified under the current identified ore-controlling structure conditions.

[0055] The small-scale evidence vector and the small-scale weight template are multiplied element-wise. The small-scale evidence vector is 266-dimensional, denoted as vector G, and the small-scale weight template is 266-dimensional, denoted as vector F. After element-wise multiplication, a new 266-dimensional vector H is obtained, where Hi = Gi × Fi. This operation achieves secondary dynamic weighting modulation of the small-scale evidence vector, which strengthens the prospecting indicator features that match the current ore-controlling structure and filters out the mismatched features. For example, the original value of the dimension of "kaolinite alteration spectral anomaly peak" in the original small-scale evidence vector is 0.82, which, after weighting... After being modulated by a weight of 0.72, the value becomes 0.59, maintaining a relatively high intensity. The original value of the "layered lead-zinc mineral reflectance characteristics" dimension, 0.68, is attenuated to 0.054 after being modulated by a weight of 0.08, almost completely filtered out. Although the original observed value of the latter is not low, it is correctly suppressed because it does not match the current ore-controlling structure type. The modulated vector H is named the preliminary fusion vector. This vector is the final feature expression after two layers of causal guidance and step-by-step regulation. The weights of each dimension feature have been reasonably allocated according to the geological logic from regional background to ore-controlling structure to prospecting indicators. The initial values ​​of the preset geological inference matrix and the preset local indicator matrix can be set by experts and used as learnable parameters for end-to-end fine-tuning and optimization during the overall training of the model, to achieve data-driven adaptive adjustment of the knowledge matrix.

[0056] S104: Construct a knowledge graph related to the target area based on geological exploration reports and mineral deposit survey information, and use graph neural networks to encode the knowledge graph to generate knowledge embedding vectors.

[0057] In S104 above, after completing the causal hierarchical fusion of multi-scale features, it is necessary to introduce prior knowledge from the field of geological exploration to constrain and guide the training process of the deep learning model. This is because although deep neural networks can automatically learn complex nonlinear patterns from a large amount of data, the learned patterns may violate the basic laws of geological mineralization. The field of geological exploration has accumulated rich mineralization theories, ore deposit models and exploration experience over hundreds of years. This knowledge exists in the form of geological exploration reports, ore deposit survey data, academic literature and so on. If this structured or semi-structured domain knowledge can be effectively embedded into the deep learning model, it can enhance interpretability and geological rationality while maintaining the model's powerful learning ability. Therefore, it is necessary to construct a systematic knowledge graph to organize this domain knowledge and transform the knowledge graph into a vector representation that can be fused with the deep learning model through graph neural networks.

[0058] Based on geological exploration reports and mineral deposit survey information, a knowledge graph related to the target area is constructed. A graph neural network is then used to encode the knowledge graph and generate knowledge embedding vectors. Specifically, this involves: extracting geological knowledge triples from the geological exploration reports and mineral deposit survey information, storing these triples in a graph database to construct a global knowledge graph related to the target area; dividing the target area into multiple spatial units, identifying the core geological entities contained within each spatial unit, and performing a correlation retrieval within the global knowledge graph centered on these core geological entities to extract local knowledge subgraphs corresponding to the target spatial units. The current spatial unit is any grid unit among the multiple spatial units; performing multi-round iterative aggregation encoding on each node in the local knowledge subgraph using a graph neural network to obtain the node embedding vectors for each node in the local knowledge subgraph; performing graph pooling on the node embedding vectors of all nodes in the local knowledge subgraph to obtain the initial knowledge embedding vector corresponding to the target spatial unit; and summing the initial knowledge embedding vectors corresponding to all spatial units to obtain the final knowledge embedding vector.

[0059] Specifically, knowledge in the field of geological exploration is collected from multiple sources and a knowledge graph is constructed. The knowledge sources mainly include two categories of data. The first category is structured or semi-structured data, such as the basic information table of mineral deposits stored in the mineral deposit database. This table contains fields such as mineral deposit name, mineral deposit type, metallogenic epoch, tectonic location, main ore minerals, host rock type, ore-controlling structure, and mineralization and alteration characteristics. These fields themselves implicitly contain the structure of entities and relationships. For example, a record states that the Lima River-Lala North area in Huili, Sichuan is located in the middle section of the Kang-Dian axis and east of the Anning River fault, superimposed with the metallogenic background of the Emei Mountain igneous province. Two completely different types of copper-nickel related deposits are developed in the area.

[0060] The Limahe mining area is based on Hercynian mafic-ultramafic complexes, belonging to magmatic segregation type copper-nickel deposits. The rock mass intrudes into pre-Sinian metamorphic carbonate rocks and quartz schist. The ore bodies occur in stratiform and lenticular forms in peridotite, pyroxene facies, and fault zones. The ore has a typical sponge meteorite texture. The main metallic minerals are nickel pyrrhotite, pyrrhotite, and chalcopyrite. The surrounding rocks are often serpentinized and talcified. The copper and nickel grades are stable, and the magnetic and induced polarization anomalies are obvious. The ore-bearing strata in the Lalabei area are Proterozoic Hekou Formation metamorphic volcanic sedimentary rocks, which are volcanic sedimentary metamorphic superimposed hydrothermal iron-copper deposits. Nickel is only an associated element. The ore bodies occur along the bedding planes, in stratiform and lenticular forms. The ore develops banded and disseminated structures. Chalcopyrite and magnetite are the main useful minerals. They show albite alteration and biotite alteration. Geochemical exploration shows Cu, Co, Ni, and Mo assemblage anomalies. The deep north-south faults in the region play a prominent role in controlling rocks and mineralization. The Lima River relies on basic and ultrabasic rock bodies and geophysical and geochemical anomalies for mineral exploration, while the Lala North focuses on tracking the volcanic rock strata of the Hekou Formation and the sodic alteration zone. The two types of deposits are controlled by two major mineralization mechanisms: magmatic differentiation and stratigraphic metamorphic hydrothermal alteration.

[0061] For knowledge extraction from structured data, a pre-defined mapping script is used for rule-based transformation. This script predefines transformation rules from fields to knowledge graph triples based on the semantics of the database table fields. For example, for the ore deposit information table, the mapping rules include "ore deposit entity - belongs to type - ore deposit type entity", "ore deposit entity - formed in era - geological age entity", "ore deposit entity - located in tectonic unit - tectonic entity", "ore deposit entity - hosted in surrounding rock - rock type entity", "ore deposit entity - controlled by structure - tectonic type entity", "ore deposit entity - associated alteration - alteration type entity", etc. In actual execution, the script iterates through each record in the ore deposit information table, filling in triple templates based on field values. For example, when the ore deposit name field of a record is read as "Limahe Mine", the script will perform a triple transformation. When the field "area or nickel as associated" and the type field is "Hessean magmatic segregation type copper-nickel deposit or nickel as associated element" are read, a ternary set is generated: <Limahe mining area or nickel as associated, type, Hercynian magmatic segregation type copper-nickel deposit or nickel as associated element deposit>. When the mineralization age field is read as "Hessean or Proterozoic", a ternary set is generated: <Limahe mining area or nickel as associated, formed in age, Hercynian or Proterozoic>. This process continues, parsing all entities and relationships involved in the record into a set of ternary sets. After traversing the entire database table, the first geological knowledge ternary set extracted from the structured data is obtained. This set contains a large amount of attribute information of known deposits and the correlation of mineralization elements. For example, after processing a database containing 500 deposit records, approximately 5,000 ternary sets may be generated.

[0062] For knowledge extraction from unstructured text data, a deep learning-based natural language processing technique is employed. First, a named entity recognition (NER) task is performed. The goal of this task is to automatically label geological entities belonging to predefined categories from the text. These predefined entity categories include deposit names, mineral types, geological ages, tectonic units, rock types, mineral names, structural types, alteration types, geochemical anomalies, and geophysical anomalies. The NER model uses a pre-trained BERT-CRF architecture, where BERT acts as the feature encoder to extract contextual semantic representations of the text, and CRF acts as the sequence labeling layer for entity boundary recognition and category classification. During the training phase, the model uses manually labeled geological literature corpora for supervised learning. The labeled corpus contains key paragraphs from approximately 2000 geological reports and academic papers, including geological entities... The data is labeled according to the BIO annotation format. For example, in the sentence "The study area is located in the Lima River-Lala North area of ​​Huili City, Sichuan Province", "Lima River of Huili City, Sichuan Province" is labeled as B-tectonic unit, I-tectonic unit, I-tectonic unit, I-tectonic unit, I-tectonic unit, and "Lala North area" is labeled as B-tectonic location. After the model is trained, it can automatically identify geological entities in new text. For example, if the input sentence is "Hercy or Proterozoic Lima River mining area surrounding rocks or Lala North area surrounding rocks intruding into the Precambrian metamorphic rock series or the Proterozoic Hekou Formation metamorphic volcanic sedimentary rock series", the model outputs the recognition results as follows: "Hercy or Proterozoic" is labeled as a geological age entity, "Lima River mining area surrounding rocks or Lala North area surrounding rocks" is labeled as a rock type entity, "Precambrian or Proterozoic" is labeled as a geological age entity, and "metamorphic rock series" is labeled as a rock type entity.

[0063] After identifying geological entities in the text, a relation extraction model is used to identify the relationships between entities. This model, also based on a pre-trained language model, employs an entity pair classification architecture. For each pair of entities in the text, the model determines whether a predefined relationship type exists and what that relationship is. Predefined relationship types include those consistent with geological semantics, such as "located in," "formed at," "intruded upon," "controlled by," "associated with," "enriched in," "indicating," and "transformed into." The relation extraction model is trained using manually annotated corpus by geological professionals who annotate entity pairs with clear relationships and their relationship types. For example, in the sentence "Hercy or Proterozoic Limahe mining area surrounding rocks or Lala North area surrounding rocks intruded into the Precambrian metamorphic rock series or the Proterozoic Hekou Formation metamorphic volcanic sedimentary rock series," the entity pair <Hercy or Proterozoic Limahe mining area surrounding rocks or Lala North area surrounding rocks, Precambrian metamorphic rock series or Proterozoic Hekou Formation metamorphic volcanic sedimentary rock series> is represented by <Hercy or Proterozoic Limahe mining area surrounding rocks or Lala North area surrounding rocks, Precambrian metamorphic rock series or Proterozoic Hekou Formation metamorphic volcanic sedimentary rock series>. The entity pair <regional north-south trending deep faults, major metallic minerals> is labeled with the relation "intrusion into". In the sentence "regional north-south trending deep faults control the distribution of major metallic minerals", the entity pair <regional north-south trending deep faults, major metallic minerals> is labeled with the relation "controls". After the model is trained, it can automatically extract entity relations from new text. For example, if the input sentence contains the entities "study area" and "Lima River-Lala North area of ​​Huili City, Sichuan Province", the model outputs the relation triple <study area, located in, Lima River-Lala North area of ​​Huili City, Sichuan Province>. If the input sentence contains the entity "mineralization type" and "magmatic segregation type copper-nickel mineralization or volcanic sedimentary metamorphic superposition hydrothermal iron-copper mineralization", the model outputs the relation triple <mineralization type, belonging to, magmatic segregation type copper-nickel mineralization or volcanic sedimentary metamorphic superposition hydrothermal iron-copper mineralization>. By combining the two steps of named entity recognition and relation extraction, a second set of geological knowledge triples can be automatically constructed from unstructured text. For example, from a geological exploration report containing 50 pages of text, approximately 800 triples may be extracted.

[0064] The first set of geological knowledge triplets extracted from structured data and the second set of geological knowledge triplets extracted from unstructured text are aggregated and merged, and deduplication and consistency checks are performed. The deduplication operation identifies triplets with the same semantics but slightly different expressions and merges them into a unified form. For example, <Limahe mining area or nickel is associated, belonging to type, porphyry copper deposit> and <Limahe mining area or nickel is associated, type, Hercynian magmatic segregation type copper-nickel deposit or nickel is associated element> are identified as the same relationship and unified into the standard form. The consistency check identifies contradictory triplets and marks them for manual confirmation. For example, if the same deposit is labeled with different mineralization in different sources. The era then triggers conflict warnings. The cleaned set of triples is called the geological knowledge triple. This set of triples is stored in a graph database to construct a global knowledge graph. The graph database uses Neo4j or a similar graph storage system. Entities in the triples are stored as nodes, and relations are stored as edges. Nodes carry attribute information such as entity name, entity type, and confidence level, while edges carry attribute information such as relation type, source literature, and extraction time. The final global knowledge graph contains tens of thousands of nodes and hundreds of thousands of relation edges, covering multiple levels of content, including geological background knowledge of the study area and its surroundings, typical mineral deposit case knowledge, and metallogenic regularity theory.

[0065] After constructing the global knowledge graph, local knowledge subgraphs related to the target region are extracted for specific prediction tasks. Because the global knowledge graph contains a wide range of knowledge but contains a large amount of irrelevant information for prediction tasks at specific spatial locations, inputting the entire global graph into the model would introduce noise and increase computational burden. Therefore, it is necessary to select relevant knowledge subsets based on the spatial scope of the prediction task and perform spatial grid partitioning on the target region. The entire study area is divided into regular grids with the same spatial resolution as the feature extraction stage, e.g., 500m × 500m. Each grid cell serves as an independent spatial prediction unit. After grid partitioning, multiple spatial units are obtained. For example, a 100 square kilometer study area divided into 500m grids contains 400 spatial units. For any given spatial unit, its corresponding local knowledge subgraph needs to be extracted. The extraction process first identifies the core geological entities contained within the spatial unit. The identification of core geological entities is based on spatial location matching, i.e., checking which entity nodes in the global knowledge graph have spatial coordinates falling within the geographical range of the current grid unit. These entities may include the area covered by the grid. The knowledge graph encompasses known mineral deposits, geological observation points, sampling points, geological boundary intersections, and structural segments within the grid. It also includes larger-scale geological unit entities to which the grid belongs, such as the lithostratigraphic unit, tectonic unit, and metallogenic belt. These spatial relationships are established during the construction of the global knowledge graph. For example, spatial overlay analysis determines which rock body and tectonic belt each mineral deposit is located within, and these spatial inclusion relationships are stored in the graph as triples. Therefore, once the geographical extent of the current grid unit is determined, spatial queries can be performed using the graph database. The function quickly retrieves all entity nodes that are spatially related to the grid and uses these entities as the core geological entity set. For example, for grid cell numbered G245, the search found that it contains one known copper deposit and two geochemical sampling points. Moreover, the grid is located within the range of the Hercynian or Proterozoic Limahe mining area or the surrounding rocks of the Lalabei area and is traversed by the NE-trending fault. Therefore, the core geological entity set includes entity nodes such as "Limahe mining area or nickel as associated point", "Hercynian or Proterozoic Limahe mining area or the surrounding rocks of the Lalabei area", and "NE-trending fault F1".

[0066] After identifying the core geological entities, an extended search is performed on the global knowledge graph centered on these entities to extract local knowledge subgraphs. This extended search employs a K-hop neighbor extraction strategy, meaning that starting from each core entity node, a breadth-first traversal is performed along the relational edges in the graph, incorporating all nodes and edges within the K-hop range into the local subgraph. The value of K needs to balance knowledge coverage and computational efficiency, and is typically set to 2 or 3 hops. For example, when K is set to 2, for the core entity "Limahe mining area or nickel as associated point", the nodes and edges directly related to it are extracted first. The system extracts one-hop neighbor nodes, which may include the deposit type node "Hessen-era magmatic segregation type copper-nickel deposit or nickel-associated element deposit," the metallogenic epoch node "Hessen or Proterozoic," the ore-controlling structural node "ring fault," and the alteration type node "potassic alteration." Then, it continues to extract two-hop neighbors from these one-hop neighbors. For example, starting from the "Hessen-era magmatic segregation type copper-nickel deposit or nickel-associated element deposit" node, it may be associated with the "island arc environment" node, representing the typical tectonic setting of this type of deposit. From the "Hessen or Proterozoic" node... Starting from the "Pacific Plate Subduction" node, which represents the tectonic event of that period, and starting from the "Ring Fault" node, which may be associated with the "Magma Intrusion Channel" node, representing the mineralization significance of the tectonic structure, after two hops of expansion, a subgraph containing more than ten nodes and more than twenty edges is extracted, centered on "Lima River mining area or nickel as an associated point". This subgraph contains knowledge of the metallogenic model of the deposit. Similarly, the core entity "Heravician or Proterozoic Lima River mining area host rocks or Lala North area host rocks" is expanded to extract the associated magma evolution sequence. Knowledge nodes such as source region characteristics and enrichment of ore-forming elements are used to expand the core entity "NE-trending fault F1" and extract related knowledge nodes such as tectonic stress field, ore-controlling mechanism, and associated alteration. After merging and deduplicating the expanded subgraphs of all core entities, a local knowledge subgraph corresponding to the current grid unit G245 is obtained. This subgraph may contain 30-50 nodes and 50-100 relational edges. Compared with the global map containing tens of thousands of nodes, the scale is greatly reduced, but the knowledge content most relevant to the geological background and ore-forming potential evaluation of this spatial unit is retained.

[0067] After extracting the local knowledge subgraph, the graph structure needs to be transformed into a vector form that can be processed by deep learning models. This transformation process is carried out through graph neural networks. The core idea of ​​graph neural networks is to learn the vector representation of nodes by iteratively aggregating the neighbor information of nodes, so that the topological and semantic relationships in the graph structure are encoded into the vector space. In specific implementation, graph convolutional networks (GCN) or graph attention networks (GAT) are used as encoding models. Taking graph attention networks as an example, we will explain in detail. The advantage of GAT is that it can adaptively learn the importance weights of different neighbor nodes instead of treating all neighbors equally. This is especially important for geological knowledge graphs because different relationship types have very different indicative significance for mineralization prediction. For example, the importance of the "ore-controlling structure" relationship is usually higher than that of the "associated mineral" relationship. GAT can automatically learn this differentiated importance through the attention mechanism.

[0068] Before performing graph neural network encoding, the nodes in the local knowledge subgraph need to be initialized. The initialization method assigns an initial feature vector to each node, generated based on the node's type and attributes. Different initialization strategies are used for different types of entity nodes. For example, for a ore deposit entity node, its initial feature vector is formed by concatenating the unique thermal encoding of the ore deposit type, the multi-thermal encoding of the ore-forming elements, and the numerical features of the ore deposit scale, resulting in a 128-dimensional initial vector. For a geological age entity node, its initial feature vector consists of the absolute geological age value of the age and the unique thermal encoding of the epoch to which the age belongs. The initial feature vector is composed of a 64-dimensional code. For rock type entity nodes, the initial feature vector consists of the unique thermal code of the rock type, the multi-thermal code of the main mineral composition, and the typical values ​​of the rock chemical composition, forming a 96-dimensional initial vector. For other types of entities such as tectonic types and alteration types, initial vectors of corresponding dimensions are constructed based on their attribute features. To unify the feature dimensions of different types of nodes for subsequent network processing, a linear projection layer is used to map the initial feature vectors of all nodes to the same embedding dimension, such as 256 dimensions. After mapping, each node has a 256-dimensional initial embedding vector denoted as h. v (0) The superscript (0) indicates the 0th layer, i.e., the initial layer.

[0069] When starting the multi-round iterative aggregation process of the graph attention network, each round of iteration is called a layer. In the aggregation process of the l-th layer, for each node v in the local knowledge subgraph, its embedding vector needs to be updated by aggregating the embedding vectors of its neighboring nodes in the (m-1)-th layer. The core of the aggregation process is to calculate the attention weights. The specific calculation steps include the following mathematical expressions.

[0070] First, for node v and each of its neighboring nodes u, their embedding vectors at layer m-1 are linearly transformed using the weight matrix W. The mathematical expression for the linear transformation is: Z v (m-1) =W (m) h v (m-1) ; Z u (m-1) =W (m) h u (m-1) ; Among them, W (m) R d×d Here, d is the learnable weight matrix of the m-th layer, d is the embedding vector dimension (e.g., 256-dimensional), and h is the weight matrix of the m-th layer. v (m-1) and h u (m-1) Z represents the embedding vectors of nodes v and u at the (m-1)th layer, respectively. v (m-1) and Z u (m-1) It is the transformed feature vector.

[0071] The two transformed feature vectors are concatenated to form a 512-dimensional vector. Assuming the embedding dimension is 256, the concatenated vector is 512-dimensional. The mathematical expression for the concatenation operation is: C vu =[Z v (m-1) ||Z u (m-1) ], where || denotes vector concatenation operation, C vu R 2d It is the concatenated vector.

[0072] The concatenated vector is input into a single-layer attention function, which is typically a learnable weight vector multiplied by the concatenated vector and then activated by LeakyReLU, resulting in a scalar attention score, mathematically expressed as: E vu =LeakyReLU(a T ×[Z v (m-1) ||Z u (m-1) ]), where a R 2d Let E be the attention parameter vector. vu The scalar attention score measures the importance of neighbor node u to the central node v.

[0073] After calculating a set of attention scores for all neighboring nodes of node v, the attention weights are obtained by normalizing them using the softmax function. The normalization formula is as follows: ; Where N(u) represents the set of all neighboring nodes of node v, and A vu The normalized attention weights satisfy... And 0≤A vu ≤1. For example, if node v has 3 neighboring nodes, u1, u2, and u3, the calculated value is E. vu1 =1.2、E vu2 =0.8, E vu3 =0.4, and the attention weight A is obtained after softmax normalization. vu1 =0.52, A vu2 =0.33, A vu3 =0.15.

[0074] After obtaining the attention weights, a weighted aggregation is performed. The transformed feature vectors of all neighboring nodes are summed according to their corresponding attention weights to obtain the aggregated neighborhood information vector. The mathematical expression is: ; Where σ(*) is a nonlinear activation function such as ReLU or ELU, h v (m) It is the new embedding vector of node v at the m-th layer.

[0075] To enhance the expressive power of the model, multi-head attention mechanisms are often used in practical applications. This involves computing K independent attention heads in parallel and then averaging the results. The mathematical expression for multi-head attention is: ; Where K is the number of attention heads, || k k=1 This represents the output concatenation operation of K heads, W （m，k） A is the weight matrix of the k-th attention head in the m-th layer. vu k It is the attention weight calculated by the kth attention head.

[0076] Alternatively, use an averaging operation in the last layer: ; Let σ(*) represent the node embedding vector output by node v after processing in the Lth (i.e., the last) layer of the graph attention network; σ(*) is the non-linear activation function; K represents the number of attention heads in the multi-head attention mechanism; 1 / k is the normalization coefficient for the averaging operation, used to average and fuse the outputs of the K attention heads; ∑ K k=1 This represents the summation of the computation results from K attention heads; this is the first-level summation operation in multi-head attention fusion. uєN(u) This represents summing the results of all neighboring nodes u of node v, where N(u) is the set of neighboring nodes of node v in the knowledge graph; A vu k W is the normalized attention weight coefficient from neighbor node u to center node v in the k-th attention head; (L,k) h represents the learnable weight matrix corresponding to the k-th attention head in the L-th layer graph attention network; u (L-1) This represents the node embedding vector of the neighbor node u in the L-1 layer (i.e., the previous layer). This vector is the intermediate representation of node u after propagation through the graph attention network of the previous L-1 layers, and serves as the input feature for the aggregation operation in the current layer.

[0077] After completing the iterative aggregation of layer L, each node obtains its final node embedding vector h. v (L) This vector aggregates information from all nodes within a node's L-hop neighborhood, and then graph pooling is used to obtain a graph-level knowledge embedding vector. The mathematical expression for average pooling is: ; The mathematical expression for max pooling is: ; i = 1, 2, ..., d; where V represents the set of all nodes in the local knowledge subgraph, |V| is the total number of nodes, and g is the graph-level knowledge embedding vector. i h is the i-th dimension of the knowledge embedding vector. (L) v,i It is the i-th dimension of the embedding vector of node v at the L-th layer.

[0078] S105: Input the knowledge embedding vector and the preliminary fusion vector into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit in the target area.

[0079] In S105 above, the knowledge embedding vector and the preliminary fusion vector are input into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit in the target area. First, an evidence reasoning framework simulating the reasoning process of geologists is constructed. This framework can logically associate and mutually verify the observational evidence from multimodal data with the theoretical knowledge from the knowledge graph, just like experts analyze the genesis of mineral deposits. This results in a mineralization judgment that is both supported by data and conforms to geological laws, and the credibility of this judgment can be quantified.

[0080] The knowledge reasoning model includes an evidence questioning module, an evidence construction module, and an evidence evaluation module. Knowledge embedding vectors and preliminary fusion vectors are input into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value for each spatial unit within the target area. Specifically, this includes: deconstructing the preliminary fusion vector into a multi-level data evidence set; using the evidence questioning module, guided by the knowledge embedding vector, performing correlation questioning on the data evidence set to obtain a reasoned evidence vector; concatenating the reasoned evidence vector with the knowledge embedding vector using the evidence construction module to generate a comprehensive evidence vector; inputting the comprehensive evidence vector into the evidence network of the evidence evaluation module, mapping it through the output layer of the evidence network and a non-negative activation function to output the first evidence strength supporting the mineralization hypothesis and the second evidence strength supporting the non-mineralization hypothesis; calculating 1 plus the sum of the first and second evidence strengths to obtain the evidence sum, and dividing 1 by the evidence sum to obtain the prediction uncertainty value; calculating the sum of 1 and the first evidence strength to obtain the numerator evidence value; calculating the sum of the first and second evidence strengths with the preset values ​​corresponding to the number of categories in the classification problem to obtain the denominator evidence value; and dividing the numerator evidence value by the denominator evidence value to obtain the mineralization probability value.

[0081] Specifically, the initial fusion vector obtained from the multimodal fusion network is deconstructed to form a multi-level data evidence set. This deconstruction process is necessary because although the initial fusion vector contains comprehensive information from multiple sources such as remote sensing images, geophysical data, geochemical data, and geological maps, this information is highly compressed and mixed, making it impossible to directly and precisely correspond to discrete geological concepts in the knowledge graph. Therefore, the initial fusion vector needs to be decomposed into evidence subsets with different abstract levels and different geological semantics through deconstruction operations. For example, in a specific implementation case, assuming the initial fusion vector has a dimension of 512, the deconstruction operation can map it to M different evidence spaces using a set of learnable projection matrices {P1, P2, ..., PM}. Each evidence space corresponds to a type of geological observation dimension. Specifically, M = 4 evidence levels can be set, corresponding to geochemical anomaly evidence, geophysical anomaly evidence, tectonic background evidence, and lithological assemblage evidence, respectively. For the initial fusion vector Ff with a dimension of 512, through the projection matrix P1... R 128×512The geochemical evidence vector ec = p1 × Ff is obtained. This vector has 128 dimensions, of which the first 64 dimensions may capture anomalous patterns of ore-forming elements such as Au and Cu, and the latter 64 dimensions capture anomalous patterns of indicator elements such as As and Sb. Similarly, the geophysical evidence vector eg, the tectonic evidence vector es, and the lithological evidence vector el are obtained through projection matrices P2, P3, and P4, respectively. These four evidence vectors together constitute the data evidence set εd = {ec, eg, es, el}. Each evidence vector after deconstruction has a clearer geological semantics, which facilitates subsequent association and reasoning with corresponding concepts in the knowledge graph.

[0082] After obtaining the data evidence set, the evidence questioning module uses the knowledge embedding vector as a guide to perform correlation questioning on the data evidence set to obtain the inferred evidence vector. Specifically, this includes: projecting the knowledge embedding vector into the first linear transformation layer of the evidence questioning module to generate a query vector; projecting each sub-feature vector in the data evidence set into the second and third linear transformation layers of the evidence questioning module to generate key vectors and value vectors respectively; calculating the inner product of the query vector and each key vector, and normalizing it using a normalized exponential function to obtain the corresponding attention weights; and using the attention weights to perform a weighted summation of all value vectors to generate the inferred evidence vector.

[0083] Specifically, embedding knowledge into vector g R 256 The input is projected into the first linear transformation layer of the evidence challenge module to generate a query vector q. The first linear transformation layer consists of a weight matrix W. Q R dk×256 and bias vector b Q R dk The structure consists of dk, where dk is the target dimension of the query vector, for example, set to 128 dimensions to balance computational efficiency and expressive power. The mathematical expression for the linear transformation is q=W. Q g+b Q In a specific embodiment, it is assumed that the first 8 dimensions of the knowledge embedding vector g are [0.82, -0.45, 1.23, 0.67, -0.91, 0.34, 1.05, -0.58]. After projection through the first linear transformation layer, the first 8 dimensions of the query vector q are [1.15, -0.72, 0.89, 1.43, -0.67, 0.91, -1.22, 0.55]. This query vector encodes prior information in the knowledge graph about favorable metallogenic zones, ore-controlling structures, and combinations of ore-forming elements.

[0084] Each sub-feature vector in the data evidence set is projected into the second and third linear transformation layers of the evidence challenge module to generate corresponding key and value vectors. The data evidence set contains multiple sub-feature vectors {f1, f2, ..., fN}, where N is the number of sub-features, including geochemical anomaly features, geophysical anomaly features, remote sensing alteration features, and tectonic buffer features, etc. Each sub-feature vector fi R df Where df is the original dimension of the sub-feature, for example, 64 dimensions, and the second linear transformation layer consists of the weight matrix W. K R dk×df and bias vector b k R dk The third linear transformation layer, which is used to generate key vectors, consists of a weight matrix W. v R dv×df and bias vector b v R dv The structure is used to generate value vectors, where dv is the target dimension of the value vector, for example, set to 128 dimensions to match the query vector dimension for easier subsequent calculations. For the i-th sub-feature vector fi, the mathematical expression for the key vector is ki = W. K fi+b K The mathematical expression for the value vector is vi = W v fi+b v， In a specific embodiment, it is assumed that the data evidence set contains four sub-feature vectors corresponding to geochemical Au element anomalies f1, geochemical As element anomalies f2, aeromagnetic anomalies f3, and fault structure buffers f4, respectively. The first eight dimensions of the geochemical Au element anomaly feature vector f1 are [2.34, 0.87, -0.45, 1.62, 0.23, -1.11, 0.76, 1.39]. After projection through the second linear transformation layer, the first eight dimensions of the bond vector k1 are [1.87, -0.34, 1.25, 0.68, -0.92, 1.43, 0.51, -0.76]. After projection through the third linear transformation layer, the first 8 dimensions of the value vector v1 are [2.15, 0.93, -0.67, 1.48, 0.34, -1.26, 0.89, 1.52]. Similarly, the other sub-feature vectors are transformed to obtain the corresponding key vectors {k2, k3, k4} and value vectors {v2, v3, v4}. The significance of doing this is to map the original data evidence with different physical meanings and dimensions to a unified semantic representation space. The key vectors act as feature indexes that can be queried and retrieved, while the value vectors carry the effective information content of the evidence, laying the foundation for subsequent association matching and information extraction.

[0085] The dot product of the query vector q and each key vector {k1, k2, ..., kN} is calculated and normalized using a normalized exponential function to obtain the corresponding attention weights. The mathematical expression for the dot product calculation is: Si = q T ki, yielding the original similarity score. This score reflects the semantic relevance between the knowledge-guided challenge vector and each data evidence key vector. To avoid the inner product result being too large and causing the gradient vanishing problem in the subsequent softmax function, the similarity score is usually scaled to obtain Si1=q. T ki / (√dk), where √dk is the scaling factor. For example, when dk=128, the scaling factor is 11.31. The scaled similarity scores are converted into attention weights through a softmax normalized exponential function. The resulting attention weights satisfy the condition that the sum of all weights is 1 and each weight value is between 0 and 1.

[0086] The evidence vector e after inference is generated by weighting all value vectors using attention weights. The mathematical expression is as follows: ; where a i Represents attention weights, v iRepresents a value vector. For example, by weighting and summing the four value vectors {v1, v2, v3, v4} according to their corresponding weights {a1=0.52, a2=0.11, a3=0.04, a4=0.33}, the first dimension value of the inferred evidence vector e is calculated as e1=0.52×2.15+0.11×v2,1+0.04×v3,1+0.33×v4,1. Assuming (v2,1)=0.87, (v3,1)=-0.45, (v4,1)=1.62, we get e1=0.52×2.15+0.11×0.87+0.04×(-0.45)+0.33×1.62=1.731. Similarly, calculating the other dimension values ​​of the inferred evidence vector ultimately yields a complete 128-dimensional vector. This inferred evidence vector e is part of the data evidence set. The refined representation of all sub-feature vectors after knowledge-guided association questioning and weighted selective aggregation is not a simple splicing or averaging of all original data. Instead, it achieves differentiated integration of evidence through an attention mechanism. Key evidence features that are highly related to the mineralization mechanism are incorporated into the final representation with greater weight, while weakly related evidence features are marginalized with less weight. This achieves the goal of extracting effective mineralization signals from massive multi-source data. Compared with the traditional approach of treating all features equally or relying on human experience for feature selection, this evidence questioning module automatically learns the knowledge-guided evidence importance assessment criteria through end-to-end deep learning. This makes the evidence vector after reasoning contain both data-driven observational information and knowledge-driven prior constraints, laying a high-quality information foundation for subsequent evidence network construction and fusion reasoning.

[0087] After obtaining the inferred evidence vector, the evidence construction module concatenates it with the knowledge embedding vector to generate a comprehensive evidence vector. The design philosophy of this step is to explicitly combine the data evidence after knowledge verification with the knowledge itself, forming a complete evidentiary body that simultaneously contains "observed facts" and "theoretical basis." This is analogous to legal reasoning, where both on-site physical evidence (data) and legal provisions (knowledge) are required to reach a judgment. In practice, the 128-dimensional inferred evidence vector *er* and the 256-dimensional knowledge embedding vector *gk* are concatenated along the feature dimension to obtain the comprehensive evidence vector *e*. com=[er||gk], the concatenated vector has a dimension of 128+256=384. The first 128 dimensions carry multimodal observation information after knowledge review and screening, while the last 256 dimensions directly embed structured geological laws extracted from the knowledge graph. For example, in a real case, the first 128 dimensions of the comprehensive evidence vector of a certain spatial unit may contain features after knowledge weighting, such as the geochemical element content anomaly index, gravity and magnetic anomaly intensity, distance from known fault zones, and unique thermal encoding of lithology type of the unit. The knowledge-embedded vector of the last 256 dimensions may encode prior information such as "the mineralization type of this area is hydrothermal, the mineralization age is Hercynian or Proterozoic, and the ore-controlling structure is NE-trending fault". This concatenation method maintains the relative independence of the two types of information, avoids information loss that may be caused by premature feature mixing, and provides rich input signals for the subsequent evidence evaluation module.

[0088] The comprehensive evidence vector is input into the evidence network of the evidence evaluation module. Mapping is performed through the output layer of the evidence network and a non-negative activation function to output the first evidence strength supporting the mineralization hypothesis and the second evidence strength supporting the non-mineralization hypothesis. The core objective of this step is to quantitatively extract the evidence strengths supporting and opposing mineralization from the comprehensive evidence, and to ensure that these strength values ​​are non-negative to conform to subjective logic. In Logic, the mathematical constraints of evidence quality measurement are used. The evidence network adopts a multi-layer fully connected neural network architecture. In one specific implementation, the evidence network consists of three fully connected layers. The first layer maps the 384-dimensional comprehensive evidence vector to a 256-dimensional hidden representation using the ReLU activation function. The second layer further maps to 128 dimensions and also uses ReLU activation. The third layer, the output layer, maps the 128 dimensions to 2 dimensions, corresponding to the evidence strength supporting mineralization and the evidence strength supporting non-mineralization, respectively. The output layer is followed by a non-negative activation function to ensure that the output value is positive. Commonly used non-negative activation functions include the Softplus function f(x) = log(1 + exp(x)) or the positive half-branch plus offset of the exponential linear unit ELU. In this embodiment, the Softplus function is used, mathematically expressed as b = Softplus(W3σ(W2σ(W1e)). com +b1)+b2)+b3), where W1∈R 256×384 W2∈R 128×256 W3∈R 2×128 It is a three-layer weight matrix, b1, b2, b3 are bias vectors, σ(×) represents the ReLU activation function, and the final output is a two-dimensional vector b=[b min b bar ] T , where b min To support the first strength of evidence for the mineralization hypothesis, b bar To support the second piece of evidence for the non-mineralization hypothesis, 𝑒𝑐𝑜𝑚 This is a comprehensive evidence vector. For example, in a certain computational instance, b might be obtained. min =3.8 and b bar =1.2, which means that the combined evidence supports the mineralization hypothesis much more strongly than the non-mineralization hypothesis. The evidence network learns how to identify and quantify the degree of support for different hypotheses from the combined evidence through an end-to-end training process. The training loss function includes not only the prediction accuracy constraint but also the evidence strength regularization constraint to prevent the network from outputting extreme evidence values.

[0089] After obtaining two levels of evidence strength, the sum of 1 and the first and second levels of evidence strength is first calculated to obtain the evidence sum. Then, 1 is divided by the evidence sum to obtain the predicted uncertainty value. This calculation step originates from the definition of uncertainty in subjective logic theory. Within the framework of subjective logic, for a binary hypothesis space (mineralization or non-mineralization), the total evidence quality can be measured by the belief quality, where the belief quality supporting mineralization is denoted as b. m The quality of the belief that the minerals cannot be mined is denoted as b. b The quality of belief that fails to be assigned to any hypothesis characterizes the uncertainty u, and the three satisfy the normalization constraint b. m +b b +u=1, in the implementation of evidence theory, the first evidentiary strength b output by the evidence network. min Second evidence strength b bar It is not directly equivalent to the quality of belief, but requires normalization transformation, specifically assuming the evidence sum is S=b. min +b bar Uncertainty is defined as u = 1 / (S+1). The rationale behind this formula is that the larger the evidence and S, the stronger the overall support for the two hypotheses, and the more confident the model's judgment, thus resulting in a smaller uncertainty u. Conversely, when S approaches 0, the evidence is very weak, the model cannot make a reliable judgment, and the uncertainty u approaches 1. For example, S = 3.8 + 1.2 = 5.0, therefore the uncertainty u = 1 / (5.0 + 1) ≈ 0.167, indicating that the prediction for this spatial unit has approximately 16.7% uncertainty. In other words, the model has approximately 83.3% confidence in the prediction. This quantification of uncertainty provides a quantitative basis for decision-makers to assess the exploration risks of different target areas. For example, among two target areas with the same high probability of mineralization, the target area with lower uncertainty should be prioritized for exploration work because the prediction results are more reliable.

[0090] The numerator evidence value is obtained by summing 1 with the first evidence strength. The denominator evidence value is obtained by summing the first evidence strength, the second evidence strength, and the preset values ​​corresponding to the number of categories in the classification problem. The numerator evidence value is divided by the denominator evidence value to obtain the mineralization probability value. The preset value corresponding to the number of categories in the classification problem is set to 2. This formula originates from the conversion of Dirichlet distribution parameters to expected probabilities. In the evidence deep learning framework, evidence strength is modeled as the concentration parameter of the Dirichlet distribution. For binary classification problems, the expected probability of the Dirichlet distribution can be calculated through the concentration parameter. Specifically, let the evidence strength supporting mineralization be a1 = b. min The strength of evidence supporting the non-mineralization is a² = b. bar Then the expected probability of mineralization is P. min =(a1+1) / (a1+a2+2), where the operations of adding 1 and adding 2 correspond to the hyperparameters of the Dirichlet prior. The preset value of 2 comes from the number of categories K=2 in the binary classification problem. Adding 1 to the numerator is the pseudo-count of the corresponding category, and adding K to the denominator is the sum of the pseudo-counts of all categories. In the above example, the evidence value of the numerator is 1+3.8=4.8, and the evidence value of the denominator is 3.8+1.2+2=7.0. Therefore, the mineralization probability value P is... min =4.8 / 7.0=0.686, meaning that this spatial unit has a mineralization probability of about 68.6%. This probability value not only reflects the model's tendency to judge mineralization, but also implies information about the sufficiency of evidence.

[0091] In one possible implementation, the goal of iteratively training the knowledge reasoning model is to fully utilize the synergistic effect of strong and weak supervision signals under the realistic constraint of extremely scarce known ore deposit samples, while leveraging the geoscientific prior constraints provided by the knowledge graph. This enables the knowledge reasoning model to learn reliable mineralization prediction patterns from limited labeled data and a large amount of regional background information, and to achieve dynamic evaluation of the reliability of the prediction results through an uncertainty quantification mechanism. The iterative training of the knowledge reasoning model involves acquiring known ore deposit sample data within the target region as strong supervision labels, and generating weak regional background labels based on the known metallogenic belt range of the target region. Knowledge embedding vectors are injected into the knowledge reasoning model as prior constraints. The known ore deposit sample data, weak regional background labels, mineralization probability values ​​of spatial units, and prediction uncertainty values ​​are input into a weakly supervised adaptive loss function to calculate a comprehensive loss value. The weakly supervised adaptive loss function includes adaptive weight coefficients that dynamically adjust the loss weights of the weak regional background labels based on the prediction uncertainty values. Based on the comprehensive loss value, the parameters of the knowledge reasoning model and the graph neural network are iteratively updated multiple times using a backpropagation algorithm until the knowledge reasoning model converges.

[0092] Specifically, known mineral deposit sample data within the target area are acquired as strongly supervised labels. This data originates from mineral deposit databases or historical exploration reports published by geological survey departments. Each sample contains precise geographic coordinates (latitude, longitude, or projected coordinates), deposit type (e.g., Hercynian magmatic segregation type copper-nickel deposit, nickel-associated copper deposit, or Carlin-type gold deposit), mineralization scale (e.g., large, medium, or small), and major ore-forming elements. For example, if the target area is a metallogenic belt in the western part of a province covering approximately 50,000 square kilometers, querying the province's geological and mineral resources database yields 87 known gold deposit samples within this area. The coordinates of these deposit points are then matched with the unified spatial grid used by the system. If a 500m × 500m grid is used, there are a total of 200,000 grid cells containing known mineral deposits. The grid cells containing the points are labeled as positive samples and assigned a label value of 1. These 87 ore deposit points are distributed across 82 different grid cells. Since 5 grid cells contain multiple neighboring ore deposit points, 82 strongly supervised positive samples are obtained. These strongly supervised labels are highly deterministic, meaning that ore deposits do exist at these locations. Therefore, they should be given high weights in the loss function to ensure that the model can accurately learn the mineralization feature patterns of these known ore deposits. However, since the number of positive samples accounts for only 0.041% of the total number of grid cells, or 0.041%, simply treating all the remaining unlabeled grid cells as negative samples would cause an extreme class imbalance problem. Furthermore, this labeling method is logically flawed because unlabeled areas may contain undiscovered hidden ore deposits. Therefore, although the strongly supervised labels are accurate, their coverage is limited and cannot provide sufficient training signals for the model.

[0093] To alleviate the scarcity of strongly supervised samples, weak regional background labels are generated based on the known metallogenic belt range of the target area. The known metallogenic belt range is a favorable geological zone with mineralization potential delineated by geologists through long-term regional geological surveys, comprehensively analyzing various geological factors such as regional tectonic evolution history, distribution of magmatic activity zones, stratigraphic lithology, and the spatial distribution patterns of known mineral deposits. The boundary information of these metallogenic belts is usually stored in a geological GIS database in the form of vector polygons. In the above embodiment, it is assumed that the geological data of the western metallogenic belt of the province shows three main gold metallogenic belts, named the northern ductile shear zone type gold metallogenic belt, the central Carlin type gold metallogenic belt, and the southern epithermal low-temperature gold metallogenic belt. After the vector boundary data of these three metallogenic belts are superimposed onto a unified spatial grid, the total number of grid cells covered is approximately 45,000, accounting for 22.5% of the total number of grid cells in the study area. For grid cells falling within these metallogenic belt ranges but for which no mineral deposits have yet been discovered, the system marks them as weakly supervised positive samples and assigns a label value of 0.5. This value, between 0 and 1, reflects that these areas have mineralization potential. The generation logic of these weak labels, which represent uncertainties that have not yet been verified, is based on the premise that known metallogenic belts have similar geological backgrounds and metallogenic conditions to known deposits. Therefore, the probability of new deposits appearing in these areas is significantly higher than in areas outside the metallogenic belts. However, since these locations are not exact deposit points, their label confidence is lower than that of strongly supervised samples. Furthermore, for approximately 155,000 grid cells outside the metallogenic belts, due to the lack of clear metallogenic geological background support, the system labels them as weakly supervised negative samples and assigns them label values ​​close to 0, such as 0.1. This value reflects that while the probability of metallization in these areas is low, it is not absolutely impossible for deposits to exist, especially when exploration is insufficient. In this way, weak labels provide the model with a wide range of training signals covering all grid cells in the study area, enabling the model to learn the macroscopic geological differences between metallogenic and non-metallogenic areas. This compensates for the limitations of strongly supervised samples, which only focus on deposit locations. However, it should be noted that the uncertainty of weak labels is high; therefore, the loss function design must assign dynamically adjusted weights to weakly labeled samples to avoid inaccurate weak labels misleading model training.

[0094] During model training, knowledge embedding vectors need to be injected into the knowledge reasoning model as prior constraints. These knowledge embedding vectors are high-dimensional vector representations generated by multi-layer message passing and feature aggregation of the knowledge graph through a graph neural network. They encode prior information such as the metallogenic geological background knowledge of the region, including ore-controlling structural types, favorable lithological combinations, ore-forming element enrichment patterns, and metallogenic processes. Embedding this structured geoscientific knowledge into a data-driven deep learning model can function on two levels. First, mutual attention calculation is performed between the knowledge embedding vector and the preliminary fusion vector to perform geoscientific semantic correction on the data features in the preliminary fusion vector. The preliminary fusion vector is a comprehensive data representation obtained by a multimodal fusion network after preliminary fusion of remote sensing features, geophysical features, geochemical features, and geological vector features. The vector dimension is assumed to be 512, while the knowledge embedding vector dimension is assumed to be 256. The mutual attention calculation process involves mapping the preliminary fusion vector to a query matrix Qdata∈R through a linear transformation layer. N×dk Where N is the number of grid cells and dk is the dimension of the query vector. The knowledge embedding vector is then mapped to a key matrix K through two additional linear transformation layers. know ∈R M×dk Sum matrix V know ∈R M×dv Where M is the number of knowledge nodes related to the region in the knowledge graph, for example, including 5 main geological entity nodes and 8 relation nodes, totaling 13 nodes; dv is the dimension of the value vector, for example, 256 dimensions. Then, the similarity between the query matrix and the key matrix is ​​calculated to obtain the attention weight matrix A = softmax((Qdata×K)). T know ) / √dk)∈R N×M Each row of the attention weight matrix corresponds to a grid cell, and each column corresponds to a knowledge node. The matrix element Aij represents the semantic relevance strength between the data features of the i-th grid cell and the j-th knowledge node. Then, the attention weights are used to perform a weighted summation of the value matrix to obtain the knowledge correction vector C. know =AV kow ∈R N ×dvThe knowledge-corrected vector and the initial fusion vector are integrated through residual connections or gated fusion mechanisms to obtain a fusion representation vector after geoscientific semantic correction. For example, the initial fusion vector of a certain grid cell, without knowledge correction, shows that the cell has a high geochemical gold anomaly value of 3.2 times the background value, but a weak geophysical aeromagnetic anomaly of only 1.1 times the background value. Furthermore, the cell is located on the edge of the surrounding rocks in the Lima River mining area or the Lala North area. After mutual attention calculation with the knowledge graph, it is found that the data characteristics of this cell are consistent with the knowledge node "contact zone of epithermal gold deposit - ore-controlling structure - subvolcanic rock mass". The attention weight value of 0.68 is the highest among all knowledge nodes, indicating that the geological background of this unit is highly consistent with the epithermal mineralization. The knowledge correction vector integrates this prior information into the data feature representation, enabling the model to not only rely on the numerical intensity of geochemical anomalies but also comprehensively consider the rationality of the geological genesis of the anomalies when making predictions, thereby improving the geological reliability of the predictions. This mutual attention mechanism realizes the two-way interaction between data features and prior knowledge, allowing knowledge to guide the interpretation of data features while activating relevant knowledge nodes, avoiding the problem that overly rigid knowledge constraints prevent the model from discovering new patterns from the data.

[0095] The second approach to injecting knowledge embedding vectors into the model involves introducing the difference between the knowledge embedding vector and the initial fusion vector as a geoscientific prior constraint into a predefined loss function. The logic behind this approach is that if the data feature representation learned by the model deviates significantly from the mineralization patterns encoded in the knowledge graph, the model's predictions are considered to potentially violate geological principles. Therefore, the loss function is used to penalize the model, prompting it to adjust its parameters to align its predictions with prior knowledge. First, a knowledge consistency metric function is defined, such as using cosine similarity or Euclidean distance to calculate the difference between the initial fusion vector and the knowledge embedding vector. If cosine similarity is used as the consistency metric, for the i-th grid cell, whose initial fusion vector is fi and corresponding knowledge embedding vector is gi, the knowledge consistency score is calculated as si = fi × gi / (||fi||||gi||). This score ranges from -1 to 1. A score close to 1 indicates a high degree of consistency between the data features and the prior knowledge, while a score close to -1 indicates complete opposites. Then, a knowledge constraint loss term is defined. ; Here, m is a preset consistency threshold, for example, set to 0.6, and si represents the knowledge consistency score. This loss term means that when the consistency score is lower than the threshold, the model is penalized, and the lower the consistency, the greater the penalty. Finally, the knowledge constraint loss term is weighted by a certain coefficient. In this way, the knowledge constraint loss term prevents the model from making predictions that do not conform to geological laws based solely on the statistical correlation of the data surface, thereby enhancing the interpretability of the model and the credibility of the prediction results.

[0096] After completing the design of the knowledge injection mechanism, a weakly supervised adaptive loss function needs to be constructed to drive the model training process. This loss function needs to simultaneously process four types of information: strong supervision labels, weak regional background labels, predicted mineralization probability values, and prediction uncertainty values. The loss weights for weak regional background labels are dynamically adjusted based on the prediction uncertainty values. The weakly supervised adaptive loss function consists of three main parts. The first part is the strong supervision loss term, which constrains the model's prediction behavior at known mineral deposit locations. Assuming the set of known mineral deposit points contains 82 strongly labeled sample points, the cross-entropy between the predicted mineralization probability and the true label value 1 for each strongly labeled sample constitutes the loss value for that sample. The second part is the background constraint loss term, which uses weak regional background labels to expand the training signal. Assuming the weak label sample set contains multiple samples, where... For each j-th grid cell, the weak label value corresponding to the input feature is, for example, 0.5 within the ore-forming zone or 0.1 outside the ore-forming zone. This sample represents the uncertainty value predicted by the model for that grid cell. The basic form of the background constraint loss term is a weighted binary cross-entropy, where the cross-entropy loss between the predicted probability and the weak label value of the j-th weak label sample is multiplied by the adaptive weight coefficient of that sample. This weight coefficient is dynamically adjusted based on the prediction uncertainty. The design principle is to reduce the weight of a weak label sample in the loss function when the model's prediction of a weak label sample has high uncertainty, avoiding unreliable weak labels that force the model to train in the wrong direction; and to increase the weight of a weak label sample when the model's prediction of a weak label sample has low uncertainty, allowing the model to obtain stronger learning signals from high-confidence predictions. Therefore, the specific formula for calculating the weight coefficient is: W j =exp(-βu j ), where β is the rate at which the hyperparameter control uncertainty decays with respect to the weights, u j It is the prediction uncertainty, W j These are adaptive weighting coefficients. For example, if β is set to 2.0, when the prediction uncertainty u... j When W is 0, j The weight equals 1 and reaches its maximum value; when the prediction uncertainty value u j When W is 1.0, j The weights decay significantly to 0.135. The third part is a regularization term used to prevent overfitting, for example, L2 weight decay towards Lrge=λreg∑. l ||W l || 2 Where λreg is the regularization coefficient, such as 0.0001, W l It is the weight matrix of the l-th layer network.

[0097] Integrating the above three parts yields the complete weakly supervised adaptive loss function, which is: Weakly supervised adaptive loss function = Strong supervision loss + α × Weak supervision loss + Knowledge constraint loss × Weight coefficient + Regularization coefficient × Regularization loss. Here, α is the weight coefficient that balances the strong supervision loss and the weak supervision loss. Considering that the number of strong supervision samples is much smaller than the number of weak supervision samples, in order to prevent the weak supervision loss from overwhelming the strong supervision signal, α is usually set to a value less than 1, such as α = 0.1. Assume that the average loss of 82 strong supervision samples in a certain iteration is 0. The weighted average loss of 324,199,918 weakly supervised samples is 0.586, the knowledge constraint loss is 0.142, the regularization loss is 0.008, the weight coefficient is 0.3, and the regularization coefficient is 0.0001. Therefore, the total loss is 0.324 + 0.1 × 0.586 + 0.3 × 0.142 + 0.0001 × 0.008 = 0.426. This comprehensive loss value reflects the model's performance in four aspects: fitting known ore deposits, learning regional background, knowledge consistency constraints, and parameter complexity control.

[0098] Based on the calculated comprehensive loss value, the parameters of the knowledge reasoning model and the graph neural network need to be iteratively updated multiple times using the backpropagation algorithm until the model converges. The implementation process of the backpropagation algorithm is as follows: first, calculate the gradient of the loss function with respect to the output layer of the model; then, propagate forward layer by layer to calculate the gradient of the loss with respect to the parameters of each layer; finally, use a gradient descent optimization algorithm to update the parameters. The Adam optimizer is usually used because it combines the advantages of momentum method and adaptive learning rate, resulting in good training stability and fast convergence speed. The main hyperparameters of the Adam optimizer include an initial learning rate set to 0.001, an exponential decay rate of 0.9 for the first moment estimate, and an exponential decay rate of 0.999 for the second moment estimate. For example, the total number of parameters in the model is approximately 850. The dataset contains approximately 3.2 million parameters for the multimodal feature extractor, 2.8 million parameters for the multimodal fusion network, 1.5 million parameters for the graph neural network, and 1 million parameters for the prediction head. The training dataset includes 82 strongly supervised samples and 199,918 weakly supervised samples, divided into training, validation, and test sets in an 8:1:1 ratio. The training set contains 66 strongly supervised samples and 159,934 weakly supervised samples for model parameter updates. The validation set contains 8 strongly supervised samples and 19,992 weakly supervised samples to monitor overfitting and perform early stopping during training. The test set contains 8 strongly supervised samples and 19,992 weakly supervised samples for final model performance evaluation. The training process employs... The mini-batch stochastic gradient descent strategy uses 32 samples per batch, including at least 4 strongly supervised samples to ensure a clear positive signal in each batch. Training lasts for 200 epochs, meaning the training set is traversed 200 times. Within each epoch, the training set is first randomly shuffled, then input into the model batch by batch for forward propagation to calculate the predicted mineralization probability and prediction uncertainty for each sample. The prediction uncertainty is calculated using the Monte Carlo Dropout method: during inference, the Dropout layer remains active, and the same sample is forward-propagated T times (e.g., T=10 times) to obtain T different predicted probability values. The standard deviation of these T predicted values ​​is then calculated as the uncertainty measure. The results and true labels are input into a weakly supervised adaptive loss function to calculate the comprehensive loss value. Then, the backpropagation algorithm is called to calculate the gradient of the loss with respect to all parameters. When calculating the gradient, it is important to note that the gradient of the knowledge constraint loss term will not only be backpropagated to the parameters of the knowledge inference model but also to the parameters of the graph neural network. This means that the knowledge embedding representation of the graph neural network will also be fine-tuned according to the feedback from downstream tasks, thereby achieving end-to-end joint optimization of knowledge graph encoding and mineralization prediction tasks. After obtaining the gradient, the Adam optimizer is used to update the parameters. In the early stage of training, i.e., the first 50 epochs, the learning rate is kept at 0.001 to quickly reduce the loss. From the 51st epoch onwards, a cosine annealing learning rate scheduling strategy is used to gradually reduce the learning rate to 0.To achieve fine-tuning of model parameters, the model performance is evaluated using a validation set after each epoch. The average loss and AUC area under the ROC curve on the validation set are calculated. If the validation set loss does not decrease for 20 consecutive epochs, an early stopping mechanism is triggered to terminate training and prevent overfitting.

[0099] The final model, after training, was evaluated on the test set. Of the 8 strongly supervised test samples, 7 were correctly predicted as high-probability mineralization areas, with mineralization probabilities all greater than 0.85. Only one sample had a predicted probability of 0.67, slightly lower than expected but still above the threshold. Among the 19,992 weakly supervised test samples, the average predicted probability of samples within the metallogenic belt was 0.58, while the average predicted probability of samples outside the metallogenic belt was 0.13, showing a significant difference. This indicates that the model effectively learned the correlation between regional geological background and mineralization potential. The overall AUC of the test set reached 0.915, indicating that the model has excellent ability to distinguish between positive and negative samples. Further analysis of prediction uncertainty revealed that the average uncertainty of the strongly supervised test samples was 0.08, indicating that the model was very confident in predicting known mineral deposits. The average uncertainty of the weakly supervised samples within the metallogenic belt was 0.31, indicating that the model's predictions of these potential mineralization areas have some uncertainty, which is consistent with the fact that these areas have not yet been verified. The average uncertainty of the weakly supervised samples outside the metallogenic belt was 0.42, indicating that the model's prediction uncertainty for low-probability areas was higher.

[0100] S106: Determine the preset probability threshold and preset uncertainty threshold based on the geological conditions of the target area, and compare the mineralization probability value of each spatial unit with the preset probability threshold.

[0101] In S106 above, the core objective of determining the preset probability threshold and preset uncertainty threshold based on the geological conditions of the target area and comparing the mineralization probability value of each spatial unit with the preset probability threshold is to achieve spatial adaptive adjustment of the target area delineation standard. This allows the target area screening strategy to fully consider the differences in geological background of different regions, thus avoiding the omission of high-potential target areas in areas with favorable geological conditions due to excessively low thresholds, and preventing the misjudgment of low-reliability areas as mineral exploration target areas in areas with unfavorable geological conditions due to excessively high thresholds. Determining the preset probability threshold and preset uncertainty threshold based on the geological conditions of the target area specifically includes: acquiring the ore-controlling element characteristics of each spatial unit within the target area, including fault structural density, known mineral deposit distribution density, and stratigraphic mineralization favorability; calculating the geological mineralization favorability index of each spatial unit by weighted summation based on the fault structural density, known mineral deposit distribution density, and stratigraphic mineralization favorability; calculating the preset probability threshold corresponding to each spatial unit based on the geological mineralization favorability index; and calculating the preset uncertainty threshold corresponding to each spatial unit based on the geological mineralization favorability index.

[0102] Specifically, the ore-controlling element characteristics of each spatial unit within the target area are obtained. Ore-controlling element characteristics refer to the spatial distribution characteristics of geological elements that have controlling or indicative significance for mineralization. These mainly include three key elements: fault structural density, known mineral occurrence distribution density, and stratigraphic mineralization favorability. Fault structural density reflects the intensity of regional tectonic activity and the degree of development of ore-guiding and ore-hosting spaces. Known mineral occurrence distribution density reflects the accumulated results of historical mineral exploration work in the region and the actual manifestation of mineralization. Stratigraphic mineralization favorability reflects the material basis supporting mineralization from different stratigraphic lithological combinations. These three types of ore-controlling elements comprehensively characterize the geological and metallogenic background of each spatial unit from three dimensions: tectonic conditions, mineralization verification, and material source. For example, a metallogenic belt in the western part of a province, covering an area of ​​50,000 square kilometers, is divided into 200,000 grid units of 500m x 500m each. First, the fault line density is calculated. Vector data of all fault lines in the study area are extracted from the geological vector database, including regional deep faults, secondary faults, and brittle shear zones, totaling 327 fault lines with a total length of approximately 15,800 kilometers. For each grid unit, a circular buffer zone with a radius of 2.5 kilometers is established with the center point of the unit as the center. The area of ​​this buffer zone is approximately 19.63 square kilometers. The total length of all fault line segments within the buffer zone is counted and then divided by the buffer zone area to obtain the fault line density value of that unit, expressed in kilometers per square kilometer (km / km). 2 In the specific calculation, assuming a grid cell located within the northern ductile shear zone is numbered U001, and its 2.5 km buffer zone contains three fracture segments with lengths of 4.2 km, 1.8 km, and 2.3 km respectively, totaling 8.3 km in length, then the fracture structure line density of this cell is approximately 8.3 / 19.63 ≈ 0.423 km / km. 2 The spatial distribution statistics of fracture structural line density, obtained by calculating all 200,000 grid cells, show that the average fracture structural line density across the entire region is 0.158 km / km. 2 The standard deviation is 0.127 km / km 2 The maximum value reached 0.895 km / km 2 The minimum value appearing at the intersection of the main fault zone and the secondary fault in the north is 0.002 km / km. 2 It appears within a stable block in the southwest.

[0103] The distribution density of known mineral deposits was calculated by extracting the coordinate information of all known mineral deposits and mineral points within the study area from the mineral deposit database. This included the aforementioned 87 gold deposits and other types of mineral deposits and mineral points such as copper and lead-zinc deposits, totaling 135 mineral deposits and mineral points. For each grid cell, a circular buffer with a radius of 5 kilometers was established with the center point of the grid cell as the center. The area of ​​this buffer is approximately 78.54 square kilometers. The reason for setting the buffer radius to 5 kilometers is that the spatial influence range of mineral deposits is usually on a scale of several kilometers. A buffer that is too small would result in a mineral point density of zero in many cells, while a buffer that is too large would smooth out spatial differences. The number of known mineral deposits and mineral points within the buffer was counted and then divided by the buffer area to obtain the distribution density value of known mineral points in that cell, in units of 1 / km². 2 Referring to the example grid cell U001 above, which contains 6 known gold deposits within its 5-kilometer buffer zone, the distribution density of known deposits in this cell is approximately 6 / 78.54 ≈ 0.076 deposits / km. 2 Since the 5-kilometer buffer zone of grid cell U150000 does not contain any known mineral deposits, the distribution density of known mineral deposits in this cell is 0 / 78.54 = 0 per km. 2 After calculating the distribution density of known mineral deposits across all grid cells, the statistical results show that the average density for the entire area is 0.014 deposits / km². 2 The standard deviation is 0.028 units / km. 2 The maximum value reached 0.153 per km. 2 It appears in the core mineralization area of ​​the central Carlin-type gold metallogenic belt.

[0104] The calculation of stratigraphic mineralization favorability is a quantitative evaluation index of the mineralization potential of different stratigraphic lithological combinations. It is determined based on geologists' understanding of regional metallogenic regularities and statistical results from historical mineral exploration practices. Within this study area, 12 major stratigraphic lithological units were divided according to geological maps, including Precambrian metamorphic rocks, Paleozoic carbonate rocks, Mesozoic volcanic rocks, and Mesozoic intrusive rocks. Based on the spatial correlation analysis between each stratigraphic unit and known mineral deposits, and the lithological characteristics of the strata influencing the enrichment of ore-forming elements and the degree of mineralization and alteration, a team of geological experts assigned a mineralization favorability score to each stratigraphic unit. This score uses a normalized value from 0 to 1. Specifically, the Paleozoic Devonian shallow metamorphic clastic rocks, as the main host rocks of Carlin-type gold deposits, were assigned the highest favorability value of 1.0. The Mesozoic Hercynian or Proterozoic host rocks of the Limahe mining area or the Lalabei area were also assigned a favorable value due to their association with marine environments. Western magmatic segregation-type copper-nickel deposits, or those where nickel is an associated element and closely related to hydrothermal mineralization, are assigned a favorable value of 0.85. Precambrian gneiss is assigned a favorable value of 0.70 because it can serve as an initial source layer for gold. Paleozoic Carboniferous limestone is assigned a favorable value of 0.65 because it can serve as a chemical barrier and ore-hosting space for hydrothermal fluids. Quaternary loose sediments are assigned the lowest favorable value of 0.10 because they are not related to endogenous mineralization. For each grid cell, the stratigraphic lithology unit at its location is queried, and the corresponding mineralization favorable score is extracted as the stratigraphic mineralization favorable value of that unit. In specific implementation, assuming that the geological map of the location of grid cell U001 shows that its lithology is Paleozoic Devonian slate, then the stratigraphic mineralization favorableness of this unit is 1.0, while the location of grid cell U150000 is covered by Quaternary alluvial deposits, then its stratigraphic mineralization favorableness is 0.10.

[0105] After obtaining the characteristics of three ore-controlling elements—fracture structural line density, known mineral occurrence distribution density, and stratigraphic mineralization favorability—it is necessary to calculate the geological mineralization favorability index for each spatial unit by weighted summation based on these three characteristics. The geological mineralization favorability index is a quantitative indicator that comprehensively reflects the supporting capacity of the geological background of each grid unit for mineralization. The calculation formula is F. i =w1×D f +w2×D d +w3×S s , where F i D is the geological mineralization favorable index of the i-th grid cell. f D is the fracture structure line density of this unit. d S is the known mineral deposit distribution density of this unit. s It represents the mineralization favorability of the stratigraphy in this unit. w1, w2, and w3 are the weight coefficients of the three types of mineralization control elements and satisfy w1+w2+w3=1.

[0106] The determination of weighting coefficients requires comprehensive consideration of the relative importance of different ore-controlling elements to mineralization prediction. In this embodiment, the weighting coefficients are determined using the analytic hierarchy process (AHP) based on regional metallogenic regularity analysis and the opinions of multiple geological experts. Since stratigraphic mineralization favorability represents the material basis and ore-bearing space of mineralization, which is a primary condition for mineralization, it is assigned the highest weight, w3 = 0.45. The density of known mineral occurrences reflects the actual manifestation of mineralization and the historical verification of prospecting work, thus having high indicative significance; therefore, it is assigned a weight, w2 = 0.35. Although fault line density plays an important role in controlling and guiding ore, its indicative role is relatively indirect; therefore, it is assigned a weight, w1 = 0.20. However, due to the significant differences in the dimensions and numerical ranges of the three types of ore-controlling elements, the numerical range of fault line density is from 0.002 to 0.895 km / km. 2 Between these values, the known mineral deposit density ranges from 0 to 0.153 per km. 2 Between these values, the numerical range of the mineralization favorability of the stratigraphy is between 0.10 and 1.0. Directly performing weighted summation will lead to the dominant factor in the calculation result. Therefore, it is necessary to standardize the three types of ore-controlling factors before weighted summation. The maximum-minimum standardization method is used to map the values ​​of each factor to the interval between 0 and 1. The standardization formula for the fault structural line density is (D f1 i) = (D f ,i)-(D f ,min) / (D f ,max)-(D f ,min), where (D f (max) and (D) f (min) represents the maximum and minimum values ​​of the fault structural line density in the entire region, D f1 , where i represents the processed fault structural line density. Then, referring to the above-mentioned standardized formula for fault structural line density, the known mineral deposit distribution density and stratigraphic mineralization favorability are standardized to obtain the processed known mineral deposit distribution density and processed stratigraphic mineralization favorability. Finally, the standardized ore-controlling element values ​​are weighted and summed to obtain the geological mineralization favorability index.

[0107] For example, the fracture structure line density of grid cell U001 is 0.423 km / km. 2 The standardized value is 0.472, and the known mineral deposit density is 0.076 per km. 2After standardization, the value is 0.497, and the stratigraphic mineralization favorability is 1.0. Therefore, the geological mineralization favorability index of this unit is 0.20×0.472+0.35×0.497+0.45×1.0=0.718. 0.718 indicates that unit U001 has a high geological mineralization favorable background. After calculating all grid units, the statistical distribution of the geological mineralization favorability index was obtained. The average value of the geological mineralization favorability index for the entire area is 0.285, the standard deviation is 0.218, the maximum value is 0.891, which appears in the core area of ​​the central metallogenic belt, and the minimum value is 0.003, which appears in the stable block in the southwest. The spatial distribution of this index clearly outlines the favorable geological background areas of the three main metallogenic belts in the study area.

[0108] After obtaining the geological mineralization favorability index of each spatial unit, a preset probability threshold is calculated for each spatial unit based on the index. The basic logic of the adaptive adjustment strategy of the preset probability threshold is that a relatively high probability threshold should be used in areas with a high geological mineralization favorability index to screen out areas with relatively high mineralization probability in the area, avoiding the delineation of too many target areas due to the threshold being too low, which would lead to the dispersion of subsequent verification resources. In contrast, the probability threshold requirement should be appropriately reduced in areas with a low geological mineralization favorability index to discover potential mineral exploration opportunities and avoid completely ignoring the mineralization possibility of these areas due to the threshold being too high.

[0109] The first step is to calculate the ratio of the geological mineralization favorability index to the sum of the index plus 1 to obtain the normalized favorability value. The purpose of this normalization operation is to map the geological mineralization favorability index to the range of 0 to 1 and to avoid excessive influence on the threshold adjustment when the index value is too large. For example, for grid cell U001, its geological mineralization favorability index is 0.718, then the normalized favorability value is 0.718 / (1+0.178) = 0.418.

[0110] The second step is to obtain the probability scaling factor by subtracting the product of the preset probability adjustment coefficient and the normalized favorability value from 1. The preset probability adjustment coefficient ranges from greater than 0 to less than 1. Its function is to control the intensity of the geological background's adjustment of the probability threshold. Based on the spatial differences in geological conditions of the study area and previous mineral exploration experience, the preset probability adjustment coefficient is set to 0.3. The probability scaling factor ranges from the difference between 1 and the preset probability adjustment coefficient to 1. When the normalized favorability value is 0, the scaling factor reaches its maximum value of 1. When the normalized favorability value is at its maximum and close to 1, the scaling factor reaches its minimum value of 1 minus the preset probability adjustment coefficient. The scaling factor means that a proportional adjustment is made based on the base probability threshold. In areas with low favorability, the scaling factor is close to 1, meaning the threshold is close to the base value and does not decrease too much. In areas with high favorability, the scaling factor is less than 1, meaning the threshold is relatively higher than the base value. For example, for grid cell U001, its normalized favorability value is 0.418, so the probability scaling factor is 0.875.

[0111] The third step is to multiply the probability scaling factor by the preset base probability threshold to obtain the preset probability threshold for each cell. The preset base probability threshold is determined based on the statistical characteristics of the mineralization probability distribution across the entire area and the resource input capacity for mineral exploration. The preset base probability threshold can be set to 0.70, which corresponds to the delineation threshold when a unified standard is applied across the entire area. Applying the probability scaling factor to the base probability threshold yields a spatially adaptive preset probability threshold. For example, if the probability scaling factor for grid cell U001 is 0.875 and the preset base probability threshold is 0.70, then the preset probability threshold is 0.613.

[0112] A similar method is used to calculate the preset uncertainty threshold for each spatial unit based on the geological mineralization favorability index. The basic logic of the adaptive adjustment strategy for the preset uncertainty threshold is that a relatively low uncertainty threshold should be used in areas with a high geological mineralization favorability index, meaning a higher requirement for the certainty of the prediction results is needed because these areas have a good geological background and the model prediction should be more reliable. Conversely, in areas with a low geological mineralization favorability index, the uncertainty tolerance can be appropriately relaxed, meaning a higher uncertainty threshold can be used to avoid completely excluding these areas due to overly strict certainty requirements. The first step is to calculate the negative geological mineralization favorability index power of the natural constant e to obtain the exponential decay value. The second step is to add 1 to the product of the preset uncertainty adjustment coefficient and the exponential decay value to obtain the uncertainty scaling coefficient. The preset uncertainty adjustment coefficient has a value greater than 0, and its role is to control the intensity of the geological background's adjustment of the uncertainty threshold. The third step is to subtract the preset uncertainty scaling coefficient from 2 and multiply it by the basic uncertainty threshold to obtain the preset uncertainty threshold for each unit. The basic uncertainty threshold is determined based on the statistical distribution of the model uncertainty quantification results and the overall requirements for prediction reliability. To ensure that the delineated target area has high prediction reliability, the basic uncertainty threshold is set to 0.30. For example, for grid cell U001, its geological mineralization favorability index is 0.718, and the index decay value is 0.488. When the index decay value is 0.488, the uncertainty scaling factor is 1.244. When the uncertainty scaling factor is 1.244 and the basic uncertainty threshold is 0.30, the preset uncertainty threshold is 0.227.

[0113] After calculating the preset probability threshold and preset uncertainty threshold for each spatial unit, the mineralization probability value of each spatial unit is compared with the corresponding preset probability threshold, and the predicted uncertainty value is also compared with the preset uncertainty threshold. Only grid units that simultaneously satisfy the condition that the mineralization probability value is greater than or equal to the preset probability threshold and the predicted uncertainty value is less than or equal to the preset uncertainty threshold are initially identified as candidate target areas. For example, grid unit U001 has a mineralization probability value of 0.812 and a predicted uncertainty value of 0.18 obtained through model prediction, a preset probability threshold of 0.788, and a preset uncertainty threshold of 0.227. Since both conditions of 0.812 ≥ 0.788 and 0.18 ≤ 0.227 are met, unit U001 is identified as a candidate target area. Another grid cell U002, located in the same metallogenic belt, has a predicted mineralization probability of 0.765 and a predicted uncertainty of 0.15. Although the uncertainty meets the threshold requirement, the mineralization probability of 0.765 is lower than the preset probability threshold of 0.788. Because the geological favorableness index of this cell is also high, the threshold is raised. Therefore, cell U002 was not delineated.

[0114] S107: When the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, the target spatial unit within the target area is determined to be delineated as the mineral target area.

[0115] In step S107 above, the mineralization probability value and predicted uncertainty value of each spatial unit within the target area are obtained. Based on the above, a preset probability threshold and a preset uncertainty threshold corresponding to each spatial unit are determined. One spatial unit is randomly selected from multiple spatial units in the target area as the target spatial unit, and the preset probability threshold and preset uncertainty threshold corresponding to the target spatial unit are retrieved. The mineralization probability value of the target spatial unit is then compared with the preset probability threshold. If the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, it is further determined whether the predicted uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold. Only target spatial units that simultaneously meet both conditions are marked as part of the candidate mineralization target area. For example, the study area is divided into 200,000 spatial units, each 500 meters by 500 meters. After the model completes the prediction, it first queries the knowledge graph for the corresponding threshold setting rules based on the geological structural unit type to which each spatial unit belongs. For the target spatial unit numbered A12345, the preset probability threshold for the target spatial unit is 0.6 and the preset uncertainty threshold is 0.25. The model calculates that the mineralization probability value of the target spatial unit is 0.72, which is greater than the preset probability threshold of 0.6. At the same time, the prediction uncertainty value of the unit is 0.18, which is less than the preset uncertainty threshold of 0.25. Since both the mineralization probability and prediction uncertainty are met, the target spatial unit is determined to have the basic qualifications to be designated as a mineral target area.

[0116] However, relying solely on the threshold judgment of a single spatial unit is insufficient to form a target area with practical exploration significance, because real mineral deposits usually have a certain spatial extension rather than isolated point distributions. Therefore, further spatial continuity analysis is performed, and a region growing algorithm is used to search for adjacent spatial units that also meet the dual threshold constraints in the direction of the target spatial unit as the seed point. When the number of adjacent units that continuously meet the conditions reaches the preset minimum target area threshold, for example, at least 16 consecutive units corresponding to 2 square kilometers, this continuous area is delineated as an independent mineral target area and assigned a unique target area number. In the above example, after expanding with unit A12345 as the center using the region growing algorithm, a total of 23 adjacent spatial units that meet the conditions were identified, forming a continuous target area with an area of ​​approximately 5.75 square kilometers. This target area is named target area T001, and comprehensive evaluation indicators are calculated, including an average mineralization probability of 0.68, a standard deviation of 0.05, and an average uncertainty of 0.16 for all units within the target area. These statistical indicators further verify that the target area as a whole has high mineralization potential and low prediction risk.

[0117] Furthermore, another target spatial unit B67890 located on the edge of the study area, although its mineralization probability value reached 0.70, and the preset probability threshold was set at 0.75 with a preset uncertainty threshold of 0.15, although the predicted uncertainty value of this unit was 0.12, which met the uncertainty requirements, the mineralization probability value of 0.70 failed to reach the more stringent preset probability threshold of 0.75. Therefore, it was determined that this unit did not meet the conditions for delineating as a mineralization target area. This differentiated threshold setting strategy effectively reflects the flexibility of risk control under different geological backgrounds. Through the intelligent delineation process of dual threshold constraints and spatial continuity analysis, it can effectively avoid the risk of misjudgment in high uncertainty areas while making full use of the model prediction results. Finally, candidate mineralization target areas were delineated in the entire study area. This delineation strategy provides users with differentiated exploration deployment suggestions, enabling limited exploration resources to be prioritized for target areas with the lowest risk and highest return.

[0118] This application also provides a mineralization prediction system based on multimodal geological data and knowledge graphs. Figure 2 This is a schematic diagram of the structure of a mineralization prediction system based on multimodal geological data and knowledge graphs provided in an embodiment of this application. (Refer to...) Figure 2 The system includes an acquisition unit 201, an extraction and fusion unit 202, a construction unit 203, and a prediction unit 204. Acquisition unit 201 acquires multimodal geological data, geological exploration reports, and mineral deposit survey information for the target area; Extraction and fusion unit 202 is used to extract multi-scale features from multimodal geological data to obtain multiple feature vectors. Each feature vector corresponds to a type of scale feature in the multi-scale features. Based on the causal hierarchy relationship, the multiple feature vectors are processed and fused to obtain a preliminary fusion vector. The causal hierarchy relationship is the relationship from low-level geological phenomena to high-level metallogenic models. Construction unit 203 constructs a knowledge graph related to the target area based on geological exploration reports and mineral deposit survey information, and uses a graph neural network to encode the knowledge graph and generate knowledge embedding vectors; Prediction unit 204 inputs the knowledge embedding vector and the preliminary fusion vector into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit in the target area; based on the geological state of the target area, it determines the preset probability threshold and preset uncertainty threshold, and compares the mineralization probability value of each spatial unit with the preset probability threshold; when the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, it determines that the target spatial unit in the target area is delineated as the mineral target area, and the target spatial unit is any one of the spatial units.

[0119] In one possible implementation, the acquisition unit 201 is used to acquire remote sensing image data, geophysical and geochemical data, and geological vector map data from multimodal geological data; to perform multi-level feature extraction on the remote sensing image data to obtain small-scale, mesoscale, and large-scale remote sensing feature vectors, respectively; the extraction and fusion unit 202 is used to perform multi-band signal decomposition on the geophysical and geochemical data using wavelet transform to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors, respectively; and to perform multi-level semantic parsing on the geological vector map data based on geological levels to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors, respectively. Large-scale, meso-scale, and large-scale geological feature vectors are generated. These large-scale remote sensing feature vectors, geophysical and geochemical feature vectors, and geological feature vectors are concatenated to obtain a large-scale evidence vector. Similarly, meso-scale remote sensing feature vectors, geophysical and geochemical feature vectors, and geological feature vectors are concatenated to obtain a meso-scale evidence vector. Small-scale remote sensing feature vectors, geophysical and geochemical feature vectors, and geological feature vectors are concatenated to obtain a small-scale evidence vector. Finally, the small-scale, meso-scale, and large-scale evidence vectors are output as feature vectors.

[0120] In one possible implementation, the extraction and fusion unit 202 is used to calculate the similarity between the large-scale evidence vector and each background pattern in the preset metallogenic background pattern library, and to select the target background pattern corresponding to the maximum value from multiple similarity results. The confidence vector corresponding to the target background pattern is used as the regional background diagnosis vector. Based on the preset geological inference matrix and the regional background diagnosis vector, matrix multiplication is performed to generate a mesoscale attention weight template. The preset geological inference matrix is ​​an n×m dimensional matrix, where n represents the n rows of the preset geological inference matrix corresponding to n mesoscale features, and m represents the preset geological inference matrix. The m columns of the matrix correspond to m regional mineralization modes; the mesoscale evidence vector is multiplied element-wise with the mesoscale attention weight template to obtain the ore-controlling structure judgment vector; matrix multiplication is performed based on the preset local indicator matrix and the ore-controlling structure judgment vector to obtain the small-scale weight template. The preset local indicator matrix is ​​a p×q dimensional matrix, where p represents the p rows of the preset local indicator matrix corresponding to p small-scale mineral exploration marker features, and q represents the q columns of the preset local indicator matrix corresponding to q mesoscale ore-controlling structure types; the small-scale evidence vector is multiplied element-wise with the small-scale weight template to obtain the preliminary fusion vector.

[0121] In one possible implementation, the acquisition unit 201 is used to extract geological knowledge triples from geological exploration reports and mineral deposit survey information, and store the geological knowledge triples in a graph database to construct a global knowledge graph related to the target area; the construction unit 203 is used to divide the target area into multiple spatial units, identify the core geological entities contained in the current spatial unit, and perform association retrieval in the global knowledge graph with the core geological entities as the center to extract the local knowledge subgraph corresponding to the target spatial unit, where the current spatial unit is any grid unit among the multiple spatial units; based on a graph neural network, each node in the local knowledge subgraph is iteratively aggregated and encoded in multiple rounds to obtain the node embedding vector of each node in the local knowledge subgraph; the node embedding vectors of all nodes in the local knowledge subgraph are graph pooled to obtain the initial knowledge embedding vector corresponding to the target spatial unit; the initial knowledge embedding vectors corresponding to all spatial units are summarized to obtain the knowledge embedding vector.

[0122] In one possible implementation, the prediction unit 204 is used to deconstruct the initial fusion vector into a multi-level data evidence set; through the evidence questioning module, guided by the knowledge embedding vector, the data evidence set is subjected to correlation questioning processing to obtain the inferred evidence vector; through the evidence construction module, the inferred evidence vector and the knowledge embedding vector are concatenated to generate a comprehensive evidence vector; the comprehensive evidence vector is input into the evidence network of the evidence evaluation module, and mapped through the output layer of the evidence network and the non-negative activation function to output the first evidence strength supporting the mineralization hypothesis and the second evidence strength supporting the non-mineralization hypothesis; the sum of 1 and the first evidence strength and the second evidence strength is calculated to obtain the evidence sum, and 1 is divided by the evidence sum to obtain the prediction uncertainty value; the sum of 1 and the first evidence strength is calculated to obtain the numerator evidence value, and the sum of the first evidence strength, the second evidence strength and the preset values ​​corresponding to the number of categories of the classification problem are calculated to obtain the denominator evidence value, and the numerator evidence value is divided by the denominator evidence value to obtain the mineralization probability value.

[0123] In one possible implementation, the prediction unit 204 is used to project the knowledge embedding vector into the first linear transformation layer of the evidence questioning module to generate a query vector; to project each sub-feature vector in the data evidence set into the second and third linear transformation layers of the evidence questioning module to generate key vectors and value vectors respectively; to calculate the inner product of the query vector and each key vector, and to normalize it using a normalized exponential function to obtain the corresponding attention weights; and to use the attention weights to perform a weighted summation of all value vectors to generate the inference evidence vector.

[0124] In one possible implementation, the acquisition unit 201 is used to acquire the ore-controlling element characteristics of each spatial unit within the target area. The ore-controlling element characteristics include fault structural line density, known mineral point distribution density, and stratigraphic mineralization favorability. The prediction unit 204 is used to calculate the geological mineralization favorability index of each spatial unit by performing a weighted summation based on the fault structural line density, known mineral point distribution density, and stratigraphic mineralization favorability. Based on the geological mineralization favorability index, a preset probability threshold corresponding to each spatial unit is calculated. Based on the geological mineralization favorability index, a preset uncertainty threshold corresponding to each spatial unit is calculated.

[0125] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0126] This application also discloses an electronic device. (See reference...) Figure 3 , Figure 3 This application provides a schematic diagram of the structure of an electronic device. The electronic device 300 may include: at least one processor 301, at least one network interface 304, a user interface 303, a memory 302, and at least one communication bus 305.

[0127] The communication bus 305 is used to enable communication between these components.

[0128] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0129] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0130] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 302, and by calling data stored in memory 302. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and application requests; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0131] The memory 302 may include random access memory (RAM) or read-only memory. Optionally, the memory 302 may include a non-transitory computer-readable storage medium. The memory 302 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 302 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc. The data storage area may store data involved in the various method embodiments described above. Optionally, the memory 302 may also be at least one storage device located remotely from the aforementioned processor 301.

[0132] like Figure 3 As shown, the memory 302, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for mineralization prediction based on multimodal geological data and knowledge graphs.

[0133] exist Figure 3 In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call the application program stored in the memory 302 that is based on multimodal geological data and knowledge graph for mineralization prediction. When executed by one or more processors, the electronic device performs one or more of the methods described in the above embodiments.

[0134] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0135] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0136] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between devices or units may be electrical or other forms.

[0137] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0139] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0140] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and practical application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure.

Claims

1. A mineralization prediction method based on multimodal geological data and knowledge graphs, characterized in that, The method includes: Acquire multimodal geological data, geological exploration reports, and mineral deposit survey information for the target area; Multi-scale feature extraction is performed on the multimodal geological data to obtain multiple feature vectors. Each feature vector corresponds to a type of scale feature in the multi-scale features. The multi-scale feature extraction of the multimodal geological data to obtain multiple feature vectors specifically includes: Remote sensing image data, geophysical and geochemical data, and geological vector map data are obtained from the multimodal geological data. Multi-level feature extraction is performed on the remote sensing image data to obtain small-scale, mesoscale, and large-scale remote sensing feature vectors, respectively. The small-scale remote sensing feature vector reflects the local texture intensity of the input image under a specific convolution kernel response. The mesoscale remote sensing feature vector represents medium-scale mineralization control elements, including rock mass boundary morphology, fault zone strike extension, and the intersection relationships of multiple structural groups. The large-scale remote sensing feature vector… Vector representations encompass factors that macroscopically control mineralization, including tectonic location, regional magmatic belt distribution, and the dominant tectonic system. Wavelet transform is used to perform multi-band signal decomposition on the geophysical and geochemical data to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors. The small-scale feature vectors reflect locally enriched mineralization centers or anomalies; the mesoscale feature vectors characterize the migration paths and alteration zoning of the ore-forming hydrothermal system; and the large-scale feature vectors reflect spatial... The overall characteristics of the regional geophysical field of the element are analyzed. Multi-level semantic parsing is performed on the geological vector map data based on geological levels to obtain small-scale, mesoscale, and large-scale geological feature vectors. The small-scale geological feature vectors characterize the direct geological environment of the spatial unit; the mesoscale geological feature vectors characterize the position of the spatial unit within the regional stratigraphic framework and tectonic system; and the large-scale geological feature vectors characterize the regional tectonic background and large-scale ore-controlling environment of the spatial unit. The large-scale remote sensing feature vectors, large-scale geophysical and geochemical feature vectors, and large-scale geological feature vectors are concatenated to obtain a large-scale evidence vector. The mesoscale remote sensing feature vectors, mesoscale geophysical and geochemical feature vectors, and mesoscale geological feature vectors are concatenated to obtain a mesoscale evidence vector. The small-scale remote sensing feature vectors, small-scale geophysical and geochemical feature vectors, and small-scale geological feature vectors are concatenated to obtain a small-scale evidence vector. The small-scale evidence vectors, mesoscale evidence vectors, and large-scale evidence vectors are then output as the feature vectors. Based on a causal hierarchy, multiple feature vectors are processed and fused to obtain a preliminary fused vector. The causal hierarchy represents the relationship from low-level geological phenomena to high-level metallogenic models. Specifically, this process includes: calculating the similarity between the large-scale evidence vector and various background models in a pre-defined metallogenic background model library; selecting the target background model corresponding to the maximum similarity result from multiple similarity results; and using the confidence vector corresponding to the target background model as the regional background diagnostic vector. The pre-defined metallogenic background model library is a knowledge base constructed by summarizing historical metallogenic data and geological literature, storing standard feature templates for various typical regional metallogenic models. Furthermore, matrix multiplication is performed based on a pre-defined geological inference matrix and the regional background diagnostic vector to generate a mesoscale attention weight template. The pre-defined geological inference matrix is ​​an n×m dimensional matrix, where n represents the n rows of the pre-defined geological inference matrix corresponding to n mesoscale features, and m represents... The m columns of the preset geological inference matrix correspond to m regional metallogenic models, where the mesoscale features represent the feature values ​​in the mesoscale evidence vector. The mesoscale evidence vector is multiplied element-wise with the mesoscale attention weight template to obtain the ore-controlling structure judgment vector. A small-scale weight template is obtained by matrix multiplication based on the preset local indicator matrix and the ore-controlling structure judgment vector. The preset local indicator matrix is ​​a p×q dimensional matrix, a knowledge matrix constructed based on geological prospecting experience. Here, p represents the p rows of the preset local indicator matrix corresponding to p small-scale prospecting marker features, and q represents the q columns of the preset local indicator matrix corresponding to q mesoscale ore-controlling structure types. p represents the feature dimension of the small-scale evidence vector, and q represents the number of types of mesoscale ore-controlling structures, determined by cluster analysis of the mesoscale evidence vector. The small-scale evidence vector is multiplied element-wise with the small-scale weight template to obtain the preliminary fusion vector. Based on the geological exploration report and the mineral deposit survey information, a knowledge graph related to the target area is constructed, and a graph neural network is used to encode the knowledge graph to generate knowledge embedding vectors; The knowledge embedding vector and the preliminary fusion vector are input into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit in the target area; Based on the geological conditions of the target area, a preset probability threshold and a preset uncertainty threshold are determined, and the mineralization probability value of each spatial unit is compared with the preset probability threshold. When the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, the target spatial unit in the target area is determined to be delineated as a mineral target area, and the target spatial unit is any one of the spatial units.

2. The method according to claim 1, characterized in that, The process of constructing a knowledge graph related to the target area based on the geological exploration report and the mineral deposit survey information, and encoding the knowledge graph using a graph neural network to generate knowledge embedding vectors, specifically includes: Geological knowledge triplets are extracted from the geological exploration report and the mineral deposit survey information, and the geological knowledge triplets are stored in the graph database to construct a global knowledge graph related to the target area; The target area is divided into multiple spatial units. The core geological entity contained in the current spatial unit is identified. The core geological entity is used as the center to perform an association search on the global knowledge graph to extract the local knowledge subgraph corresponding to the current spatial unit. The current spatial unit is any one of the multiple spatial units. Based on the graph neural network, each node in the local knowledge subgraph is iteratively aggregated and encoded in multiple rounds to obtain the node embedding vector of each node in the local knowledge subgraph. Graph pooling is performed on the node embedding vectors of all nodes in the local knowledge subgraph to obtain the initial knowledge embedding vector corresponding to the current spatial unit. The initial knowledge embedding vectors corresponding to all spatial units are summarized to obtain the knowledge embedding vector.

3. The method according to claim 1, characterized in that, The knowledge reasoning model includes an evidence questioning module, an evidence construction module, and an evidence evaluation module. The knowledge embedding vector and the preliminary fusion vector are input into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit within the target area. Specifically, this includes: The initial fusion vector is deconstructed into a multi-level data evidence set; The evidence questioning module, guided by the knowledge embedding vector, performs correlation questioning on the data evidence set to obtain the inferred evidence vector. The evidence construction module concatenates the reasoned evidence vector with the knowledge embedding vector to generate a comprehensive evidence vector. The comprehensive evidence vector is input into the evidence network of the evidence evaluation module, and mapped through the output layer and non-negative activation function of the evidence network to output the first evidence strength supporting the mineralization hypothesis and the second evidence strength supporting the non-mineralization hypothesis. Calculate 1 with the sum of the first evidence strength and the second evidence strength to obtain the evidence sum, and divide 1 by the evidence sum to obtain the prediction uncertainty value; Calculate the sum of 1 and the first evidence strength to obtain the numerator evidence value. Calculate the sum of the first evidence strength, the second evidence strength and the preset values ​​corresponding to the number of categories of the classification problem to obtain the denominator evidence value. Divide the numerator evidence value by the denominator evidence value to obtain the mineralization probability value.

4. The method according to claim 3, characterized in that, The step of using the evidence questioning module, guided by the knowledge embedding vector, to perform correlation questioning on the data evidence set to obtain the inferred evidence vector specifically includes: The knowledge embedding vector is input into the first linear transformation layer of the evidence questioning module for projection to generate a query vector; Each sub-feature vector in the data evidence set is input into the second linear transformation layer and the third linear transformation layer of the evidence challenge module for projection, thereby generating a key vector and a value vector. Calculate the inner product of the query vector and each of the key vectors, and normalize it using a normalized exponential function to obtain the corresponding attention weights; The attention weights are used to perform a weighted summation of all the value vectors to generate the inference-based evidence vector.

5. The method according to claim 1, characterized in that, The determination of the preset probability threshold and preset uncertainty threshold based on the geological state of the target area specifically includes: The ore-controlling element characteristics of each spatial unit within the target area are obtained. These ore-controlling element characteristics include fault structural density, known mineral occurrence distribution density, and stratigraphic mineralization favorability. The geological mineralization favorability index of each spatial unit is calculated by weighted summation based on the density of the fracture structure line, the density of the known mineral occurrence distribution, and the mineralization favorability of the stratigraphy. The preset probability threshold corresponding to each spatial unit is calculated based on the geological mineralization favorability index. The preset uncertainty threshold corresponding to each spatial unit is calculated based on the geological mineralization favorability index.

6. A mineralization prediction system based on multimodal geological data and knowledge graphs, characterized in that, The system includes an acquisition unit, an extraction and fusion unit, a construction unit, and a prediction unit. The acquisition unit acquires multimodal geological data, geological exploration reports, and mineral deposit survey information for the target area; The extraction and fusion unit performs multi-scale feature extraction on the multimodal geological data to obtain multiple feature vectors. Each feature vector corresponds to a type of scale feature in the multi-scale features. Specifically, the multi-scale feature extraction of the multimodal geological data to obtain multiple feature vectors includes: acquiring remote sensing image data, geophysical and geochemical exploration data, and geological vector map data from the multimodal geological data; performing multi-layer feature extraction on the remote sensing image data to obtain small-scale, mesoscale, and large-scale remote sensing feature vectors, respectively. The small-scale remote sensing feature vector reflects the input image at a specific volume... The local texture intensity under the nucleus response, the mesoscale remote sensing feature vector represents medium-scale mineralization control elements including rock mass boundary morphology, fault zone strike extension, and the intersection relationships of multiple structures, while the large-scale remote sensing feature vector represents factors that macroscopically control mineralization, including tectonic location, regional magmatic belt distribution, and dominant tectonic systems. Wavelet transform is used to perform multi-band signal decomposition on the geophysical and geochemical data to obtain small-scale, mesoscale, and large-scale geophysical and geochemical feature vectors, respectively. The small-scale geophysical and geochemical feature vector reflects locally enriched mineralization centers or anomalies. The mesoscale geophysical and geochemical feature vectors characterize the migration paths and alteration zoning characteristics of the ore-forming hydrothermal system, while the large-scale geophysical and geochemical feature vectors reflect the overall characteristics of the regional geophysical field in which the spatial unit is located. Multi-level semantic parsing is performed on the geological vector map data based on geological hierarchies to obtain small-scale, mesoscale, and large-scale geological feature vectors, respectively. The small-scale geological feature vectors characterize the direct geological environment of the spatial unit, the mesoscale geological feature vectors characterize the position of the spatial unit within the regional stratigraphic framework and tectonic system, and the large-scale geological feature vectors characterize the regional tectonics of the spatial unit. Background and large-scale ore-controlling environment; the large-scale remote sensing feature vector, the large-scale geophysical and geochemical feature vector, and the large-scale geological feature vector are concatenated to obtain a large-scale evidence vector; the mesoscale remote sensing feature vector, the mesoscale geophysical and geochemical feature vector, and the mesoscale geological feature vector are concatenated to obtain a mesoscale evidence vector; the small-scale remote sensing feature vector, the small-scale geophysical and geochemical feature vector, and the small-scale geological feature vector are concatenated to obtain a small-scale evidence vector; the small-scale evidence vector, the mesoscale evidence vector, and the large-scale evidence vector are output as the feature vector;Based on a causal hierarchy, multiple feature vectors are processed and fused to obtain a preliminary fused vector. The causal hierarchy represents the relationship from low-level geological phenomena to high-level metallogenic models. Specifically, this process includes: calculating the similarity between the large-scale evidence vector and various background models in a pre-defined metallogenic background model library; selecting the target background model corresponding to the maximum similarity result from multiple similarity results; and using the confidence vector corresponding to the target background model as the regional background diagnostic vector. The pre-defined metallogenic background model library is a knowledge base constructed by summarizing historical metallogenic data and geological literature, storing standard feature templates for various typical regional metallogenic models. Furthermore, matrix multiplication is performed based on a pre-defined geological inference matrix and the regional background diagnostic vector to generate a mesoscale attention weight template. The pre-defined geological inference matrix is ​​an n×m dimensional matrix, where n represents the n rows of the pre-defined geological inference matrix corresponding to n mesoscale features, and m represents... The m columns of the preset geological inference matrix correspond to m regional metallogenic models, where the mesoscale features represent the feature values ​​in the mesoscale evidence vector. The mesoscale evidence vector is multiplied element-wise with the mesoscale attention weight template to obtain the ore-controlling structure judgment vector. A small-scale weight template is obtained by matrix multiplication based on the preset local indicator matrix and the ore-controlling structure judgment vector. The preset local indicator matrix is ​​a p×q dimensional matrix, a knowledge matrix constructed based on geological prospecting experience. Here, p represents the p rows of the preset local indicator matrix corresponding to p small-scale prospecting marker features, and q represents the q columns of the preset local indicator matrix corresponding to q mesoscale ore-controlling structure types. p represents the feature dimension of the small-scale evidence vector, and q represents the number of types of mesoscale ore-controlling structures, determined by cluster analysis of the mesoscale evidence vector. The small-scale evidence vector is multiplied element-wise with the small-scale weight template to obtain the preliminary fusion vector. The construction unit constructs a knowledge graph related to the target area based on the geological exploration report and the mineral deposit survey information, and encodes the knowledge graph using a graph neural network to generate knowledge embedding vectors; The prediction unit inputs the knowledge embedding vector and the preliminary fusion vector into the knowledge reasoning model for processing to obtain the mineralization probability value and prediction uncertainty value of each spatial unit within the target area; based on the geological state of the target area, it determines a preset probability threshold and a preset uncertainty threshold, and compares the mineralization probability value of each spatial unit with the preset probability threshold; when the mineralization probability value of the target spatial unit is greater than or equal to the preset probability threshold, and the prediction uncertainty value of the target spatial unit is less than or equal to the preset uncertainty threshold, it determines that the target spatial unit within the target area is delineated as a mineral target area, where the target spatial unit is any one of the spatial units.

7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-5.

Images