Intelligent prediction method and system for tungsten ore metallogenic target area

Abstract

The invention discloses an intelligent prediction method and system for a tungsten ore metallogenic target area, and relates to the technical field of metallogenic target area prediction. An intelligent prediction system for a tungsten ore metallogenic target area comprises a data mapping module, a metallogenic characteristic module, a metallogenic map module, a prediction feedback module, an ore deposit prediction module and a target area output module. According to the method, a prediction-discovery feedback loop is designed under the support of a mineralization mechanism knowledge graph, and a model blind area and a knowledge conflict area are identified by utilizing hierarchical prediction of first-round multi-scale interaction and an ore deposit scale prediction result of cross-scale collaborative optimization; and mining geologic feature space association rules by adopting a constraint-based causal relationship discovery algorithm in a blind area of the model, and updating map inference rules after confirmation by geological experts, so that dynamic evolution of the mineralization mechanism knowledge map is realized, and the predictive ability for new-type ore deposits and atypical mineralization is improved.

Description

Technical Field

[0001] This invention relates to the field of mineralization target area prediction technology, and in particular to an intelligent prediction method and system for tungsten mineralization target areas. Background Technology

[0002] Tungsten resources are an important strategic mineral resource. The tungsten mineralization process is comprehensively controlled by multiple factors, including regional tectonic setting, intrusive evolution of rock masses, fault tectonic activity, and geochemical fields. With the advancement of geological exploration, multi-source prospecting data, including geological data, fault tectonic data, and geochemical data, have been gradually accumulated in the target study area. Researchers typically use geographic information systems, statistical analysis, or machine learning methods to integrate this multi-source prospecting information to delineate tungsten mineralization target areas. Existing research commonly employs techniques such as prospecting prediction based on the weighted evidence method, favorable area delineation based on the superposition of mineralization elements, and qualitative evaluation based on empirical prospecting indicators. Some studies have also attempted to introduce new methods such as knowledge graphs and deep learning to express mineralization mechanisms and improve prediction accuracy.

[0003] Current tungsten mineralization target area prediction technologies still suffer from several shortcomings. On the one hand, the treatment of tectonic ore control largely remains at the level of static geometric indicators such as fracture density and fracture intersections, lacking simulations of ore-forming fluid transport based on paleofractal networks during the mineralization period. This makes it difficult to physically characterize the control effect of fracture networks on the transport and accumulation of ore-forming fluids. On the other hand, knowledge related to mineralization mechanisms exists primarily in the form of empirical rules and textual descriptions. Even when knowledge graphs are introduced, they are only used as static prior libraries, lacking a closed-loop interaction mechanism with prediction results and failing to self-update under the drive of new data and mineralization discoveries. Furthermore, multi-scale mineral exploration prediction typically employs a simple top-down cascade model, lacking bidirectional information constraints and collaborative optimization between regional, orefield, and deposit scales, leading to inconsistencies in prediction results across different scales. Simultaneously, there is a lack of unified quantitative evaluation of the uncertainty and data quality of prediction results, and the output target areas lack clear reliability indicators, hindering subsequent exploration deployment. Summary of the Invention

[0004] This invention proposes an intelligent prediction method that, under the constraints of a knowledge graph of mineralization mechanisms, organically integrates simulation of ore-forming fluid migration in paleofault networks during the mineralization period, hierarchical prediction with multi-scale interaction, prediction-discovery feedback loops, and cross-scale collaborative optimization. It also quantifies the uncertainty of prediction results and data quality, enabling the method to reflect the genetic mechanisms of tectonic control of ore, lithological assemblage, and geochemical anomalies. By driving the update of the knowledge graph of mineralization mechanisms through prediction results, the dynamic evolution of mineralization models is realized, and finally, candidate mineralization target areas with credibility indicators and target area scores are output.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent prediction of tungsten ore-forming target areas includes: Acquire multi-source mineral exploration data for the target study area, including geological data, fault structure data, and geochemical data; establish multi-scale spatial units; map multi-source mineral exploration data to each spatial unit; form multi-scale characteristic data; and calculate data quality indicators. Based on fracture tectonic data and geological data, a paleofractal network of mineralization period was constructed to simulate the migration of ore-forming fluids, and the fluid convergence intensity and mineralization channel accessibility of each spatial unit were obtained as structural ore-controlling characteristics. A knowledge graph of mineralization mechanism is constructed. Based on structural mineralization control characteristics, geological data and geochemical data, each spatial unit is mapped into a spatial unit subgraph. Pre-set graph reasoning rules are run to generate an initial mechanism conformity vector. With the support of the metallogenic mechanism knowledge graph, a prediction-discovery feedback loop is executed to update the metallogenic mechanism knowledge graph and the mechanism conformity vector; By utilizing the knowledge graph of mineralization mechanism and the mechanism conformity vector, a hierarchical prediction with multi-scale interaction is performed to obtain the ore deposit scale prediction results; cross-scale collaborative optimization is then performed on the ore deposit scale prediction results to update the ore deposit scale prediction results. Uncertainty indices are calculated for the ore deposit scale prediction results. These indices are then fused with data quality indices and mechanism conformity vectors to generate credibility indices. Target area scores are determined based on the ore deposit scale prediction results, mechanism conformity vectors, and uncertainty indices. Connected regions whose scores meet preset conditions are extracted as candidate ore-forming target areas for output.

[0006] As a preferred embodiment of the present invention, the mineralization period paleofault network includes: a set of nodes consisting of fault intersections, fault endpoints, and fault turning points, and a set of edges consisting of fault segments; each fault segment is associated with and stored geometric attributes, structural attributes, ore-conducting attribute parameters, and ore-blocking attribute parameters; the geometric attributes include strike, dip, and length; the structural attributes include fault level and lithological contact relationship, determined by fault interpretation and fault period division of fault structural data; the ore-conducting attribute parameters and ore-blocking attribute parameters are assigned by lithological and stratigraphic information in geological data.

[0007] As a preferred technical solution of the present invention, the ore-forming fluid migration simulation includes: discretizing the ancient fault network during the ore-forming period into migration paths composed of fault segments, taking known mineral occurrences and geochemical anomaly nodes as injection locations of virtual mineralization sources, and, considering the ore-conducting attribute parameters and ore-resisting attribute parameters of each fault segment as well as preset attenuation parameters, gradually propagating and accumulating the ore-forming fluid flux along the migration path. The fluid convergence intensity is determined by statistically analyzing the cumulative value of ore-forming fluid flux within spatial units at various scales that intersect with the migration path, and the accessibility of mineralization channels is determined by the comprehensive resistance and path connectivity along the fault segments between the virtual mineralization source and each spatial unit.

[0008] As a preferred technical solution of the present invention, the mineralization mechanism knowledge graph includes: a set of nodes for representing geological entities and a set of edges for representing mineralization relationships; the set of nodes includes rock mass nodes, stratigraphic nodes, structural nodes, alteration zone nodes, and geochemical anomaly nodes; the set of edges includes ore-controlling relationship edges, ore-supplying relationship edges, sealing relationship edges, and enrichment relationship edges; the nodes and edges are associated with and stored with spatial location attributes, scale attributes, and mapping relationship attributes with spatial units of each scale; based on structural ore-controlling characteristics, geological data, and geochemical data, the attribute information corresponding to fault structures, lithology, stratigraphy, alteration zones, and geochemical anomalies within each spatial unit of each scale is associated with the corresponding types of nodes and mineralization relationship edges in the set of nodes and the set of edges according to a preset mapping rule, forming a spatial unit subgraph corresponding to each spatial unit of each scale.

[0009] As a preferred technical solution of the present invention, the preset graph reasoning rules include: performing subgraph matching reasoning and path reasoning on each spatial unit subgraph based on the mineralization mechanism knowledge graph. The graph reasoning rules include rules for determining whether the tectonic control relationship is satisfied, rules for determining whether the lithological combination is conducive to mineralization, and rules for determining the synergistic characteristics of multi-element geochemical anomalies. Based on the degree to which each spatial unit subgraph satisfies the graph reasoning rules, the tectonic conformity component, the lithological conformity component, and the geochemical conformity component are calculated respectively. The initial mechanism conformity vector is composed of the tectonic conformity component, the lithological conformity component, and the geochemical conformity component.

[0010] As a preferred embodiment of the present invention, the execution prediction-discovery feedback loop includes: utilizing the initial mechanism conformity vector to perform the first round of multi-scale interactive hierarchical prediction and cross-scale collaborative optimization to obtain the first round of deposit-scale prediction results; comparing and analyzing the spatial distribution of the first round of deposit-scale prediction results with known mineral occurrences, geochemical anomalies, and structural and lithological features characterized by geological data to identify model blind spots and knowledge conflict areas; and, for model blind spots, using a constraint-based causal relationship discovery algorithm based on multi-source prospecting data to mine spatial associations of geological features not encoded by the current metallogenic mechanism knowledge graph. The rules generate candidate new metallogenic models, whose geological features include structural and lithological features characterized by geological data, as well as their spatial combination with geochemical anomalies. For knowledge conflict areas, the confidence of the map reasoning rules related to the area is evaluated and adjusted. The candidate new metallogenic models are submitted to geological experts in the form of rules for confirmation or correction. After confirmation or correction by geological experts, the verified candidate new metallogenic models are added to or replaced with new map reasoning rules, the metallogenic mechanism knowledge map is updated, and an updated mechanism conformity vector is generated based on the updated metallogenic mechanism knowledge map.

[0011] As a preferred embodiment of the present invention, the hierarchical prediction of multi-scale interaction includes: dividing spatial units into regional scale spatial units, ore field scale spatial units, and ore deposit scale spatial units according to scale type; inputting the multi-scale feature data corresponding to each scale spatial unit into the corresponding regional scale prediction sub-model, ore field scale prediction sub-model, and ore deposit scale prediction sub-model; using the regional scale prediction sub-model to predict the regional scale spatial units to obtain regional scale prediction results; generating regional scale guidance information based on the regional scale prediction results and the multi-scale feature data of the ore field scale spatial units; inputting the regional scale guidance information and the mechanism conformity vector together into the ore field scale prediction sub-model to obtain the ore field scale prediction results; generating ore field scale guidance information based on the ore field scale prediction results and the multi-scale feature data of the ore deposit scale spatial units; inputting the ore field scale guidance information, the mechanism conformity vector, and the structural ore-controlling features together into the ore deposit scale prediction sub-model; and outputting the ore deposit scale prediction results from the ore deposit scale prediction sub-model.

[0012] As a preferred embodiment of the present invention, the cross-scale collaborative optimization includes: constructing a cross-scale collaborative optimization model based on the regional-scale prediction results and the ore-field-scale prediction results obtained from hierarchical predictions performed with multi-scale interaction, and using the ore-deposit-scale prediction results as optimization variables; in the cross-scale collaborative optimization model, introducing a global constraint term to constrain the consistency between the ore-deposit-scale prediction results and the regional-scale prediction results, introducing a local constraint term to constrain the consistency between the ore-deposit-scale prediction results and the ore-field-scale prediction results and the mechanism conformity vector, and combining the deviations between each constraint term and the original ore-deposit-scale prediction results to form an objective function; and correcting the ore-deposit-scale prediction results under multi-scale constraints by iteratively solving the objective function, and obtaining the updated ore-deposit-scale prediction results when the preset convergence conditions are met.

[0013] As a preferred technical solution of the present invention, the generation of the credibility index includes: performing multiple inferences on the updated deposit scale prediction results under the condition of introducing random inactivation or parameter perturbation to obtain multiple deposit scale prediction results, and determining an uncertainty index based on the degree of dispersion among the multiple deposit scale prediction results; normalizing the uncertainty index, data quality index, and mechanism conformity vector respectively, and weighting and combining the normalized uncertainty index, data quality index, and mechanism conformity vector according to a preset fusion function to obtain the credibility index.

[0014] A smart prediction system for tungsten ore-forming target areas includes: Data mapping module: acquires multi-source mineral exploration data of the target study area, establishes multi-scale spatial units, maps multi-source mineral exploration data to each spatial unit, forms multi-scale feature data, and calculates data quality indicators; Mineralization Feature Module: Based on the fracture structure data and geological data, the ancient fracture network of the mineralization period is constructed to simulate the migration of mineralizing fluids, and the fluid convergence intensity and mineralization channel accessibility of each spatial unit are obtained as structural mineralization control features; Metallogenic Map Module: Constructs a knowledge graph of metallogenic mechanisms, maps each spatial unit into a spatial unit subgraph based on structural ore-controlling features, geological data, and geochemical data, and runs preset map reasoning rules to generate an initial mechanism conformity vector; Prediction Feedback Module: With the support of the metallogenic mechanism knowledge graph, it executes a prediction-discovery feedback loop to update the metallogenic mechanism knowledge graph and the mechanism conformity vector; Ore deposit prediction module: Utilizing the knowledge graph of mineralization mechanism and the mechanism conformity vector, it performs multi-scale interactive hierarchical prediction to obtain ore deposit scale prediction results; and performs cross-scale collaborative optimization on the ore deposit scale prediction results to obtain cross-scale collaboratively optimized ore deposit scale prediction results. Target area output module: Calculates uncertainty index for ore deposit scale prediction results, integrates uncertainty index with data quality index and mechanism conformity vector to generate credibility index; determines target area score value based on ore deposit scale prediction results, mechanism conformity vector and uncertainty index, and extracts connected regions that meet preset conditions as candidate areas for mineralization target area for output.

[0015] The present invention has the following advantages: This invention constructs an ancient fault network for the mineralization period based on fault structure data and geological data, and performs ore-forming fluid migration simulation on this network to obtain the fluid convergence intensity and mineralization channel accessibility of each spatial unit. These are used as structural ore-controlling features, realizing the transformation from static geometric fault indicators to ore-forming process constraints. This gives the role of fault-controlled ore in prediction a clear physical orientation and stronger genetic explanation.

[0016] This invention constructs a knowledge graph of mineralization mechanisms, consisting of rock mass nodes, stratigraphic nodes, tectonic nodes, alteration zone nodes, geochemical anomaly nodes, as well as edges representing ore-controlling relationships, ore-supplying relationships, sealing relationships, and enrichment relationships. It then runs pre-defined graph reasoning rules on the spatial unit subgraphs to generate a mechanism conformity vector composed of tectonic conformity components, lithological conformity components, and geochemical conformity components. This achieves a structured and quantitative expression of traditional qualitative mineralization models, improving the precision and computability of mechanism constraints in tungsten mineralization target area prediction.

[0017] This invention designs a prediction-discovery feedback loop supported by a knowledge graph of mineralization mechanisms. It utilizes the results of hierarchical prediction through multi-scale interaction in the first round and deposit-scale prediction through cross-scale collaborative optimization to identify blind spots and knowledge conflict areas in the model. Within the blind spots, a constraint-based causal relationship discovery algorithm is used to mine spatial association rules of geological features. After confirmation by geological experts, the graph reasoning rules are updated, realizing the dynamic evolution of the knowledge graph of mineralization mechanisms. In long-term application, it has the ability to discover new mineralization models and correct old rules, thereby improving the prediction ability for new types of deposits and atypical mineralization.

[0018] This invention constructs regional-scale, ore-field-scale, and ore-deposit-scale prediction sub-models on regional-scale spatial units, ore-field-scale spatial units, and ore-deposit-scale spatial units, respectively. It introduces regional-scale and ore-field-scale guiding information into the prediction process and combines mechanism conformity vectors and structural ore-controlling characteristics to form a multi-scale interactive hierarchical prediction. This realizes the information linkage between regional background mineralization favorability, ore-field-scale favorable areas, and ore-deposit-scale local enrichment areas, improving the consistency of prediction results across different spatial scales and the rationality of the overall mineral exploration layout.

[0019] This invention constructs a cross-scale collaborative optimization model based on hierarchical prediction with multi-scale interaction. It introduces a global constraint term to constrain the consistency between the ore deposit-scale prediction results and the regional-scale prediction results, as well as a local constraint term to constrain the consistency between the ore deposit-scale prediction results and the ore field-scale prediction results and the mechanism conformity vector. The ore deposit-scale prediction results are iteratively corrected, achieving the technical effect of keeping the ore deposit-scale prediction results in harmony with the large-scale metallogenic laws and mechanism constraints without sacrificing the sensitivity to local anomalies. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort. Figure 1 This is a schematic diagram of the structure of an intelligent prediction system for tungsten ore formation target areas used in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] Example 1: A smart prediction method for tungsten ore-forming target areas, comprising the following steps: Step S1: Obtain multi-source mineral exploration data for the target study area, including geological data, fault structure data, and geochemical data; establish multi-scale spatial units; map the multi-source mineral exploration data to each spatial unit; form multi-scale feature data; and calculate data quality indicators. In this embodiment, the multi-source mineral exploration data refers to a collection of various spatialized geoscientific information formed within the target study area, focusing on mineralization and mineral exploration activities. Multi-source mineral exploration data mainly includes the following three categories: Geological data includes regional geological mapping results and basic data generated from mineral exploration work. Sources of geological data include regional geological maps, mining area geological maps, borehole logging data, tunnel logging data, and field geological route survey records generated from geological survey projects. Geological data is stored in planar units, and each geological data (planar unit) is associated with at least the following attribute fields: lithology type field, used to indicate the rock type of the unit, such as granite, diorite, gneiss, slate, etc., which can be represented by rock codes; stratigraphic code field, used to indicate the stratigraphic layer or geological age to which the unit belongs, such as Sinian, Paleozoic, Mesozoic, etc.; rock mass type and genesis field, used to distinguish different types of intrusive rock masses, such as Type I granite, Type II granite, and their typical relationship with mineralization; contact zone marker field, used to mark whether the unit is located inside an intrusive rock mass, in a contact zone area, or in a surrounding rock area; alteration type and alteration intensity field, used to record alteration types such as sericitization, chloritization, silicification, etc., and their weak, medium, and strong grades. Geological data is uniformly converted to a plane coordinate system consistent with the target study area during import, and attribute fields are standardized and coded.

[0023] Fault structure data includes regional tectonic interpretation results and detailed tectonic survey results for mining areas. Sources include remote sensing image interpretation results, digital elevation model (DEM) analysis results, fault lines from geological mapping reports, and measured profile data. Fault structure data is stored in linear vector form. Each fault line or segment is associated with the following attribute fields: fault strike, dip, and dip angle fields, used to characterize the fault's geometric orientation in space; fault length field, used to characterize the fault's extension scale in the plane; fault level field, used to distinguish regional primary faults, secondary faults, and local fracture zones; fault nature field, used to label faults as extensional, compressive, or strike-slip; fault-lithological contact relationship field, used to record whether the fault cuts through or runs parallel to a specific lithological contact zone; and fault period field, used to indicate the main tectonic period of fault formation or activity, facilitating subsequent comparison and screening with mineralization periods.

[0024] Geochemical data includes regional geochemical measurements, rock geochemical analyses during mining area mapping, and soil geochemical measurements from the detailed exploration phase. Geochemical data is stored as point records, with each sampling point containing the following fields: sampling point coordinates (indicating its location within the study area); sampling medium (identifying sample type, such as rock, soil, or sediment); elemental content (recording the content of elements relevant to tungsten mineralization, such as W, Sn, Mo, Bi, As, Cu, Pb, and Zn); detection limits and analytical methods (describing the technical conditions of the elemental analysis); and anomaly markers (marking single-element and multi-element anomalies based on regional background values ​​and anomaly thresholds, for example, defining anomaly ranges by adding a certain number of standard deviations to the mean). Geochemical data undergoes coordinate transformation and anomaly detection processing during import.

[0025] After acquiring and preprocessing multi-source mineral exploration data, the target study area is divided into multi-scale spatial units. A multi-scale spatial unit refers to a nested spatial partitioning system within the same study area, constructed using different grid scales or zonal scales based on the requirements of mineral exploration prediction accuracy and data resolution. For example, a larger grid size is used for coarse partitioning at the regional scale, while a smaller grid size or fine partitioning based on known mineral cluster outlines is used at the ore field and deposit scales. Each spatial unit has a unique spatial unit number, spatial extent, and scale identifier, used to represent the geological, structural, and geochemical features summarized within that unit.

[0026] The process of mapping multi-source mineral exploration data to various spatial units includes: spatial overlay analysis of geological data, overlaying each spatial unit with its covered lithological units, stratigraphic units, rock mass units, and alteration zones, statistically analyzing attributes such as the proportion of various lithologies, the distance to the contact zone with important rock masses, and the area ratio of alteration zones, and recording the statistical results as the geological characteristic components of the spatial unit; spatial overlay analysis of fault structure data, associating each spatial unit with the fault lines and fault intersections passing through the unit, statistically analyzing structural features such as fault length density, number of fault intersections, and minimum distance to major faults, and using these as the structural characteristic components of the spatial unit; and spatial overlay analysis of geochemical data, matching each spatial unit with the geochemical sampling points within its range, statistically analyzing the average content, maximum content, number of anomalies, and multi-element anomaly superposition index of key elements related to tungsten mineralization, and using these as the geochemical characteristic components of the spatial unit.

[0027] For each spatial unit, a feature vector containing geological feature components, tectonic feature components, and geochemical feature components is constructed. In the embodiments, this feature vector is collectively referred to as multi-scale feature data, which is used to uniformly characterize the mineralization favorability and geoscientific background information of the spatial unit at different spatial scales.

[0028] Data quality indicators are a set of metrics that quantitatively characterize the completeness and reliability of multi-source mineral exploration data within each spatial unit. For a single spatial unit, data quality indicators should include at least the following: geochemical sampling density, representing the ratio of the number of geochemical sampling points to the area of ​​the spatial unit, reflecting the degree of geochemical information support for that unit; data missing rate, representing the proportion of missing or null values ​​in the geological, fault structure, and geochemical attribute fields of the spatial unit, reflecting the completeness of the data in that unit; data source credibility level, reflecting whether the data in the spatial unit mainly comes from measured data or interpolated inference data, for example, assigning a higher credibility level to measured data and a relatively lower credibility level to data obtained through interpolation or model inference; and multi-source consistency, assessing the degree of consistency between geological interpretation results and geochemical anomaly distribution within the spatial unit, for example, assigning higher consistency scores to areas where favorable lithology, structural locations, and elemental anomalies overlap.

[0029] Step S2: Based on the fracture structure data and geological data, simulate the migration of ore-forming fluids in the ancient fracture network constructed during the metallogenic period to obtain the fluid convergence intensity and mineralization channel accessibility of each spatial unit, which serve as structural ore-controlling characteristics. In this embodiment, the "mineralization period paleofault network" refers to a set of faults that were active during the mineralization period or had a controlling effect on the migration of ore-forming fluids, selected from the fault structure data based on the current fault structure framework and combined with information on geological age and tectonic phase. These faults are then organized into a network structure according to their topological relationships. This network is used to describe the main migration channels of ore-forming fluids in three-dimensional space and their connectivity, and serves as the basis for subsequent quantitative characterization of tectonic control over ore deposits.

[0030] The paleofault network of the metallogenic period includes: a set of nodes consisting of fault intersections, fault endpoints, and fault turning points, and a set of edges consisting of fault segments; each fault segment is associated with and stored geometric attributes, structural attributes, ore-conducting attribute parameters, and ore-blocking attribute parameters; the geometric attributes include strike, dip, and length; the structural attributes include fault level and lithological contact relationship, determined by fault interpretation and fault period division of the fault structural data; the ore-conducting attribute parameters and ore-blocking attribute parameters are assigned by lithological and stratigraphic information in the geological data.

[0031] Specifically, constructing the paleofault network during the mineralization period includes the following processing steps: Based on the fault structure data formed in step S1, the main fault lines are subjected to unified geometric normalization, dividing long faults into several fault segments with clear start and end points. Each fault segment has a unique line segment geometry on the plane. Based on this, by analyzing the intersection relationships and endpoint positions between fault lines, fault intersections, fault endpoints, and fault turning points are extracted and used as nodes in the network; the fault line segments between any two adjacent nodes are used as edges in the network.

[0032] By combining the comparative results of geological data on tectonic periods, intrusion ages of rock masses, and mineralization ages, the fault structure data is divided into fault periods. Fault segments that overlap with the mineralization period and have evidence of activity during that period are selected into the paleofault network of the mineralization period. Fault segments that are unrelated to the mineralization period or have been dulled in the early stages are removed or given lower weights, thus obtaining a tectonic framework that is more consistent with the mineralization process.

[0033] In terms of attribute assignment, the strike, dip and length in geometric attributes are directly calculated from the fault structure data, the fault level in structural attributes is filled by the grading results in the fault interpretation results, and the contact relationship with lithology is determined by overlay analysis of the spatial relationship between the fault line and the rock mass and stratigraphic boundary in the geological data. For example, it records whether a certain fault segment cuts the granite body, extends along the rock mass boundary or cuts a specific stratum.

[0034] Ore-conducting and ore-resisting properties are used to quantitatively represent the degree to which a fault segment facilitates or hinders the migration of ore-forming fluids. Based on lithological and stratigraphic information from geological data, fault segments that traverse fractures with high openability, brittle surrounding rocks, and a tendency to form tensile fractures are assigned higher ore-conducting properties, while fault segments that traverse ductile surrounding rocks, exhibit argillaceous infill, and possess sealing characteristics are assigned higher ore-resisting properties. For example, a fault segment traversing the contact zone between granite porphyry and the surrounding rock, accompanied by shear alteration, can be considered highly ore-conducting, while a fault segment traversing thick mudstone layers and showing argillaceous infill can be considered highly ore-resisting. Through this method, a comprehensive description of geometric, structural, and ore-conducting / ore-resisting properties is formed for each fault segment.

[0035] The simulation of ore-forming fluid migration includes: discretizing the paleofault network during the mineralization period into migration paths composed of fault segments; using known mineral occurrences and geochemical anomaly nodes as injection locations of virtual mineralization sources; and, considering the ore-conducting and ore-resisting properties of each fault segment as well as preset attenuation parameters, gradually propagating and accumulating the flux of ore-forming fluid along the migration path. The fluid convergence intensity is determined by statistically analyzing the cumulative value of ore-forming fluid flux within spatial units at various scales that intersect with the migration path, and the accessibility of mineralization channels is determined by the comprehensive resistance and path connectivity along the fault segments between the virtual mineralization source and each spatial unit.

[0036] In this embodiment, the simulation of ore-forming fluid migration refers to simulating the migration and diffusion of ore-forming fluids along fracture channels on a paleofault network during the mineralization period, starting from known mineral occurrences or obvious geochemical anomaly nodes. This is used to obtain the strength of fluid transport effects on spatial units at various scales during the mineralization process.

[0037] "Virtual mineralization source" refers to selecting a point in space that corresponds to a known mineralization location or a strong geochemical anomaly location, and treating it as the injection source of ore-forming fluid. A certain initial fluid flux is allocated to each virtual mineralization source to start the migration simulation process.

[0038] In the discretization process, each fault segment in the paleofault network during the mineralization period is considered as a transport path unit through which ore-forming fluids can pass, and the nodes at both ends of the fault segment are considered as distribution and convergence nodes of the fluid in the network. For each virtual mineralization source, a connection is established with the nearest fault segment or node based on its location, and the initial ore-forming fluid flux is injected into the corresponding fault segment or node. Subsequently, considering the ore-conducting and ore-resisting properties of each fault segment, multi-step propagation is carried out along the transport path according to a preset propagation rule. The propagation rule includes: in each step, the fluid flux at the node is weighted according to the ore-conducting properties of adjacent fault segments, and the fluid flux is attenuated according to the ore-resisting and attenuation properties of each fault segment to reflect the degree of favorability of different fault segments for fluid transport and the energy loss.

[0039] As the propagation process proceeds, the ore-forming fluid flux gradually expands in space along the network, and each fault segment traversed by the ore-forming fluid flux and the spatial unit it passes through will record the corresponding flux contribution.

[0040] To calculate fluid convergence intensity, for each spatial unit, the cumulative flux of ore-forming fluids along all migration paths intersecting that unit is statistically analyzed. This cumulative value is used as the fluid convergence intensity index for that spatial unit. Fluid convergence intensity reflects the degree of convergence of ore-forming fluids at the location of that spatial unit during the mineralization process.

[0041] To calculate the accessibility of mineralization pathways, a comprehensive resistance index and connectivity index are constructed between the virtual mineralization source and spatial units at various scales. This is based on the ore-conducting and ore-resisting properties of fault segments on the paleofault network during the metallogenic period, taking into account the magnitude of resistance along the path, the total path length, and the number of alternative paths. Spatial units with lower comprehensive resistance and multiple independent pathways are assigned higher mineralization pathway accessibility scores, while spatial units that require crossing high-resistivity fault segments or have poor connectivity are assigned lower mineralization pathway accessibility scores.

[0042] Step S3: Construct a knowledge graph of mineralization mechanism. Based on structural mineralization control features, geological data, and geochemical data, map each spatial unit into a spatial unit subgraph. Run the preset graph reasoning rules to generate an initial mechanism conformity vector. The mineralization mechanism knowledge graph includes: a set of nodes representing geological entities and a set of edges representing mineralization relationships; the set of nodes includes rock mass nodes, stratigraphic nodes, tectonic nodes, alteration zone nodes, and geochemical anomaly nodes; the set of edges includes ore-controlling relationship edges, ore-supplying relationship edges, sealing relationship edges, and enrichment relationship edges; the nodes and edges are associated with and stored spatial location attributes, scale attributes, and mapping relationship attributes with spatial units of each scale; based on structural ore-controlling characteristics, geological data, and geochemical data, the attribute information corresponding to fault structures, lithology, stratigraphy, alteration zones, and geochemical anomalies within each spatial unit of each scale is associated with the corresponding types of nodes and mineralization relationship edges in the set of nodes and the set of edges according to preset mapping rules, forming a spatial unit subgraph corresponding to each spatial unit of each scale.

[0043] In this embodiment, the "metallogenic mechanism knowledge graph" refers to a metallogenic mechanism knowledge base organized in the form of a graph theory structure. Its essence is a directed or undirected weighted graph composed of various types of geological entity nodes and metallogenic relationship edges. The metallogenic mechanism knowledge graph is used to uniformly express the spatial combination and causal relationships between rock masses, strata, fault structures, alteration zones, and geochemical anomalies in the metallogenic process, transforming traditional text- and experience-based metallogenic models into a computable and reasonable data structure.

[0044] At the node level, rock mass nodes are used to represent intrusive bodies of different lithologies and genesis, such as granite, granite porphyry, and diorite, and record the geological age, lithological assemblage, geochemical characteristics, and known coupling relationships with tungsten mineralization of the rock mass; stratigraphic nodes are used to represent different stratigraphic units, including stratigraphic codes, ages, lithological assemblages, and whether they are favorable host rocks; tectonic nodes are used to represent fault or fault assemblage units, maintaining a correspondence with the paleofault network of the mineralization period in step S2, and recording the nature, level, and contact relationship of the faults with the lithology; alteration zone nodes are used to represent different types of alteration forms and their intensity levels; geochemical anomaly nodes are used to represent single-element anomalies and multi-element combination anomalies, recording the types of anomalous elements, anomaly intensity, and anomaly superposition.

[0045] All the above nodes are associated with spatial location attributes to indicate the node's location or spatial range within the study area, associated with scale attributes to identify the regional scale, mining field scale, or ore deposit scale to which the node belongs, and associated with mapping relationship attributes with spatial units of each scale to indicate which spatial units the node has spatial overlap or proximity relationship with.

[0046] At the boundary level, ore-controlling relationships are used to indicate the control effect of a certain type of geological entity on the distribution of ore bodies or the migration of ore-forming fluids, such as "fracture-controlled ore" and "rock contact zone-controlled ore." Ore-supplying relationships are used to indicate the supply relationship of materials and heat, such as deep magma chambers supplying ore-forming materials to upper fault zones. Blocking relationships are used to indicate the obstruction effect of low-permeability strata or closed structures on the migration of ore-forming fluids. Enrichment relationships are used to indicate the promoting effect of favorable surrounding rocks or specific pore structures on the enrichment of ore-forming materials. Relationship edges can be accompanied by a confidence weight field to reflect the degree of support for the ore-forming relationship in existing geological studies and measured data, such as assigning a high confidence level to a certain ore-controlling combination based on statistical results of typical deposits.

[0047] Based on structural ore-controlling characteristics, geological data, and geochemical data, the attribute information corresponding to fault structures, lithology, strata, alteration zones, and geochemical anomalies within each spatial unit at each scale is associated with nodes and mineralization relationships of the corresponding types in the node set and edge set according to preset mapping rules, forming a spatial unit subgraph corresponding to each spatial unit at each scale.

[0048] Specifically, for a spatial unit of a certain scale, firstly, based on the data mapping results in step S1 and the structural ore-controlling characteristics in step S2, the lithological units with the highest proportion within the spatial unit, the spatial relationships with the contact zones of important intrusive rock masses, and the types of alteration zones developed in the surrounding rocks are selected from the geological data and mapped as rock mass nodes, stratigraphic nodes, and alteration zone nodes, respectively. Then, based on the fault structure data and the local structure of the paleofault network during the mineralization period, the main fault segments passing through the spatial unit are mapped as one or more structural nodes. Structural-structural ore-controlling edges are established between structural nodes according to their intersection relationships, and ore-controlling relationship edges or sealing relationship edges are established between structural nodes and rock mass nodes and stratigraphic nodes according to their contact relationships. At the same time, based on the types, intensities, and superposition characteristics of anomalous elements within the spatial unit in the geochemical data, geochemical anomalous nodes are constructed and connected to adjacent rock mass nodes, stratigraphic nodes, and alteration zone nodes through enrichment relationship edges.

[0049] The construction of geochemical anomaly nodes refers to creating knowledge graph nodes based on the analysis results of geochemical data within a spatial unit, used to characterize the geochemical anomaly features of that unit. The attributes of these nodes include at least: anomaly element combination encoding, recording the types of elements exhibiting significant anomalies in structured string or array form (e.g., "W-Sn-Mo"), used to characterize the element co-occurrence combination characteristics; anomaly intensity level, classified into "weak, medium, strong" levels based on the contrast value (ratio of element content to background value) or standardized score of each anomaly element within the unit, or represented by continuous numerical values; anomaly superposition index, formed by statistically analyzing the number of overlapping or spatially adjacent single anomalies within the unit, or calculating the spatial coupling strength of multi-element anomalies, forming a comprehensive quantitative index reflecting the complexity of the anomalies; and spatial distribution characteristics, an optional attribute used to describe the distribution pattern of the anomalies within the unit, such as "area-like," "point-cluster-like," or "band-like." After the geochemical anomaly nodes are formed, they are connected by enrichment relationship edges based on their spatial positional relationship and geological genetic association with adjacent rock mass nodes, stratigraphic nodes, and alteration zone nodes. If a geochemical anomaly node is located within the contact zone of a rock mass node, or highly overlaps spatially with a specific stratigraphic node, then an enrichment relationship edge is established between the anomaly node and the rock mass or stratigraphic node. If the distribution of an anomaly node coincides with the range of an alteration zone node, or there is a significant spatial co-occurrence relationship, then an enrichment relationship edge is established between the two.

[0050] Through the above mapping process, each spatial unit can obtain a local subgraph corresponding to its spatial location and data attributes, namely the spatial unit subgraph, which is used to express the mineralization mechanism structure inside the spatial unit and its neighborhood at the map level.

[0051] The preset graph reasoning rules include: performing subgraph matching reasoning and path reasoning on each spatial unit subgraph based on the mineralization mechanism knowledge graph. The graph reasoning rules include rules for determining whether the tectonic control relationship is satisfied, rules for determining whether the lithological combination is conducive to mineralization, and rules for determining the synergistic characteristics of multi-element geochemical anomalies. Based on the degree to which each spatial unit subgraph satisfies the graph reasoning rules, the tectonic conformity component, the lithological conformity component, and the geochemical conformity component are calculated respectively. The initial mechanism conformity vector is composed of the tectonic conformity component, the lithological conformity component, and the geochemical conformity component.

[0052] In this embodiment, the “graph reasoning rule” is a type of pattern matching and path constraint rule abstracted from geological metallogenic mechanisms and statistical characteristics of typical mineral deposits. It is used to identify favorable metallogenic structure combinations on the metallogenic mechanism knowledge graph.

[0053] Rules used to determine whether a structural ore-controlling relationship is satisfied include, for example: When there is a connecting path of “regional-level fault node – secondary fault node – geochemical anomaly node” in the spatial unit sub-map, and the fault structures at both ends of the path correspond to the high ore-conducting fault segments in the paleo-fault network during the metallogenic period, the path is identified as a favorable tectonic ore-controlling path, and a corresponding score is added to the tectonic conformity component. When the spatial unit subgraph shows that there is a contact zone controlling ore-bearing edge between the fault node and the rock mass node, and the location of the contact zone highly overlaps with the distribution of known ore bodies or high anomalies, the ore-bearing combination is regarded as part of a typical tectonic ore-bearing mode.

[0054] Rules used to determine whether a lithological assemblage is favorable for mineralization include, for example: When both "highly differentiated granite body nodes" and "favorable surrounding rock strata nodes" exist simultaneously in the spatial unit sub-graph, and there is a rock body-surrounding rock contact relationship edge between the two, the lithological combination is regarded as a favorable lithological combination; When alteration zone nodes show strong silicification or chloritization alteration and are adjacent to a specific lithological assemblage, an additional positive weight is applied to the lithological consistency component.

[0055] Rules for determining the synergistic characteristics of multi-element geochemical anomalies include, for example: When anomalies of elements such as W, Sn, and Mo appear simultaneously in a geochemical anomaly node, and the anomaly superposition index reaches a preset level, it is considered a multi-element synergistic anomaly. If the multi-element synergistic anomaly node forms a closed structure with the tectonic node and the rock mass node through enrichment relationship edges and ore-controlling relationship edges, then a higher score is given to the geochemical consistency component.

[0056] Subgraph matching reasoning refers to searching for typical structural patterns corresponding to the above rules in the spatial unit subgraph, such as specific connection patterns with three or more nodes, and accumulating or weighting the structural consistency component, lithological consistency component, and geochemical consistency component based on the matching results. Path reasoning refers to traversing the path from deep source nodes to surface mineralization or anomalous nodes in the metallogenic mechanism knowledge graph, checking whether the path meets constraints such as continuous ore-guiding, favorable lithology, and anomalous synergy, in order to evaluate the mechanistic consistency between the metallogenic channel and the metallogenic site.

[0057] Based on the degree to which each spatial unit sub-map satisfies the map inference rules, the structural consistency component, lithological consistency component, and geochemical consistency component are calculated separately. For example, when a spatial unit sub-map satisfies multiple structural ore-controlling rules but only a few lithological favorable rules, the structural consistency component of that unit is high, while the lithological consistency component is relatively low. When the multi-element geochemical anomaly synergistic characteristic rules are satisfied and multiple enrichment relationship edges are supported, the geochemical consistency component of that unit is high. Through appropriate normalization, the above three components are made comparable within a unified numerical range and combined into an initial mechanism consistency vector. This vector is used for fusion and feedback with the data-driven prediction results in subsequent steps, thereby reflecting the a priori constraint effect of the ore-forming mechanism on mineral exploration prediction.

[0058] Step S4: Execute the prediction-discovery feedback loop with the support of the ore-forming mechanism knowledge graph to update the ore-forming mechanism knowledge graph and the mechanism conformity vector; In this embodiment, the "prediction-discovery feedback loop" refers to starting with the metallogenic mechanism knowledge graph and the initial mechanism conformity vector. First, a round of mineral exploration prediction based on multi-scale interaction and cross-scale collaborative optimization is performed. Then, the prediction results are systematically compared with known mineralization evidence and multi-source mineral exploration data to identify areas with significant prediction deviations. These areas are used as observation objects for "knowledge gaps" or "regular conflicts." Causal relationship discovery methods are used to mine new spatial association rules of geological features from the multi-source mineral exploration data. These new rules are then verified by geological experts and fed back and solidified into the metallogenic mechanism knowledge graph, thus forming a closed-loop process of "prediction—deviation identification—new pattern discovery—knowledge update—re-prediction." Through this closed loop, the metallogenic mechanism knowledge graph is no longer a static set of prior knowledge, but a dynamic mechanism model that continuously evolves and improves in actual prediction applications.

[0059] The execution prediction-discovery feedback loop includes: using the initial mechanism conformity vector, performing the first round of multi-scale interactive hierarchical prediction and cross-scale collaborative optimization to obtain the first round of deposit-scale prediction results; comparing and analyzing the spatial distribution of the first round of deposit-scale prediction results with known mineral occurrences, geochemical anomalies, and structural and lithological features characterized by geological data to identify model blind spots and knowledge conflict areas; for model blind spots, using a constraint-based causal relationship discovery algorithm based on multi-source mineral exploration data to mine spatial association rules of geological features not encoded by the current metallogenic mechanism knowledge graph, and generating candidate... The new metallogenic model, whose geological features include structural and lithological features characterized by geological data, as well as spatial combination relationships with geochemical anomalies; for knowledge conflict areas, the confidence of the map reasoning rules related to the area is evaluated and adjusted; the candidate new metallogenic model is submitted to geological experts in the form of rules for confirmation or correction. After confirmation or correction by geological experts, the verified candidate new metallogenic model is supplemented or replaced with new map reasoning rules, the metallogenic mechanism knowledge map is updated, and an updated mechanism conformity vector is generated based on the updated metallogenic mechanism knowledge map.

[0060] In practical implementation, the initial mechanism conformity vector obtained in step S3 is used as a mechanism constraint signal reflecting the degree of agreement between the tectonic, lithological, and geochemical conditions of each spatial unit and typical metallogenic models. This vector, along with the multi-scale feature data from step S1 and the structural ore-controlling features from step S2, is input into the subsequent multi-scale interactive hierarchical prediction process. The specific process of the first round of multi-scale interactive hierarchical prediction and cross-scale collaborative optimization is described in detail in step S5, and in this step, it is only used as a whole functional module. Through the first round of prediction, the first round of ore deposit-scale prediction results covering the entire study area are obtained. These results are typically expressed as the metallogenic favorability score or target area prediction score corresponding to each ore deposit-scale spatial unit.

[0061] The results of the first round of deposit-scale predictions were compared and analyzed with the spatial distribution of known mineral occurrences, geochemical anomalies, and structural and lithological features characterized by geological data. Known mineral occurrence data were derived from existing mineral exploration results, including discovered tungsten deposits, mineralization points, and mineralization clues. Their coordinate locations and corresponding attributes such as ore body size and mineralization type were standardized during the initial data processing. Geochemical anomaly data were derived from regional and mining area geochemical measurements. Following the anomaly determination method in step S1, element concentrations exceeding the background threshold were assigned anomaly markers. Structural and lithological features characterized by geological data included information such as fault density within the aforementioned spatial units, the number of fault intersections, distances to major faults, favorable lithological types, favorable surrounding rock assemblages, and the distribution of alteration zones. This information collectively constitutes the "real-world evidence" basis for assessing the rationality of the prediction results.

[0062] By spatially overlaying the first-round deposit-scale prediction results with the aforementioned real-world evidence, and comparing the consistency between the prediction scores of each spatial unit and the mineralization evidence, model blind spots and knowledge conflict areas are identified.

[0063] The model blind zone refers to areas where mineralization or high anomalies actually exist but the prediction results score is low. That is, in these spatial units, known mineral occurrences or strong geochemical anomalies indicate the existence of significant mineralization activities, but the first round of prediction results did not give corresponding high scores, reflecting that the current metallogenic mechanism knowledge map or prediction model does not fully understand or express this type of metallogenic model.

[0064] The knowledge conflict zone refers to the area where the prediction results are highly rated but lack existing mineralization indicators. In other words, in these spatial units, the first round of prediction results show a high mineralization favorability, but there is a lack of traditional mineral exploration indicators such as known mineral occurrences, obvious geochemical anomalies, or typical favorable structural and lithological combinations. This suggests that some reasoning rules in the current metallogenic mechanism knowledge graph have been given too high a weight or have failed under specific geological backgrounds.

[0065] To address model blind spots, this embodiment utilizes a constraint-based causal relationship discovery algorithm based on multi-source mineral exploration data to uncover spatial association rules of geological features not encoded in the current metallogenic mechanism knowledge graph. The constraint-based causal relationship discovery algorithm is a type of structure learning method centered on conditional independence testing. It gradually constructs directed or undirected dependency structures between variables by jointly analyzing multiple variables in multi-source mineral exploration data. In this embodiment, the variables involved in causal relationship discovery include: lithological type of spatial units, favorable surrounding rock indicators, fault density, fault-rock contact type, fluid convergence intensity, mineralization channel accessibility, elemental content, and anomaly superposition index, etc.

[0066] Within the model blind zone, spatial units with significantly low predicted scores but possessing mineral occurrences or strong anomalies are selected, and local data subsets are constructed using these units as the sample set. Within this subset, by progressively examining the conditional independence relationships of different variables on "existence of mineralization" or "occurrence of anomalies," feature combinations that remain significantly correlated with mineralization results while controlling for other major variables are identified, and then abstracted into spatial association rules for geological features. For example, in a certain model blind zone, causal relationship findings indicate that spatial units "located at the edge of a specific type of granite body, close to late-stage extensional faults, and accompanied by moderate to strong silicification alteration" are significantly positively correlated with the occurrence of tungsten mineralization. Since this combination is not explicitly expressed as a rule in the existing mineralization mechanism knowledge map, it can be considered as one of the core contents of a candidate new mineralization model.

[0067] In this embodiment, the geological features include structural and lithological features characterized by geological data, as well as the spatial combination relationship with geochemical anomalies. That is, the synergistic combination pattern formed in space by simultaneously considering factors such as fault geometry, tectonic phases, favorable lithology or surrounding rocks, alteration features, and superposition of multi-element anomalies.

[0068] For knowledge conflict areas, this embodiment evaluates and adjusts the confidence of the graph reasoning rules related to these areas. Specifically, it checks the feature combinations of spatial units in the knowledge conflict area against the rules driving high mechanism conformity scores in the mineralization mechanism knowledge graph one by one, and counts the support frequency and counterexample frequency of these rules in mineralized and non-mineralized areas. Rules with insufficient supporting evidence and many counterexamples have their confidence weight reduced, and are marked as rules to be corrected if necessary.

[0069] Candidate new metallogenic models are submitted to geological experts in the form of rules for confirmation or correction. After confirmation or correction by geological experts, the verified candidate new metallogenic models are supplemented or replaced with new map reasoning rules, the metallogenic mechanism knowledge map is updated, and an updated mechanism conformity vector is generated based on the updated metallogenic mechanism knowledge map.

[0070] When candidate new mineralization models are expressed in the form of rules, they typically adopt a structure that states "the favorable conditions for mineralization increase when certain geological characteristics are met simultaneously." For example, "When a spatial unit is located within a certain distance of the contact zone outside a highly differentiated granite body, intersects with a late-stage extensional fault, and is accompanied by moderate to strong silicification alteration and superimposed W-Sn multi-element anomalies, the structural and geochemical consistency components of the unit increase." When reviewing these rules, geological experts will consider the regional geological background, existing mineralization models, and drilling verification results to judge the geological rationality and genetic interpretation of the rules. They will modify inaccurate combinations of conditions or eliminate rules lacking genetic support.

[0071] Candidate new metallogenic models, confirmed by experts, are formally incorporated into the metallogenic mechanism knowledge graph, serving as new or revised graph inference rules in subsequent inference processes. After the metallogenic mechanism knowledge graph is updated, the graph inference rules are re-run for each spatial unit subgraph according to the mechanism conformity calculation process in step S3, yielding updated tectonic conformity components, lithological conformity components, and geochemical conformity components, which are then combined to form an updated mechanism conformity vector. This updated mechanism conformity vector will replace the initial mechanism conformity vector in subsequent steps, participating in new multi-scale interactive predictions and cross-scale collaborative optimization. This allows the prediction results to gradually reflect the newly introduced metallogenic mechanism models, thereby achieving a closed-loop iteration and mutual reinforcement between metallogenic mechanism knowledge and data-driven prediction.

[0072] Step S5: Utilize the mineralization mechanism knowledge graph and mechanism conformity vector to perform multi-scale interactive hierarchical prediction to obtain deposit-scale prediction results; perform cross-scale collaborative optimization on the deposit-scale prediction results and update the deposit-scale prediction results. In this embodiment, "multi-scale interactive hierarchical prediction" refers to constructing prediction sub-models at three spatial scales: regional scale, ore field scale, and ore deposit scale. While ensuring that each scale prediction model has independent inputs and outputs, cross-scale guiding information and mechanism conformity vectors are introduced to enable information interaction between different scales during the prediction process. This allows for the representation of mineralization regularities at a large scale and the preservation of responses to local anomalies at a small scale. Compared to single-scale prediction methods, multi-scale interactive hierarchical prediction provides a more complete representation of the coupling relationship between regional metallogenic background, favorable ore field areas, and local enrichment areas of ore deposits.

[0073] The “regional-scale prediction sub-model” is used to characterize the overall mineralization favorability and mineralization background zoning of the target study area at a relatively coarse spatial resolution. Generally, it adopts a larger grid size or a zoning result based on tectonic units as the regional-scale spatial unit. The features at this scale mainly reflect information such as the distribution of large-scale rock masses, the framework of deep and large fault structures, and the regional geochemical field. The “mineral field-scale prediction sub-model” is used to further distinguish favorable areas within the mineral field in the already identified favorable zones or key exploration areas at the regional scale. The spatial unit size at this scale is small, and the key features reflect the combination of small and medium-sized faults, favorable lithological combinations, and superposition of multi-element anomalies within the mineral cluster area. The “Ore Deposit-Scale Prediction Sub-Model” is used to make detailed predictions of ore deposits or target areas within favorable areas at the ore field scale. The spatial unit size at this scale is smaller, and the key features reflect the subtle changes in fluid convergence intensity and mineralization accessibility in local fault intersections, lithological transition zones, alteration zones, and structural ore-controlling features.

[0074] The “mechanism conformity vector” is a combination of the structural conformity component, lithological conformity component, and geochemical conformity component obtained in steps S3 and S4. It is used to introduce prior constraints at the level of metallogenic mechanism in the multi-scale prediction process, so that the prediction results not only depend on statistical characteristics, but also take into account the degree of fit with typical metallogenic models.

[0075] In this embodiment, the "mechanism conformity vector" is a three-dimensional real-valued vector formed by sequentially combining the tectonic conformity component, the lithological conformity component, and the geochemical conformity component. Each component of this vector ranges from 0 to 1, where 0 indicates a complete non-compliance with known metallogenic models, 1 indicates a complete conformity with typical metallogenic models, and values ​​in between indicate partial conformity or a certain degree of matching. When calculating each component, the original scores are first accumulated based on the matching results of the map inference rules. Then, a linear normalization method is used to scale the original scores according to the minimum and maximum values ​​of all spatial units within the study area for that component, mapping them to a unified interval of 0 to 1. If a component has no difference in score across the entire study area, it is uniformly set to the median value of 0.5 to maintain the validity and comparability of the vector.

[0076] The hierarchical prediction process involving multi-scale interaction includes: dividing spatial units into regional scale spatial units, ore field scale spatial units, and ore deposit scale spatial units according to scale type; inputting the multi-scale feature data corresponding to each scale spatial unit into the corresponding regional scale prediction sub-model, ore field scale prediction sub-model, and ore deposit scale prediction sub-model; using the regional scale prediction sub-model to predict the regional scale spatial units to obtain regional scale prediction results; generating regional scale guidance information based on the regional scale prediction results and the multi-scale feature data of the ore field scale spatial units; inputting the regional scale guidance information and the mechanism conformity vector into the ore field scale prediction sub-model to obtain the ore field scale prediction results; generating ore field scale guidance information based on the ore field scale prediction results and the multi-scale feature data of the ore deposit scale spatial units; inputting the ore field scale guidance information, the mechanism conformity vector, and the structural ore-controlling features into the ore deposit scale prediction sub-model; and outputting the ore deposit scale prediction results from the ore deposit scale prediction sub-model.

[0077] In practical implementation, firstly, based on the multi-scale spatial units established in step S1, the spatial units are divided into three levels according to their scale identifiers: regional scale spatial units, ore field scale spatial units, and ore deposit scale spatial units. For each scale spatial unit, multi-scale feature data containing geological feature components, tectonic feature components, and geochemical feature components have been calculated in step S1, and structural ore-controlling feature information has been supplemented in step S2. Mechanism conformity vector information has been supplemented in steps S3 and S4.

[0078] For the regional-scale prediction sub-model, its input features mainly include multi-scale feature data corresponding to regional-scale spatial units and comprehensive indicators related to mineralization mechanisms, such as the distribution of rock mass types over a large area, the density and extension direction of deep and large faults, regional geochemical background values ​​and their anomalies. The regional-scale prediction sub-model employs a supervised learning method based on deep neural networks. During the training phase, regional-scale units containing existing mineralized clusters or typical deposits are used as positive samples, while regional-scale units lacking mineralization evidence are used as negative samples, learning the correspondence between multi-scale feature data and the favorable conditions for regional mineralization. In the prediction phase, inference is performed on all regional-scale spatial units to obtain regional-scale prediction results, which reflect the favorable distribution pattern of mineralization across the entire study area on a large scale.

[0079] Based on this, regional-scale guiding information is generated using regional-scale prediction results and multi-scale feature data of mineral field-scale spatial units. Regional-scale guiding information refers to projecting the regional-scale prediction results onto mineral field-scale spatial units according to their spatial location. For each mineral field-scale spatial unit, a weighted statistical value of the prediction scores of several nearby regional-scale spatial units is calculated, using methods such as neighborhood average, weighted average, or maximum value, to form guiding features reflecting the mineralization favorability of the regional background. Through these guiding features, the mineral field-scale prediction sub-model, when judging the mineralization favorability of a mineral field-scale spatial unit, considers not only local features but also the strength of the mineralization background of the region on a large scale.

[0080] When regional-scale guiding information and mechanism conformity vectors are input into the orefield-scale prediction sub-model, the mechanism conformity vector is incorporated as an additional input dimension, ensuring that the prediction results are simultaneously constrained by data-driven features and prior knowledge of mineralization mechanisms. During training, the orefield-scale prediction sub-model can select data from known orefield boundaries, the interior and exterior of typical mineralized areas, and learn the relationship between orefield-scale characteristics and mineralization favorability under regional-scale background and mechanism conformity conditions. In the prediction phase, inference is performed on all orefield-scale spatial units to obtain orefield-scale prediction results, which are used to characterize a more refined range of prospecting favorable areas within the regional favorable zone.

[0081] Mineral field-scale guidance information is generated based on the prediction results at the mineral field scale and the multi-scale characteristic data of the mineral deposit-scale spatial units. The generation method for mineral field-scale guidance information is similar to that of regional-scale guidance information. This involves mapping the mineral field-scale prediction results to mineral deposit-scale spatial units according to their spatial location. For each mineral deposit-scale spatial unit, the prediction score statistics of its own mineral field-scale unit and neighboring mineral field-scale units are calculated, thus forming background constraints reflecting the favorable distribution of mineralization within the mineral field. When the mineral field-scale guidance information, along with the mechanism conformity vector and structural ore-controlling features, are input into the mineral deposit-scale prediction sub-model, the mechanism conformity vector provides mechanistic constraints in terms of structure, lithology, and geochemistry; the structural ore-controlling features provide two types of quantitative structural ore-controlling indicators: fluid convergence intensity and mineralization channel accessibility; and the multi-scale characteristic data provides local geological and geochemical background information. During the training phase, the ore deposit scale prediction sub-model uses known ore deposit target areas, mineralized points, and unmineralized units as samples to construct a mapping relationship between the spatial unit characteristics of ore deposit scale and the mineralization favorability score through supervised learning. During the prediction phase, the ore deposit scale prediction sub-model outputs ore deposit scale prediction results for all ore deposit scale spatial units, thereby achieving fine prediction of the ore deposit or target area scale.

[0082] The cross-scale collaborative optimization includes: constructing a cross-scale collaborative optimization model based on the regional-scale prediction results and the ore field-scale prediction results obtained from hierarchical predictions performed with multi-scale interactions, using the ore deposit-scale prediction results as optimization variables; introducing global constraint terms to constrain the consistency between the ore deposit-scale prediction results and the regional-scale prediction results, and introducing local constraint terms to constrain the consistency between the ore deposit-scale prediction results and the ore field-scale prediction results and the mechanism conformity vector, and combining the deviations between each constraint term and the original ore deposit-scale prediction results to form an objective function; and correcting the ore deposit-scale prediction results under multi-scale constraints by iteratively solving the objective function, obtaining the updated ore deposit-scale prediction results when the preset convergence conditions are met.

[0083] After the hierarchical prediction at multiple scales is completed, in order to avoid contradictions between prediction results at different scales and improve the consistency of the final deposit-scale prediction results at the regional, ore field, and mechanistic levels, this embodiment performs cross-scale collaborative optimization on the deposit-scale prediction results. "Cross-scale collaborative optimization" refers to making overall adjustments to the deposit-scale prediction results under the condition of integrating regional-scale prediction results, ore field-scale prediction results, and mechanistic consistency vectors, so that the prediction distribution at the deposit-scale level follows both the large-scale metallogenic background and the distribution of favorable areas at the mesoscale, while maintaining consistency with local mechanistic information.

[0084] Specifically, based on the regional-scale and ore-field-scale prediction results obtained from hierarchical predictions involving multi-scale interactions, a cross-scale collaborative optimization model is constructed, using the ore-deposit-scale prediction results as the optimization variables to be adjusted. The cross-scale collaborative optimization model sets two main types of constraints: One type is the global constraint term, which is used to ensure that the prediction results at the deposit scale are consistent with those at the regional scale in space. The global constraint term compares the prediction score of the regional scale spatial unit where the deposit-scale spatial unit is located with the prediction score of the deposit-scale spatial unit, and applies a penalty term to spatial units with large differences. This guides the deposit-scale prediction results to follow the regional scale mineralization favorability distribution over a large area, and avoids high-value or low-value anomalies that deviate significantly from the regional background.

[0085] Another type is the local constraint term, used to ensure consistency between the deposit-scale prediction results and the field-scale prediction results, as well as the mechanism conformity vector. The local constraint term includes two parts: one part guides the changes in the deposit-scale prediction results within the field to conform to the favorable area distribution framework at the field scale by comparing the prediction scores of the deposit-scale spatial unit with its superior field-scale spatial unit; the other part enhances the prediction score in spatial units with high mechanism conformity and suppresses the prediction score in spatial units with low mechanism conformity or even those that contradict typical metallogenic models, thereby achieving coordination between data-driven prediction and mechanism constraints.

[0086] In the cross-scale collaborative optimization model, the deviations between global constraints, local constraints, and the original deposit-scale prediction results are comprehensively constructed into an objective function. This objective function measures the overall inconsistency between the current deposit-scale prediction results and the prediction results at each scale, as well as the mechanistic constraints. The objective function is solved iteratively, for example using gradient-based optimization algorithms or other numerical iterative methods. In each iteration, the prediction scores of the deposit-scale spatial units are adjusted, gradually reducing the objective function. During the iteration process, to avoid over-smoothing and loss of local anomaly information, a certain weight of the original deposit-scale prediction result term can be retained in the objective function to maintain sensitivity to local high-value anomalies.

[0087] When the change in the objective function falls below a preset threshold or the number of iterations reaches a preset upper limit, the cross-scale collaborative optimization process is considered to have converged. The ore deposit scale prediction result obtained at this point is used as the updated ore deposit scale prediction result. The updated ore deposit scale prediction result numerically integrates the distribution information at the regional and ore field scales, and spatially strengthens areas with high mechanism consistency and suppresses areas with low mechanism consistency.

[0088] Step S6: Calculate the uncertainty index for the ore deposit scale prediction results, and fuse the uncertainty index with the data quality index and the mechanism conformity vector to generate a credibility index; determine the target area score value based on the ore deposit scale prediction results, the mechanism conformity vector and the uncertainty index, and extract the connected regions that meet the preset conditions as candidate areas for mineralization target areas for output.

[0089] In this embodiment, the deposit-scale prediction result in step S6 is the updated deposit-scale prediction result obtained after cross-scale collaborative optimization in step S5. This result has been coordinated and corrected under the combined effect of regional scale, ore field scale, and mechanistic constraints. However, due to the incompleteness of the input data, the parameter uncertainty of the prediction model, and the existence of certain residuals between multi-scale constraints, it is still necessary to conduct an uncertainty assessment on the prediction result and comprehensively consider the uncertainty with data quality and mechanistic conformity to generate a credibility index, so as to improve the reliability of the metallogenic target area evaluation result.

[0090] The generation of the credibility index includes: performing multiple inferences on the updated deposit scale prediction results under the condition of introducing random inactivation or parameter perturbation to obtain multiple deposit scale prediction results, and determining the uncertainty index based on the degree of dispersion among the multiple deposit scale prediction results; normalizing the uncertainty index, data quality index, and mechanism conformity vector respectively, and weighting and combining the normalized uncertainty index, data quality index, and mechanism conformity vector according to a preset fusion function to obtain the credibility index.

[0091] In practical implementation, "random deactivation" refers to randomly masking a portion of neurons or feature channels within the model during the inference phase while maintaining the model structure. This results in slightly different effective structures in different inference rounds, reflecting the uncertainty of model parameters and structure. "Parameter perturbation" refers to applying small perturbations to some weight parameters or minor changes to input features within the existing error range during the inference phase, based on the model's training, to simulate the impact of training and input errors on the prediction results. By performing multiple inferences on the updated deposit-scale prediction results under random deactivation or parameter perturbation conditions, a set of prediction scores is obtained for each deposit-scale spatial unit. The dispersion of these prediction scores reflects the model's prediction stability within that spatial unit.

[0092] The degree of dispersion is used to quantify the uncertainty index, which can be characterized by statistical measures such as standard deviation and range. When the predicted score of a spatial unit changes little in multiple inferences, it indicates that the prediction result of that unit is not sensitive to internal model disturbances, has good prediction stability, and the corresponding uncertainty index is low. When the predicted score of a spatial unit varies greatly between different inference rounds, it indicates that the prediction result of that unit is more sensitive to model structure or parameter disturbances, and the corresponding uncertainty index is high.

[0093] The term "multiple inferences" refers to quantifying the uncertainty of prediction results by repeatedly executing the model's inference process. In practice, a fixed number of inference iterations, N, is set, where N is an integer greater than or equal to 3 and less than or equal to 10. For example, N=5. This range of iterations aims to ensure statistical stability while also considering computational efficiency. In practical applications, the curve of uncertainty indicators (such as the standard deviation of prediction scores) changing with the number of inference iterations is observed through preliminary experiments; the number of iterations at which the curve flattens out is selected as N. During the uncertainty calculation process, for the same spatial unit, after introducing different random inactivation modes or parameter perturbations, N inference iterations are performed independently to obtain N deposit-scale prediction results for that unit, and then its dispersion is calculated.

[0094] The data quality indicators are the set of indicators calculated in step S1, including geochemical sampling density, data missing rate, data source reliability level, and multi-source consistency. The mechanism consistency vector is a combination of the structural consistency component, lithological consistency component, and geochemical consistency component formed in steps S3 and S4, used to characterize the degree of consistency between each spatial unit and typical metallogenic mechanism models. Since the uncertainty indicators, data quality indicators, and mechanism consistency vector have different numerical ranges and dimensions, they need to be normalized separately to map them to a unified numerical range for easy integration within the same framework. The normalization process uses interval scaling to ensure that the comparison between different indicators has a consistent numerical scale.

[0095] After normalization, the normalized uncertainty index, data quality index, and mechanism conformity vector are weighted and combined according to a preset fusion function to obtain the credibility index. The "credibility index" is a comprehensive score reflecting the reliability of the prediction results at the level of a single ore deposit-scale spatial unit, taking into account the completeness of the data foundation, the consistency of the mineralization mechanism, and the stability of the model output. In this embodiment, the fusion function is set to highlight spatial units with high data quality, high mechanism conformity, and low uncertainty, while applying appropriate penalty weights to spatial units with weak data quality, low mechanism conformity, or high uncertainty.

[0096] The fusion function is configured to increase the weight of spatial units with high data quality indicators and high consistency of the mechanism consistency vector, and decrease the confidence index of spatial units with large uncertainty indicators. Specifically, when the data quality indicators of a spatial unit show that the unit has a high sampling density, a low data missing rate, and a high multi-source consistency score, and the structural consistency component, lithological consistency component, and geochemical consistency component in the mechanism consistency vector are all at a high level, the spatial unit will receive a high positive weight in the fusion function, thus making its confidence index close to the upper limit. When a spatial unit exhibits a high uncertainty index in multiple inferences, that is, the prediction result is very sensitive to model perturbations, the fusion function will lower its confidence index, and even if its prediction score is high, it will not be directly regarded as a high-confidence target area.

[0097] After obtaining the credibility index, the target area score is determined based on the deposit-scale prediction results, mechanism consistency vector, and uncertainty index. The target area score is the final evaluation score obtained for each deposit-scale spatial unit after comprehensively considering mineralization favorability, mechanism consistency, and prediction reliability. Typically, the updated deposit-scale prediction results are used as the base score when constructing the target area score. The structural consistency, lithological consistency, and geochemical consistency in the mechanism consistency vector are used as mechanism correction factors, and the credibility index is used as a reliability weighting factor, thus forming a comprehensive scoring framework that reflects both mineralization probability and identification credibility. For spatial units with high mechanism consistency, high credibility index, and high prediction score, the target area score will be significantly higher than other areas; for spatial units with high prediction score but low mechanism consistency or high uncertainty, the target area score will be reduced due to mechanism penalty or credibility penalty.

[0098] Spatially, the target area score is mapped back to the spatial location of ore deposit-scale spatial units, and units whose score values ​​meet preset conditions are selected. These preset conditions include a minimum target area score threshold, a minimum confidence index threshold, and a minimum area requirement. For example, a spatial unit is considered a favorable ore-forming unit if its target area score is greater than a given threshold and its confidence index is higher than a certain limit.

[0099] Based on this, spatial connectivity analysis is used to aggregate adjacent mineralization-favorable units and extract connected regions whose score values ​​meet preset conditions. A "connected region" refers to a spatial cluster composed of continuously adjacent mineralization-favorable units under a given spatial adjacency rule (e.g., using a four-adjacency or eight-adjacency rule). Such connected regions spatially reflect areas representing concentrated mineralization distributions or favorable mineralization zones.

[0100] When extracting connected regions, further screening is performed based on area, shape, spatial relationship with known ore deposit locations, and correspondence with deep structural channels. Scattered high-value units with excessively small areas are excluded or merged, while connected regions that are continuously distributed and consistent with the regional structural pattern are retained and marked. Finally, connected regions that meet the preset conditions are output as candidate areas for mineralization targets, providing spatial indications for subsequent field verification, geophysical exploration deployment, and engineering validation.

[0101] Example 2, an intelligent prediction system for tungsten ore-forming target areas, see [link to example]. Figure 1 As shown, it includes the following modules: Data mapping module: acquires multi-source mineral exploration data of the target study area, establishes multi-scale spatial units, maps multi-source mineral exploration data to each spatial unit, forms multi-scale feature data, and calculates data quality indicators; Mineralization Feature Module: Based on the fracture structure data and geological data, the ancient fracture network of the mineralization period is constructed to simulate the migration of mineralizing fluids, and the fluid convergence intensity and mineralization channel accessibility of each spatial unit are obtained as structural mineralization control features; Metallogenic Map Module: Constructs a knowledge graph of metallogenic mechanisms, maps each spatial unit into a spatial unit subgraph based on structural ore-controlling features, geological data, and geochemical data, and runs preset map reasoning rules to generate an initial mechanism conformity vector; Prediction Feedback Module: With the support of the metallogenic mechanism knowledge graph, it executes a prediction-discovery feedback loop to update the metallogenic mechanism knowledge graph and the mechanism conformity vector; Ore deposit prediction module: Utilizing the knowledge graph of mineralization mechanism and the mechanism conformity vector, it performs multi-scale interactive hierarchical prediction to obtain ore deposit scale prediction results; and performs cross-scale collaborative optimization on the ore deposit scale prediction results to obtain cross-scale collaboratively optimized ore deposit scale prediction results. Target area output module: Calculates uncertainty index for ore deposit scale prediction results, integrates uncertainty index with data quality index and mechanism conformity vector to generate credibility index; determines target area score value based on ore deposit scale prediction results, mechanism conformity vector and uncertainty index, and extracts connected regions that meet preset conditions as candidate areas for mineralization target area for output.

[0102] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of tungsten ore-forming target areas, characterized in that, include: Acquire multi-source mineral exploration data for the target study area, including geological data, fault structure data, and geochemical data; establish multi-scale spatial units; map multi-source mineral exploration data to each spatial unit; form multi-scale characteristic data; and calculate data quality indicators. Based on fracture tectonic data and geological data, a paleofractal network of mineralization period was constructed to simulate the migration of ore-forming fluids, and the fluid convergence intensity and mineralization channel accessibility of each spatial unit were obtained as structural ore-controlling characteristics. A knowledge graph of mineralization mechanism is constructed. Based on structural mineralization control characteristics, geological data and geochemical data, each spatial unit is mapped into a spatial unit subgraph. Pre-set graph reasoning rules are run to generate an initial mechanism conformity vector. With the support of the metallogenic mechanism knowledge graph, a prediction-discovery feedback loop is executed to update the metallogenic mechanism knowledge graph and the mechanism conformity vector; By utilizing the knowledge graph of mineralization mechanism and the mechanism conformity vector, a hierarchical prediction with multi-scale interaction is performed to obtain the ore deposit scale prediction results; cross-scale collaborative optimization is then performed on the ore deposit scale prediction results to update the ore deposit scale prediction results. The uncertainty index is calculated for the ore deposit scale prediction results, and the uncertainty index is fused with the data quality index and the mechanism conformity vector to generate a credibility index. Based on the ore deposit scale prediction results, mechanism conformity vector and uncertainty index, the target area score value is determined, and the connected regions whose score values ​​meet the preset conditions are extracted as candidate areas for ore-forming targets and output.

2. The intelligent prediction method for tungsten ore-forming target areas according to claim 1, characterized in that, The paleofault network of the metallogenic period includes: a set of nodes consisting of fault intersections, fault endpoints, and fault turning points, and a set of edges consisting of fault segments; each fault segment is associated with and stored geometric attributes, structural attributes, ore-conducting attribute parameters, and ore-blocking attribute parameters; the geometric attributes include strike, dip, and length; the structural attributes include fault level and lithological contact relationship, determined by fault interpretation and fault period division of the fault structural data; the ore-conducting attribute parameters and ore-blocking attribute parameters are assigned by lithological and stratigraphic information in the geological data.

3. The intelligent prediction method for tungsten ore-forming target areas according to claim 2, characterized in that, The simulation of ore-forming fluid migration includes: discretizing the paleofault network during the mineralization period into migration paths composed of fault segments; using known mineral occurrences and geochemical anomaly nodes as injection locations of virtual mineralization sources; and, considering the ore-conducting and ore-resisting properties of each fault segment as well as preset attenuation parameters, gradually propagating and accumulating the flux of ore-forming fluid along the migration path. The fluid convergence intensity is determined by statistically analyzing the cumulative value of ore-forming fluid flux within spatial units at various scales that intersect with the migration path, and the accessibility of mineralization channels is determined by the comprehensive resistance and path connectivity along the fault segments between the virtual mineralization source and each spatial unit.

4. The intelligent prediction method for tungsten ore-forming target areas according to claim 1, characterized in that, The mineralization mechanism knowledge graph includes: a set of nodes representing geological entities and a set of edges representing mineralization relationships; the set of nodes includes rock mass nodes, stratigraphic nodes, tectonic nodes, alteration zone nodes, and geochemical anomaly nodes; the set of edges includes ore-controlling relationship edges, ore-supplying relationship edges, sealing relationship edges, and enrichment relationship edges; the nodes and edges are associated with and stored spatial location attributes, scale attributes, and mapping relationship attributes with spatial units of each scale; based on structural ore-controlling characteristics, geological data, and geochemical data, the attribute information corresponding to fault structures, lithology, stratigraphy, alteration zones, and geochemical anomalies within each spatial unit of each scale is associated with the corresponding types of nodes and mineralization relationship edges in the set of nodes and the set of edges according to preset mapping rules, forming a spatial unit subgraph corresponding to each spatial unit of each scale.

5. The intelligent prediction method for tungsten ore-forming target areas according to claim 4, characterized in that, The preset graph reasoning rules include: performing subgraph matching reasoning and path reasoning on each spatial unit subgraph based on the mineralization mechanism knowledge graph. The graph reasoning rules include rules for determining whether the tectonic control relationship is satisfied, rules for determining whether the lithological combination is conducive to mineralization, and rules for determining the synergistic characteristics of multi-element geochemical anomalies. Based on the degree to which each spatial unit subgraph satisfies the graph reasoning rules, the tectonic conformity component, the lithological conformity component, and the geochemical conformity component are calculated respectively. The initial mechanism conformity vector is composed of the tectonic conformity component, the lithological conformity component, and the geochemical conformity component.

6. The intelligent prediction method for tungsten ore-forming target areas according to claim 4, characterized in that, The execution prediction-discovery feedback loop includes: using the initial mechanism conformity vector, performing the first round of multi-scale interactive hierarchical prediction and cross-scale collaborative optimization to obtain the first round of deposit-scale prediction results; comparing and analyzing the spatial distribution of the first round of deposit-scale prediction results with known mineral occurrences, geochemical anomalies, and structural and lithological features characterized by geological data to identify model blind spots and knowledge conflict areas; for model blind spots, using a constraint-based causal relationship discovery algorithm based on multi-source mineral exploration data to mine spatial association rules of geological features not encoded by the current metallogenic mechanism knowledge graph, and generating candidate... The new metallogenic model, whose geological features include structural and lithological features characterized by geological data, as well as spatial combination relationships with geochemical anomalies; for knowledge conflict areas, the confidence of the map reasoning rules related to the area is evaluated and adjusted; the candidate new metallogenic model is submitted to geological experts in the form of rules for confirmation or correction. After confirmation or correction by geological experts, the verified candidate new metallogenic model is supplemented or replaced with new map reasoning rules, the metallogenic mechanism knowledge map is updated, and an updated mechanism conformity vector is generated based on the updated metallogenic mechanism knowledge map.

7. The intelligent prediction method for tungsten ore-forming target areas according to claim 1, characterized in that, The hierarchical prediction process involving multi-scale interaction includes: dividing spatial units into regional scale spatial units, ore field scale spatial units, and ore deposit scale spatial units according to scale type; inputting the multi-scale feature data corresponding to each scale spatial unit into the corresponding regional scale prediction sub-model, ore field scale prediction sub-model, and ore deposit scale prediction sub-model; using the regional scale prediction sub-model to predict the regional scale spatial units to obtain regional scale prediction results; generating regional scale guidance information based on the regional scale prediction results and the multi-scale feature data of the ore field scale spatial units; inputting the regional scale guidance information and the mechanism conformity vector into the ore field scale prediction sub-model to obtain the ore field scale prediction results; generating ore field scale guidance information based on the ore field scale prediction results and the multi-scale feature data of the ore deposit scale spatial units; inputting the ore field scale guidance information, the mechanism conformity vector, and the structural ore-controlling features into the ore deposit scale prediction sub-model; and outputting the ore deposit scale prediction results from the ore deposit scale prediction sub-model.

8. The intelligent prediction method for tungsten ore-forming target areas according to claim 7, characterized in that, The cross-scale collaborative optimization includes: constructing a cross-scale collaborative optimization model based on the regional-scale prediction results and the ore field-scale prediction results obtained from hierarchical predictions performed with multi-scale interactions, using the ore deposit-scale prediction results as optimization variables; introducing global constraint terms to constrain the consistency between the ore deposit-scale prediction results and the regional-scale prediction results, and introducing local constraint terms to constrain the consistency between the ore deposit-scale prediction results and the ore field-scale prediction results and the mechanism conformity vector, and combining the deviations between each constraint term and the original ore deposit-scale prediction results to form an objective function; and correcting the ore deposit-scale prediction results under multi-scale constraints by iteratively solving the objective function, obtaining the updated ore deposit-scale prediction results when the preset convergence conditions are met.

9. The intelligent prediction method for tungsten ore-forming target areas according to claim 1, characterized in that, The generation of the credibility index includes: performing multiple inferences on the updated deposit scale prediction results under the condition of introducing random inactivation or parameter perturbation to obtain multiple deposit scale prediction results, and determining the uncertainty index based on the degree of dispersion among the multiple deposit scale prediction results; normalizing the uncertainty index, data quality index, and mechanism conformity vector respectively, and weighting and combining the normalized uncertainty index, data quality index, and mechanism conformity vector according to a preset fusion function to obtain the credibility index.

10. A smart prediction system for tungsten ore-forming target areas, characterized in that, The system applies the intelligent prediction method for tungsten ore-forming target areas according to any one of claims 1 to 9, including: Data mapping module: acquires multi-source mineral exploration data of the target study area, establishes multi-scale spatial units, maps multi-source mineral exploration data to each spatial unit, forms multi-scale feature data, and calculates data quality indicators; Mineralization Feature Module: Based on the fracture structure data and geological data, the ancient fracture network of the mineralization period is constructed to simulate the migration of mineralizing fluids, and the fluid convergence intensity and mineralization channel accessibility of each spatial unit are obtained as structural mineralization control features; Metallogenic Map Module: Constructs a knowledge graph of metallogenic mechanisms, maps each spatial unit into a spatial unit subgraph based on structural ore-controlling features, geological data, and geochemical data, and runs preset map reasoning rules to generate an initial mechanism conformity vector; Prediction Feedback Module: With the support of the metallogenic mechanism knowledge graph, it executes a prediction-discovery feedback loop to update the metallogenic mechanism knowledge graph and the mechanism conformity vector; Ore deposit prediction module: Utilizing the knowledge graph of mineralization mechanism and the mechanism conformity vector, it performs multi-scale interactive hierarchical prediction to obtain ore deposit scale prediction results; and performs cross-scale collaborative optimization on the ore deposit scale prediction results to obtain cross-scale collaboratively optimized ore deposit scale prediction results. Target area output module: Calculates uncertainty index for ore deposit scale prediction results, integrates uncertainty index with data quality index and mechanism conformity vector to generate credibility index; determines target area score value based on ore deposit scale prediction results, mechanism conformity vector and uncertainty index, and extracts connected regions that meet preset conditions as candidate areas for mineralization target area for output.

Images