Tungsten deposit deep edge metallogenic prediction method based on machine learning

Abstract

The invention discloses a tungsten deposit deep edge metallogenic prediction method based on machine learning, and particularly relates to the technical field of geological prediction and data processing. The method comprises the following steps: acquiring multi-source geological data, constructing a multi-dimensional space grid, and dividing candidate prediction units; quartz vein density, alteration indexes, geochemical combination anomaly weights and high-frequency magnetic anomaly waveform features are extracted, and an ore control feature tensor matrix is constructed; extracting a spatial dependency relationship among the candidate units based on a graph neural network, constructing a metallogenic potential prediction model, and outputting a metallogenic probability graph and a metallogenic hot spot graph; model parameters and feature weights are dynamically optimized through spatial superposition analysis and an abnormal response recognition mechanism; according to the method, the precision and efficiency of deep edge tungsten ore prospecting can be remarkably improved, and good self-adaptability and popularization and application value are achieved.

Description

Technical Field

[0001] This invention relates to the field of geological prediction and data processing technology, specifically to a machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits. Background Technology

[0002] Tungsten, as a strategic and critical mineral resource, is characterized by complex occurrence patterns and strong spatial variability. Particularly in deep and peripheral areas of ore bodies, geological information is sparse, tectonic structures are intensely superimposed, and the mineralization process is unclear, posing significant challenges to traditional prediction methods. Currently, the prediction of deep and peripheral tungsten deposits relies on costly methods such as drilling and trenching for verification, which suffers from coarse prediction granularity, low identification accuracy, and poor exploration efficiency, making it difficult to support precise mineral exploration.

[0003] Existing methods mostly employ linear superposition of geological and geochemical factors for modeling, which fails to capture the nonlinear, implicit ore-controlling characteristics inherent in complex tectonic-hydrothermal interactions. Furthermore, in deep, peripheral regions, intense tectonic activity and uneven distribution of geological information sampling lead to models prone to overfitting or poor generalization ability, particularly in prediction gaps or high-risk blind areas.

[0004] Furthermore, current models struggle to collaboratively model with historical sparse data and current high-resolution remote sensing or geophysical data, lack dynamic evolution capabilities, and fail to construct a spatiotemporal coupling mechanism in the deep-edge mineralization process, thus limiting prediction accuracy and application value. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based method for predicting mineralization in the deep edges of tungsten deposits, in order to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits, comprising: Acquire multi-source geological data for the target area, including surface, borehole, and tunnel geological survey data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data; A multi-dimensional spatial grid is constructed based on the multi-source geological data to divide the target area into multiple candidate prediction units; Clustering algorithms were used to divide the sample space of candidate prediction units and extract mineralization indicator factors for each unit, including quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform characteristics. The mineralization indicator factors are recalibrated to construct a deep-edge mineralization control feature tensor matrix, and a graph neural network is used to extract the non-Euclidean spatial dependencies between candidate units. A mineralization potential prediction model was constructed, mineralization probability scores were given for each candidate unit, and mineralization hotspot maps of key target areas in the deep edge were output. By spatially overlaying the prediction results with existing tungsten ore distribution points, abnormal response patterns of the mineralization potential prediction model in uncontrolled areas are identified, and input feature weights and model parameters are optimized.

[0007] Preferably, the target region is divided into multiple candidate prediction units, including: Spatial registration was performed on geological survey data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data. Under the registered unified spatial reference system, the target region is divided into regular grid cells using a gridding algorithm; Based on the data point density, independent variable completeness and information entropy index within each unit, effective units that meet the data quality threshold are selected as candidate prediction units. A unique spatial code is attached to the candidate prediction unit, and a mapping relationship between the candidate unit and the geological attributes is established.

[0008] Preferably, a clustering algorithm is used to divide the sample space of candidate prediction units and extract the mineralization indicator factors for each unit, including: The standardized feature matrix of the candidate prediction unit is first reduced in dimensionality and denoised, and principal component analysis is used to retain the principal components with a cumulative contribution rate of not less than 90%. Based on the principal components, density peak clustering or DBSCAN is used to divide the sample space, and the clustering parameters are automatically selected by the silhouette coefficient to identify geological isoprotic groups. For each subgroup, the quartz vein density is calculated as the sum of the lengths of quartz veins per unit volume, the alteration index is the weighted average of the remote sensing band ratio and the content of altered minerals in the field, and the geochemical assemblage anomaly weight is calculated by Bayesian weighted summation based on the elemental synergistic anomaly probability. High-frequency magnetic anomaly waveform features are extracted, and the energy spectral density is calculated within a sliding window using short-time Fourier transform and normalized to eigenvalues. Each indicator factor forms the final mineralization indicator vector.

[0009] Preferably, the method for calculating the density of quartz veins is as follows: First, extract all quartz vein segments that traverse candidate units in three-dimensional space and project their two-dimensional spatial shapes onto a three-dimensional coordinate system; then, calculate the total length L of all quartz vein segments intersecting with the volume of each unit; finally, divide the total quartz vein length L by the spatial volume V of the unit to obtain the quartz vein density S, expressed as: Quartz Vein Density .

[0010] Preferably, the calculation process for the geochemical assemblage anomaly weight includes: determining the element combination closely related to tungsten mineralization; performing Z-score standardization on the content data of each element in the candidate prediction unit; constructing the joint anomaly probability relationship between elements using a Bayesian network model, whereby the Bayesian network evaluates the probability of multiple elements simultaneously exhibiting anomalies by setting prior probability and conditional probability distributions; and finally, summing the joint probabilities of the coordinated anomalies of each element according to element importance to obtain the comprehensive geochemical assemblage anomaly weight for that unit.

[0011] Preferably, the deep edge ore-controlling feature tensor matrix is ​​constructed, including: The mineralization indicator factors extracted from each candidate prediction unit are subjected to multi-scale normalization. The normalized quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform features are encoded into four feature channels, and a three-dimensional ore-controlling feature tensor is constructed using the spatial location of candidate units as an index. The tensor dimensions correspond to the unit ID, number of channels, and feature value, respectively. An adjacency matrix between candidate units is constructed using a spatial connectivity graph, and the weights of the edges are generated by fusing three factors: spatial topological distance, fracture connectivity, and geological unit similarity.

[0012] Preferably, a graph neural network is used to extract the non-Euclidean spatial dependencies between candidate units, including: Based on the spatial location, structural connectivity, and geological attribute similarity of candidate prediction units, a graph structure is constructed to form a weighted undirected graph with nodes as units and edges as spatial associations. An adjacency matrix A is generated from this graph, where the edge weights are fused with topological distance and attribute similarity using a Gaussian kernel function. The control feature tensor of each candidate unit is used as the initial feature of the input node of the graph neural network. The graph convolutional neural network is used to perform feature propagation and update the node embedding representation according to the hierarchical propagation rule. Each convolutional layer achieves context information aggregation through feature weighted averaging. The graph embedding vector of each node in the final output is used as a spatial dependency enhancement feature for deep edge candidate prediction units.

[0013] Preferably, a mineralization potential prediction model is constructed, and a mineralization probability score is given to each candidate unit, including: The candidate unit embedding vector processed by the graph neural network is used as the input feature to construct a multilayer perceptron mineralization potential scoring model with a softmax output layer, and outputs the mineralization probability value of each unit. The model was trained under supervision using the cross-entropy loss function. The training samples were derived from the labeled data of known mining sites and background points. The optimization method was the Adam algorithm, and the learning rate was set to 0.001. Based on the mineralization probability value, a confidence threshold is set to divide key prediction units, and high-value cluster areas are generated by aggregating spatially adjacent high-potential units. The probability distribution map is rasterized to output the mineralization hot spot map of the key target area in the deep edge.

[0014] Preferably, the prediction results are spatially overlaid with existing tungsten ore distribution points to identify anomalous response patterns of the mineralization potential prediction model in uncontrolled areas, including: The predicted mineralization probability map is spatially overlaid with the known tungsten deposit distribution points in the region, and the coverage rate and confidence index of each actual mineral point falling into the high predicted probability area are calculated. Identify candidate prediction units corresponding to the areas where the mineral deposits are not covered, analyze the differences in mineralization indicator factors between them and areas with low scores in the mineralization potential prediction model, and construct anomaly response pattern labels; Based on the anomaly response unit, the feature importance analysis method is used to evaluate the degree of influence of various input factors on prediction bias; Adjust the normalization method or weight ratio for the feature channels with significant impact, and retrain the mineralization potential model.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention, by constructing a mineralization control feature tensor and a graph neural network model, can effectively identify non-Euclidean spatial dependencies between candidate prediction units, solving the problem that traditional linear modeling methods struggle to characterize complex tectonic-ore-forming coupling mechanisms. Compared to existing technologies, this invention introduces a multi-channel spatial tensor structure at the data representation layer and a graph structure learning mechanism at the model layer, significantly improving prediction accuracy, spatial resolution, and the ability to identify ore-forming target areas.

[0016] 2. This invention establishes a closed-loop optimization path between model prediction results and actual mineral deposit distribution by introducing spatial overlay analysis and anomaly response feedback mechanisms. It can dynamically adjust feature weights and model parameters based on identifying model blind spots, exhibiting good adaptability and generalization ability. This method is not only applicable to tungsten prospecting in deep, complex tectonic zones, but can also be extended to the prediction of other highly concealed and data-sparse metal minerals. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] For examples, please refer to Figure 1 As shown in this embodiment, the machine learning-based method for predicting mineralization in the deep edges of tungsten deposits includes: Acquire multi-source geological data for the target area, including surface, borehole, tunnel data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data; A multi-dimensional spatial grid is constructed based on the multi-source geological data to divide the target area into multiple candidate prediction units; Clustering algorithms were used to divide the sample space of candidate prediction units and extract mineralization indicator factors for each unit, including quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform characteristics. The mineralization indicator factors are recalibrated to construct a deep-edge mineralization control feature tensor matrix, and a graph neural network is used to extract the non-Euclidean spatial dependencies between candidate units. A mineralization potential prediction model was constructed, mineralization probability scores were given for each candidate unit, and mineralization hotspot maps of key target areas in the deep edge were output. By spatially overlaying the prediction results with existing tungsten ore distribution points, abnormal response patterns of the mineralization potential prediction model in uncontrolled areas are identified, and input feature weights and model parameters are optimized.

[0021] In this invention, the first step is to systematically acquire multi-source geological data of the target tungsten ore prediction area to support subsequent mineralization prediction modeling. Specifically, this includes: First, acquiring geological survey data through existing geological exploration data or surface, borehole, and tunnel data, recording lithological assemblage, quartz vein occurrence, and mineralization alteration characteristics. The sampling interval is recommended to be controlled at 100–200 meters to ensure uniform spatial coverage. Second, using high-resolution airborne geophysical techniques, such as airborne magnetometry and avionics, to systematically scan the target area, extracting high-frequency magnetic anomalies and electrical response anomalies, and performing standardized raster processing. Then, acquiring medium- to high-resolution remote sensing images (such as Landsat-8 and Sentinel-2), and extracting remote sensing mineralization indicators such as alteration information indices and vegetation obscuration indices through principal component analysis and band ratio methods. Finally, based on regional geochemical measurement data, extracting anomaly values ​​and their combined anomaly areas for mineralization-related elements such as tungsten, molybdenum, arsenic, and bismuth.

[0022] A multi-dimensional spatial grid is constructed based on the multi-source geological data, dividing the target area into multiple candidate prediction units. This division includes: spatial registration of geological survey data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data; using a gridding algorithm to divide the target area into regular raster units under a unified spatial reference system after registration; selecting effective units that meet data quality thresholds as candidate prediction units based on data point density, independent variable completeness, and information entropy indices within each unit; assigning a unique spatial code to each candidate prediction unit and establishing a mapping relationship between the candidate unit and geological attributes.

[0023] Data preprocessing involved projecting all data using the unified WGS-84 geographic coordinate system and performing necessary spatial resampling. Spatial resampling used a standard raster unit resolution of 30m x 30m to ensure a uniform raster size across different data sources within the same spatial framework, facilitating subsequent gridding and model input. Furthermore, missing and outlier values ​​were imputed using spatial neighborhood mean interpolation to ensure data continuity and usability.

[0024] After unifying the data space, a multi-dimensional spatial grid model containing three-dimensional information is constructed. Traditional two-dimensional grids only consider planar coordinate information, while tungsten deposits in deep and peripheral regions often exhibit obvious zoning of ore-bearing strata and interpenetration of fault structures. Therefore, this invention introduces elevation values ​​as a third-dimensional spatial variable in grid construction, enabling each spatial unit to have a three-dimensional spatial expression of "longitude-latitude-elevation".

[0025] Specifically, the target area is divided into regular grid cells using spatial grid partitioning algorithms (such as regular grid partitioning). The grid side length is set to 30 meters, and the vertical layer thickness is set to 10 meters based on geological survey data and DEM (Digital Elevation Model) information, with each layer representing an elevation zone. During the vertical stratification process, stratigraphic interface depth, quartz vein density distribution, and alteration zone location are introduced as reference indicators to construct stratification rules and achieve a stratified structure with geological indicative significance.

[0026] Through the above method, a three-dimensional spatial grid structure is formed, in which each unit represents a small volume geological unit and has an identifiable location code, three-dimensional coordinate information and corresponding data attributes.

[0027] Because real-world data may contain areas with insufficient geological information or extremely complex terrain, directly using all spatial units for model training would lead to noise and overfitting issues. Therefore, this invention performs effectiveness screening on the constructed spatial units, with the screening criteria primarily based on the following three indicators: Data point density index: The number of effective geological sampling points contained in each spatial unit is counted. If it is less than a certain threshold (e.g., less than 5), the data in that unit is considered sparse and not useful for training.

[0028] Variable integrity score: assess the integrity of various geological attributes (such as geochemical element content, remote sensing index, geophysical response value, etc.) within each spatial unit, set an integrity threshold (such as 80%), and units below this value will be removed.

[0029] Information entropy index: Calculate the information entropy value of all variables in the unit. If the information entropy is extremely low, it means that the unit has redundant information or poor discrimination and cannot provide effective differential features for the model.

[0030] After normalization, the three indicators mentioned above are combined and weighted to construct a spatial unit effectiveness scoring function. The weight allocation can be set as 40% for data density, 30% for variable integrity, and 30% for information entropy. Units whose scoring function results are greater than a certain preset threshold (such as 0.7) will be identified as candidate prediction units and participate in the subsequent modeling process.

[0031] To facilitate data management and the correspondence between model inputs and outputs, each selected candidate prediction unit is assigned a unique spatial code ID. This ID consists of the latitude and longitude center point, elevation layer number, and grid index value. For example, "E115.2345_N29.875_L3_ID0874" represents a spatial unit numbered 874 in the third layer. Simultaneously, a mapping relationship is established between the candidate unit and its contained geological attributes. Specifically, this involves: reading all data points covered by the unit; performing feature aggregation processing on each type of geological factor (e.g., calculating the mean, median, range, and standard deviation); and assigning the aggregated statistical values ​​to the unit, forming a structured feature vector input set.

[0032] For all candidate prediction units, a standardized feature matrix with a unified dimension is constructed. Each row of this matrix corresponds to a candidate prediction unit, and each column represents a spatial variable (i.e., feature factor), including structural parameters (such as quartz vein density), remote sensing information (such as principal component scores), geochemical indicators (such as elemental content), geophysical features (such as magnetic anomaly values), etc.

[0033] To eliminate the influence of dimensions and differences in data distribution, Z-score standardization is used to normalize the original feature data. The standardization formula is: subtract the global mean of the feature from the original value, and then divide by the standard deviation. The result satisfies a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0034] To reduce noise interference caused by dimensionality redundancy, Principal Component Analysis (PCA) is used to reduce the dimensionality of the standardized feature matrix. Specifically, the sample covariance matrix is ​​calculated, eigenvalues ​​and eigenvectors are obtained, and the eigenvalues ​​are sorted from largest to smallest. The top principal components with a cumulative contribution rate of at least 90% are selected as the feature inputs after dimensionality reduction, thus preserving the main variation information of the data while removing redundant noise variables.

[0035] Based on the reduced feature space described above, sample space clustering analysis is performed to identify geological units with consistent spatial characteristics. To accommodate the non-spherical distribution and uneven sample density of geological data, this invention preferably employs density peak clustering or the DBSCAN algorithm for unsupervised clustering.

[0036] Density peak clustering calculates the local density of each sample and its distance from high-density points, identifying density abrupt changes as cluster centers. DBSCAN, on the other hand, automatically clusters high-density regions into a single class by setting a neighborhood radius ε and a minimum number of points MinPts, capable of identifying clusters of arbitrary shapes. The key parameters of both clustering algorithms are automatically selected using the silhouette coefficient, which measures cluster compactness and separation; a value closer to 1 indicates a more reasonable clustering.

[0037] After clustering, each candidate prediction unit is assigned to a geological isoprotic group, laying the data hierarchical foundation for subsequent factor statistics and modeling within the group.

[0038] Quartz vein density reflects the degree of control exerted by tectonic activity on mineralization during the mineralization period. Its calculation method is as follows: First, extract all quartz vein segments traversing candidate units in three-dimensional space and project their two-dimensional spatial shapes onto a three-dimensional coordinate system. Then, calculate the total length L of all quartz vein segments intersecting the volume of each unit. Finally, divide the total quartz vein length L by the spatial volume V of the unit to obtain the quartz vein density S, expressed as: Quartz Vein Density The unit is meters per cubic meter ( ).

[0039] The alteration index is an important remote sensing mineralization indicator reflecting the intensity of hydrothermal alteration. The extraction process includes: selecting sensitive bands (such as red-edge, near-infrared, and shortwave infrared) from remote sensing satellite data; calculating the ratios of alteration-sensitive band combinations, such as SWIR1 / SWIR2 and red-edge / NIR; performing principal component analysis on these band ratios and extracting the scores of the first two principal components as the composite alteration factor; simultaneously, combining in-situ measured alteration mineral content (such as kaolinite and sericite) data to construct a linear regression model between the remote sensing index and mineral abundance; using this model to estimate the remote sensing abundance of each prediction unit and normalizing it to between 0 and 1, the final value being the alteration index. A higher alteration index indicates stronger hydrothermal alteration of the area, and is an important indicator of potential mineralization zones.

[0040] The geochemical assemblage anomaly weight is used to measure the contribution of synergistic anomalies of multiple elements to mineralization potential. Its calculation process includes: identifying elemental combinations closely related to tungsten mineralization, such as tungsten (W), molybdenum (Mo), arsenic (As), and bismuth (Bi); performing Z-score standardization on the content data of each element in candidate prediction units to measure its deviation from the local background; constructing joint anomaly probability relationships among elements using a Bayesian network model, where the Bayesian network assesses the probability of multiple elements simultaneously exhibiting anomalies by setting prior and conditional probability distributions; and finally, weighting and summing the joint probabilities of synergistic anomalies of each element according to element importance to obtain the comprehensive geochemical assemblage anomaly weight for that unit. This weight can reveal the synergistic response characteristics of multiple elemental anomaly combinations to the mineralization environment.

[0041] To capture the irregular magnetic response characteristics of concealed underground structures, high-frequency magnetic anomaly waveform features were extracted from airborne magnetic data. The steps are as follows: bandpass filtering was applied to the original magnetic anomaly curve data to extract high-frequency components with frequencies higher than a set threshold (e.g., 0.05 Hz); short-time Fourier transform (STFT) was used to perform time-frequency analysis of the high-frequency signals within a sliding window (e.g., a 300-meter window with a step size of 30 meters); the peak energy spectral density, rate of change of dominant frequency (frequency drift rate), and spectral skewness were extracted within each window. These parameters reflect the irregularity of magnetic field disturbances and the variability of anomaly sources; after normalization, all parameters were combined into a magnetic anomaly waveform feature vector, which served as the space geophysical feature term for subsequent model input.

[0042] Since the extracted mineralization indicator factors originate from different geoscientific data (such as structural information, remote sensing information, geochemical indicators, geophysical responses, etc.), their numerical ranges and statistical distributions differ significantly. Directly inputting them into the model would cause an imbalance in feature dominance, affecting the model's robustness. Therefore, multi-scale normalization processing is required for various factors.

[0043] Normalization methods include, but are not limited to, the following three: Max-Min Normalization: Linearly scales each feature to the [0,1] interval; Local Z-score standardization: Standardizes feature values ​​using the mean and standard deviation within a sub-region, preserving local outliers; Quantile normalization: restricts outliers to a distribution range, improving robustness.

[0044] For each type of factor (quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform characteristics), a suitable normalization method is selected to ensure that subsequent processing is carried out at the same scale.

[0045] Each type of normalized mineralization indicator factor is regarded as an independent feature channel. The four types of factors correspond to four channels, which are named as follows: Channel 1: Quartz vein density; Channel 2: Alteration index; Channel 3: Geochemical assemblage anomaly weight; Channel 4: High-frequency magnetic anomaly waveform characteristics. Using the spatial location (e.g., longitude, latitude, and elevation coding) of the candidate prediction units as an index, the above four channels are combined to form a three-dimensional ore-controlling feature tensor. The dimensions of this tensor are defined as follows: First dimension (dimension...) ): Unit ID or spatial index number; Second dimension (dimension) ): Number of feature channels (4 in total); Third dimension (dimension) The feature dimension within each channel is as follows: for example, the density of quartz veins is 1-dimensional, while magnetic anomaly features are multi-dimensional (such as spectral peaks, frequency offsets, and sharpness). The constructed tensor is stored in sparse matrix form and can be input into the graph model as the initial attribute representation of the nodes.

[0046] To model dependencies between spatial units, a graph structure reflecting regional connectivity needs to be constructed. The specific method is as follows: Node definition: Each node in the graph corresponds to a candidate prediction unit; Edge definition: An edge between nodes represents their spatial adjacency relationship or attribute correlation; Construction of Adjacency Matrix A: Generate a symmetric weighted adjacency matrix A, representing the edge weight between any two elements. The weight value is defined as follows: Edge weight = Gaussian kernel function (topological distance + attribute similarity + construction connectivity); specifically expressed as: edge weight_ij = Where: d_ij represents the Euclidean spatial distance between element i and element j; s_ij represents the cosine distance between their mineralization indicator vectors; c_ij indicates whether the two share the same fault system (value is 0 or 1); α, β, and γ are adjustable coefficients that adjust the degree of influence of each factor, which can be determined through model training or cross-validation, and are set by default to α=1.0, β=0.5, and γ=2.0. The map constructed in this way has a non-Euclidean spatial connectivity structure, which can reflect the spatial heterogeneity and mineralization correlation in complex geological environments.

[0047] The ore-controlling feature tensor constructed above is used as the initial feature input of the nodes in the graph neural network, and the adjacency matrix A is used as the input for defining the graph structure to form a graph G = (V, E, X), where: V represents the set of nodes (candidate prediction units); E represents the set of edges (spatial and attribute connections); X represents the node feature tensor (multi-channel ore-forming indicator factor); the graph structure is fed into the graph neural network for information propagation and spatial feature enhancement.

[0048] The Graph Convolutional Network (GCN) is used as the base model, and the propagation mechanism of each layer is as follows: ;in: This represents the feature representation of node i at layer l. This represents the weight between nodes i and j in the adjacency matrix; Let be the trainable weight matrix of the l-th layer; σ be the activation function (such as ReLU); and N(i) represent the set of neighboring nodes of node i. This mechanism realizes the weighted aggregation of features among nodes, and can extract the contextual relationship of candidate units in the spatial structure and their impact on mineralization potential.

[0049] After multi-layer graph convolution propagation, a graph embedding vector is obtained for each candidate prediction unit, which is a spatially enhanced representation that integrates its own features and neighborhood information. This embedding vector has the following characteristics: it contains non-Euclidean spatial relationships; it integrates local tectonic information and long-range mineralization clues; and it can serve as input for a mineralization potential scoring model or as a basis for visual prediction.

[0050] The final output graph embedding vector serves as the spatial context representation of the candidate prediction unit and is input into the subsequent classifier or regression model to generate a deep-edge mineralization probability score map, thereby achieving intelligent prediction of deep-edge tungsten deposits.

[0051] This invention employs a multi-layer perceptron (MLP) as the core mineralization potential scoring model. This model possesses strong nonlinear fitting capabilities and is suitable for fusing multiple high-dimensional mineralization factors and outputting mineralization probabilities.

[0052] The model structure includes: Input layer: dimension equal to the dimension of the embedding vector output by the graph neural network (e.g., 64-dimensional); Hidden layer: 2-3 fully connected hidden layers, with the number of nodes decreasing sequentially in each layer, such as 128, 64, and 32; Activation function: ReLU activation function is used between layers to introduce a non-linear mapping; Output layer: 1 output node is set, using the Sigmoid activation function, with the output range limited to 0 to 1, representing the probability of mineralization in the unit. The model output is: the probability value of mineralization corresponding to each candidate prediction unit; the closer the value is to 1, the higher the probability of mineralization.

[0053] To train the above mineralization potential model, a set of positive and negative samples is needed for supervised learning. The specific method is as follows: Positive Samples (Metallogenic Units): Candidate prediction units are selected from mineralization exposure points of known tungsten deposits or the extent of proven ore bodies. Negative samples (background units): Units without obvious anomalies are selected from areas far from the ore body, with weak tectonic activity and stable geological background as background controls; All samples were labeled: mineralized units were labeled as 1, and background units were labeled as 0.

[0054] To address the issue of imbalanced samples, a class weighting mechanism or undersampling technique is introduced during training to adjust negative samples, thereby improving the model's sensitivity to identifying mineralization categories.

[0055] The model training uses the Binary Cross-Entropy Loss function to measure the error between the predicted probability and the actual label: The loss function is expressed as: Where: N is the total number of samples; The true label (0 or 1) of the i-th sample; This represents the mineralization probability value predicted by the model.

[0056] The model optimization algorithm employs Adam adaptive gradient descent, an optimizer that converges quickly in the early stages of training and is suitable for optimization problems in complex, high-dimensional feature spaces. Key parameters are set as follows: initial learning rate: 0.001; momentum parameters beta1 = 0.9, beta2 = 0.999; maximum learning epochs: 100 epochs; and a validation set is used to monitor overfitting.

[0057] During training, K-Fold Cross Validation can be used to evaluate the robustness of the model, ensuring that the model has good generalization ability in different spatial regions.

[0058] After training, the model is applied to all unlabeled candidate prediction units, taking their graph embedding features as input and outputting the mineralization probability value of each unit, ranging from 0 to 1.

[0059] To facilitate target region identification, this invention sets a probability threshold. = 0.8 (adjustable), units with scores higher than this threshold are considered "high potential units". Further, the following spatial aggregation strategy is used to identify mineralized hotspots: spatial adjacency matrices are used to identify the connectivity between high potential units; continuously distributed high potential units are clustered (e.g., using DBSCAN or spatial connectivity analysis); the boundaries, area, and mean mineralization value of each cluster are output, and clusters that meet the minimum area and average score thresholds are selected as candidate hotspot areas.

[0060] The model prediction results are rasterized and output as mineralization probability maps and heatmaps through a Geographic Information System (GIS) platform. The visualization methods are as follows: Mineralization probability score map: using a continuous color mapping from 0 to 1 to reflect the gradient of mineralization potential; Heatmap: highlighting the outlines of high-potential areas after clustering and filtering, and overlaying a geological base map for spatial verification; Output format: exportable to data formats supported by common exploration platforms such as GeoTIFF and Shapefile, used to guide field prospecting deployment. This heatmap not only provides an intuitive basis for determining target area locations but can also be used to guide geological work such as drilling layout, trenching, and geochemical supplementary surveys, effectively improving the efficiency and accuracy of deep-edge mineral exploration.

[0061] The probability raster map output by the mineralization potential prediction model is spatially overlaid with the distribution points of known tungsten deposits within the region. The prediction result outputs a mineralization probability value for each candidate prediction unit, ranging from 0 to 1, indicating the likelihood that the unit is predicted to be a mineralization area.

[0062] All known tungsten deposit locations (such as the center of the proven ore body and the location of mineralized boreholes) were loaded into the GIS platform and overlaid onto the mineralization probability map. The following indicators were used for evaluation: Coverage index: defined as the proportion of the number of mining sites that fall within the high predicted probability zone (e.g., greater than 0.8) to the total number of mining sites; Mean confidence score: defined as the average of the predicted probability values ​​of all units containing mining sites, used to evaluate the overall sensitivity of the prediction model to known mining sites; Omission identification: Identify mineral deposits that do not fall into high-value areas as potential areas of prediction bias.

[0063] The spatial overlay analysis described above can effectively evaluate the model's response accuracy to known mineralized regions and identify potential missed areas.

[0064] For the actual mineral deposit areas that were not predicted with high confidence by the model in the above-mentioned omission identification, the corresponding candidate prediction units are further identified, and the input mineralization indicator factors (such as quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly characteristics) values ​​of such units are statistically analyzed.

[0065] Simultaneously, by comparing with other regional units that have significantly lower mineralization probability scores, their average factor performance is extracted. Through the statistical differences between the two, abnormal response patterns are identified, namely, feature combinations that "actually contain minerals but have low model scores." These patterns are assigned "abnormal response labels" to construct new training sample sets.

[0066] To explain the prediction bias of the model in the aforementioned anomalous units, this invention introduces a feature importance analysis method to assess the influence of various input factors on the prediction results. One of the following two methods can be used: SHAP (SHapley Additive exPlanations): Based on game theory, each input factor is assigned a "marginal contribution value" to measure its positive or negative impact on the model output; LIME (Local Interpretable Model-agnostic Explanations): A local linear model is constructed within the unit's neighborhood to evaluate the fit between local features and the prediction results.

[0067] In practice, abnormal response units and normal high-scoring units are used as control groups. The average SHAP value or LIME weight of each input factor is calculated and sorted by their absolute value to screen out key influencing factors that cause model prediction bias in abnormal regions. For example, if the SHAP value of the alteration index in an abnormal unit is negative and large, it indicates that the model relies too heavily on this feature and may misjudge under specific geological backgrounds.

[0068] Based on the above feature importance ranking results, the following optimization strategies are adopted for key influencing factors: Normalization adjustment: For features with distribution shifts or extreme value interference (such as geochemical anomaly weights), the original maximum and minimum value normalization is changed to quantile normalization or local Z-score normalization to enhance the ability to perceive local changes; Feature channel weight reallocation: Feature channel weight parameters are introduced into the model input layer or embedding layer, and lower or higher learning weights are assigned to channels with greater influence, and the weights are optimized through gradient descent; Data augmentation and label update: Anomaly response units are included in the training set and labeled with "enhanced supervision" to improve the model's ability to learn such marginal feature patterns; Model fine-tuning training: Based on the original model parameters, the model is fine-tuned for 5 to 10 rounds using a small learning rate (such as 0.0005) to maintain the original learning structure while absorbing anomaly response information.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits, characterized by: include: Acquire multi-source geological data for the target area, including surface, borehole, and tunnel geological survey data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data; A multi-dimensional spatial grid is constructed based on the multi-source geological data to divide the target area into multiple candidate prediction units; Clustering algorithms were used to divide the sample space of candidate prediction units and extract mineralization indicator factors for each unit, including quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform characteristics. The mineralization indicator factors are recalibrated to construct a deep-edge mineralization control feature tensor matrix, and a graph neural network is used to extract the non-Euclidean spatial dependencies between candidate units. A mineralization potential prediction model was constructed, mineralization probability scores were given for each candidate unit, and mineralization hotspot maps of key target areas in the deep edge were output. By spatially overlaying the prediction results with existing tungsten ore distribution points, abnormal response patterns of the mineralization potential prediction model in uncontrolled areas are identified, and input feature weights and model parameters are optimized.

2. The method for predicting mineralization in the deep periphery of tungsten deposits based on machine learning according to claim 1, characterized in that: The target region is divided into multiple candidate prediction units, including: Spatial registration was performed on geological survey data, high-resolution airborne geophysical data, remote sensing spectral data, and geochemical anomaly data. Under the registered unified spatial reference system, the target region is divided into regular grid cells using a gridding algorithm; Based on the data point density, independent variable completeness and information entropy index within each unit, effective units that meet the data quality threshold are selected as candidate prediction units. A unique spatial code is attached to the candidate prediction unit, and a mapping relationship between the candidate unit and the geological attributes is established.

3. The method for predicting mineralization in the deep periphery of tungsten deposits based on machine learning according to claim 1, characterized in that: Clustering algorithms are used to partition the sample space of candidate prediction units, and mineralization indicator factors for each unit are extracted, including: The standardized feature matrix of the candidate prediction unit is first reduced in dimensionality and denoised, and principal component analysis is used to retain the principal components with a cumulative contribution rate of not less than 90%. Based on the principal components, density peak clustering or DBSCAN is used to divide the sample space, and the clustering parameters are automatically selected by the silhouette coefficient to identify geological isoprotic groups. For each subgroup, the quartz vein density is calculated as the sum of the lengths of quartz veins per unit volume, the alteration index is the weighted average of the remote sensing band ratio and the content of altered minerals in the field, and the geochemical assemblage anomaly weight is calculated by Bayesian weighted summation based on the elemental synergistic anomaly probability. High-frequency magnetic anomaly waveform features are extracted, and the energy spectral density is calculated within a sliding window using short-time Fourier transform and normalized to eigenvalues. Each indicator factor forms the final mineralization indicator vector.

4. The machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits according to claim 3, characterized in that: The method for calculating quartz vein density is as follows: First, extract all quartz vein segments that traverse candidate cells in three-dimensional space and project their two-dimensional spatial shapes onto a three-dimensional coordinate system; then, calculate the total length L of all quartz vein segments intersecting with the volume of each cell; finally, divide the total quartz vein length L by the spatial volume V of the cell to obtain the quartz vein density S, expressed as: Quartz Vein Density .

5. The machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits according to claim 3, characterized in that: The calculation process for the geochemical assemblage anomaly weights includes: determining the elemental assemblage closely related to tungsten mineralization; performing Z-score standardization on the content data of each element in the candidate prediction unit; constructing the joint anomaly probability relationship between elements using a Bayesian network model, where the Bayesian network evaluates the probability of multiple elements simultaneously exhibiting anomalies by setting prior and conditional probability distributions; and finally, summing the joint probabilities of the coordinated anomalies of each element according to element importance to obtain the comprehensive geochemical assemblage anomaly weight for that unit.

6. The machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits according to claim 1, characterized in that: Constructing the deep edge ore-controlling feature tensor matrix, including: The mineralization indicator factors extracted from each candidate prediction unit are subjected to multi-scale normalization. The normalized quartz vein density, alteration index, geochemical assemblage anomaly weight, and high-frequency magnetic anomaly waveform features are encoded into four feature channels, and a three-dimensional ore-controlling feature tensor is constructed using the spatial location of candidate units as an index. The tensor dimensions correspond to the unit ID, number of channels, and feature value, respectively. An adjacency matrix between candidate units is constructed using a spatial connectivity graph, and the weights of the edges are generated by fusing three factors: spatial topological distance, fracture connectivity, and geological unit similarity.

7. The machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits according to claim 6, characterized in that: A graph neural network is used to extract non-Euclidean spatial dependencies between candidate units, including: Based on the spatial location, structural connectivity, and geological attribute similarity of candidate prediction units, a graph structure is constructed to form a weighted undirected graph with nodes as units and edges as spatial associations. An adjacency matrix A is generated from this graph, where the edge weights are fused with topological distance and attribute similarity using a Gaussian kernel function. The control feature tensor of each candidate unit is used as the initial feature of the input node of the graph neural network. The graph convolutional neural network is used to perform feature propagation and update the node embedding representation according to the hierarchical propagation rule. Each convolutional layer achieves context information aggregation through feature weighted averaging. The graph embedding vector of each node in the final output is used as a spatial dependency enhancement feature for deep edge candidate prediction units.

8. The method for predicting mineralization in the deep periphery of tungsten deposits based on machine learning according to claim 1, characterized in that: A mineralization potential prediction model is constructed, and the mineralization probability of each candidate unit is scored, including: The candidate unit embedding vector processed by the graph neural network is used as the input feature to construct a multilayer perceptron mineralization potential scoring model with a softmax output layer, and outputs the mineralization probability value of each unit. The model was trained under supervision using the cross-entropy loss function. The training samples were derived from the labeled data of known mining sites and background points. The optimization method was the Adam algorithm, and the learning rate was set to 0.

001. Based on the mineralization probability value, a confidence threshold is set to divide key prediction units, and high-value cluster areas are generated by aggregating spatially adjacent high-potential units. The probability distribution map is rasterized to output the mineralization hot spot map of the key target area in the deep edge.

9. The machine learning-based method for predicting mineralization in the deep periphery of tungsten deposits according to claim 8, characterized in that: Spatial overlay analysis was performed with existing tungsten deposit distribution points to identify anomalous response patterns of the mineralization potential prediction model in uncontrolled areas, including: The predicted mineralization probability map is spatially overlaid with the known tungsten deposit distribution points in the region, and the coverage rate and confidence index of each actual mineral point falling into the high predicted probability area are calculated. Identify candidate prediction units corresponding to the areas where the mineral deposits are not covered, analyze the differences in mineralization indicator factors between them and areas with low scores in the mineralization potential prediction model, and construct anomaly response pattern labels; Based on the anomaly response unit, the feature importance analysis method is used to evaluate the degree of influence of various input factors on prediction bias; Adjust the normalization method or weight ratio for the feature channels with significant impact, and retrain the mineralization potential model.

Images