A method for predicting a concealed rare metal ore body

Abstract

The present application relates to the technical field of mineral exploration, and particularly relates to a rare metal concealed ore body prediction method, comprising collecting multi-modal geoscience data; determining a weak mineralization anomaly area based on a comparison result of a sampling point singularity index and a preset singularity threshold; determining a favorable space range of mineralization in the weak mineralization anomaly area based on a cross gradient function; adjusting the preset singularity threshold under the condition that the favorable space range of mineralization is unqualified based on a residual spectrum similarity; adjusting a weight coefficient of the cross gradient function when an element combination anomaly is unqualified; constructing a generative adversarial network model and expanding a training data set; constructing a three-dimensional convolutional neural network discriminant model, optimizing the three-dimensional convolutional neural network discriminant model based on a mineralization suspected area, and outputting a three-dimensional space occurrence probability distribution of a concealed ore body. The present application solves the problems of insufficient utilization of multi-source geoscience information and difficulty in meeting the precise positioning requirements of deep concealed ore bodies in the spatial structure of geophysical and geochemical exploration.

Description

Technical Field

[0001] This invention relates to the field of mineral exploration technology, and in particular to a method for predicting concealed rare metal ore bodies. Background Technology

[0002] my country's demand for key metallic mineral resources continues to increase, while easily discoverable surface deposits are decreasing, making deep and concealed deposits the focus of mineral exploration. As exploration depth increases, surface geochemical signals are weakened and interference factors increase, making conventional anomaly extraction methods difficult to distinguish weak mineralization from differences in geological background, easily overlooking weak anomalies caused by concealed ore bodies. In recent years, nonlinear methods such as fractals, multifractals, and local singularities have been gradually applied to geochemical data processing, effectively identifying local enrichment and structural changes in element spatial distribution. However, most methods still remain at a single scale and single data level, lacking sufficient quantitative characterization, spatial constraints, and automated mineralization prediction capabilities for weak anomalies. This makes it difficult to meet the high-precision, quantitative exploration needs of deep concealed deposits, and single-scale analysis cannot reflect the multi-scale aggregation characteristics of element spatial distribution, resulting in poor anomaly identification accuracy.

[0003] Chinese Patent Application Publication No. CN117951551A discloses a mineral resource prediction method based on singularity index and self-organizing neural network, including the following steps: obtaining the spatial distribution of catchment basins in the study area; statistically analyzing the singularity index of each catchment basin; extracting geological ore-controlling elements in the study area; obtaining the mineralization favorability of each geological ore-controlling element to the catchment basin; classifying the catchment basins using a self-organizing neural network; and identifying mineralization favorable areas in the study area. This invention, based on singularity index and self-organizing neural network, addresses the difficulty in identifying mineralization information for key metallic mineral resources due to its weak characteristics. Based on the sampling principle of stream sediments, it uses sampling units as the basic units for geochemical anomaly extraction and introduces geological ore-controlling factors within the catchment basin units as constraints to identify geochemical anomalies, providing a reference for subsequent mineral exploration.

[0004] The existing technology still has the following problems: it is impossible to achieve coordinated constraints on the spatial structure of geophysical and geochemical exploration by using only single geochemical data; multi-source geoscience information is not fully utilized; the evaluation accuracy of element combination anomalies is limited by using self-organizing neural networks for classification; and it can only achieve two-dimensional planar prediction, which is difficult to meet the needs of accurate positioning of deep concealed ore bodies. Summary of the Invention

[0005] To address these issues, this invention provides a method for predicting concealed rare metal ore bodies. This method overcomes the limitations of existing technologies, which rely solely on single geochemical data, fail to achieve coordinated constraints between geophysical and geochemical spatial structures, lack sufficient utilization of multi-source geoscience information, employ self-organizing neural networks for classification, have limited accuracy in evaluating elemental combination anomalies, and only achieve two-dimensional planar prediction, thus failing to meet the requirements for precise location of deep concealed ore bodies.

[0006] To achieve the above objectives, the present invention provides a method for predicting concealed rare metal ore bodies, comprising: Collect multimodal geoscientific data using known metallogenic areas and simulated metallogenic conditions; Based on the singularity index of each sampling point calculated by local singularity analysis of the geochemical element concentration data, and the comparison result with the preset singularity threshold, weak mineralization anomaly areas are determined. Centered on the weak mineralization anomaly region, a cross gradient function of the geophysical field and the geochemical field is constructed, and the favorable spatial range for mineralization is determined by minimizing the cross gradient function. Based on the similarity between the geophysical inversion residual spectrum of the favorable mineralization spatial range and the preset residual spectrum, it is determined whether the favorable mineralization spatial range is qualified, and the preset singularity threshold is adjusted if it is not qualified. Extract the multifractal spectrum parameters of geochemical elements under the condition that the favorable spatial range for mineralization is qualified, in order to determine whether the element combination is abnormal, and adjust the weight coefficients of the first gradient vector and the second gradient vector under abnormal conditions. Based on the multifractal spectrum parameters and the corresponding geophysical data, a generative adversarial network model is used to generate several sets of simulated mineralization samples to expand the training dataset. A three-dimensional convolutional neural network discrimination model is constructed based on the expanded training dataset. Based on the spatial occurrence probability of concealed ore bodies obtained from the input multifractal spectrum parameters and geophysical data, mineralization suspected areas are identified, and the three-dimensional convolutional neural network discrimination model is optimized based on the information of mineralization suspected areas.

[0007] Furthermore, the process of determining weakly mineralized anomaly regions based on the singularity index of sampling points includes: Multiple local square sliding windows of various scales were established with each geochemical sampling point as the center. For each sliding window, the average concentration of geochemical elements within the window range is calculated to fit the power-law relationship between the average concentration of geochemical elements and the scale of the sliding window; The exponent of the power-law relationship is determined as the singularity index of a single sampling point; The singularity index is compared with a preset singularity threshold; Based on the comparison results of the singularity index being less than the preset singularity threshold, the corresponding grid cell is determined to be a weakly mineralized anomaly region.

[0008] Furthermore, the process of determining the favorable spatial range for mineralization based on the cross gradient function includes: Obtain the first gradient vector and the second gradient vector; The magnitude of the cross product of the first gradient vector and the second gradient vector is calculated as the cross gradient function; Using a single weakly mineralized anomaly region as the initial seed point, the cross gradient function of adjacent weakly mineralized anomaly regions is calculated sequentially. Compare the obtained cross gradient function values ​​with the preset gradient threshold; Based on the comparison results where the cross gradient function value is less than the preset gradient threshold, the covered spatial range is determined to be the favorable spatial range for mineralization.

[0009] Furthermore, the process of determining whether the favorable spatial range for mineralization is qualified based on the geophysical inversion residual spectrum includes: Perform a discrete Fourier transform on the geophysical inversion residual spectrum to obtain the residual spectrum; Calculate the cosine similarity between the residual spectrum and the preset residual spectrum; The cosine similarity is compared with a preset similarity threshold; Based on the comparison results where the cosine similarity is less than a preset similarity threshold, it is determined that the favorable spatial range for mineralization is unqualified.

[0010] Furthermore, under the condition that the favorable space range for mineralization is unqualified, the process of adjusting the preset singularity threshold includes: The difference between the preset singularity threshold and the preset step size is determined as the adjusted preset singularity threshold.

[0011] Furthermore, the process of determining whether the elemental combination within the favorable spatial range for mineralization is abnormal based on multifractal spectrum parameters includes: The mass indices of all rare metal ore-forming elements and associated index elements under different moments are calculated within the qualified favorable ore-forming spatial range to obtain the singularity index spectral function. The three core parameters of the singularity index spectral function—spectral width, asymmetry coefficient, and peak position—are extracted and used as multifractal spectral parameters. The multifractal spectrum parameters are input into a pre-trained support vector machine classifier, and the classifier outputs whether the element combination is abnormal. Under the condition that the vector machine determines it to be abnormal, the combination of elements within the favorable space for mineralization is determined to be abnormal.

[0012] Furthermore, under abnormal element combination conditions, the process of adjusting the first and second weighting coefficients is as follows: The sum of the second weight coefficient and the preset weight adjustment step size is determined as the adjusted second weight coefficient; At the same time, the difference between the first weight coefficient and the preset weight adjustment step size is determined as the adjusted first weight coefficient.

[0013] Furthermore, the process of expanding the training dataset by generating an adversarial network model based on the multifractal spectrum parameters and the corresponding geophysical data includes: Multiple fractal spectrum parameters and corresponding three-dimensional geophysical data within the favorable spatial range of mineralization are selected and combined to form real mineralization sample pairs; Build a generative adversarial network model consisting of a generator and a discriminator, and set the basic structural parameters of the model; The discriminator output probability is monitored in real time, and training is stopped when the termination condition is met. The trained generator is invoked to generate simulated samples, qualified simulated samples are selected, and then merged and shuffled with the original real samples to expand the training dataset.

[0014] Furthermore, the process of determining suspected mineralization areas based on the spatial occurrence probability of concealed ore bodies includes: Extract the multifractal spectrum parameters and corresponding three-dimensional geophysical field data of the area to be tested, and perform normalization and dimension alignment on the data; The processed test data is input into the trained three-dimensional convolutional neural network discrimination model to obtain the spatial occurrence probability of the concealed ore body in the test area. The spatial occurrence probability of the concealed ore body is compared with a first probability threshold and a second probability threshold; Based on the comparison result that the spatial occurrence probability of the concealed ore body is greater than the first probability threshold and less than the second probability threshold, the area to be tested is determined to be a suspected mineralization area.

[0015] Furthermore, the rare metal includes at least one selected from lithium, beryllium, niobium, tantalum, rubidium, and cesium; Associated indicator elements include at least one of tin, tungsten, and fluorine; The geochemical element concentration data includes the concentrations of ore-forming elements and associated elements; The geophysical field data includes at least one of magnetic susceptibility data, resistivity data, or density data; The remote sensing alteration data includes hyperspectral mineral mapping data. The entire method is only applicable to the prediction and exploration of concealed ore bodies of deep rare metals, and is not applicable to the mineralization prediction of precious metals and ferrous metals.

[0016] Compared with the prior art, the beneficial effects of the present invention are that by integrating geochemical, geophysical and remote sensing alteration multimodal geoscientific data, and introducing local singularity analysis, cross-gradient spatial structure synergistic constraints, geophysical inversion residual spectrum adaptive adjustment, multifractal spectrum parameter and support vector machine joint evaluation, generative adversarial network data expansion and three-dimensional convolutional neural network probability prediction, the present invention effectively overcomes the shortcomings of insufficient utilization of single geochemical data information and the inability of two-dimensional planar prediction to meet the needs of accurate positioning of deep concealed ore bodies.

[0017] Furthermore, this invention performs local singularity analysis on geochemical element concentration data, calculates the singularity index of each sampling point and compares it with a preset threshold to determine weak mineralization anomaly regions. Then, using these regions as seed points, a spatial search is performed along the direction of decreasing gradient at the intersection of the geophysical field and the geochemical field. This achieves synergistic constraints and improves the accuracy and reliability of delineating favorable mineralization spatial ranges.

[0018] Furthermore, this invention performs geophysical three-dimensional inversion within the favorable spatial range of mineralization and calculates the geophysical inversion residual spectrum. After obtaining the residual spectrum through Fourier transform, it compares the cosine similarity with the preset residual spectrum. If it is unqualified, the singularity threshold is automatically reduced by a preset step size and the search is restarted until the residual spectrum meets the requirements. This realizes the adaptive adjustment of the preset singularity threshold, improving the accuracy of the qualified determination of the favorable spatial range of mineralization and the stability of the prediction results.

[0019] Furthermore, this invention extracts multifractal spectrum parameters within a qualified favorable mineralization space, uses a support vector machine classifier to determine whether element combination anomalies are qualified, and increases the weight coefficient of the geochemical gradient vector in the cross gradient function and redetermines the favorable mineralization range when unqualified. At the same time, a generative adversarial network is constructed using qualified multifractal spectrum parameters and corresponding geophysical data as real samples. After the generator and discriminator reach Nash equilibrium through a game, the training dataset is expanded, solving the problem of insufficient training of machine learning models and improving the accuracy of element combination anomaly evaluation and the generalization ability of the model.

[0020] Furthermore, this invention constructs a three-dimensional convolutional neural network discrimination model based on the expanded training dataset, outputting the three-dimensional spatial probability distribution of concealed ore bodies. By conducting targeted geochemical exploration of suspected mineralized areas, the verification results are fed back as new samples to the generative adversarial network for incremental learning, achieving a leap from two-dimensional planar prediction to precise three-dimensional spatial positioning. This improves the prediction efficiency of deep concealed ore bodies, the reliability of exploration target area selection, and the model's continuous learning ability. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for predicting concealed rare metal ore bodies according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of determining weakly mineralized anomaly regions in an embodiment of the present invention. Figure 3 A flowchart illustrating the determination of favorable spatial range for mineralization in an embodiment of the present invention; Figure 4 This is a flowchart for determining whether the favorable space range for mineralization is qualified in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] The present invention will now be described in a clear and detailed manner with reference to the accompanying drawings.

[0025] Please see Figures 1-4 As shown, Figure 1 This is a flowchart of a method for predicting concealed rare metal ore bodies according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of determining weakly mineralized anomaly regions in an embodiment of the present invention. Figure 3 A flowchart illustrating the determination of favorable spatial range for mineralization in an embodiment of the present invention; Figure 4 This is a flowchart for determining whether the favorable space range for mineralization is qualified in an embodiment of the present invention.

[0026] The method for predicting concealed rare metal ore bodies according to embodiments of the present invention includes: Step S1: Collect multimodal geoscientific data based on known ore-forming areas and simulated ore-forming conditions. The multimodal geoscientific data includes geochemical element concentration data, geophysical field data, and remote sensing alteration data. Step S2: Based on the comparison between the singularity index of each sampling point calculated by local singularity analysis of the geochemical element concentration data and the preset singularity threshold, the weak mineralization anomaly area is determined. Step S3: Using the weak mineralization anomaly region as the center, construct the cross gradient function of the geophysical field and the geochemical field, and determine the favorable spatial range for mineralization by minimizing the cross gradient function. Step S4: Based on the similarity between the geophysical inversion residual spectrum of the favorable mineralization spatial range and the preset residual spectrum, determine whether the favorable mineralization spatial range is qualified; if it is not qualified, adjust the preset singularity threshold. Step S5: Extract the multifractal spectrum parameters of geochemical elements under the condition that the favorable spatial range for mineralization is qualified, in order to determine whether the element combination is abnormal, and adjust the weight coefficients of the first gradient vector and the second gradient vector under abnormal conditions. Step S6: Generate an adversarial network model based on the multifractal spectrum parameters and the corresponding geophysical data, and generate several sets of simulated mineralization samples through the generative adversarial network model to expand the training dataset; Step S7: Construct a three-dimensional convolutional neural network discrimination model based on the expanded training dataset. Determine the spatial occurrence probability of concealed ore bodies based on the input multifractal spectrum parameters and geophysical data. Optimize the three-dimensional convolutional neural network discrimination model based on the information of the suspected mineralization areas.

[0027] Specifically, by setting up a series of processes—including multimodal data acquisition from known ore-forming areas to simulated ore-forming conditions, identification of weak mineralization anomalies guided by local singularity indices, search for favorable ore-forming ranges constrained by cross-gradient functions, adjustment of singularity thresholds through geophysical inversion residual spectrum similarity feedback, adjustment of weight coefficients through multifractal spectrum parameters and support vector machine feedback, expansion of training samples through generative adversarial networks, and output probability distribution of three-dimensional convolutional neural networks—the system achieves coordinated constraints on the spatial structure of geophysical and geochemical exploration and adaptive fusion of multi-source data. This improves the positioning accuracy of deep concealed ore bodies and reduces the cost of manual trial and error through automated threshold and weight adjustment, thereby enhancing exploration efficiency and target area selection reliability under complex geological conditions.

[0028] Specifically, known metallogenic areas refer to regions where industrial ore bodies have been confirmed to exist, and where there is a relatively in-depth understanding of their metallogenic geological characteristics, ore-controlling factors, and prospecting indicators; Simulated mineralization conditions refer to the mineralization physicochemical environment artificially created or calculated in the laboratory or numerical model based on mineralization theory, as well as favorable mineralization geological environments inferred from known mineralization models in geological exploration that have not yet been fully revealed. These conditions are used to compensate for the lack of field data through laboratory simulation, numerical simulation, and geological analogy simulation. Simultaneously collecting data from both sides allows for verification of whether the prospecting model established in the known area is applicable to the simulated area, thereby improving prospecting potential and accuracy.

[0029] Specifically, based on known metallogenic areas and simulated metallogenic conditions, the methods for acquiring multimodal geoscientific data include: All data acquisition is conducted using square windows with a side length of 100m, and various types of data are collected based on these windows. Based on geochemical element concentration data, soil, rock, and stream sediment samples were collected in individual grid cells of known and simulated metallogenic areas, and the corresponding coordinates and geological background of each sample were recorded. The coordinates of each geochemical sample were used as geochemical sampling points. The concentrations of rare metal ore-forming elements and associated index elements were tested to obtain the spatial distribution and combination anomaly characteristics of elements. To obtain magnetic susceptibility parameters from geophysical field data through high-precision ground-based magnetic measurements, a profile network was established, and the total intensity of the geomagnetic field was collected point by point using a proton magnetometer. Simultaneously, resistivity sounding was conducted, and the resistivity and polarizability of the medium at different depths were obtained using a symmetrical quadrupole device. Bouguer gravity anomalies were measured at grid nodes using a high-precision gravimeter, and the changes in physical properties such as density, magnetism, and conductivity of underground rocks were inverted to form geophysical field data. A conjugate gradient inversion algorithm was used to perform three-dimensional physical property inversion, supplemented by the least squares method for data fitting and correction. For remote sensing alteration data, the Gaofen-5 hyperspectral satellite imagery was mainly used. After radiometric calibration, atmospheric correction, and geometric fine correction preprocessing, characteristic bands related to hydroxyl alteration and iron staining alteration were selected to construct a principal component analysis model. The third principal component was fixedly selected to carry out principal component analysis to accurately extract surface hydroxyl alteration and iron staining alteration information. Among them, the third principal component can reflect the spectral characteristics of hydroxyl and iron staining altered minerals to the greatest extent and effectively distinguish altered and non-altered areas.

[0030] Specifically, by setting up the synchronous acquisition of multimodal geoscientific data of known metallogenic areas and simulated metallogenic conditions, and using a unified gridded sampling scheme for both known and simulated metallogenic areas, the ability to verify the applicability of the prospecting model in the known area to the simulated area was generated. This effectively made up for the shortcomings of a single data source in the field, improved the prospecting potential and accuracy of deep concealed ore bodies, and enhanced the spatial comparability and fusion efficiency of multi-source data.

[0031] Specifically, the process of determining weakly mineralized anomaly regions based on the singularity index of sampling points includes: Multiple local square sliding windows of various scales were established with each geochemical sampling point as the center. For each sliding window, the average concentration of geochemical elements within the window range is calculated to fit the power-law relationship between the average concentration of geochemical elements and the scale of the sliding window; The exponent of the power-law relationship is determined as the singularity index of a single sampling point; The singularity index is compared with a preset singularity threshold; Based on the comparison results where the singularity index is less than a preset singularity threshold, the corresponding grid cell is determined to be a weakly mineralized anomaly region. Based on the comparison results where the singularity index is greater than or equal to the preset singularity threshold, the corresponding grid cell is determined to be a non-weakly mineralized anomaly region.

[0032] Specifically, square sliding windows with different side lengths are established centered on geochemical sampling points. The side lengths of the square sliding windows are 100m, 500m, 1000m, and 2000m, respectively, to adapt to rare metal exploration scales of 1:50,000 to 1:200,000, covering a multi-scale analysis range from local enrichment to regional background. For sampling points near the boundary of the study area, the sliding window only retains the part within the study area and does not extend outward to avoid invalid calculations. If the number of sampling points in a single sliding window is less than 3, it is judged as a fitting failure. Failed sampling points are no longer directly included in the singularity index calculation. Instead, they are supplemented by assigning values ​​using the Kriging interpolation method with the surrounding effective sampling points before being included in the weak mineralization anomaly area judgment process.

[0033] Specifically, the expression for the power-law relationship is:

[0034] Where C(r) is the average element concentration corresponding to scale r, r is the sliding window side length, k is the proportionality constant, and a is the power law exponent; During fitting, the window size r and the concentration mean C(r) are respectively taken as natural logarithms to convert them into a linear relationship.

[0035] The least squares method is used to perform linear fitting on the transformed data to minimize the sum of squared residuals from all data points to the fitted line. The coefficient of determination is limited to no less than 0.85. Sampling points with a coefficient of determination lower than this value are directly marked as invalid sampling points. The absolute value of the slope of the fitted line is the singularity index of the current sampling point.

[0036] Specifically, the process of calculating the average geochemical element concentration within each sliding window includes: finding all collected geochemical sampling points within a single sliding window; for the same rare metal ore-forming element, calculating the arithmetic mean of the element concentration values ​​of all sampling points within the sliding window, where the arithmetic mean represents the average concentration level of the element in the neighborhood centered on the sampling point at the window scale; and for the same center point, calculating the corresponding element concentration average under all square sliding windows with different side lengths in sequence.

[0037] Specifically, the preset singularity threshold is set to [1.5, 2.5]. The preset singularity threshold is set based on the power-law distribution characteristics of the geochemical background field of rare metals. For study areas with uneven distribution of geological background field, geological units are divided based on regional stratigraphic distribution maps, lithological zoning maps, and tectonic unit division maps. The distribution characteristics of geochemical background field within each independent geological unit are statistically analyzed, and the corresponding preset singularity threshold is calculated and set independently. Among them, lithium ore is preferred at 1.8, beryllium ore at 2.0, and niobium-tantalum ore at 2.2. For other rare metals such as rubidium and cesium, a uniform preferred threshold of 2.0 is adopted. For study areas with uneven geological background field, a regional dynamic singularity threshold is adopted, and the threshold is calculated independently for different geological units.

[0038] Specifically, when the window is close to the boundary of the study area, only the actual sample points within the window are used to calculate the average concentration, without forcibly filling in zeros or extrapolating outwards, to ensure that the calculation results are consistent with the actual geological conditions.

[0039] Specifically, by setting up multiple sliding windows of different scales centered on each sampling point, the arithmetic mean of the geochemical element concentrations within each window is calculated. The natural logarithm of the window scale and the average concentration is then used for least squares linear fitting. This index is compared with a preset singularity threshold, effectively distinguishing between a uniform background field of elements and local enrichment anomalies. This improves the accuracy and geological rationality of identifying weak mineralization anomaly areas. At the same time, multi-scale automatic analysis reduces the subjectivity of manual window selection and improves the efficiency of local singularity analysis.

[0040] Specifically, the process of determining the favorable spatial range for mineralization based on the cross gradient function includes: Obtain the first gradient vector and the second gradient vector; The magnitude of the cross product of the first gradient vector and the second gradient vector is calculated as the cross gradient function; Using a single weakly mineralized anomaly region as the initial seed point, the cross gradient function of adjacent weakly mineralized anomaly regions is calculated sequentially. Compare the obtained cross gradient function values ​​with the preset gradient threshold; Based on the comparison results of cross gradient function values ​​being less than a preset gradient threshold, the covered spatial range is determined to be the favorable spatial range for mineralization. Based on the comparison results of cross gradient function values ​​being greater than or equal to preset gradient thresholds, the covered spatial range is determined to be a non-mineralized favorable spatial range.

[0041] Specifically, gradient vector calculation uniformly adopts a 50m grid spacing as the basic calculation scale.

[0042] Specifically, the cross gradient function is used to quantitatively characterize the consistency of the spatial variation directions of the geophysical field and the geochemical field. The smaller the function value, the stronger the spatial structural synergy between geophysical anomalies and geochemical anomalies, and the higher the favorable degree of mineralization.

[0043] Specifically, the first gradient vector is obtained by taking the preprocessed and gridded geophysical field data as a continuous spatial scalar field and uniformly using the central difference method to calculate the spatial gradient vector. When calculating the cross gradient function, since the resistivity anomaly in the rare metal mineralization area is the most significant, the resistivity field gradient is taken as the first gradient vector for subsequent calculations. The first gradient vector points to the spatial direction in which the resistivity value of the underground medium gradually increases. The second gradient vector is obtained by treating the interpolated gridded geochemical element concentration field as a continuous spatial scalar field and calculating the spatial gradient vector using the central difference method. This second gradient vector points in the direction of increasing element concentration. The interpolation uses the Kriging interpolation method. When the interpolated gradient value exceeds the range of [0,10], it is determined to be a distorted gradient and is discarded. The gradient vector calculation is only performed within the effective sampling point coverage range of ±500m. If the gradient exceeds the range, it is set to zero.

[0044] Specifically, the central difference method is calculated as follows: for the east-west gradient component, the concentration of the neighboring node to the right of the node is subtracted from the concentration of the neighboring node to the left, and then divided by twice the grid spacing. For the north-south gradient component, it is obtained by subtracting the concentration of the adjacent node below from the concentration of the node above it, and then dividing by twice the grid spacing. When the central difference calculation cannot be used for the boundary grid nodes of the study area, forward difference or backward difference calculation is used as an alternative to ensure that the gradient vector can be solved in the entire region.

[0045] Specifically, the process of calculating the cross gradient function is as follows:

[0046] in, This is the first weighting coefficient of the geophysical field gradient. This is the second weighting coefficient of the geochemical field gradient. Let be the magnitude of the first gradient vector of the resistivity field gradient vector. Let be the magnitude of the second gradient vector of the element concentration field gradient vector. The spatial angle between the two types of gradient vectors. The sine of the included angle is given.

[0047] Specifically, the first and second weighting coefficients are both set to 0.5. By using equal weights to utilize both geophysical and geochemical field data to the same extent, we can avoid artificially favoring one type of information and obtain a more objective and comprehensive cross-gradient field, thereby improving the accuracy of delineating favorable mineralization areas.

[0048] Specifically, the process of spatially searching from the initial seed point until the cross gradient function value is less than the preset gradient threshold includes: using all weak mineralization anomaly regions as initial seed points to form a seed point set, ensuring that the corresponding cross gradient function value has been calculated for each seed point; calculating the cross gradient function values ​​of the 26 regions adjacent to the seed point as the center; comparing the cross gradient function values ​​of adjacent nodes with those of the seed point; selecting adjacent nodes whose cross gradient function values ​​are less than those of the current seed point and including them in the search range as new seed points; repeating the above selection process, starting from the newly included seed point, continuing to search for adjacent nodes of the newly included seed point, selecting the node with the smallest cross gradient, until the cross gradient function values ​​of all newly included grid nodes are less than the preset gradient threshold, ensuring that the search covers all areas, stopping the search; for weak mineralization anomaly regions without adjacent valid grids, the search is determined to be invalid, and the region is directly marked as a non-mineralization favorable area.

[0049] Specifically, the preset gradient threshold is set to 0.2. Multiple sets of comparative tests were conducted on known industrial ore body sections, weak mineralization anomaly sections, and mineral-free background sections in typical rare metal mineralization areas to determine the threshold. Thirty known favorable mineralization sections and 30 mineral-free background sections were selected in the area. The cross gradient function values ​​of the geophysical field and the geochemical field were calculated respectively. The results showed that the cross gradient function values ​​of the favorable mineralization sections were concentrated in the range of 0-0.2, while those of the mineral-free background sections were concentrated in the range above 0.2. The preset gradient threshold of 0.2 was determined by the average value of multiple tests at the boundary between the two types of intervals. This effectively distinguishes between favorable mineralization space and geological background space, ensuring accurate and reliable delineation of the scope, effectively avoiding the inclusion of the mineral-free background grid in the search range, and completely covering the favorable mineralization area extending from the weak mineralization anomaly.

[0050] Specifically, the search process should follow the principle that when encountering the boundary of the study area, only the grid nodes within the study area should be searched, without extending outwards. If adjacent nodes have already been included in the search range, they should not be searched again.

[0051] Specifically, by processing the geochemical field and the geophysical field into a second gradient vector and a first gradient vector, respectively, a weighted cross gradient function is calculated. Using the weak mineralization anomaly region as a seed point, the cross gradient function values ​​of adjacent regions are continuously compared with a preset gradient threshold. Regions with function values ​​less than the preset gradient threshold are selected, thereby achieving synergistic constraints on the spatial variation direction of geophysical anomalies and geochemical anomalies, and improving the accuracy of delineating the favorable spatial range for mineralization.

[0052] Specifically, the process of determining whether the favorable spatial range for mineralization is qualified based on the geophysical inversion residual spectrum includes: Perform a discrete Fourier transform on the geophysical inversion residual spectrum to obtain the residual spectrum; Calculate the cosine similarity between the residual spectrum and the preset residual spectrum; The cosine similarity is compared with a preset similarity threshold; Based on the comparison results where the cosine similarity is greater than or equal to a preset similarity threshold, the favorable spatial range for mineralization is determined to be qualified. Based on the comparison results where the cosine similarity is less than a preset similarity threshold, it is determined that the favorable spatial range for mineralization is unqualified.

[0053] Specifically, the process of obtaining the geophysical inversion residual spectrum is as follows: Within the favorable spatial range for mineralization, based on the geological background, lithology, and structural constraints of the study area, the conjugate gradient method is used to perform three-dimensional inversion on the gridded geophysical field data. The applicable exploration depth range is from the surface to a depth of 1500m underground. The inversion operation uses regional lithological distribution, fault structures, and stratigraphic attitude as hard constraints. The maximum number of inversion iterations is set to 50. If the iteration limit is reached and convergence is not achieved, the inversion operation is terminated. A subsurface physical property model is constructed, and the theoretical geophysical response corresponding one-to-one with the measured grid nodes is obtained through forward modeling. The measured geophysical data of the same grid node is subtracted from the theoretical geophysical response to obtain the residual value. The residual values ​​of all grid nodes are arranged in an orderly manner according to three-dimensional spatial coordinates to form a spatial distribution dataset of residuals, which is the geophysical inversion residual spectrum. This is existing technology and will not be described in detail here.

[0054] Specifically, when the geophysical inversion residual spectrum is a stationary spatial signal without local distortion, a discrete Fourier transform is performed. If the geophysical inversion residual spatial signal exhibits non-stationary characteristics, a Hanning window is used for signal smoothing preprocessing before performing a discrete Fourier transform. When the residual spectrum vector is inconsistent with the preset residual spectrum vector dimension length, low-frequency effective components are uniformly retained, and high-frequency redundant components are truncated to achieve vector dimension alignment. If both sets of spectrum vectors are all zero vectors, the similarity check is directly determined to be invalid, and the corresponding favorable spatial range for mineralization is unqualified. The singularity threshold is forcibly lowered according to the unqualified process.

[0055] Specifically, the cosine similarity between the residual spectrum and the preset residual spectrum can be directly implemented using conventional software such as MATLAB, Python, Surfer, and Oasis Montaj, and will not be elaborated further.

[0056] Specifically, the preset similarity threshold is set at 0.75. This was determined through multiple sets of repeated experiments on known industrial ore body sections, weakly mineralized anomaly sections, and non-mineralized background sections within typical rare metal mineralization areas. Thirty known favorable mineralization sections and 30 non-mineralized background sections within the area were selected, and the cosine similarity between their geophysical inversion residual spectra and standard mineralization spectra was calculated. The statistical results show that the average similarity of favorable mineralization sections is concentrated in the range of 0.75-1.0, while the average similarity of non-mineralized background sections is concentrated in the range of 0-0.75. The preset similarity threshold of 0.75 was determined by the average value of multiple experiments at the boundary between the two ranges. This can accurately distinguish between qualified favorable mineralization spaces and non-mineralized background spaces, ensuring that the geophysical response within the favorable mineralization space range is significantly different from the background field while tolerating reasonable local geological heterogeneity.

[0057] Specifically, by performing three-dimensional inversion of geophysical data within the favorable mineralization spatial range based on the comprehensive geological background, theoretical responses are obtained and residuals between theoretical responses and measured data are calculated. The residuals are then arranged in an orderly manner according to the grid node coordinates to form a spatial distribution spectrum, thereby realizing the quantification and visualization of geophysical fitting errors and improving the objectivity and spatial positioning accuracy of the qualification inspection of favorable mineralization spatial range.

[0058] Specifically, under the condition that the favorable space range for mineralization is unsuitable, the process of adjusting the preset singularity threshold includes: The difference between the preset singularity threshold and the preset step size is determined as the adjusted preset singularity threshold, and the adjusted preset singularity threshold is not less than 1.5.

[0059] Specifically, a preset step size of 0.1 was used. Multiple iterative adjustment experiments were conducted on geochemical data from typical rare metal mineralization areas, using step sizes of 0.05, 0.1, 0.15, and 0.2 respectively. The singularity threshold adjustment, weak mineralization anomaly delineation, favorable mineralization space search, and residual spectrum similarity verification processes were repeatedly performed, and the iterative convergence speed and delineation accuracy were statistically analyzed. Experimental results show that a step size of 0.1 achieves the optimal balance between convergence speed and adjustment accuracy, quickly approaching the qualified threshold range with the fewest iterations without causing parameter jumps or result distortion. Therefore, the optimal balance value from multiple experiments was determined to be the preset step size of 0.1.

[0060] Specifically, by performing a Fourier transform on the geophysical inversion residual spectrum to obtain the residual spectrum, and calculating the cosine similarity between this spectrum and the preset residual spectrum obtained based on the statistics of known mineralized areas, the cosine similarity is compared with the preset similarity threshold. If the result is unqualified, the preset singularity threshold is reduced and the favorable mineralization range is searched again. This improves the automation and convergence efficiency of the qualification determination of the favorable mineralization spatial range, and reduces the time cost of repeated manual trial and error.

[0061] Specifically, the process of determining whether the elemental combination within the favorable spatial range for mineralization is abnormal based on multifractal spectrum parameters includes: The mass indices of all rare metal ore-forming elements and associated index elements under different moments are calculated within the qualified favorable ore-forming spatial range to obtain the singularity index spectral function. The spectral width, asymmetry coefficient, and peak position extracted from the singularity index spectral function are combined to form a set of multifractal spectral parameters. The multifractal spectrum parameters are input into a pre-trained support vector machine classifier, and the classifier outputs whether the element combination is abnormal. Under the condition that the vector machine determines it to be abnormal, it is determined that the element combination within the favorable mineralization space is abnormal; Under the condition that the vector machine determines that there are no anomalies, it is determined that there are no anomalies in the element combination within the favorable space range for mineralization.

[0062] Specifically, the spectral width is the difference between the maximum and minimum singularity index within the effective distribution interval of the singularity index spectral function; a larger value indicates a more significant difference in the aggregation and dispersion distribution of elements. The asymmetry coefficient is calculated by dividing the spectral curve by the peak position and separately calculating the span of the left and right intervals. The characteristic coefficient obtained by the ratio of the left and right interval spans characterizes the offset shape of the multifractal spectral curve. The peak position refers to the singularity index value corresponding to when the singularity index spectral function curve reaches its maximum value.

[0063] Specifically, the order moment takes values ​​in the range of [-5, 5], and the step size of the order moment is fixed at 0.5. The mass index is obtained by Legendre transformation to obtain the singularity index spectrum function. The calculation of the mass index, Legendre transformation, and construction of the singularity index spectrum function can be directly implemented by general software such as MATLAB, Python, GS+, and Geosoft, and will not be elaborated further.

[0064] Specifically, multifractal spectrum parameters of known ore-forming sections and non-ore-forming background sections are selected to form positive and negative training samples, respectively. After normalizing the samples, radial basis kernel functions are used for supervised learning training, and the model parameters are optimized to obtain a pre-trained classifier.

[0065] Specifically, under abnormal element combination conditions, the process of adjusting the first and second weighting coefficients is as follows: The sum of the second weight coefficient and the preset weight adjustment step size is determined as the adjusted second weight coefficient; At the same time, the difference between the first weight coefficient and the preset weight adjustment step size is determined as the adjusted first weight coefficient.

[0066] Specifically, the preset weight adjustment step size is 0.05, and the values ​​of the first and second weight coefficients are both in the range of [0.2, 0.8]. The adjustment of the first and second weight coefficients must not exceed this range. This was determined through multiple sets of weight adjustment comparative experiments conducted on typical rare metal mineralization areas. The adjustment step sizes were 0.02, 0.05, 0.10, and 0.15, respectively. The cross-gradient weight adjustment, mineralization favorable spatial range search, and element combination anomaly judgment process were repeated to statistically analyze the adjustment stability and convergence efficiency. The experimental results show that a step size of 0.05 can achieve the optimal balance between adjustment accuracy and convergence speed, and can steadily increase the contribution ratio of the geochemical field without causing parameter mutations or spatial search distortion.

[0067] Specifically, by setting up a method to extract multifractal spectrum parameters from a qualified favorable mineralization space and inputting these parameters into a pre-trained support vector machine (SVM) classifier, the SVM classifier automatically compares the input parameters with pre-learned mineralization anomaly feature standards. Based on the parameter distribution characteristics, it judges the qualification of element combination anomalies. If the parameters are not qualified, it increases the weight coefficient of the geochemical gradient vector in the cross gradient function and simultaneously decreases the weight coefficient of the geophysical gradient vector, thereby redetermining the favorable mineralization space. This improves the accuracy of element combination anomaly evaluation and the efficiency of mineralization favorable range correction, while reducing the workload of repeated manual parameter tuning.

[0068] Specifically, the process of expanding the training dataset by generating adversarial network models based on the multifractal spectrum parameters and corresponding geophysical data includes: Multiple fractal spectrum parameters and corresponding three-dimensional geophysical data within the favorable spatial range of mineralization are selected and combined to form real mineralization sample pairs; Build a generative adversarial network model consisting of a generator and a discriminator, and set the basic structural parameters of the model; The discriminator output probability is monitored in real time, and training is stopped when the termination condition is met. The trained generator is invoked to generate simulated samples, qualified simulated samples are selected, and then merged and shuffled with the original real samples to expand the training dataset.

[0069] Specifically, the termination condition is that the probability stabilizes between 0.48 and 0.52 and the fluctuation is less than 0.01 for 10 consecutive rounds.

[0070] Specifically, in this embodiment of the invention, a generative adversarial network model consisting of a generator and a discriminator is constructed. The basic structural parameters of the model are set as follows: the number of neurons in the fully connected layer of the generator is 128, 64, and 32 respectively; the kernel size of the convolutional layer is 3×3 and the stride is 1; the kernel size of the convolutional layer of the discriminator is 3×3 and the stride is 2; and the pooling layer adopts max pooling.

[0071] Specifically, in this embodiment of the invention, the generator network adopts a combination structure of three fully connected layers and two convolutional layers. The input is a random noise vector, the activation function is the ReLU function, and the output is a simulated multifractal spectrum parameter and geophysical data pair consistent with the format and dimension of the real sample. The discriminator network adopts a convolutional neural network structure of three convolutional layers, two pooling layers, and one fully connected layer. The activation function is the LeakyReLU function, where the negative slope is 0.01. The input is sample data, and the output is the probability value between [0, 1]. The overall network uses the Adam optimizer to complete parameter optimization. The optimizer learning rate is set to 0.0002, the decay coefficient is set to 0.9, and the iteration batch is set to 32 to ensure that the model can be trained stably. Further details are omitted.

[0072] Specifically, after calling the trained generator to output simulated mineralization samples, samples with a relative error of less than 8% in the multifractal spectrum parameters are selected as qualified simulated samples. After the selection is completed, the samples are mixed and shuffled in a ratio of 1:3 between real mineralization samples and simulated generated samples to form an expanded standardized training dataset.

[0073] Specifically, the alternating training process of the model is as follows: In each round of training, the parameters of the generator network are fixed to maximize the discriminator's accuracy in distinguishing between real samples and generated samples, and the parameters of the discriminator network are updated; the parameters of the discriminator network are fixed to minimize the discriminator's ability to recognize generated samples, and the parameters of the generator network are updated.

[0074] Specifically, by setting qualified multifractal spectrum parameters and corresponding geophysical data as real samples, a generator network combining fully connected layers and convolutional layers and a discriminator network with a convolutional neural network structure are constructed. The two are trained alternately until the discriminator can no longer distinguish between real samples and generated samples. Then, the qualified simulated samples generated by the generator are merged with the original real samples to form an expanded training dataset. This achieves high-quality expansion of training data under the condition of limited real samples, improves the training efficiency and generalization ability of subsequent 3D convolutional neural network models, and avoids the overfitting problem caused by small samples.

[0075] Specifically, the process of constructing a three-dimensional convolutional neural network discriminant model in this embodiment of the invention includes building the basic structure of the three-dimensional convolutional neural network discriminant model, including an input layer, three convolutional layers, two pooling layers, one fully connected layer, and an output layer; dividing the expanded training dataset into a training set and a validation set in a 7:3 ratio, setting the training batch size to 32 and the learning rate to 0.001, monitoring the model validation accuracy in real time, and stopping training when the validation accuracy is stable above 90% and there is no significant improvement after five consecutive iterations, thus completing the construction of the three-dimensional convolutional neural network discriminant model; In the convolutional layers, the kernel size is 3×3×3, and the number of neurons is 64, 128, and 256 respectively, with ReLU activation function; max pooling is used in the pooling layers, with a pooling window of 2×2×2; the number of neurons in the fully connected layers is 64, with ReLU activation function; and the activation function of the output layer is Sigmoid.

[0076] Specifically, the process of determining suspected mineralization areas based on the spatial occurrence probability of concealed ore bodies includes: Extract the multifractal spectrum parameters and corresponding three-dimensional geophysical field data of the area to be tested, and perform normalization and dimension alignment on the data; The processed test data is input into the trained three-dimensional convolutional neural network discrimination model to obtain the spatial occurrence probability of the concealed ore body in the test area. The spatial occurrence probability of the concealed ore body is compared with a first probability threshold and a second probability threshold; Based on the comparison result that the spatial occurrence probability of the concealed ore body is greater than the first probability threshold and less than the second probability threshold, the area to be tested is determined to be a suspected mineralization area. Based on the comparison results that the spatial occurrence probability of the concealed ore body is less than or equal to the first probability threshold, the area to be tested is determined to be a mineral-free background area. Based on the comparison results of the spatial occurrence probability of the concealed ore body being greater than or equal to the second probability threshold, the area to be tested is determined to be a favorable area for mineralization.

[0077] Specifically, the first probability threshold is set at 0.3, and the second probability threshold is set at 0.7. This value is applicable to 1:50,000 scale rare metal mineral exploration. During 1:50,000 scale exploration, numerous comparative experiments were conducted using measured samples from multiple mineralization groups in the study area and the three-dimensional occurrence probability distribution results. The experiments statistically analyzed the occurrence probability distribution patterns of ore bodies in known mineral-free blank areas, areas with weak mineralization potential, and areas rich in industrial ore bodies. After averaging multiple sets of experimental data, it was found that areas with probability values ​​below 0.3 showed almost no mineralization anomaly response. Areas with a stable geological background and a probability value between 0.3 and 0.7 possess weak mineralization exploration potential and are potential exploration targets. Areas with a probability value higher than 0.7 closely match the spatial location of known concealed ore bodies, indicating extremely high reliability in mineralization. When conducting 1:200,000 scale regional mineralization prospect prediction, the above experimental procedure was followed. The results show that adjusting the first probability threshold to 0.25 and the second probability threshold to 0.75 results in clear grading boundaries and stable classification accuracy, which can meet the actual exploration needs for grading and delineating concealed rare metal ore bodies.

[0078] Specifically, the process of optimizing the three-dimensional convolutional neural network discrimination model based on the actual geological and mineralization information of suspected mineralization areas includes: Targeted geochemical exploration is carried out in the suspected mineralization area to obtain the actual geological and mineralization information of the area. The verification results are used as new real samples and fed back to the generative adversarial network model for incremental learning to update the model parameters. Based on the generative adversarial network model after incremental learning, the training dataset is expanded again, and the three-dimensional convolutional neural network discrimination model is retrained using the dataset. Then, the retrained model is used to re-predict the suspected mineralization area. Repeat the above process until the termination condition is met.

[0079] Specifically, the iteration termination condition was set as the area of ​​suspected mineralization areas being less than 10%. According to the statistical analysis of multi-regional field measurement iterations, when the area of ​​suspected areas is higher than 10%, the ambiguity range of regional mineralization determination is too large, the model discrimination bias is obvious, and the exploration direction is weak. When the proportion drops to 10% or below, the model's classification accuracy for non-mineralized background areas and high-probability mineralized areas has stabilized, and the classification error is controlled within the allowable range of exploration. The remaining small number of suspected areas do not affect the overall mineral exploration target area delineation and exploration deployment. Continuing iterative optimization has low returns and will increase the computation and field survey costs. Therefore, the termination condition was determined to be an area proportion of less than 10%.

[0080] Specifically, by setting up suspected mineralization zones and conducting targeted geochemical exploration in these zones, the verification results are fed back to the generative adversarial network as new real samples for incremental learning until the termination condition is met. This enables proactive sampling and optimization of uncertain areas in the model's predictions, improving the continuous learning ability and prediction reliability of the 3D convolutional neural network model, and effectively reducing exploration risks and cost waste caused by high-confidence misjudgments.

[0081] Specifically, when the favorable mineralization space is determined to be unqualified, the singularity threshold is lowered and the weak mineralization anomaly area is redefined; when the element combination is determined to be abnormal and the weight coefficient is adjusted, the cross gradient function is recalculated and a new favorable mineralization space is defined until all verification indicators meet the preset qualified standards.

[0082] Specifically, the rare metal includes at least one of lithium, beryllium, niobium, tantalum, rubidium, and cesium; Associated indicator elements include at least one of tin, tungsten, and fluorine; The geochemical element concentration data includes the concentrations of ore-forming elements and associated elements; The geophysical field data includes at least one of magnetic susceptibility data, resistivity data, or density data; The remote sensing alteration data includes hyperspectral mineral mapping data. The entire method is only applicable to the prediction and exploration of concealed ore bodies of deep rare metals, and is not applicable to the mineralization prediction of precious metals and ferrous metals.

[0083] Specifically, it enables targeted acquisition and multimodal data fusion of key indicator elements of rare metal mineralization systems, improving the accuracy of subsequent mineralization prediction and data utilization efficiency, and avoiding interference from irrelevant elements and redundant processing of non-targeted geophysical and remote sensing information.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting concealed rare metal ore bodies, characterized in that, include: Collect multimodal geoscientific data using known metallogenic areas and simulated metallogenic conditions; Based on the singularity index of each sampling point calculated by local singularity analysis of the geochemical element concentration data, and the comparison result with the preset singularity threshold, weak mineralization anomaly areas are determined. Centered on the weak mineralization anomaly region, a cross gradient function of the geophysical field and the geochemical field is constructed, and the favorable spatial range for mineralization is determined by minimizing the cross gradient function. Based on the similarity between the geophysical inversion residual spectrum of the favorable mineralization spatial range and the preset residual spectrum, it is determined whether the favorable mineralization spatial range is qualified, and the preset singularity threshold is adjusted if it is not qualified. Extract the multifractal spectrum parameters of geochemical elements under the condition that the favorable spatial range for mineralization is qualified, in order to determine whether the element combination is abnormal, and adjust the weight coefficients of the first gradient vector and the second gradient vector under abnormal conditions. Based on the multifractal spectrum parameters and the corresponding geophysical data, a generative adversarial network model is used to generate several sets of simulated mineralization samples to expand the training dataset. A three-dimensional convolutional neural network discrimination model is constructed based on the expanded training dataset. Based on the spatial occurrence probability of concealed ore bodies obtained from the input multifractal spectrum parameters and geophysical data, mineralization suspected areas are identified, and the three-dimensional convolutional neural network discrimination model is optimized based on the information of mineralization suspected areas.

2. The method for predicting concealed rare metal ore bodies according to claim 1, characterized in that, The process of determining weakly mineralized anomaly regions based on the singularity index of sampling points includes: Multiple local square sliding windows of various scales were established with each geochemical sampling point as the center. For each sliding window, the average concentration of geochemical elements within the window range is calculated to fit the power-law relationship between the average concentration of geochemical elements and the scale of the sliding window; The exponent of the power-law relationship is determined as the singularity index of a single sampling point; The singularity index is compared with a preset singularity threshold; Based on the comparison results of the singularity index being less than the preset singularity threshold, the corresponding grid cell is determined to be a weakly mineralized anomaly region.

3. The method for predicting concealed rare metal ore bodies according to claim 2, characterized in that, The process of determining the favorable spatial range for mineralization based on the cross gradient function includes: Obtain the first gradient vector and the second gradient vector; The magnitude of the cross product of the first gradient vector and the second gradient vector is calculated as the cross gradient function; Using a single weakly mineralized anomaly region as the initial seed point, the cross gradient function of adjacent weakly mineralized anomaly regions is calculated sequentially. Compare the obtained cross gradient function values ​​with the preset gradient threshold; Based on the comparison results where the cross gradient function value is less than the preset gradient threshold, the covered spatial range is determined to be the favorable spatial range for mineralization.

4. The method for predicting concealed rare metal ore bodies according to claim 3, characterized in that, The process of determining whether a favorable spatial range for mineralization is qualified based on the geophysical inversion residual spectrum includes: Perform a discrete Fourier transform on the geophysical inversion residual spectrum to obtain the residual spectrum; Calculate the cosine similarity between the residual spectrum and the preset residual spectrum; The cosine similarity is compared with a preset similarity threshold; Based on the comparison results where the cosine similarity is less than a preset similarity threshold, it is determined that the favorable spatial range for mineralization is unqualified.

5. The method for predicting concealed rare metal ore bodies according to claim 4, characterized in that, When the favorable mineralization space range is unsuitable, the process of adjusting the preset singularity threshold includes: The difference between the preset singularity threshold and the preset step size is determined as the adjusted preset singularity threshold.

6. The method for predicting concealed rare metal ore bodies according to claim 5, characterized in that, The process of determining whether the elemental combination within a favorable mineralization spatial range is abnormal based on multifractal spectral parameters includes: The mass indices of all rare metal ore-forming elements and associated index elements under different moments are calculated within the qualified favorable ore-forming spatial range to obtain the singularity index spectral function. The three core parameters of the singularity index spectral function—spectral width, asymmetry coefficient, and peak position—are extracted and used as multifractal spectral parameters. The multifractal spectrum parameters are input into a pre-trained support vector machine classifier, and the classifier outputs whether the element combination is abnormal. Under the condition that the vector machine determines it to be abnormal, the combination of elements within the favorable space for mineralization is determined to be abnormal.

7. The method for predicting concealed rare metal ore bodies according to claim 6, characterized in that, Under abnormal element combination conditions, the process of adjusting the first and second weighting coefficients is as follows: The sum of the second weight coefficient and the preset weight adjustment step size is determined as the adjusted second weight coefficient; At the same time, the difference between the first weight coefficient and the preset weight adjustment step size is determined as the adjusted first weight coefficient.

8. The method for predicting concealed rare metal ore bodies according to claim 7, characterized in that, The process of expanding the training dataset by using the multifractal spectrum parameters and corresponding geophysical data to generate adversarial network models includes: Multiple fractal spectrum parameters and corresponding three-dimensional geophysical data within the favorable spatial range of mineralization are selected and combined to form real mineralization sample pairs; Build a generative adversarial network model consisting of a generator and a discriminator, and set the basic structural parameters of the model; The discriminator output probability is monitored in real time, and training is stopped when the termination condition is met. The trained generator is invoked to generate simulated samples, qualified simulated samples are selected, and then merged and shuffled with the original real samples to expand the training dataset.

9. The method for predicting concealed rare metal ore bodies according to claim 8, characterized in that, The process of determining suspected mineralization areas based on the spatial occurrence probability of concealed ore bodies includes: Extract the multifractal spectrum parameters and corresponding three-dimensional geophysical field data of the area to be tested, and perform normalization and dimension alignment on the data; The processed test data is input into the trained three-dimensional convolutional neural network discrimination model to obtain the spatial occurrence probability of the concealed ore body in the test area. The spatial occurrence probability of the concealed ore body is compared with a first probability threshold and a second probability threshold; Based on the comparison result that the spatial occurrence probability of the concealed ore body is greater than the first probability threshold and less than the second probability threshold, the area to be tested is determined to be a suspected mineralization area.

10. The method for predicting concealed rare metal ore bodies according to claim 9, characterized in that, The rare metals include at least one of lithium, beryllium, niobium, tantalum, rubidium, and cesium; Associated indicator elements include at least one of tin, tungsten, and fluorine; The geochemical element concentration data includes the concentrations of ore-forming elements and associated elements; The geophysical field data includes at least one of magnetic susceptibility data, resistivity data, or density data; Remote sensing alteration data includes hyperspectral mineral mapping data. The entire method is only applicable to the prediction and exploration of concealed ore bodies of deep rare metals, and is not applicable to the mineralization prediction of precious metals and ferrous metals.

Images