Modeling and ore body positioning method, system and device of gravity-magnetic-electric multi-physical field coupling

Abstract

The application provides a heavy magnetic electric multi-physical field coupling modeling and ore body positioning method, system and device, and relates to the technical field of physical exploration modeling. In the method, observation data, historical measurement data and physical geological information measured in a target detection area are obtained; an initial three-dimensional space model is divided into grids to obtain a plurality of hexahedral grids, and initial physical property parameters are initialized and configured for each hexahedral grid; the preprocessed observation data is input into a target deep learning model for multi-modal feature extraction and fusion inversion, and predicted physical property parameters are output; the initial physical property parameters are adjusted by using the predicted physical property parameters, and a three-dimensional coupled physical property model is output; the three-dimensional coupled physical property model is subjected to multi-scale segmentation and stratum calibration to obtain a three-dimensional geological structure model; and the three-dimensional geological structure model is subjected to ore-bearing evaluation to delineate the spatial position of a regular ore body, so that the spatial position of the ore body is accurately delineated.

Description

Technical Field

[0001] This application relates to the field of physical exploration modeling technology, specifically to a method, system, and equipment for modeling and ore body location using multi-physics coupling of gravity, magnetism, and electricity. Background Technology

[0002] With the increasing depletion of surface and shallow mineral resources, deep mineral exploration has become a major trend in global mineral resource exploration. Geophysical exploration is the core technology for probing deep geological structures and mineral resources. Gravity exploration infers the distribution of geological bodies by measuring changes in the gravity field caused by differences in rock density; magnetic exploration detects magnetic geological bodies based on differences in rock magnetic properties; and electrical exploration uses differences in rock electrical properties (such as resistivity and polarizability) to identify mineralization zones or structures. Under complex geological conditions at depth, a single geophysical method is insufficient to accurately characterize ore-controlling structures and delineate the spatial location of ore bodies. Developing multi-method joint inversion and interpretation techniques to achieve complementary advantages of data from different physical fields is a key technological development direction for improving the success rate of deep mineral exploration, and it has enormous economic value and strategic significance.

[0003] In related technologies, conventional dual-field joint inversion is usually adopted, which uses gravity and magnetic, gravity and electric, or magnetoelectric data for simple combination. For the coupling of two physical fields, it relies on the initial model and prior information. When constructing the initial model, it often relies on subjective experience or simple uniform background assumptions. When performing inversion based on the constructed model, most methods use linear or quasi-linear iterative optimization, which is difficult to adaptively overcome the heterogeneous differences in the physical dimensions of gravity, magnetic and electric methods. It has limited ability to characterize complex geological bodies (strong heterogeneity, blurred boundaries of anomalous bodies, etc.), and it is difficult to obtain high-resolution three-dimensional physical property models. Therefore, it is impossible to clearly and accurately delineate the location of ore bodies to support precise drilling deployment. Summary of the Invention

[0004] This application provides a method, system, and equipment for modeling and ore body location using multi-physics field coupling of gravity, magnetism, and electricity. This method enables three-dimensional modeling of deep geological structures and accurate spatial delineation of ore bodies, providing a basis for physical exploration and drilling deployment.

[0005] The technical solution of this application embodiment is as follows: In a first aspect, embodiments of this application provide a method for modeling and ore body localization using multiphysics coupling of gravity, magnetism, and electricity, the method comprising: Acquire observational data of the target detection area, historical measurement data of the target detection area, and physical geological information, wherein the observational data includes gravity data, magnetic data, and electrical data, and preprocess the gravity data, magnetic data, and electrical data to obtain corresponding gravity data volumes, magnetic data volumes, and electrical data volumes; Based on the spatial range and depth of the target detection area, the preset initial three-dimensional spatial model is divided into multiple hexahedral grids. The historical measurement data and the physical geological information are used to initialize and configure the initial physical property parameters, including density, magnetic susceptibility and resistivity, for each hexahedral grid. The gravity data volume, the magnetic data volume, and the electrical data volume are input into a preset target deep learning model for multimodal feature extraction and fusion inversion. The model outputs predicted physical property parameters including density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. The initial physical property parameters are adjusted using the predicted physical property parameters to output a three-dimensional coupled physical property model. The target deep learning model is used to spatially and consistently couple the gravity data volume, the magnetic data volume, and the electrical data volume. The three-dimensional coupled physical property model is segmented at multiple scales and its stratigraphic calibration is performed to obtain a three-dimensional geological structure model; Based on a preset set of physical property index rules, the three-dimensional geological structure model is evaluated for mineralization to delineate the spatial location of ore bodies that conform to the rules.

[0006] In the aforementioned technical solution, multi-source heterogeneous data is formed by acquiring gravity data, magnetic data, electrical data, historical measurement data, and physical geological information, and preprocessing the data to support subsequent multimodal data processing. Hexahedral fine-grid partitioning is used to achieve a digital representation of complex underground spaces, providing a stable initial model foundation that fits the regional geological reality for subsequent multiphysics coupled inversion. Coupled inversion using a target deep learning model effectively suppresses inversion ambiguities under complex deep geological conditions, improving the extraction accuracy of deep stratigraphic interfaces, fault structures, and mineralization anomalies, achieving consistency in three-dimensional data spatial coupling. Predicted physical property parameters are adjusted against initial physical property parameters to obtain a high-resolution three-dimensional physical property model, providing a highly reliable digital foundation for subsequent geological semantic modeling and accurate orebody identification. Based on the three-dimensional coupled physical property model, multi-scale segmentation and stratigraphic calibration are performed to construct a three-dimensional geological structural model. This three-dimensional coupled physical property model is then converted into a three-dimensional geological structural model entity, and combined with a set of physical property index rules, the location of orebodies is clearly and accurately delineated, supporting precise drilling deployment. By using multi-rule coupling constraints, mineral-induced anomalies can be effectively distinguished from non-mineral water-bearing and rock mass anomalies, significantly improving the success rate and efficiency of deep mineral exploration and providing reliable data support for the precise deployment of drilling projects.

[0007] Secondly, embodiments of this application provide a modeling and ore body location system based on the coupling of gravity, magnetism, and electricity in multiple physics fields, the system comprising: The data acquisition module is used to acquire observation data of the target detection area, historical measurement data of the target detection area, and physical geological information. The observation data includes gravity data, magnetic data, and electrical data. The module preprocesses the gravity data, magnetic data, and electrical data to obtain corresponding gravity data volumes, magnetic data volumes, and electrical data volumes. The mesh generation module is used to divide the preset initial three-dimensional spatial model into multiple hexahedral meshes according to the spatial range and detection depth of the target detection area. The module uses the historical measurement data and the physical geological information to initialize and configure the initial physical property parameters, including density, magnetic susceptibility and resistivity, for each hexahedral mesh. The inversion prediction module is used to input the gravity data volume, the magnetic data volume, and the electrical data volume into a preset target deep learning model for multimodal feature extraction and fusion inversion, and output predicted physical property parameters including density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. The initial physical property parameters are adjusted using the predicted physical property parameters, and a three-dimensional coupled physical property model is output. The target deep learning model is used to perform spatially consistent coupling of the gravity data volume, the magnetic data volume, and the electrical data volume. The model construction module is used to perform multi-scale segmentation and stratigraphic calibration on the three-dimensional coupled physical property model to obtain a three-dimensional geological structure model. The spatial delineation module is used to assess the mineralization of the three-dimensional geological structure model according to a preset set of physical property index rules, so as to delineate the spatial location of ore bodies that conform to the rules.

[0008] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, a user interface, a communication bus, and a network interface. The processor, the memory, the user interface, and the network interface are respectively connected to the communication bus. The memory is used to store instructions. The user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as provided in any one of the first aspects.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing instructions that, when executed, perform the method described in any one of the methods provided in the first aspect above.

[0010] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By acquiring multi-source heterogeneous data from gravity, magnetic, electrical, historical measurement, and physical geological information, and performing data preprocessing, support is provided for subsequent multimodal data processing. Furthermore, hexahedral fine-grid partitioning enables a digital representation of complex underground spaces, providing a stable initial model foundation that aligns with regional geological realities for subsequent multiphysics coupled inversion. Coupled inversion using a target deep learning model effectively suppresses inversion ambiguities under complex deep geological conditions, improving the extraction accuracy of deep stratigraphic interfaces, fault structures, and mineralization anomalies, achieving consistency in 3D data spatial coupling. Predicted physical property parameters are adjusted against initial physical property parameters to obtain a high-resolution 3D physical property model, providing a highly reliable digital foundation for subsequent geological semantic modeling and accurate orebody identification. Based on the 3D coupled physical property model, multi-scale segmentation and stratigraphic calibration are performed to construct a 3D geological structural model. This 3D coupled physical property model is then converted into a 3D geological structural model entity, and combined with a set of physical property index rules, the location of orebodies is clearly and accurately delineated, supporting precise drilling deployment. By employing multi-rule coupling constraints, this method effectively distinguishes between mineralized anomalies and non-mineralized water-bearing and rock mass anomalies, significantly improving the success rate and efficiency of deep mineral exploration and providing reliable data support for precise deployment of drilling projects. Therefore, it effectively solves the problem in related technologies of limited ability to characterize complex geological bodies, difficulty in obtaining high-resolution three-dimensional physical property models, and thus the inability to clearly and accurately delineate the location of ore bodies.

[0011] 2. Under the different resolution levels of the multi-scale physical property spatial pyramid, the coarse scale ensures that the overall geological framework is not distorted, while the fine scale accurately depicts the fine mineralization boundaries and alteration structures. The progressive refinement method greatly improves the fit and authenticity of the geological boundaries, so that the segmented blocks completely fit the actual underground physical property distribution and geological structure.

[0012] 3. The implicit feature alignment network solves the problem of spatial misalignment of deep anomalies caused by different physical dimensions of multi-source heterogeneous data such as gravity, magnetic field and electrical methods, and realizes the substantial coupling of "mathematics and physics". Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a method for modeling and ore body location using multiphysics coupling of gravity, magnetism, and electricity, provided in one embodiment of this application. Figure 2 This is a schematic diagram of the target deep learning model of the modeling and ore body location method of multiphysics coupling of gravity, magnetoelectricity and gravity provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a multi-physics coupling modeling and ore body positioning system provided in one embodiment of this application; Figure 4This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0015] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0016] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0017] This application provides a method, system, electronic device, and readable storage medium for modeling and ore body location using multiphysics coupling of gravity, magnetics, and electricity. The method first acquires gravity data, magnetic data, electrical data, historical measurement data, and physical geological information to form multi-source heterogeneous data, and performs data preprocessing to support subsequent multimodal data processing. Then, through hexahedral fine-grid subdivision, it achieves a digital representation of complex underground space, providing a stable initial model base that fits the regional geological reality for subsequent multiphysics coupling inversion. Using a target deep learning model for coupling inversion effectively suppresses inversion ambiguities under complex deep geological conditions, improving the extraction accuracy of deep stratigraphic interfaces, fault structures, and mineralization anomalies, achieving consistency in three-dimensional data spatial coupling. Finally, it adjusts the initial physical property parameters with predicted physical property parameters to obtain a high-resolution three-dimensional physical property model, providing a highly reliable digital base for subsequent geological semantic modeling and accurate ore body identification. Then, based on the 3D coupled physical property model, multi-scale segmentation and stratigraphic calibration are performed to construct a 3D geological structure model. The 3D coupled physical property model is converted into a 3D geological structure model entity, and combined with the physical property index rule set, the location of the ore body is clearly and accurately delineated, supporting precise drilling deployment. Through multi-rule coupling constraints, mineralized anomalies are effectively distinguished from non-mineralized water-bearing and rock mass anomalies, significantly improving the success rate and exploration efficiency of deep mineral exploration, and providing reliable data support for the precise deployment of drilling projects.

[0018] It should be noted that this modeling and ore body location method based on the coupling of gravity, magnetism, and electricity is mainly applicable to scenarios such as deep prospecting prediction of metallic minerals (such as vanadium-titanium magnetite, copper, gold, etc.), deep structural analysis of metallogenic belts, location of concealed ore bodies, and evaluation of resource potential in the deep and peripheral parts of mines.

[0019] The technical solutions provided in the embodiments of this application will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating the gravity, magnetoelectric, and multiphysics field coupling modeling and ore body location method provided in this application embodiment. The gravity, magnetoelectric, and multiphysics field coupling modeling and ore body location method is applied to a gravity, magnetoelectric, and multiphysics field coupling modeling and ore body location system. The method is executed by a processor in an electronic device or a readable storage medium. The gravity, magnetoelectric, and multiphysics field coupling modeling and ore body location method includes steps S100, S200, S300, S400, and S500.

[0021] Step S100: Obtain observation data of the target detection area, historical measurement data of the target detection area, and physical geological information. The observation data includes gravity data, magnetic data, and electrical data. The gravity data, magnetic data, and electrical data are preprocessed to obtain the corresponding gravity data volume, magnetic data volume, and electrical data volume.

[0022] In one embodiment, the target exploration area refers to a designated work area for deep mineral exploration and geological structure analysis, which can be defined according to exploration needs. The exploration depth is no less than 2000m, covering target areas such as metallogenic belts and deep edges of mines. Observational data refers to the observational data obtained from measurements in the target exploration area, including gravity data, magnetic data, and electrical data. Electrical data includes resistivity, polarizability, etc., and carries raw information such as the coordinates, elevation, and frequency of the observation points. Historical measurement data refers to the geophysical measurement data and engineering measurement data accumulated from past explorations in the target exploration area, including historical gravity, magnetic, and electrical observation data, tunnel measurement data, etc., used to constrain the initial parameters of the model and improve the model's fit. Physical geological information refers to the basic geological data of the target exploration area, including geological maps, structural outline maps, borehole core data, rock property test data, and spatial distribution data of strata / rock masses / fractures, used to construct geological prior constraint rules. The data volume is a three-dimensional gridded data matrix with unified coordinates, matched precision, and regular format after standardized preprocessing. It is a standardized data format that is suitable for the input of deep learning models.

[0023] A reconnaissance survey was conducted in the target detection area, and a unified design for the reconnaissance and survey network was completed. A unified coordinate system for gravity, magnetic, and electrical resistivity survey networks was established, and field data acquisition was carried out simultaneously. Magnetic data acquisition was synchronized with diurnal variation monitoring, gravity data acquisition was synchronized with elevation measurements, and electrical resistivity data acquisition avoided human-induced electromagnetic interference. The coordinates, elevations, and geophysical parameters of each observation point were fully recorded to obtain the raw observation dataset. Historical measurement data, borehole core physical property data, geological structure vector data, and statistical data on upper and lower limits of rock physical properties in the target area were comprehensively collected to establish a regional geological prior knowledge database.

[0024] The raw observation data, namely gravity data, magnetic data, and electrical data, are acquired using the same coordinate system. Then, the gravity data, magnetic data, and electrical data are preprocessed. The specific preprocessing process is as follows: The three types of raw observation data were corrected step by step, sequentially performing diurnal variation correction, height correction, and topographic correction to eliminate background interference from non-target geological bodies. Diurnal variation correction: For magnetic and electrical data, the diurnal variation field value at the base station at the same moment was subtracted from the flow observation values ​​at each measuring point based on the change in field value over time recorded synchronously by diurnal variation base stations (or reference stations) set up in or near the survey area. For gravity data, equivalent solid tide correction and instrument zero-point drift correction were implemented. Height correction: The observation values ​​at each measuring point were normalized to a unified reference elevation surface. For example, gravity height correction used the formula g = 0.3086 × h, where h is the height difference of the measuring point relative to the reference surface; magnetic data underwent vertical gradient correction based on the height difference of the measuring points and the normal geomagnetic gravitational gradient. Topographic correction: Based on a high-precision digital elevation model of the survey area, the influence of topographic undulations around the measuring points (i.e., redundant mass or missing mass) on the distortion of the observed field values ​​was eliminated. The field response of the topographic undulations in different radius areas around the central measuring point to the central point is calculated and subtracted from the observation data to obtain the gravity anomaly, magnetic anomaly, and topographically corrected apparent resistivity / phase parameters.

[0025] During field data acquisition, human interference or sudden instrument malfunctions may lead to a few isolated distortion anomalies. A spatial moving window is used to remove these anomalies. A spatial neighborhood window with radius R is set centered on the current measurement point. The median and median difference of all similar physical data within the window are calculated, and an approximate standard deviation is calculated. If the absolute value of the difference between the observed value and the median at the current measurement point is greater than the approximate standard deviation, it is identified as a distortion anomaly and removed. The mean of the normal data within the neighborhood window is then calculated and used to replace the removed anomaly. A low-pass filtering algorithm is used to suppress high-frequency noise, preserving effective deep-seated anomalous signals.

[0026] Due to the varying resolutions, uneven distribution, and non-overlapping grid points in actual field measurements of gravity, magnetic, and electrical methods, it is impossible to directly fuse the data spatially. This paper addresses the issues of uneven distribution and non-overlapping grid points by using a Kriging interpolation algorithm to achieve data gridding. Specifically: a unified target grid region boundary and spatial resolution are set; for each type of geophysical data, its experimental semivariogram function is calculated, and a Gaussian model is used to fit the theoretical semivariogram function to characterize the spatial correlation of the physical field; based on the theoretical semivariogram function, the Kriging weight coefficient sequence is solved with the objective of minimizing the variance of the interpolation point estimation; the solved weight coefficients are then used to perform a weighted summation of the known discrete measurement point data, calculating the values ​​at all target regular grid points, thereby obtaining gravity, magnetic, and electrical grid data with completely consistent spatial coordinate alignment. The aligned gravity, magnetic, and electrical grid data were Z-score normalized, mapping the values ​​to the [-1, 1] interval. The normalized 2D gravity, magnetic, and electrical grid maps were treated as multiple "channels" of an image and stacked and recombined in the depth dimension to construct a 3D data tensor, resulting in gravity, magnetic, and electrical data volumes. This approach overcomes the limitations of single-source data, forming multi-source heterogeneous data, which is then mapped to preprocess scattered measurement data into a 3D data volume. This ensures the fusion of multimodal features and provides a reliable geological basis for subsequent model initialization, inversion constraints, and geological calibration.

[0027] Step S200: Based on the spatial range and depth of the target detection area, the preset initial three-dimensional spatial model is divided into multiple hexahedral grids. Historical measurement data and physical geological information are used to initialize and configure initial physical property parameters, including density, magnetic susceptibility and resistivity, for each hexahedral grid.

[0028] In one embodiment, the initial 3D spatial model is based on a pre-built and pre-stored standard 3D geological spatial model. After being called, trimmed, and meshed, it is constructed to fit the actual terrain and depth of the target detection area, thus restoring the surface undulations and overcoming the shortcomings of traditional flat-ground modeling. The hexahedral mesh is the smallest computational unit for discretizing the underground space, generated using high-precision uniform meshing. It serves as the basic carrier for 3D physical property modeling and inversion iteration, with each mesh possessing unique (x, y, z) 3D spatial coordinates. The initial physical property parameters are basic physical attributes configured for each mesh based on prior regional geological information, including density, magnetic susceptibility, resistivity, and four core parameters such as polarizability, corresponding to the physical characteristics of underground rocks and ore bodies.

[0029] A pre-built initial 3D spatial model adapted to the full-area exploration standards is obtained and used as a preset benchmark model. This benchmark model is a general standard model adapted to the real terrain and the maximum depth of standard deep exploration in the digital elevation model. Based on the observation data, spatial range, and actual exploration depth of the current target exploration area, the model is spatially trimmed, boundary matched, and depth truncated to accurately extract the 3D spatial range that completely covers the current exploration area, resulting in an initial 3D spatial model adapted to the target exploration area. Based on the spatial range and exploration depth of the target exploration area, a 3D envelope space is constructed by extending it outwards and into the depths by a preset distance to eliminate boundary effects. This 3D envelope space is then divided into several uniformly sized initial coarse-grained hexahedral grids. For each coarse-grained hexahedral grid, its grid subdivision indicator factor is calculated. This indicator factor is jointly determined by the depth of the grid center point from the surface terrain and the spatial distance to the nearest observation point (measuring point), ensuring that the indicator factor value is larger in areas closer to the surface, with greater terrain undulation, and closer to the observation point. A threshold of 0.5 is set for the indicator factor. Traversing all grids in space, if the indicator factor value of the current grid is greater than the indicator factor threshold, an octree splitting algorithm is used to divide the single hexahedral grid into 8 smaller sub-hexahedral grids. This partitioning process is repeated iteratively on the new grids until the grid size reaches the set minimum physical resolution or the indicator factor value is less than or equal to the indicator factor threshold. The indicator factor value is calculated using the formula: λ = α·e -βd +γ·1 / r 2 In this context, α, β, and γ are preset weighting coefficients, set by professionals based on experience. When the shallow terrain is highly undulating, γ is increased; when it is necessary to detect deep macroscopic structures, α is increased. d is the depth of the grid center point from the surface terrain, and r is the spatial distance to the nearest observation point. The smaller d and r are, the larger the value of the indicator factor, indicating that the corresponding number of points is in one grid, thus achieving grid division.

[0030] After completing the spatial adaptive mesh generation, the spatial mesh is assigned physical property values ​​using historical measurement data and physical geological information. Specifically, this involves converting the extracted historical measurement data and physical geological information into discrete physical property point sets with three-dimensional spatial coordinates, and determining whether each adaptive hexahedral mesh contains points from the discrete physical property point set. If it contains one or more prior known points, the arithmetic mean or volume-weighted average of the parameter values ​​of these points is calculated, and the average value is directly used as the initial physical property parameters of the hexahedral mesh; simultaneously, the mesh is marked as a "prior constraint mesh." For unknown hexahedral meshes without known internal data, a three-dimensional kriging interpolation algorithm is used to estimate the parameters based on historical measurement data, and the estimated values ​​are used as the initial values ​​for the unknown mesh parameters, thus completing the initial configuration of the three-dimensional spatial model. Through refined hexahedral mesh generation, the digital characterization of deep, minute geological structures and fine vein-like ore bodies is achieved, and the parameters are initialized based on real geological prior information, completely avoiding the problems of slow model iteration convergence, computational crashes, and result distortion caused by traditional zero-value initialization.

[0031] Step S300: Input the gravity data volume, magnetic data volume, and electrical data volume into the preset target deep learning model for multimodal feature extraction and fusion inversion, and output predicted physical property parameters including density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. Use the predicted physical property parameters to adjust the initial physical property parameters and output a three-dimensional coupled physical property model. The target deep learning model is used to perform spatially consistent coupling of the gravity data volume, magnetic data volume, and electrical data volume.

[0032] In one embodiment, the target deep learning model structure includes a three-branch encoder, an implicit feature alignment network, and a feature fusion layer. Each encoder includes a feature extraction layer, a pooling layer, and a fully connected layer, used to spatially and consistently couple gravity data volumes, magnetic data volumes, and electrical data volumes. Multimodal feature fusion refers to extracting high-dimensional spatial structure features from geophysical data of three different dimensions and physical mechanisms (gravity, magnetic, and electrical) and completing cross-modal fusion to achieve complementary features across the three fields. The predicted physical property parameters are the optimal physical property parameters for each grid output after model optimization, including density distribution, magnetic susceptibility distribution, resistivity distribution, and precise values ​​of coordinate parameters. The three-dimensional coupled physical property model is a three-dimensional grid model with highly matched spatial structures of gravity, magnetism, and electricity, and precisely converged physical property parameters, output by deep learning multimodal fusion inversion and optimized with multi-field structural consistency constraints.

[0033] The trained target deep learning model is invoked to process gravity, magnetic, and electrical data volumes. The model extracts spatial structure features from each individual field using a three-branch encoder, aligns the three-field features spatially using an implicit feature alignment network and a feature fusion layer, and iteratively optimizes using a joint objective function to continuously correct the initial physical property parameters of the mesh. The initial physical property parameters are then comprehensively updated using the optimal predicted physical property parameters, outputting a three-dimensional coupled physical property model with unified spatial structure and optimal accuracy. This approach overcomes the shortcomings of traditional cross-gradient constraint inversion, such as incomplete coupling of the three fields, strong ambiguity, and poor adaptability to complex geology. It achieves integrated quantitative inversion of gravity, magnetic, electrical, and depth through deep learning adaptive feature fusion; it forces a high degree of consistency in the spatial structure of the three fields, effectively suppressing ambiguity in deep exploration inversion and significantly improving the characterization accuracy of deep stratigraphic interfaces, fault zones, and mineralization anomalies; the output three-dimensional coupled physical property model solves the problems of structural misalignment and insufficient deep accuracy in traditional dual-field inversion, providing a highly reliable digital foundation for subsequent geological modeling and ore body location.

[0034] In one embodiment, gravity data volume, magnetic data volume, and electrical data volume are input into a preset target deep learning model for multimodal feature extraction and fusion inversion, and predicted physical property parameters are output, including but not limited to the following steps: Step S310: Input the gravity data volume, magnetic data volume and electrical data volume into the corresponding encoders for three-dimensional feature extraction to obtain the initial gravity spatial features, magnetic spatial features and electrical spatial features.

[0035] Specifically, the encoder employs a three-layer feature extraction layer, using a 3×3×3 three-dimensional convolutional kernel to mine high-dimensional geometric features such as spatial edges, corners, and thickness of the 3D data. The feature extraction layers are as follows: the first 3D convolutional layer has a 3×3×3 kernel, 32 output channels, a stride of 1, and padding of 1, followed by BatchNorm and ReLU activation functions; the second 3D convolutional layer has a 3×3×3 kernel, 64 output channels, a stride of 2 (downsampling), followed by BatchNorm and ReLU activation functions; and the third 3D convolutional layer has a 3×3×3 kernel, 128 output channels, a stride of 2, followed by BatchNorm and ReLU activation functions. Core spatial features such as abrupt boundary changes in physical properties, geological body outlines, and structural orientations of the underground space are extracted layer by layer. Through processing by the three feature extraction layers, initial gravity spatial features, magnetic spatial features, and electrical spatial features are obtained.

[0036] Step S320: The gravity space features, magnetic space features and electric space features are mapped to a unified high-dimensional implicit space using an implicit feature alignment network to perform tensor alignment and output standardized extracted features.

[0037] Specifically, the initial gravity spatial features, magnetic spatial features, and electrical spatial features with 128 channels output in step S310 are transformed dimension-wise using independent projection layers containing 1×1×1 three-dimensional convolutional kernels. The 1×1×1 convolution enables cross-channel information interaction without changing the spatial scale resolution of the three-dimensional feature map, mapping heterogeneous features caused by differences in the physical dimensions of gravity, magnetism, and electricity to an initial implicit space with the same semantic dimension. To eliminate spatial displacement bias in the response of different physical background fields to the same geological body, a three-dimensional spatial attention module is constructed. The three projected feature tensors are initially concatenated along the channel dimension. Global extrema and mean spatial prior distributions are extracted using max pooling and average pooling. Then, a 7×7×7 large receptive field three-dimensional convolution is performed, and a three-dimensional spatial attention weight mask is output using a Sigmoid activation function, with values ​​ranging from [0,1]. The three-dimensional spatial attention weight mask is then element-wise multiplied with the projected gravity, magnetism, and electrical feature tensors, respectively. By dynamically enhancing regions with high overlap in gravity, magnetism, and electrical response locations (such as the center of a high-fidelity anomaly) through mask weighting, regions with unrelated or conflicting responses (such as single-domain noise or false anomaly boundaries) are suppressed. The final output consists of three standardized extracted features (i.e., aligned gravity, magnetism, and electrical implicit feature tensors) with a unified number of 128 channels, consistent spatial scale, and mutual alignment of geological semantics.

[0038] Step S330: The standardized extracted features are spliced ​​and reconstructed using the feature fusion layer to output the predicted physical property parameters.

[0039] Specifically, the three standardized extracted features output from step S320 are concatenated and stitched together along the depth channel dimension. This concatenation generates a high-dimensional joint feature tensor, which integrates high-order semantic information about density, magnetic, and electrical variations under unified spatial coordinates. The concatenated joint feature tensor is then input into a 3×3×3 three-dimensional convolutional fusion layer with 256 output channels and a stride of 1. This is followed by BatchNorm and ReLU activation functions to eliminate inter-channel redundancy caused by direct concatenation, completing the deep nonlinear interweaving and dimensionality reduction compression of multi-physics information. Finally, the fused features are input into a three-dimensional deconvolutional decoder for spatial reconstruction. The first deconvolution layer uses a 3×3×3 deconvolution kernel with 64 output channels and a stride of 2 (achieving a 2x upsampling amplification for spatial resolution), matching its size to the encoder's second-layer output, followed by BatchNorm and ReLU. The second deconvolution layer uses a 3×3×3 deconvolution kernel with 32 output channels and a stride of 2 (again upsampling amplification by 2x), restoring the original 3D spatial grid size to the same as the initial input multi-source data volume, followed by BatchNorm and ReLU. The deconvolutioned tensor is then input into the property mapping layer. This layer uses a 1×1×1 3D convolution kernel, compressing its channel count to the number of target prediction parameters (e.g., 1 output channel for single-parameter inversion prediction of density; 3 output channels for joint inversion prediction of density, magnetic susceptibility, and resistivity). Based on the prior upper and lower limits of actual rock physics parameters and sigmoid function constraints, the output includes predicted property parameters containing density distribution, magnetic susceptibility distribution, and resistivity distribution. The discrete indices of each voxel unit in the 3D mesh data volume within the tensor space represent a one-to-one hard-coded mapping relationship between the 3D spatial coordinate parameters in the initial 3D mesh partitioning model. By fusing the three-channel prediction tensor output from the forward inference of the fusion network, the prior 3D spatial coordinate parameters can be directly combined to output predicted physical property parameters containing these coordinate parameters. For the specific model structure, please refer to [reference needed]. Figure 2 .

[0040] The aforementioned deep learning model overcomes the bottlenecks of traditional single-source geophysical inversion, which suffers from strong ambiguity and lacks substantial physical constraints in conventional multi-source data stitching. It constructs a data-driven implicit structural coupling mechanism: through a cross-modal spatial attention alignment module and unified convolutional dimensionality reduction fusion of aligned multi-domain features, it ensures that any predicted physical property parameter at the output is generated by a nonlinear mapping of the spatial responses of the gravitational, magnetic, and electric fields. This achieves information stitching and cross-coupling of structural coupling and physical property mapping in three-dimensional spatial distribution, significantly improving the fidelity and geological interpretability of inversion results in complex geological environments.

[0041] In one embodiment, a comprehensive three-dimensional geological-geophysical property model is reconstructed based on predicted physical property parameters. This model includes three-dimensional spatial coordinates and corresponding density, magnetic susceptibility, and resistivity values ​​at those locations. The model is then spatially compared with the three-dimensional spatial model. The predicted physical property parameters are used to adjust the initial physical property parameters, resulting in a three-dimensional coupled physical property model for subsequent ore body evaluation.

[0042] In one embodiment, the modeling and ore body localization method based on the coupling of gravity, magnetism, and electricity in multiple physics fields further includes a training process for a target deep learning model. The training process includes, but is not limited to, the following steps: Step S610: Obtain training observation data, training historical data, standard observation data, and training geological information for multiple training detection areas.

[0043] Specifically, the training exploration area consists of several standard exploration zones with known geological conditions and clearly defined ore body distributions, serving as sample areas for model training and covering different ore deposit types and topographical scenarios. The training observation data comprises standardized gravity, magnetic, and electrical data volumes measured in the training area, with format and accuracy completely consistent with actual exploration input data. The training historical data includes past exploration data on physical properties, engineering data, and structural statistics for the training area. The standard observation data consists of real observation data from drilling and geological verification within the training area, serving as ground truth labels for model training. The training geological information includes precise spatial distribution and physical property threshold rules for strata, rock masses, faults, and ore bodies within the training area.

[0044] By reading training observation data, corresponding historical training data, standard observation data, and training geological information from multiple training exploration areas stored in the database, a foundation is provided for subsequent training. More than 10 mature exploration areas with different mineralization types and structural complexities are selected to ensure diverse geological scenarios covered by the samples, thus improving the model's generalization ability. Gravity, magnetic, and electrical measured observation data, historical exploration data, drilling verification standard data, and complete geological data are collected for each training area. Following the S100 preprocessing workflow, all training observation data are corrected, denoised, gridded, and spatially registered to generate a standardized training data volume. Historical training data, standard ground truth data, and geological information are collected and organized to construct a complete training sample set. A training sample library with diverse scenarios and reliable ground truth data is built to avoid the problems of poor generalization ability and insufficient adaptability caused by training the model in a single scenario, significantly improving the model's adaptability and prediction accuracy.

[0045] Step S620: For each training detection area, the preset training three-dimensional spatial model is divided into meshes according to the spatial range and detection depth of the training detection area to obtain multiple training hexahedral meshes. Training physical property parameters are initialized and configured for each training hexahedral mesh using training historical data and training geological information.

[0046] Specifically, the training 3D spatial model is a 3D spatial model adapted to the training exploration area and subsequently adjusted. The training hexahedral mesh is the smallest computational unit for discretizing underground space, generated using a partitioning algorithm. It serves as the basic carrier for 3D physical property modeling and inversion iteration, with each mesh possessing unique (x, y, z) 3D spatial coordinates. The training physical property parameters are the basic physical property parameters initialized for the training mesh, configured based on real geological data from the training area.

[0047] For each training detection area, network training is performed. Similar to step S200, the preset training 3D spatial model is divided into multiple training hexahedral meshes based on the spatial range and detection depth of the training detection area. Training physical property parameters are initialized and configured for each training hexahedral mesh using training historical data and training geological information. The specific process is not described here. These parameters are used to provide a spatial consistency reference for coupling during the training phase, provide a precise reference for model iterative training, and effectively improve the accuracy of model feature learning.

[0048] Step S630: Input the training observation data into the preset initial deep learning model to perform multimodal feature extraction and fusion inversion to obtain the predicted inverted physical property parameters.

[0049] Specifically, the initial deep learning model, which has not yet completed parameter iterative optimization, possesses basic feature extraction and fusion capabilities. Its parameters are initially randomly configured, and its structure is consistent with that of the target deep learning model, which will not be elaborated upon here. Training observation data is input into the preset initial deep learning model for multimodal feature extraction and fusion inversion to obtain predicted inverted physical property parameters. These predicted inverted physical property parameters are the output physical property parameters of the initial deep learning model from the training data, including the inverted density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. This provides data support for subsequent loss value calculation and model parameter optimization, completing the forward feature computation for model training.

[0050] It should be noted that in the forward processing of the initial deep learning model, in addition to outputting the predicted inversion material property parameters, the feature vectors extracted by the three-branch encoder are also output to calculate the distance between the feature vectors, thereby indicating whether the gravity, magnetic and electrical data represent the same space, thus realizing data coupling, so as to calculate the loss function and achieve accurate inversion.

[0051] Step S640: With the constraint that the predicted inversion physical property parameters have a high fit on the trained three-dimensional spatial model, the loss value of the joint objective function is calculated based on the three-channel feature vector extracted from the initial deep learning model, the predicted inversion physical property parameters, the training geological information, and the standard observation data.

[0052] Specifically, the joint objective function is a weighted fusion of data fitting terms, deep learning structure constraints, and model regularization constraints, and is the core basis for model optimization. The high degree of fit between the predicted and inverted physical property parameters and the trained 3D spatial model is used as a constraint condition, and this high degree of fit is a global pre-constraint condition. During loss calculation, it is ensured that the predicted physical property parameters output by all grids strictly conform to the grid topology, spatial scale, stratigraphic boundaries, and 3D spatial distribution patterns of the trained 3D spatial model. This prevents spatial boundary violations, grid misalignment, cross-layer disorder, and spatial topological distortion in the predicted and inverted physical property parameters. Therefore, when the calculated loss function is used to adjust the model in reverse, the resulting target deep learning model ensures that the generated predicted physical property parameters conform to the 3D spatial model, achieving spatial consistency of multi-source data coupling, thus enabling accurate inversion of the ore body location and morphology.

[0053] The process iterates through all hexahedral meshes of the training 3D spatial model, matching and verifying the predicted physical property parameters of each mesh output by the model inversion with the spatial topology, layer distribution, and mesh coordinate boundaries of the training 3D spatial model. All predicted inversion physical property parameters are constrained to act only on the mesh cells corresponding to the spatial coordinates, ensuring that the 3D spatial distribution, property variation trends, and anomaly spatial locations of the predicted inversion physical property parameters are completely dependent on the physical spatial structure of the training 3D spatial model. This achieves a high degree of spatial fit between the predicted inversion physical property parameters and the training 3D model, filtering out invalid predicted inversion physical property parameters due to spatial misalignment, topological failure, or model out-of-bounds errors, and retaining only valid parameters with compliant spatial fit for subsequent loss calculations.

[0054] In one embodiment, based on the three-channel feature vectors extracted from the initial deep learning model, the predicted inverted physical property parameters, the training geological information, and the standard observation data, the loss value of the joint objective function is calculated, including but not limited to the following steps: Step S641: Calculate the similarity between the predicted observation data and the standard observation data to obtain the data fitting similarity. The predicted observation data is obtained by performing forward modeling on the physical property distribution data in the predicted inversion physical property parameters.

[0055] Specifically, the density distribution matrix, magnetic susceptibility distribution matrix, and resistivity distribution matrix from the predicted inverted physical property parameters are input into the corresponding forward modeling operators. Specifically, the predicted gravity anomaly data is calculated using a three-dimensional gravitational potential integral operator, the predicted magnetic anomaly data is calculated using a Poisson equation operator, and the predicted electromagnetic field response is calculated using a three-dimensional finite electromagnetic difference operator. These are combined to form the predicted observation data. The mean square error (MSE) is calculated between the predicted observation data and the standard observation data to obtain the data fitting similarity. The smaller the data fitting similarity value, the closer the inverted physical property volume is to the actual subsurface medium response in terms of physical magnitude. Predicted observation data is used instead of directly using training observation data and standard observation data for calculation because the training and standard observation data cannot reflect the accuracy of the underlying inverted data. By combining forward and inversion, the loss function can be adjusted to improve the accuracy of the inversion process.

[0056] Step S642: Calculate the spatial similarity of each pair of feature vectors in the three-channel feature vectors, and fuse the similarity values ​​to obtain the spatial fitting similarity.

[0057] Specifically, during training, the model extracts high-dimensional feature vectors of gravity, magnetism, and electricity from the initial deep learning model's three-channel encoder output through the feature vectors output by the three-branch encoder. The cosine similarity algorithm is then used to calculate the pairwise similarity of these three sets of feature vectors: gravity-magnetism, magnetism-electricity, and gravity-electricity. To ensure high consistency of the feature vectors in spatial structure (boundaries, trends), a penalty function is constructed (maximizing similarity is equivalent to minimizing loss). The three similarity results are weighted, fused, and averaged. By subtracting the average result from 1, this loss forces the network to capture the same geological edges (i.e., density abrupt change boundaries, magnetic abrupt change boundaries, and electrical abrupt change boundaries mutually corroborate each other in the latent space) within a three-dimensional local window, obtaining the overall fitting similarity of the three-field spatial structure, i.e., spatial fitting similarity. This quantifies the consistency differences of the multi-physics field structure. Based on structural deviations, deep learning structural constraint terms are generated to accurately penalize multi-field structural misalignment problems, thereby ensuring the consistency of multi-source data coupling in space.

[0058] Step S643: Based on the training coordinate parameters, training geological information and training historical data in the predicted inversion physical property parameters, calculate the gradient change of the physical property distribution data in the predicted inversion physical property parameters between adjacent grids, and obtain the value of the model constraint term used to constrain spatial continuity.

[0059] Specifically, a three-dimensional spatial geological prior weight matrix is ​​constructed based on training geological information (such as the distribution location of well logging faults and lithofacies boundaries) and training historical data. Within the same lithology, the weight value approaches 1; at fault or lithological contact zone boundaries, the weight value approaches 0 (allowing for drastic jumps in physical quantities). Using the extracted training grid coordinate parameter information, such as the step size of the extracted grid subdivision, and based on the training coordinate parameters in the inverted physical property parameters, the discrete partial derivatives of the predicted physical property distribution data between adjacent grids in three-dimensional space are calculated. A total variation or roughness matrix is ​​introduced to construct a model constraint term. This model constraint term is calculated by multiplying the geological prior weight matrix by the sum of the squares of the discrete partial derivatives in the predicted inverted physical property parameters. This constraint term forces a smooth and gradual change in physical properties within the same geological unit, while the three-dimensional spatial geological prior weight matrix prevents the blurring of the true geological interface caused by excessive smoothing.

[0060] Step S644: The values ​​of data fitting similarity, spatial fitting similarity, and model constraint terms are weighted and fused to obtain the loss value of the joint objective function.

[0061] Specifically, preset weight coefficients are obtained, with the sum of the corresponding weight coefficients a, b, and c being 1. The values ​​of data fitting similarity, spatial fitting similarity, and model constraint terms are then weighted and fused using these weight coefficients to obtain the loss value of the joint objective function. This is specifically expressed as L = aL1 + bL2 + cL3, where L is the value of the joint objective function, L1 is the data fitting similarity, L2 is the spatial fitting similarity, and L3 is the value of the model constraint term. During training iterations, a, b, and c are dynamically and adaptively adjusted based on the gradient norm of each loss term. This ensures that the three mechanisms—data fidelity, cross-domain structural coupling, and geological prior spatial constraints—can work synergistically to optimize both forward inference and backward updates, avoiding gradient vanishing or exploding caused by a single loss term dominating the network. This loss value reflects the comprehensive coverage of four dimensions—model space rationality, data fitting accuracy, multi-field coupling consistency, and geological law adaptability—through the four-dimensional linkage of model space topological constraints, data truth constraints, multi-field structure alignment constraints, and geological prior constraints. This ensures that model training not only closely matches the measured data but also rigorously matches the three-dimensional geological spatial structure and real mineralization laws, significantly improving the stability of model training and the accuracy of deep geological structures and ore body identification.

[0062] Step S650: The parameters of the initial deep learning model are adjusted using the loss value of the joint objective function to obtain the target deep learning model.

[0063] Specifically, model parameters are adjusted based on backpropagation of the loss value, updating core parameters such as convolutional weights, biases, and fusion weights using the gradient descent algorithm. The target deep learning model is the optimal model that completes multi-batch sample iterative training, loss value convergence, and accuracy achievement, possessing multi-field adaptive coupling and high-precision inversion capabilities. During training, the batch size and number of training iterations are set according to requirements. For example, a batch size of 2 and 1000 training iterations are not limited. Adaptive learning rate algorithms (Adam, SGD) are used to optimize the model, aiming to minimize the joint objective function loss value. Error signals are backpropagated, and all parameters of the model encoder, implicit feature alignment network, and feature fusion layer are updated layer by layer. The inversion, loss value calculation, and parameter update process is repeated across all training sample regions, continuously iterating and optimizing model parameters. Changes in the training loss value are monitored in real time; when the loss value stabilizes and no longer fluctuates significantly, the model is considered converged. All parameters of the converged model are locked, completing model solidification and generating a target deep learning model that can be directly used for actual exploration inversion. Through multi-scenario and multi-dimensional iterative optimization, the problems of inaccurate feature extraction, multi-field coupling imbalance, and poor geological adaptability of the initial model are completely solved. The trained target model has a strong adaptive learning ability, can adapt to different complex geological scenarios, accurately realize the deep coupling inversion of gravity, magnetism, and electricity multi-physics fields, significantly improve the accuracy of deep structure and ore body property characterization, and greatly enhance its practicality and versatility.

[0064] Step S400: Perform multi-scale segmentation and stratigraphic calibration on the three-dimensional coupled physical property model to obtain a three-dimensional geological structure model.

[0065] In one embodiment, multi-scale segmentation employs a hierarchical segmentation strategy of "coarse-scale macroscopic division + fine-scale refinement" for the 3D physical property model. This involves progressively breaking down geological units from global structures to local subtle anomalies, achieving precise segmentation and extraction of geological structures at different scales. Stratigraphic calibration involves semantically matching and assigning values ​​to purely digital physical property segmented blocks with real geological lithology, stratigraphic units, and structural units, assigning geological attribute labels to meaningless digital meshes. The 3D geological structural model, after multi-scale segmentation, stratigraphic lithology calibration, and topological mesh smoothing, is a physical geological model possessing complete stratigraphic information, structural boundaries, rock mass contact relationships, and spatial topological relationships, exhibiting both geological semantics and 3D physical characteristics.

[0066] The three-dimensional coupled physical property model is obtained in step S300. The integrity of the physical property parameters, consistency of spatial coordinates, and alignment accuracy of the three-field structure of all entity meshes are verified. Invalid cells with abnormal parameters or mesh failures are removed, retaining the entire effective three-dimensional entity mesh array. Then, multi-scale segmentation and stratigraphic calibration are performed. Multi-scale segmentation takes into account both macroscopic structural frameworks and fine mineralization anomalies. Intelligent stratigraphic calibration replaces subjective manual interpretation, significantly improving the standardization and efficiency of modeling, providing a precise and compliant three-dimensional geological carrier for subsequent quantitative orebody location.

[0067] In one embodiment, the three-dimensional coupled physical property model is segmented at multiple scales and its stratigraphic calibration is performed to obtain a three-dimensional geological structure model, including but not limited to the following steps: Step S410: Perform bottom-up spatial downsampling on multiple entity meshes in the three-dimensional coupled physical property model to generate a multi-scale physical property spatial pyramid from fine resolution to coarse resolution.

[0068] Specifically, the solid mesh is the smallest hexahedral computational unit of the three-dimensional coupled physical property model. Each mesh carries four types of physical property parameters—normalized density, magnetic susceptibility, resistivity, and polarizability—along with unique three-dimensional spatial coordinates, serving as the basic unit for segmented calculations. Spatial downsampling is a sampling method that performs scale compression and aggregation on high-resolution fine mesh models, aggregating multiple adjacent fine meshes into a single coarse mesh, thus progressively reducing the model resolution. The multi-scale physical property spatial pyramid is a hierarchical structure composed of multiple layers of models with different spatial resolutions. The top layer is a coarse-resolution global approximate model, and the bottom layer is the original fine-resolution model. Each layer retains the spatial distribution characteristics of physical properties at its corresponding scale.

[0069] Based on the overall mesh size of the model, a pre-defined five-level multi-scale pyramid structure is adopted, with a downsampling factor set to 2x for progressive downsampling, achieving a gradual halving of resolution and progressive expansion of model scale. Using the original finest resolution 3D coupled physical property model as the base of the pyramid, spatial downsampling is performed layer by layer upwards: at each level, spatially adjacent 2×2×2 fine-resolution solid meshes are aggregated into a coarse-resolution mesh using a weighted average algorithm, with weights adaptively allocated based on mesh spatial distance and physical property reliability. During downsampling, the density, magnetic susceptibility, resistivity, and polarizability parameters of the coarse mesh are calculated by weighting the corresponding fine mesh clusters, while preserving core spatial features such as physical property abrupt change locations, boundary contours, and anomaly centers in each layer of the model to avoid structural feature loss due to downsampling. This process sequentially completes the construction of five resolution scales, forming a complete multi-scale physical property spatial pyramid that progressively transitions from a coarse-scale global overview model to a fine-scale refined model, with complete alignment of spatial coordinates and one-to-one correspondence of topological relationships between each level of the model. By constructing a multi-scale structured expression system, the coarse level can quickly locate the overall strata and the macroscopic framework of large structures, while the fine level can accurately depict the fine mineralization boundaries, providing a complete multi-scale data carrier for subsequent step-by-step boundary optimization and segmentation from coarse to fine, ensuring that the segmentation results have both global rationality and local precision.

[0070] Step S420: Under different resolution levels of the multi-scale physical property space pyramid, three-dimensional region connectivity segmentation is performed based on the coupling physical property change gradient between adjacent entity meshes, and the boundary is refined step by step from coarse to fine to generate multiple three-dimensional segmentation blocks.

[0071] Specifically, the coupled physical property change gradient refers to the calculated comprehensive change rate of the four types of physical properties of adjacent grids. The larger the gradient value, the more abrupt the boundary between strata, rock masses, and mineralization bodies exists between grids. The 3D region connectivity segmentation algorithm, based on differences in physical property gradients, divides spatially continuous, similarly materially, and consistently changing grid clusters into the same independent region. The coarse-to-fine progressive refinement uses the coarse-scale macroscopic segmentation results as a constraint basis, iteratively correcting the boundaries at the fine-scale level to achieve a macroscopic framework that remains unchanged while continuously refining and optimizing local boundaries.

[0072] At different resolution levels of the multi-scale physical property spatial pyramid, for each level of the pyramid model, a three-dimensional first-order difference algorithm is used to calculate the individual physical property gradients of density, magnetic susceptibility, resistivity, and polarizability in the x, y, and z directions for each grid. A comprehensive coupled physical property gradient field is generated through weighted fusion. For example, the fusion weights are set to 0.35 for magnetic susceptibility, 0.35 for resistivity, 0.2 for density, and 0.1 for polarizability. At the coarse scale, a fixed gradient threshold of 0.18 is used to filter large-scale structures and macroscopic stratigraphic boundaries; at the fine scale, a fixed micro-gradient threshold of 0.08 is used to capture subtle alterations and weak mineralization edges. Gradients greater than the threshold are considered boundary grids, while those less than the threshold are considered homogeneous regions. At the coarse-scale level, a three-dimensional watershed segmentation algorithm is used to cut macroscopic boundaries and divide large geological units; at the fine-scale level, a three-dimensional region growing algorithm is used to fill homogeneous regions and optimize local connectivity, ensuring that the global region is free of voids and fragmentation. The boundaries of coarse-scale segmented blocks are projected onto the next level of fine-scale model through coordinate mapping. Using the coarse boundaries as constraints, the boundary positions are iteratively corrected by combining fine-scale micro-gradient features, eliminating the smoothing distortion problem of coarse segmentation layer by layer. Finally, multiple 3D segmented blocks with accurate boundaries, spatial connectivity, and scale hierarchy are generated. The segmentation method of hierarchical gradient thresholding combined with dual algorithms completely solves the problems of over-segmentation, under-segmentation, and blurred boundaries of traditional single algorithms. The coarse scale ensures that the overall geological framework does not shift, while the fine scale restores the fine mineralization and alteration structures, so that the segmented blocks conform to the distribution characteristics of real underground geological units.

[0073] In one embodiment, three-dimensional region connectivity segmentation is performed based on the gradient of coupling property changes between adjacent entity meshes, and the boundaries are refined step by step from coarse to fine to generate multiple three-dimensional segmentation blocks, including but not limited to the following steps: Step S421: Based on the coarse resolution level of the multi-scale physical property spatial pyramid, calculate the weighted fusion gradient of multiple physical properties as edge weights, and use a preset region segmentation algorithm to extract coarse-scale initial blocks.

[0074] Specifically, the multi-physical attribute weighted fusion gradient is a comprehensive gradient obtained by fusing four types of single-property gradients—density gradient, magnetic susceptibility gradient, resistivity gradient, and polarizability gradient—according to geological sensitivity weights. The numerical value characterizes the salience of geological boundaries, comprehensively reflecting the abrupt changes in lithology, structure, and mineralization. Edge weights represent the probability that the grid boundary belongs to a geological interface, tectonic interface, or lithological contact surface; higher weights indicate more significant boundary characteristics. The preset region segmentation algorithm is a joint segmentation algorithm combining the 3D watershed algorithm and the region growing algorithm. The coarse-scale initial block is a macroscopic geological unit block obtained from coarse-resolution hierarchical segmentation, used to constrain the overall structural framework.

[0075] At the coarse-resolution level of the multi-scale pyramid, all coarse grids are traversed, and a three-dimensional first-order difference algorithm is used to calculate the first-order spatial gradients of adjacent grids in four dimensions: density, magnetic susceptibility, resistivity, and polarizability. The gradient fields of density, magnetic susceptibility, resistivity, and polarizability in the x, y, and z dimensions are calculated grid-by-grid, accurately capturing the abrupt changes in physical properties at macroscopic structural boundaries, resulting in four independent physical property gradient fields. Based on the sensitivity of different physical properties to strata, rock masses, and structures, adaptive weighting coefficients are configured, and the four gradients are weighted and fused to obtain the comprehensive physical property gradient field.

[0076] For example, the fusion process is Gtotal = 0.35Gm + 0.35Gr + 0.2Gd + 0.1Gp, where Gm, Gr, Gd, and Gp represent the corresponding density, magnetic susceptibility, resistivity, and polarizability gradient fields, respectively, and Gtotal represents the comprehensive physical property gradient field.

[0077] The comprehensive gradient value is used as the edge weight of each grid. A coarse-scale boundary gradient threshold is set to distinguish between geological boundary grids and homogeneous internal grids within the stratigraphy. Grids with gradients greater than the threshold are identified as abrupt boundary changes in physical properties, belonging to geological boundary areas such as stratigraphic interfaces, rock mass contact surfaces, and fault structures, and participate in boundary cutting and segmentation. Grids with gradients less than the threshold are identified as homogeneous internal regions, belonging to the same type of complete geological unit, and are subject to regional growth and merging. The coarse-scale boundary gradient threshold ranges from 0.15 to 0.20, and can be 0.18, representing a dimensionless gradient value.

[0078] Grids with a comprehensive gradient value greater than a threshold are identified as geological boundary grids, while regions with a gradient less than the threshold are identified as homogeneous internal regions of the stratigraphy. A 3D region growing algorithm is used for homogeneous regions, growing connected regions based on similar physical property mean values ​​and spatial connectivity. For regions with complex boundaries, a 3D watershed algorithm is used to cut the boundaries, resulting in a coarse-scale initial block set representing the macroscopic structural framework. Through the above multi-property weighted gradient fusion, hidden and weakly anomalous geological boundaries can be comprehensively identified, avoiding the omission of structures by a single property. The combined watershed and region growing algorithm balances boundary cutting accuracy and regional connectivity, ensuring that the coarse-scale blocks completely restore the overall macroscopic framework of the regional strata, rock masses, and faults, providing a stable and accurate global constraint basis for subsequent fine-scale optimization.

[0079] Step S422: Project the boundary of the coarse-scale initial block onto the adjacent fine-resolution layer as the initial seed surface, calculate the property change gradient under the adjacent fine-resolution layer, and perform super-voxel clustering shrinkage or expansion iteration on the initial seed surface based on the property change gradient until the bottom fine-resolution layer is reached.

[0080] Specifically, the initial seed surface, obtained by projecting the boundary of the previous coarse-scale block, serves as the baseline boundary for fine-scale iterative optimization. The property change gradient is a small, localized abrupt property gradient captured at the fine-resolution level, corresponding to subtle geological structures such as small alteration zones, mineralization edges, and micro-fractures. The supervoxel clustering shrinkage or expansion iteration: using supervoxels as the smallest clustering unit, an adaptive iterative correction strategy is employed to shrink the initial boundary inwards or expand it outwards based on local micro-gradient characteristics.

[0081] Based on the coarse-scale and fine-scale model mesh indices, a 3D mapping matrix is ​​constructed, establishing a one-to-one correspondence between the coarse mesh and the corresponding covering fine mesh. Using this mapping matrix, the 3D boundaries, spatial contours, and surface coordinates of the initial coarse-scale block are precisely projected to the next adjacent fine-resolution layer. The complete initial segmentation boundary is replicated in the fine-scale model, forming the initial seed surface and ensuring complete correspondence between the coarse and fine-scale hierarchical structures. Specifically, at the fine-scale level, a 5×5×5 neighborhood difference operator is used, a micro-gradient judgment threshold is set, low-frequency background noise is filtered out, and high-frequency micro-property abrupt change signals from mineralization edges and alteration zones are amplified to calculate the fine-scale property micro-gradient field. The micro-gradient judgment threshold is 0.08; micro-gradients greater than this threshold are considered true property abrupt change edges, allowing for iterative boundary correction. Eight adjacent meshes at the fine scale are aggregated into one supervoxel unit, with the supervoxel as the smallest iterative unit, avoiding boundary fragmentation and excessive computation caused by single-mesh iteration. A supervoxel clustering iterative algorithm is employed with an iteration step size of 0.5 and a maximum of 20 iterations. When the micro-gradient on the outer side of the seed surface is greater than that on the inner side, the boundary is driven to expand outward to capture external mineralized and altered regions; when the micro-gradient on the inner side of the seed surface is greater than that on the outer side, the boundary is driven to contract inward to remove invalid background regions. The process of "boundary projection - micro-gradient calculation - clustering iterative optimization" is repeated to iterate the multi-scale pyramid from the top coarse resolution level to the bottom fine resolution level.

[0082] Addressing the micro-gradient differences between the inner and outer sides of the initial seed surface, abrupt inward gradients drive boundary contraction, while gradual outward gradients drive boundary expansion. The boundary position is iteratively corrected to match the local fine-grained physical property abrupt changes. Repeated projection, micro-gradient calculation, and clustering iterations refine the structure layer by layer, correcting boundary deviations until the finest resolution level of the pyramid is reached. Cross-level projection ensures the overall structural framework remains unchanged, while fine-scale micro-gradient iterations accurately recreate local fine geological boundaries, achieving a segmentation effect of "unchanged macroscopic structure and ultimate microscopic detail reproduction," significantly improving the ability to characterize deep, subtle, concealed structures and mineralization anomalies.

[0083] Step S423: When the contraction or expansion iterations tend to converge, output the formed three-dimensional spatial connected body as a three-dimensional segmentation block.

[0084] Specifically, iterative convergence is defined as a stable state where the magnitude of boundary contraction and expansion, the average physical property of blocks, and the extent of connected regions no longer change significantly. A three-dimensional spatially connected volume refers to a spatially continuous, homogeneous, closed-boundary, and topologically independent cluster of three-dimensional meshes, which serves as the final geological unit segmentation block. The boundary offset, the deviation of the average physical property of blocks, and the change in the extent of regional connectivity are recorded in real time after each iteration. A convergence threshold is set; when the boundary offset and physical property deviation of multiple consecutive iterations are all less than a preset minimum value, the iteration is considered to be converging. For example, five consecutive iterations satisfying the following conditions—maximum boundary offset < 0.01 mesh size and physical property deviation all less than a preset minimum value—represent a convergent state. After iterative convergence, isolated, sporadic small meshes and invalid fragmented blocks are removed, retaining valid three-dimensional connected regions with complete spatial connectivity and stable physical property characteristics. All converged and regularized independent spatially connected volumes are assigned a unique ID, and the spatial extent, boundary coordinates, and physical property parameters of each connected volume are recorded. Finally, a refined and standardized set of three-dimensional segmented blocks is output. The segmentation results closely match the distribution of real underground geological units, providing a high-precision unit basis for subsequent accurate stratigraphic calibration, geological model reconstruction, and ore body anomaly identification.

[0085] Step S430: Obtain prior geological data within the target detection area and extract the rock physical characteristic thresholds of each stratum from the prior geological data.

[0086] Specifically, the prior geological data consists of verified authoritative geological data for the target exploration area, including stratigraphic division data, borehole core logging data, rock physical property test data, rock mass and structural statistics, and mineral deposit geological characteristic data. Rock physical characteristic thresholds are the standard ranges, extreme ranges, and typical mean values ​​of density, magnetic susceptibility, resistivity, and polarizability corresponding to each stratum, lithology, and structural unit, serving as quantitative criteria for distinguishing different geological units. Complete stratigraphic layering data, borehole measured lithological data, indoor rock physical property test statistics, and regional geological structural data for the target exploration area are retrieved. Using the 3σ statistical anomaly removal criterion, abnormal noise data deviating from three times the standard deviation of the mean are removed, forming a standardized prior geological dataset. The data is grouped according to stratigraphic age, lithological type, structural attribute, and mineralization alteration type. The maximum, minimum, mean, and standard deviation of the four types of physical parameters for each group are statistically analyzed. A 95% confidence interval is selected as the effective threshold interval to exclude interference from extreme and accidental data. A directly accessible stratigraphic property threshold reference table was compiled, with the core thresholds as follows: normal sedimentary strata (density 2.3–2.6 g / cm³, magnetic susceptibility 0–0.02 SI, resistivity 500–5000 Ω·m, polarizability 0–3%); igneous intrusive bodies (density 2.6–2.9 g / cm³, magnetic susceptibility 0.04–0.1 SI, resistivity 300–1000 Ω·m, polarizability 3–6%); mineralized alteration zones (density ≥2.75 g / cm³, magnetic susceptibility ≥0.04 SI, resistivity ≤300 Ω·m, polarizability ≥6%). ​​This establishes a quantitative and standardized basis for stratigraphic lithology determination, providing accurate prior constraints for subsequent automatic similarity matching and intelligent stratigraphic calibration, significantly reducing the probability of lithological misjudgment and stratigraphic misclassification, and ensuring that geological calibration results conform to regional metallogenic and stratigraphic regularities.

[0087] Step S440: Calculate the similarity between the physical properties in each three-dimensional segmented block and the corresponding rock physical feature threshold, select the three-dimensional segmented block with the highest similarity to label the geological lithology, and obtain the stratigraphic label.

[0088] Specifically, the physical attributes of a 3D segmented block are the average of the four types of physical property parameters of all grids within a single 3D segmented block, forming a 4D block feature vector that characterizes the overall physical property features of the block. Similarity is the degree of matching between the current block's physical property distribution and the standard stratigraphic physical property threshold range; higher similarity indicates a higher degree of confidence in the block's affiliation to the corresponding stratum. Geological and lithological tags include standardized geological semantic tags containing stratigraphic age, lithological name, alteration type, and mineralization attributes.

[0089] The algorithm iterates through all 3D segmented blocks, calculating the global average values ​​of density, magnetic susceptibility, resistivity, and polarizability within each block to characterize its overall physical properties. The block's physical properties are then compared one by one with the physical thresholds of each stratigraphic rock extracted by S430. A Euclidean distance similarity algorithm is used to calculate the multi-parameter comprehensive similarity between the block and each type of stratum, with values ​​between 0 and 1. All stratigraphic similarity results for a single block are sorted, and the stratigraphic, lithological, and structural labels corresponding to the highest similarity are selected as the final geological attributes of that block, completing the single-block calibration. Blocks with similarity values ​​below a preset threshold (e.g., 0.7) are calibrated as having unknown lithology. By iterating through all 3D segmented blocks and performing similarity matching and label assignment for each, the entire model achieves fully automated stratigraphic and lithological calibration. It achieves fully automated, quantitative, and intelligent calibration of strata, lithology, and structures. Compared with single-parameter judgment, multi-parameter comprehensive similarity matching has higher recognition accuracy and stronger anti-interference ability. It can accurately distinguish complex geological units with similar physical properties but different lithologies, greatly improving the accuracy and reliability of geological models.

[0090] Step S450: Extract the contact interfaces between the three-dimensional segmented blocks after the strata are calibrated, perform spatial topology meshing and smoothing, and generate a three-dimensional geological structure model.

[0091] Specifically, the contact interface represents the spatial boundary between different strata, lithologies, and structurally segmented blocks, signifying the true contact boundary, bedding interface, and structural fault surface of a geological body. Spatial topology meshing reconstructs discrete interface points, lines, and surfaces into a continuous, closed, and conflict-free three-dimensional curved surface mesh topology. Surface smoothing is an optimization process that eliminates mesh jaggedness, local abrupt changes, and interface distortion, restoring the natural geological surface morphology. The Marching Cube 3D isosurface extraction algorithm is used to traverse all 3D segmented blocks with completed stratigraphic calibration, extracting the contact interfaces, boundary vertices, and contour segments between different geological units. A standardized topology verification algorithm is used to perform 3D topology meshing on the extracted interface elements, reconstructing the continuous topological relationships of stratigraphic interfaces, rock mass contact surfaces, and fault interfaces, and correcting topological errors such as model holes, spatial overlaps, and interface fractures. The Laplace smoothing algorithm is used to iteratively smooth all geological interfaces, with a fixed number of iterations of 15 and a smoothing coefficient of 0.2, eliminating jagged distortion and local abrupt deformation caused by meshing while preserving the original structural attitude and boundary morphology. All optimized geological interfaces and geological body units are integrated holistically to unify spatial topological relationships, ultimately generating a three-dimensional geological structural model with complete structure, smooth boundaries, geological rationality, and complete semantics. This model fully recreates the true spatial morphology and contact relationships of deep, complex strata, rock masses, and faults. The model conforms to geological laws and has high accuracy, providing a standard and reliable three-dimensional geological carrier for subsequent quantitative delineation of ore bodies.

[0092] Step S500: Based on the preset set of physical property index rules, the three-dimensional geological structure model is evaluated for ore-bearing potential in order to delineate the spatial location of ore bodies that conform to the rules.

[0093] In one embodiment, the physical property index rule set is a multi-parameter, multi-dimensional, and quantitative rule system for determining mineralization anomalies based on regional metallogenic regularities. It includes physical property thresholds, weighting coefficients, and metallogenic constraints. The spatial location of the ore body is a mineralized enrichment area underground that has mineralization potential, meets the rules for mineralization anomalies, and is spatially connected, and is characterized by a three-dimensional closed spatial range.

[0094] In one embodiment, a mineralization assessment is performed on a three-dimensional geological structure model based on a preset set of physical property index rules to delineate the spatial locations of ore bodies that conform to the rules, including but not limited to the following steps: Step S510: Spatial grid division is performed on the three-dimensional geological structure model to obtain multiple three-dimensional grid units. Based on the calibration data and predicted physical property parameters, a three-dimensional attribute matrix of each three-dimensional grid unit is generated.

[0095] Specifically, the 3D grid unit is a standardized regular grid unit derived from the secondary subdivision of a 3D geological structure model, serving as the smallest computational unit for quantitative assessment of orebody anomalies. The 3D attribute matrix is ​​a 3D data matrix indexed by grid spatial coordinates and containing four types of physical property parameters as attribute values, enabling structured and digital storage of the geological model's physical property information. For the 3D geological structure model after stratigraphic calibration and topology optimization, following a similar approach to step S200, an octree adaptive subdivision algorithm is used to spatially divide the 3D geological structure model into multiple 3D grid units; the specific process is not detailed here. For the subdivided regular grid, a 3D bilinear interpolation algorithm is used to accurately map the original model's predicted physical property parameters (density, magnetic susceptibility, resistivity, polarizability), as well as stratigraphic, lithological, alteration, and mineralization calibration labels, to each 3D grid unit. Based on the calibration data and predicted physical property parameters, a unique 3D spatial index coordinate is assigned to each 3D grid unit, establishing a one-to-one correspondence between "spatial coordinates - lithological calibration label - four-dimensional physical property value," thus completing the structured attribute binding of the grid unit. Based on the bound attributes of all 3D mesh units, a structured 3D attribute matrix is ​​generated. This matrix includes density values, magnetic susceptibility values, resistivity values, polarizability values, and geological lithology labels. The matrix dimensions strictly match the number of meshes across the entire domain. Through mesh partitioning and interpolation mapping, the visualized, unstructured geological entity model is transformed into a structured digital matrix that can be iteratively processed in batches by a computer. This provides standardized, accurate, and unified underlying data support for subsequent ore-forming rule matching and ore-bearing probability quantification calculations.

[0096] Step S520: Obtain prior geological data of the target detection area, and construct a preset set of physical property index rules based on the mineralization anomaly thresholds and multi-parameter aggregation weights of various physical property parameters set by the prior geological data.

[0097] Specifically, the mineralization anomaly threshold is used to determine the critical value range of various physical property parameters indicating the presence of mineralization, alteration, and favorable mineralization anomalies in a region. It serves as a quantitative standard for distinguishing between mineralized and non-mineralized areas. The multi-parameter aggregate weight is an importance coefficient assigned to each parameter when evaluating mineralization by integrating multiple physical property parameters. It reflects the indicative significance of different physical property indicators for a specific mineral type, and the weight values ​​are determined based on the genetic theory of mineral deposits and regional metallogenic laws.

[0098] Qualitative information such as the host strata, ore-bearing lithology, and ore-controlling structural types of the target mineral type is extracted from regional geological data. Characteristic descriptions such as ore mineral assemblage, wall rock alteration type, and ore body occurrence elements are extracted from ore deposit research data. Quantitative parameters such as ore body density range, magnetic mineral content, and resistivity variation patterns of mineralized zones are extracted from data of adjacent known mining areas. The extracted information is categorized and organized according to mineral type, deposit type, occurrence depth, and scale level to establish a priori knowledge data table. Global statistical analysis is performed on each physical property parameter in the three-dimensional attribute matrix to calculate the mean and standard deviation, determining the regional background characteristics. Then, based on the ore body physical property characteristics recorded in the prior geological data and combined with metallogenic geological theories, anomaly thresholds are set. For density parameters, for metallic sulfide deposits, the mineralization anomaly density threshold is set to the background mean plus 1.5 to 2 times the standard deviation. For different geological types, the multiple is adjusted based on the mean and standard deviation to set different anomaly thresholds. Based on the varying sensitivity of each physical property to mineralization, weights are assigned accordingly. For example, the weights are configured as follows: polarizability weight 0.4, resistivity weight 0.3, magnetic susceptibility weight 0.15, and density weight 0.15; the total weight is 1, perfectly aligning with the geological pattern of this region where "polarizability and resistivity are the main controlling factors for mineralization, while density and magnetization are auxiliary identification factors." Different weight values ​​can be set for different ore bodies. The mineralization anomaly threshold range, multi-parameter aggregated weights, and mineralization probability scores are integrated and encapsulated to form a pre-defined set of physical property indicators that can be directly called upon, serving as the basis for subsequent grid-based mineralization assessment. Multi-parameter weighted aggregation can maximize the highlighting of comprehensive mineralization anomaly characteristics, significantly improving the targeting and accuracy of mineralization anomaly identification.

[0099] Step S530: Based on the preset set of physical property index rules and the three-dimensional attribute matrix, calculate the mineralization probability score of each three-dimensional grid unit, and mark the three-dimensional grid units with mineralization probability scores higher than the preset confidence threshold as a set of candidate mineralization units.

[0100] Specifically, the mineralization probability score is a quantitative score calculated by comprehensively considering the matching degree of multiple physical property anomalies, weight contribution, and geological adaptability. The higher the score, the greater the mineralization probability. The preset confidence threshold is the critical score for distinguishing between favorable mineralization units and background units, with a value of 0.65, used to screen effective mineralization units. The candidate mineralization unit set refers to the cluster of all three-dimensional grid units with a mineralization probability higher than the confidence threshold, representing potential favorable mineralization areas. The preset physical property index rule set is retrieved, and all three-dimensional grid units within the three-dimensional attribute matrix are traversed one by one. Each grid unit is judged to determine whether its four types of physical property parameters meet the mineralization anomaly threshold conditions fixed by S520. The single-parameter anomaly matching status is recorded, and grid units with basic mineralization anomaly characteristics are screened out. A weighted scoring formula is used to calculate the ore-bearing probability score for each grid in the 0-1 interval. Based on set weights, a weighted sum is performed, and physical properties such as polarizability and resistivity are normalized. Then, the score is multiplied by the normalized polarizability (0.4), resistivity (0.3), magnetic susceptibility (0.15), and density (0.15) to obtain the ore-bearing probability score. The closer the parameters are to the optimal ore-bearing interval, the higher the score of each individual parameter and the higher the overall probability score. 3D grid cells with ore-bearing probability scores higher than a pre-set confidence threshold are selected. All qualified and valid mineralized grids are uniformly marked, numbered, and summarized. Scattered, low-matching abnormal grids are removed to generate a global candidate mineralized unit set. High-confidence mineralized anomaly units are selected using prior knowledge, eliminating a large amount of background noise and false anomalies, providing clean and effective preliminary data for subsequent accurate ore body clustering and delineation.

[0101] Step S540: Based on the spatial topological connectivity of each three-dimensional grid unit in the candidate mineralization unit set, perform three-dimensional spatial clustering, extract three-dimensional closed surfaces based on the main connected components after clustering, and use the spatial range of the three-dimensional closed surface envelope as the spatial location of the ore body that conforms to the rules.

[0102] Specifically, spatial topological connectivity refers to the three-dimensional spatial connectivity between grid cells in terms of vertical, horizontal, and front-back directions. Mineralized cells corresponding to the same ore body possess continuous connectivity characteristics. Three-dimensional spatial clustering is a spatial aggregation algorithm that aggregates spatially connected candidate mineralized cells with similar characteristics into independent mineralized clusters. A three-dimensional closed surface refers to a continuous, closed, and smooth three-dimensional outer surface that encloses a mineralized cluster, used to characterize the boundary morphology and spatial extent of the ore body.

[0103] A three-dimensional six-neighborhood connected clustering algorithm is adopted to traverse all grids in the candidate mineralization unit set. Grids that are in contact with each other on the six positive surfaces (top, bottom, left, right, front, and back) are identified as connected and associated units. Spatially continuous and interconnected discrete mineralization grids are aggregated into independent three-dimensional mineralization clusters. At the same time, a minimum unit filtering threshold is set to remove isolated and scattered pseudo-anomaly clusters with less than 10 grids, thus completely eliminating single-point noise anomaly interference.

[0104] Three-dimensional quantitative indicators were used to evaluate all effective clustered mineralization clusters. The total number of grids, average mineralization probability, and spatial continuity were used to score and rank the clusters. The total number of grids, average mineralization probability (calculated from the mineralization probability score), and spatial continuity of each cluster were statistically analyzed. The cluster with the largest grid size, the highest overall mineralization probability, and the most complete spatial morphology was selected as the main mineralization interconnection.

[0105] Among them, the degree of spatial continuity is quantified by the three-dimensional spatial compact connectivity rate. Specifically, the actual number of effective six-neighbor connected pairs within a single cluster is counted and compared with the theoretical maximum number of connected pairs of the cluster to obtain the quantified value of spatial continuity. The value of spatial continuity ranges from 0 to 1. The closer the value is to 1, the more complete the spatial morphology of the mineralized body, the fewer the fractures, and the lower the degree of dispersion.

[0106] A Poisson surface reconstruction algorithm is employed to perform global surface fitting on the spatial contour point cloud and boundary mesh of the main connected body. The algorithm possesses strong hole repair and boundary optimization capabilities, automatically generating a smooth 3D enclosing surface without gaps, distortions, or topological closure, accurately conforming to the true spatial morphology of the mineralized body. The entire 3D spatial range enveloped by the 3D closed surface is defined as the final spatial location of the ore body conforming to the mineralization rules. Simultaneously, quantitative parameters such as the 3D morphology, spatial depth, boundary coordinates, and spatial volume range of the ore body are output, completing fully automated ore body delineation. 3D spatial clustering thoroughly eliminates scattered interference anomalies, preserving the true mineralized enrichment connected bodies and effectively avoiding false anomaly interference. The surface accurately restores the true spatial morphology and boundary range of the ore body, achieving fully automated, high-precision, and quantitative delineation of deep concealed ore bodies, significantly improving the accuracy of deep mineral exploration and providing core technical support for subsequent resource evaluation, engineering layout, and breakthroughs in deep mineral exploration.

[0107] It should be noted that, in order to ensure that data with different dimensions can be calculated, data mapping was used to eliminate the difference in dimensions during the calculation process mentioned above. Any calculations not mentioned above will be explained here.

[0108] like Figure 3As shown, this application provides a gravity, magnetic, and electric multi-physics field coupling modeling and ore body positioning system 100. This system 100 acquires observational data, historical measurement data, and physical geological information of the target detection area through a data acquisition module 110. The observational data includes gravity data, magnetic data, and electrical data. The gravity, magnetic, and electrical data are preprocessed to obtain corresponding gravity, magnetic, and electrical data volumes. A mesh generation module 120 divides the preset initial three-dimensional spatial model into multiple hexahedral meshes based on the spatial range and detection depth of the target detection area. Historical measurement data and physical geological information are used to initialize and configure each hexahedral mesh, including density, magnetic susceptibility, and... The initial physical property parameters of resistivity are obtained. The gravity data volume, magnetic data volume, and electrical data volume are input into a preset target deep learning model using the inversion prediction module 130 for multimodal feature extraction and fusion inversion. The output includes predicted physical property parameters containing density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. The initial physical property parameters are adjusted using these predicted parameters to output a three-dimensional coupled physical property model. The target deep learning model is used to spatially and consistently couple the gravity data volume, magnetic data volume, and electrical data volume. The three-dimensional coupled physical property model is multi-scale segmented and stratigraphically calibrated using the model construction module 140 to obtain a three-dimensional geological structure model. The spatial delineation module 150 assesses the ore-bearing potential of the three-dimensional geological structure model according to a preset set of physical property index rules, thereby delineating the spatial locations of ore bodies that conform to the rules.

[0109] It should be noted that the data acquisition module 110 is connected to the grid division module 120, the grid division module 120 is connected to the inversion prediction module 130, the inversion prediction module 130 is connected to the model construction module 140, and the model construction module 140 is connected to the spatial delineation module 150.

[0110] It should also be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0111] This application also discloses an electronic device. (See reference...) Figure 4 , Figure 4This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0112] The communication bus 502 is used to enable communication between these components.

[0113] The user interface 503 may include a display screen and a camera. Optionally, the user interface 503 may also include a standard wired interface and a wireless interface.

[0114] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0115] The processor 501 may include one or more processing cores. The processor 501 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 505, and by calling data stored in memory 505. Optionally, the processor 501 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array. The processor 501 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and Modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 501.

[0116] The memory 505 may include random access memory (RAM) or read-only memory. Optionally, the memory 505 may include a non-transitory computer-readable storage medium. The memory 505 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 505 may also be at least one storage device located remotely from the aforementioned processor 501. (Refer to...) Figure 4 The memory 505, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a modeling and ore body location method based on multi-physics coupling of gravity, magnetoelectric, and gravity fields.

[0117] exist Figure 4 In the illustrated electronic device 500, the user interface 503 is mainly used to provide an input interface for the user and acquire user input data; while the processor 501 can be used to call an application program stored in the memory 505 for a modeling and ore body location method of gravity, magnetoelectric, and multiphysics coupling. When executed by one or more processors 501, the electronic device 500 performs one or more methods as described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0118] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0119] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0122] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0123] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will readily conceive of those skilled in the art upon consideration of the specification and the disclosure of practical truths.

[0124] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for modeling and ore body localization using multiphysics coupling of gravity, magnetism, and electricity, characterized in that, The method includes: Acquire observational data of the target detection area, historical measurement data of the target detection area, and physical geological information, wherein the observational data includes gravity data, magnetic data, and electrical data, and preprocess the gravity data, magnetic data, and electrical data to obtain corresponding gravity data volumes, magnetic data volumes, and electrical data volumes; Based on the spatial range and depth of the target detection area, the preset initial three-dimensional spatial model is divided into multiple hexahedral grids. The historical measurement data and the physical geological information are used to initialize and configure the initial physical property parameters, including density, magnetic susceptibility and resistivity, for each hexahedral grid. The gravity data volume, the magnetic data volume, and the electrical data volume are input into a preset target deep learning model for multimodal feature extraction and fusion inversion. The model outputs predicted physical property parameters including density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. The initial physical property parameters are adjusted using the predicted physical property parameters to output a three-dimensional coupled physical property model. The target deep learning model is used to spatially and consistently couple the gravity data volume, the magnetic data volume, and the electrical data volume. The three-dimensional coupled physical property model is segmented at multiple scales and its stratigraphic calibration is performed to obtain a three-dimensional geological structure model; Based on a preset set of physical property index rules, the three-dimensional geological structure model is evaluated for ore-bearing potential in order to delineate the spatial location of ore bodies that conform to the rules.

2. The method according to claim 1, characterized in that, The method further includes a training process for the target deep learning model, the training process comprising: Acquire training observation data from multiple training detection areas, training historical data corresponding to the training observation data, standard observation data, and training geological information; For each training detection area, the preset training three-dimensional spatial model is divided into multiple training hexahedral meshes according to the spatial range and detection depth of the training detection area. The training historical data and the training geological information are used to initialize and configure the training physical property parameters for each training hexahedral mesh. The training observation data is input into a preset initial deep learning model for multimodal feature extraction and fusion inversion to obtain predicted inverted physical property parameters; With the constraint that the predicted inversion physical property parameters are highly fitted to the training three-dimensional spatial model, the loss value of the joint objective function is calculated based on the three-channel feature vector extracted by the initial deep learning model, the predicted inversion physical property parameters, the training geological information, and the standard observation data. The parameters of the initial deep learning model are adjusted using the loss value of the joint objective function to obtain the target deep learning model.

3. The method according to claim 2, characterized in that, The loss value of the joint objective function is calculated based on the three-channel feature vector extracted from the initial deep learning model, the predicted inverted physical property parameters, the training geological information, and the standard observation data, including: The similarity between the predicted observation data and the standard observation data is calculated to obtain the data fitting similarity, wherein the predicted observation data is obtained by performing forward modeling on the physical property distribution data in the predicted inversion physical property parameters; The spatial similarity between each pair of feature vectors in the three channels is calculated, and the similarity values ​​are fused to obtain the spatial fitting similarity. Based on the training coordinate parameters in the predicted inversion physical property parameters, the training geological information, and the training historical data, the gradient change of the physical property distribution data in the predicted inversion physical property parameters between adjacent grids is calculated to obtain the value of the model constraint term used to constrain spatial continuity. The loss value of the joint objective function is obtained by weighting and fusing the data fitting similarity, the spatial fitting similarity, and the values ​​of the model constraint terms.

4. The method according to any one of claims 1-3, characterized in that, The structure of the target deep learning model includes a three-branch encoder, an implicit feature alignment network, and a feature fusion layer; The step of inputting the gravity data volume, the magnetic data volume, and the electrical data volume into a preset target deep learning model for multimodal feature extraction and fusion inversion, and outputting predicted physical property parameters, includes: The gravity data volume, the magnetic data volume, and the electrical data volume are respectively input into the corresponding encoder for three-dimensional feature extraction to obtain the initial gravity spatial features, magnetic spatial features, and electrical spatial features; The implicit feature alignment network is used to map the gravity space features, magnetic space features and electric space features to a unified high-dimensional implicit space for tensor alignment, and output standardized extracted features. The standardized extracted features are spliced ​​and reconstructed using the feature fusion layer to output the predicted physical property parameters.

5. The method according to claim 1, characterized in that, The process of performing multi-scale segmentation and stratigraphic calibration on the three-dimensional coupled physical property model to obtain a three-dimensional geological structure model includes: The three-dimensional coupled physical property model is subjected to bottom-up spatial downsampling of multiple solid meshes to generate a multi-scale physical property spatial pyramid from fine resolution to coarse resolution. At different resolution levels of the multi-scale physical property space pyramid, three-dimensional region connectivity segmentation is performed based on the coupling physical property change gradient between adjacent physical meshes, and the boundary is refined from coarse to fine to generate multiple three-dimensional segmentation blocks. Acquire prior geological data within the target detection area, and extract the rock physical characteristic thresholds of each stratum in the prior geological data; Calculate the similarity between the physical properties in each of the three-dimensional segmented blocks and the corresponding rock physical feature thresholds, select the three-dimensional segmented block with the maximum similarity to label the geological lithology, and obtain the stratigraphic label; The contact interfaces between the three-dimensional segmented blocks after the strata are calibrated are extracted, and spatial topology meshing and smoothing are performed to generate the three-dimensional geological structure model.

6. The method according to claim 5, characterized in that, The method involves segmenting the three-dimensional region connectivity based on the gradient of coupling property changes between adjacent entity meshes, refining the boundaries from coarse to fine, and generating multiple three-dimensional segmentation blocks, including: Based on the coarse resolution level of the multi-scale physical property spatial pyramid, the weighted fusion gradient of multiple physical properties is calculated as the edge weight, and a preset region segmentation algorithm is used to extract the coarse-scale initial block. The boundary of the coarse-scale initial block is projected onto the adjacent fine-resolution layer as the initial seed surface. The property change gradient under the adjacent fine-resolution layer is calculated. Based on the property change gradient, the initial seed surface is subjected to supervoxel clustering shrinkage or expansion iteration until the bottom fine-resolution layer is reached. When the contraction or expansion iterations tend to converge, the resulting three-dimensional spatial connected volume is output as the three-dimensional segmentation block.

7. The method according to claim 1, characterized in that, The step of assessing the ore-bearing potential of the three-dimensional geological structure model based on a preset set of physical property index rules, in order to delineate the spatial location of ore bodies that conform to the rules, includes: The three-dimensional geological structure model is divided into spatial grids to obtain multiple three-dimensional grid units. Based on the calibration data and the predicted physical property parameters, a three-dimensional attribute matrix of each three-dimensional grid unit is generated. Acquire prior geological data of the target detection area, and construct the preset set of physical property index rules based on the mineral anomaly thresholds and multi-parameter aggregation weights of various physical property parameters set by the prior geological data. Based on the preset set of physical property index rules and the three-dimensional attribute matrix, the mineralization probability score of each of the three-dimensional grid units is calculated, and the three-dimensional grid units with mineralization probability scores higher than the preset confidence threshold are marked as a set of candidate mineralization units. Based on the spatial topological connectivity of each three-dimensional grid unit within the candidate mineralization unit set, three-dimensional spatial clustering is performed. Based on the main connected body after clustering, a three-dimensional closed surface is extracted, and the spatial range of the three-dimensional closed surface envelope is used as the spatial location of the ore body that conforms to the rules.

8. A modeling and ore body positioning system using multiphysics coupling of gravity, magnetism, and electricity, characterized in that, The system includes: The data acquisition module is used to acquire observation data of the target detection area, historical measurement data of the target detection area, and physical geological information. The observation data includes gravity data, magnetic data, and electrical data. The module preprocesses the gravity data, magnetic data, and electrical data to obtain corresponding gravity data volumes, magnetic data volumes, and electrical data volumes. The mesh generation module is used to divide the preset initial three-dimensional spatial model into multiple hexahedral meshes according to the spatial range and detection depth of the target detection area. The module uses the historical measurement data and the physical geological information to initialize and configure the initial physical property parameters, including density, magnetic susceptibility and resistivity, for each hexahedral mesh. The inversion prediction module is used to input the gravity data volume, the magnetic data volume, and the electrical data volume into a preset target deep learning model for multimodal feature extraction and fusion inversion, and output predicted physical property parameters including density distribution, magnetic susceptibility distribution, resistivity distribution, and coordinate parameters. The initial physical property parameters are adjusted using the predicted physical property parameters, and a three-dimensional coupled physical property model is output. The target deep learning model is used to perform spatially consistent coupling of the gravity data volume, the magnetic data volume, and the electrical data volume. The model construction module is used to perform multi-scale segmentation and stratigraphic calibration on the three-dimensional coupled physical property model to obtain a three-dimensional geological structure model. The spatial delineation module is used to assess the mineralization of the three-dimensional geological structure model according to a preset set of physical property index rules, so as to delineate the spatial location of ore bodies that conform to the rules.

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

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images