Three-dimensional imaging exploration method and system for geological deep resources

Abstract

The invention belongs to the technical field of geological exploration, and particularly relates to a three-dimensional imaging exploration method and system for geological deep resources. The multi-source collaborative data acquisition module, the full-space velocity modeling module, the intelligent joint inversion module and the dynamic three-dimensional visualization and interpretation module are connected with the high-performance computing unit through a high-speed data network to form an integrated imaging processing closed loop. The continuity and reliability of deep geologic body spatial distribution feature description are remarkably improved, and a visual and accurate three-dimensional basis is provided for resource evaluation and exploration decision making.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of geological exploration, and particularly relates to a three-dimensional imaging exploration method and system for deep geological resources. BACKGROUND

[0002] With the increasing demand for deep mineral and energy resource exploration, high-precision and high-resolution three-dimensional imaging of deep geological structures has become a key technical direction in the field of geological exploration. Geological deep resource exploration, as a foundation for national resource security and strategic emerging industry development, is evolving its technical system from traditional point sampling or two-dimensional profile analysis to full-space, multi-physical field, and intelligent three-dimensional imaging. Currently, although multi-source detection methods such as seismic, gravity, electromagnetic, drilling, and remote sensing are widely used in deep geological research at the theoretical level, they have not yet formed a unified data collaboration mechanism and three-dimensional modeling closed loop in actual engineering, making it difficult to systematically, continuously, and accurately depict the spatial distribution characteristics, physical parameter distribution, and structural relationships of deep geological bodies.

[0003] The three-dimensional imaging-based geological deep resource exploration method aims to build a three-dimensional perception system covering the entire space, penetrating the depth, and integrating multi-physical field information, to achieve high-fidelity reconstruction of complex geological targets such as concealed ore bodies, fault zones, and fold structures. The core of this direction is to establish an integrated imaging framework with adaptive velocity modeling capability, multi-source data spatio-temporal alignment mechanism, and intelligent inversion interpretation function. However, existing technical solutions are generally limited to local optimization of a single detection method or quasi-three-dimensional processing in specific scenarios, lacking full-chain three-dimensional imaging capability for deep, wide, and heterogeneous environments.

[0004] Although some patents have improved imaging algorithms or drilling arrangements, such as using prolate spheroidal coordinates to improve seismic migration accuracy or conducting cross-hole CT detection through inclined boreholes to reduce blind areas, they have not broken through the three bottlenecks of dimension, scale and fusion. Specifically, the imaging dimension is still mainly two-dimensional profile or strip region, which cannot generate a truly three-dimensional geological model in the whole space; there is a lack of unified physical constraints and semantic association between multi-source data, and heterogeneous information such as seismic, gravity and electromagnetic is processed independently, resulting in inconsistent physical interpretation and blurred structural boundaries; field deployment relies on regular arrays or dense boreholes, which is difficult to implement in complex terrain or ecologically sensitive areas, and is costly and time-consuming; although intelligent methods such as neural networks are introduced in the inversion process, there is no adaptive fusion mechanism for multi-scale geological features, making it difficult to cope with strong deep medium heterogeneity and low signal-to-noise ratio. The above defects make the existing system have problems such as insufficient imaging resolution, poor spatial continuity and weak interpretation reliability in key tasks such as deep resource exploration, complex structure analysis and regional resource potential evaluation, and a three-dimensional imaging exploration method and system for deep geological resources are needed to break through the integration of multi-source collaborative acquisition, full-space velocity modeling, intelligent joint inversion and dynamic three-dimensional visualization. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a three-dimensional imaging exploration method and system for deep geological resources, which can effectively solve the problems in the background art.

[0006] To achieve the above purpose, the present application provides the following technical scheme: A three-dimensional imaging exploration system for deep geological resources, comprising the following components: A multi-source collaborative data acquisition module is used to deploy and control seismic, gravity, electromagnetic, drilling and remote sensing detection equipment to realize synchronous or quasi-synchronous acquisition of multi-physical field data; A full-space velocity modeling module is used to construct a three-dimensional velocity field model covering the exploration area based on the initial geological model and multi-source data; An intelligent joint inversion module is used to input multi-source data and velocity field model, and to invert the distribution of geological body physical parameters through a fusion mechanism of physical constraints and data driving; A dynamic three-dimensional visualization and interpretation module is used to reconstruct and render the inversion results, and to support geological body boundary identification and attribute labeling; each module is connected through a high-speed data network and a high-performance computing unit to form an integrated imaging processing closed loop.

[0007] The multi-source collaborative data acquisition module includes a distributed sensor network and a central scheduling unit. The distributed sensor network consists of a seismic detector array, gravity sensor nodes, electromagnetic transmitter-receiver pairs, and borehole sensors. The node spacing is adaptively adjusted according to the scale of the exploration target, ranging from 10 meters to 500 meters. The central scheduling unit adopts a time synchronization protocol to ensure that the time reference error of data acquisition from different physical fields is less than 1 millisecond, and realizes real-time data transmission through wireless self-organizing network technology. The module has a built-in terrain adaptive deployment algorithm that automatically optimizes the sensor deployment position according to the digital elevation model to avoid complex terrain obstacles.

[0008] The full-space velocity modeling module employs multi-scale grid subdivision technology. The module divides the exploration area into a background grid and a locally refined grid. The background grid is used to describe large-scale geological structures, with a grid size of 100m × 100m × 50m. The locally refined grid focuses on known anomaly areas or target bodies, with a grid size refined to 10m × 10m × 5m. The velocity modeling process is based on a strategy combining first-arrival travel-time tomography and full-waveform inversion. First, an initial velocity model is constructed using seismic first-arrival travel-time data through ray tracing and least-squares inversion. Then, full-waveform inversion is introduced to iteratively optimize the model details. The objective function is defined as the L2 norm of the residual between the observed waveform and the simulated waveform.

[0009] The intelligent joint inversion module constructs a multi-physics coupled inversion framework. This framework includes seismic elastic wave inversion sub-modules, gravity density inversion sub-modules, and electromagnetic resistivity inversion sub-modules. Each sub-module is coupled through cross-gradient constraint terms to ensure that models with different physical property parameters maintain consistency in spatial structure. A deep learning surrogate model is introduced into the inversion process to accelerate forward modeling calculations. The surrogate model adopts the U-Net architecture, taking velocity, density, and resistivity parameters as inputs and directly outputting simulated seismic records, gravity anomalies, and electromagnetic response data. The training data, exceeding 100,000 sets, is generated through numerical simulation. The objective function for the joint inversion is:

[0010] Where m = [m1, m2, m3] represents the velocity, density, and resistivity model vectors, respectively. and These are observational and forecast data, respectively. For data weighting coefficients, Here, are the cross-gradient constraint weights, where, The gradient operator is a three-dimensional space operator, and the central difference method is used for discrete calculation. The physical property parameter models m1, m2, and m3 need to be pre-normalized to the same dimension range (such as [0,1]) to ensure comparability. The optimization algorithm adopts the finite memory quasi-Newton method.

[0011] The dynamic three-dimensional visualization and interpretation module is based on volume rendering and isosurface extraction technology; the module converts the physical property parameter volume data obtained by inversion into a three-dimensional voxel model, each voxel stores a velocity value, a density value, a resistivity value and a corresponding lithology classification probability; the visualization engine supports transparency transfer function adjustment, enabling perspective observation of the internal structure of the geological body; the interpretation function integrates edge detection algorithms and region growing algorithms to automatically identify and delineate velocity sudden change interfaces, density anomaly bodies and low resistance area boundaries, and the identification results are used as candidate targets for three-dimensional labeling of potential ore bodies or structures.

[0012] A three-dimensional imaging exploration method for deep geological resources is also provided, and the specific steps are as follows: Step one, plan and implement multi-source geophysical data collaborative acquisition to obtain seismic wave field data, gravity field data, electromagnetic field data and drilling core data in the exploration area; Step two, based on the collected data and prior geological information, a full-space three-dimensional initial velocity model of the exploration area is constructed; Step three, use a deep learning agent model to accelerate multi-physical field forward modeling, and establish a joint inversion objective function for seismic, gravity and electromagnetic data; Step four, use an optimization algorithm to solve the joint inversion objective function to iteratively update the three-dimensional physical property parameter model of velocity, density and resistivity; Step five, three-dimensional reconstruction and visualization rendering are performed on the three-dimensional physical property parameter model obtained by inversion, and the spatial form and physical property characteristics of the geological body are automatically extracted and interpreted.

[0013] The multi-source data collaborative acquisition in step one uses a "star-ground" collaborative observation mode; airborne platforms carry magnetometers and gamma spectrometers for regional magnetic and radioactive measurements, and distributed nodes are deployed on the ground to receive seismic signals and controllable source electromagnetic signals excited by artificial sources; drilling data are used as "hard data" for calibration and constraint, directional drilling is implemented at key structural locations, cores are obtained and well geophysical logging is performed; all data are recorded with accurate GPS positions and UTC times during acquisition, and are converted to a unified format through a data standardization interface.

[0014] The full-space three-dimensional initial velocity model in step two is constructed using a combination of tomography and geostatistics; first, direct wave tomography is performed using seismic first arrival wave travel time to obtain shallow velocity structure; second, the density model is converted to a P-wave velocity model through the Poisson's ratio empirical relationship combined with regional gravity Bouguer anomaly data to constrain the middle-deep velocity; finally, borehole acoustic logging data are introduced as known points, and the Kriging interpolation algorithm is used for spatial interpolation and smoothing of the velocity model to generate an initial three-dimensional velocity grid model with variable resolution.

[0015] The training process of the deep learning agent model in step three includes: generating a large number of three-dimensional velocity-density-resistivity model combinations based on a known geological model library; using high-precision finite difference or finite element method forward calculation to calculate corresponding seismic records, gravity anomalies and electromagnetic responses to form a training data set; the input of the U-Net agent model is a three-dimensional data body of physical parameters, and the output is a corresponding simulated observation data body; the model loss function adopts mean square error, and the Adam optimizer is used for training until the prediction error on the validation set is less than 5%, the initial learning rate is set to 0.001 during training, the batch size is 32, and the maximum iteration number is 100,000; the validation set accounts for 20% of the total data set, and the prediction error is measured by normalized root mean square error (NRMSE), and the training is terminated when the NRMSE of the validation set is less than or equal to 5%.

[0016] In step four, the joint inversion iterative process adopts a multi-stage strategy; in the first stage, seismic data is mainly used to optimize the velocity structure of shallow high-speed and low-speed layers; in the second stage, gravity and electromagnetic data are introduced to invert the density and resistivity distribution in the middle and deep parts under the constraint of the velocity model; in the third stage, full data joint fine inversion is carried out, at this time the cross gradient constraint weight β is increased, forcing the structure of different physical property models to be consistent; after each iteration, the forward response is quickly updated using the agent model, and the residual error is calculated by comparing with the actual observation data, and the iteration is terminated when the residual error reduction rate is less than 1e-4.

[0017] In step five, three-dimensional visualization and interpretation includes volume rendering and feature extraction; volume rendering uses the ray casting algorithm, assigns colors and transparencies according to physical parameter values, and generates three-dimensional perspective images; feature extraction identifies gradient threshold τ by calculating the gradient modulus of the physical parameter body, and takes the voxel with a value range of 0.1-0.3 as the geological boundary point by calibrating the known geological model, and then connects the boundary points into a continuous three-dimensional surface through clustering algorithm, so as to delineate the geometric shape and spatial position of geological targets such as ore bodies, faults and lithological contact zones The present application has the following beneficial effects: through multi-source collaborative data acquisition and full-space velocity modeling, a high-precision, full-coverage three-dimensional initial model is constructed, which lays a reliable foundation for deep imaging, and overcomes the defects of strong model dependence and insufficient shallow constraint of traditional methods; an intelligent joint inversion framework combining physical constraints and data-driven is adopted, and a deep learning agent model is introduced to accelerate, which greatly improves the inversion efficiency and the consistency of multi-physical parameter interpretation, and solves the structural ambiguity problem caused by independent processing of heterogeneous data; dynamic three-dimensional visualization and automatic interpretation function realizes the direct conversion from data to geological cognition, significantly improves the continuity and reliability of the spatial distribution characteristics of deep geological bodies, and provides intuitive and accurate three-dimensional basis for resource evaluation and exploration decision. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1is the overall technical scheme architecture schematic diagram of the geological deep resource three-dimensional imaging exploration method and system proposed in the application; Figure 2 is the core principle framework schematic diagram of the full-space velocity modeling and intelligent joint inversion in the application; Figure 3 is the distributed sensing network and central scheduling logic framework schematic diagram of the multi-source collaborative data acquisition module in the application; Figure 4 is the volume rendering and geological feature automatic extraction logic flow framework diagram of the dynamic three-dimensional visualization and interpretation module in the application. DETAILED DESCRIPTION

[0019] In the deep prospecting exploration project of the metal ore concentration area, the multi-source collaborative data acquisition module deploys and acquires data of the distributed sensing network according to the exploration task. Referring to Figure 3 , the central scheduling unit first loads the digital elevation model of the exploration area, and the terrain adaptive deployment algorithm automatically calculates the optimal layout position of the sensor nodes according to the model, avoiding steep mountains and river obstacles. In the distributed sensing network, the seismic detector array is arranged in a grid shape on the ground surface with a spacing of 10 meters, a total of 1024 nodes, for receiving the P-wave and S-wave signals excited by the artificial seismic source; the gravity sensor nodes are arranged with a spacing of 50 meters, measuring the vertical component of the gravity field in the measurement area, and the measurement accuracy is 0.01 milligal; the electromagnetic transmitting-receiving pair is composed of one high-power controllable source audio magnetotelluric transmitter and 32 receiving stations, and the transmission frequency range is 0.1 Hz to 10 kHz; the borehole sensors are deployed in the key verification hole positions, including three-component geophones and resistivity logging instruments. The central scheduling unit adopts the precision time protocol, and provides a unified time reference for all nodes through the GPS tame clock, ensuring that the time synchronization error of seismic data and electromagnetic data acquisition is less than 0.5 milliseconds. The data is transmitted back to the central processing station in real time through the wireless ad hoc network, and the transmission rate is not less than 10 megabits per second, and the data packet is attached with accurate longitude and latitude coordinates and altitude information.

[0020] The full-space velocity modeling module receives raw data from the multi-source collaborative data acquisition module. The module first preprocesses the seismic first arrival travel time data, including denoising, gain recovery and first arrival picking, with a picking accuracy of one quarter of the sampling interval. Based on the first arrival travel time, a ray tracing tomography method is used to construct an initial shallow velocity model. Ray tracing uses the shortest path algorithm to calculate the theoretical travel time from the source to each receiver, and compares it with the observed travel time to form travel time residuals. Tomographic inversion solves the velocity model using the least squares method, discretizes the velocity model into a background grid, and sets the grid size to 100 meters x 100 meters x 50 meters. The inversion parameter is the P-wave velocity value of each grid node. After 10 iterations of inversion, the root mean square of travel time residuals decreases from 120 milliseconds to 15 milliseconds, and the velocity structure in the shallow 0-1000 meter depth range is obtained. Subsequently, the module introduces regional Bouguer gravity anomaly data, converts the density anomaly to velocity perturbation through an empirical formula, and uses it to constrain the velocity trend in the medium depth range of 500-3000 meters. Finally, the acoustic logging velocity data of the three verification drill holes are used as hard data points, and the ordinary kriging interpolation algorithm is used to spatially interpolate and smooth the full-area velocity model to generate a full-space three-dimensional initial velocity grid model. The velocity prediction error of the model at the drill hole is less than 5%.

[0021] The intelligent joint inversion module loads the initial velocity model output by the full-space velocity modeling module and the multi-source observation data. Referring to Figure 2The module first starts the deep learning surrogate model for forward calculation acceleration. The surrogate model adopts the U-Net architecture, and the input layer receives a three-dimensional velocity, density, and resistivity data body with a size of 128x128x64. After 4 times of downsampling and 4 times of upsampling operations, the output is a simulated seismic record, gravity anomaly, and electromagnetic response data body with a corresponding size. The surrogate model has been trained based on a data set generated by more than 100,000 numerical simulations. During training, the Adam optimizer is used with an initial learning rate of 0.001 and a batch size of 32. The mean square error between the predicted data on the independent validation set and the high-precision finite difference forward result is less than 3%. In joint inversion, the surrogate model replaces the traditional numerical forward, and the computational efficiency is improved by about two orders of magnitude. The joint inversion framework builds a multi-physical field coupling objective function, where the data weight coefficients are set according to the data signal-to-noise ratio. The seismic data weight α1 is 0.5, the gravity data weight α2 is 0.3, and the electromagnetic data weight α3 is 0.2. The initial value of the cross-gradient constraint weight β is set to 0.1. The inversion adopts a multi-stage strategy: in the first stage, the density and resistivity models are fixed, and only the velocity model is optimized. The velocity values of shallow high-speed mineralization bodies and low-speed fractured zones are mainly adjusted, and the iteration is performed for 20 times. In the second stage, under the constraint of the updated velocity model, the density and resistivity models are simultaneously inverted, and the iteration is performed for 30 times. In the third stage, full-parameter joint fine inversion is performed. At this time, the cross-gradient constraint weight β is increased to 0.5, forcing different physical models to maintain consistency in spatial gradients, and the iteration is performed for 50 times. The optimization algorithm uses the limited memory quasi-Newton method. After each iteration, the forward response is quickly updated by the surrogate model and the residual is calculated. When the target function value decreases by less than 1e-4 for 3 consecutive times, the inversion is terminated. Finally, a high-resolution velocity, density, and resistivity three-dimensional physical property parameter distribution model is output.

[0022] The dynamic three-dimensional visualization and interpretation module processes and displays the three-dimensional physical property parameter volume data output by the intelligent joint inversion module. Referring to Figure 4 , the module first normalizes the velocity, density, and resistivity three data volumes and fuses them into a comprehensive data volume. Each voxel contains three normalized physical values and a lithology probability index calculated according to an empirical formula. The visualization engine uses the ray casting volume rendering algorithm. From each pixel on the screen, a light ray is emitted through the three-dimensional data volume. The voxel values along the light ray path are sampled. The velocity value is mapped to the red-yellow color spectrum, the density value is mapped to the blue-green color spectrum, and the resistivity value is mapped to the gray-white color spectrum through the color transfer function. The visibility of different physical intervals is controlled through the transparency transfer function, thereby realizing the internal perspective effect of the geological structure. The interpretation function is automatically started. The edge detection algorithm calculates the three-dimensional gradient modulus of each voxel, which is:

[0023] where V represents the value of the physical property. The gradient threshold τ is set to 0.2, and voxels with gradient magnitude greater than this threshold are marked as geological boundary points. Subsequently, a density-based spatial clustering algorithm is used to cluster these boundary points, and boundary points with similar spatial positions and physical properties are merged into the same geological body. After clustering, the module uses the marching cubes algorithm to extract triangular mesh surfaces from each cluster, thereby reconstructing the geometric morphology of the geological targets such as ore body boundaries, fault surfaces, and lithological contact surfaces. The system automatically adds labels to each identified three-dimensional geological body, including its center coordinates, approximate volume, average physical property value, and geological interpretation type, and superimposes the results on the three-dimensional perspective view. The interpretation results can be interactively queried, and the user can click on any geological body to pop up a detailed list of physical property parameters and interpretation confidence.

[0024] In the exploration of deep oil and gas reservoirs in sedimentary basins, the multi-source collaborative data acquisition module implements "star-ground" collaborative observation. The airborne platform carries a magnetic gradient meter and a gamma spectrometer, flies along the planned survey line with a line spacing of 200 meters and a flight height of 100 meters, and obtains high-precision magnetic anomaly and radioactive potassium, uranium, and thorium element concentration data. Ground work is simultaneously carried out, using a controlled source vehicle to excite seismic waves with a source scanning frequency of 5-80 Hz, and a receiving array composed of 5000 wired digital geophones records the reflected wave data with an array length of 5 kilometers. At the same time, 50 electromagnetic sounding points are arranged, and the measurement frequency range is 320 Hz to 0.0001 Hz. At the structural high point, a parameter well is deployed for systematic coring and full set of logging, including acoustic, density, resistivity, and natural gamma logging, to obtain accurate longitudinal physical property profiles. All collected data is converted to SEG-Y and EDI standard formats through standardized interfaces, and unified to the WGS84 coordinate system and UTC time axis.

[0025] The full-space velocity modeling module comprehensively utilizes airborne and ground data. First, tomographic imaging is performed using seismic first arrival wave travel times to establish the velocity structure of the basin cover. Second, after processing the airborne magnetic data, the magnetic basement depth is obtained through frequency domain inversion, and combined with regional geological knowledge, the thickness and trend of the deep sedimentary sequence are inferred. Finally, using the acoustic logging velocity of the parameter well as an absolute constraint, a Kriging interpolation algorithm with external drift is used to fuse the shallow velocity structure and the deep trend model, generating an initial three-dimensional velocity model from the surface to the basement with variable vertical resolution, and the error of the model at the parameter well is less than 2%.

[0026] The intelligent joint inversion module is adjusted for the high-resolution imaging needs of hydrocarbon reservoirs. The deep learning agent model uses a version specifically trained on a library of sedimentary rock models to improve the prediction accuracy of responses in thin sand-shale interbeds. In the joint inversion, the coupling of seismic full-waveform inversion and resistivity inversion is particularly strengthened. Cross-gradient constraints are applied not only between the velocity and resistivity models, but also an empirical rock physics relationship between seismic wave impedance and resistivity is introduced as a soft constraint. The inversion process uses a multi-scale strategy from low to high frequencies, gradually inverting model details. The final velocity model can clearly distinguish sand layers with a thickness greater than 8 meters, and the resistivity model clearly delineates low-resistivity hydrocarbon accumulation areas.

[0027] The dynamic 3D visualization and interpretation module focuses on reservoir characterization. Volume rendering uses multi-attribute fusion display to highlight areas where high wave impedance values overlap with low resistivity values with a specific warm color tone, intuitively indicating potential oil and gas enrichment areas. The interpretation function uses a machine learning classifier to classify the inversion data volume voxel by voxel, dividing the voxels into categories such as shale, sand, oil-bearing sand, etc., and calculating the porosity and saturation parameters of each category. Finally, a 3D lithofacies model and attribute model are generated, which can be directly used for reserve calculation and well location design.

Claims

1. A three-dimensional imaging exploration system for deep geological resources, characterized in that it comprises: The system includes the following components: The multi-source collaborative data acquisition module is used to deploy and control seismic, gravity, electromagnetic, drilling and remote sensing detection equipment to achieve synchronous or quasi-synchronous acquisition of multi-physics field data. The full-space velocity modeling module constructs a three-dimensional velocity field model covering the exploration area based on the initial geological model and multi-source data. The intelligent joint inversion module takes multi-source data and velocity field models as input and inverts the distribution of geological body physical parameters through a fusion mechanism of physical constraints and data-driven approaches. The dynamic 3D visualization and interpretation module performs 3D reconstruction and rendering of the inversion results, and supports geological body boundary identification and attribute annotation. Each module is connected to the high-performance computing unit via a high-speed data network, forming an integrated imaging processing closed loop.

2. The three-dimensional imaging exploration system for deep geological resources according to claim 1, characterized in that: The multi-source collaborative data acquisition module includes a distributed sensor network and a central scheduling unit. The distributed sensor network consists of a seismic detector array, gravity sensor nodes, electromagnetic transmitter-receiver pairs, and borehole sensors. The node spacing is adaptively adjusted according to the scale of the exploration target, ranging from 10 meters to 500 meters. The central scheduling unit adopts a time synchronization protocol to ensure that the time reference error of data acquisition from different physical fields is less than 1 millisecond, and realizes real-time data transmission through wireless self-organizing network technology. The module has a built-in terrain adaptive deployment algorithm that automatically optimizes the sensor deployment position according to the digital elevation model to avoid complex terrain obstacles.

3. The three-dimensional imaging exploration system for deep geological resources according to claim 1, characterized in that: The full-space velocity modeling module employs multi-scale grid subdivision technology. The module divides the exploration area into a background grid and a locally refined grid. The background grid is used to describe large-scale geological structures, with a grid size of 100m × 100m × 50m. The locally refined grid focuses on known anomaly areas or target bodies, with a grid size refined to 10m × 10m × 5m. The velocity modeling process is based on a strategy combining first-arrival travel-time tomography and full-waveform inversion. First, an initial velocity model is constructed using seismic first-arrival travel-time data through ray tracing and least-squares inversion. Then, full-waveform inversion is introduced to iteratively optimize the model details. The objective function is defined as the L2 norm of the residual between the observed waveform and the simulated waveform.

4. The three-dimensional imaging exploration system for deep geological resources according to claim 1, characterized in that: The intelligent joint inversion module constructs a multi-physics coupled inversion framework. This framework includes seismic elastic wave inversion sub-modules, gravity density inversion sub-modules, and electromagnetic resistivity inversion sub-modules. Each sub-module is coupled through cross-gradient constraint terms to ensure that models with different physical property parameters maintain consistency in spatial structure. A deep learning surrogate model is introduced into the inversion process to accelerate forward modeling calculations. The surrogate model adopts the U-Net architecture, taking velocity, density, and resistivity parameters as inputs and directly outputting simulated seismic records, gravity anomalies, and electromagnetic response data. The training data, exceeding 100,000 sets, is generated through numerical simulation. The objective function for the joint inversion is:

5. The three-dimensional imaging exploration system for deep geological resources according to claim 1, characterized in that: The dynamic 3D visualization and interpretation module is based on volume rendering and isosurface extraction technology. The module converts the inverted physical property parameter volume data into a 3D voxel model. Each voxel stores velocity value, density value, resistivity value and corresponding lithology classification probability. The visualization engine supports transparency transfer function adjustment to achieve perspective observation of the internal structure of the geological body. The interpretation function integrates edge detection algorithm and region growing algorithm to automatically identify and delineate velocity change interfaces, density anomalies and low resistivity zone boundaries, and uses the identification results as candidate targets for 3D annotation of potential ore bodies or structures.

6. A three-dimensional imaging exploration method for deep geological resources, characterized in that: The method includes the following steps: Step 1: Plan and implement multi-source geophysical data collaborative acquisition to obtain seismic wavefield data, gravity field data, electromagnetic field data and drilling core data of the exploration area; Step 2: Based on the collected data and prior geological information, construct a full-space three-dimensional initial velocity model of the exploration area; Step 3: Accelerate multiphysics forward modeling by using deep learning surrogate models and establish a joint inversion objective function for seismic, gravity, and electromagnetic data; Step 4: Use an optimization algorithm to solve the joint inversion objective function and iteratively update the three-dimensional physical property parameter model of velocity, density, and resistivity; Step 5: Perform 3D reconstruction and visualization rendering on the 3D physical property parameter model obtained by inversion, and automatically extract and interpret the spatial morphology and physical property characteristics of the geological body.

7. The three-dimensional imaging exploration method for deep geological resources according to claim 6, characterized in that: Step 3, the training process of the deep learning surrogate model, includes: randomly generating a large number of three-dimensional velocity-density-resistivity model combinations based on a known geological model library; using high-precision finite difference or finite element method forward modeling to calculate the corresponding seismic records, gravity anomalies, and electromagnetic responses to form a training dataset; the input of the U-Net surrogate model is a three-dimensional data volume of physical property parameters, and the output is the corresponding simulated observation data volume; the model loss function adopts mean square error, and the Adam optimizer is used for training until the prediction error on the validation set is less than 5%.

8. The three-dimensional imaging exploration method for deep geological resources according to claim 6, characterized in that: Step five, 3D visualization and interpretation, includes volume rendering and feature extraction. Volume rendering uses a ray casting algorithm to assign color and transparency based on physical property parameter values, generating a 3D perspective image. Feature extraction involves calculating the gradient modulus of the physical property parameters, identifying the gradient threshold τ through known geological models, and using voxels with values ​​ranging from 0.1 to 0.3 as geological boundary points. Then, a clustering algorithm is used to connect the boundary points into a continuous 3D surface, thereby delineating the geometric shape and spatial location of geological targets such as ore bodies, faults, and lithological contact zones.

Images