A method for constructing and quickly inverting a geophysical gravity and magnetic anomaly source model

Abstract

The application provides a geophysical gravity anomaly source model construction and fast forward and inverse method, and relates to the technical field of intelligent three-dimensional geological modeling and simulation. According to the geological information of a research area, the method sets the source body modeling parameters of the gravity anomaly field of the research area, obtains a three-dimensional model matrix M capable of describing multiple abnormal bodies underground; a fast forward network is constructed, a nonlinear mapping from the three-dimensional model matrix to simulated anomaly data is established through the fast forward network, and the forward response of the three-dimensional model matrix is calculated; an inversion network based on a concurrent module CFTBlock, a joint CNN and a Transformer is constructed, and a nonlinear mapping from the gravity anomaly data to the three-dimensional model matrix is established. The method has the advantages of fast and accurate forward and high-resolution inversion, and is suitable for better interpretation of measured gravity data.

Description

Technical Field

[0001] This invention relates to the field of intelligent three-dimensional geological modeling and simulation technology, and in particular to a method for constructing geophysical gravity and magnetic anomaly source models and for rapid forward and inverse modeling. Background Technology

[0002] Inversion of gravity and magnetic field data can effectively guide geological prospecting and other work. However, the inversion problem is usually ill-posed, and the results are not unique. In past research, geophysicists mainly relied on Tikhonov regularization theory for solutions, such as smooth inversion based on the maximum smooth stable functional and focused inversion based on the minimum gradient support function. However, these inversion methods often require solving large systems of equations, which leads to a large number of matrix operations. Inevitably, they will get trapped in local optima during the solution process, resulting in incomplete recovery of the predicted shape and properties of anomalies. Therefore, finding a solution that approximates global optimization and reduces the ambiguity is the main task of gravity and magnetic field data inversion. In recent years, deep learning, due to its nonlinear computational characteristics, has provided a new direction for gravity and magnetic field data inversion and is one of the important development directions in geophysics today. However, it faces the following three problems:

[0003] The primary issue is the significant impact of dataset quality; the closer the simulated source bodies in the dataset are to the real source bodies, the more accurate the network's predictions. However, existing model building methods face the dual dilemma of insufficient geometric feature representation and low computational efficiency: stochastic block-level models and random walk methods struggle to accurately characterize complex geological bodies, while refined subdivision, while improving accuracy, significantly increases computational cost. Meanwhile, Berlin noise, due to its ability to generate two-dimensional continuously differentiable functions, is widely used in visual elements, reproducing natural textures, and simulating the atmosphere. Compared to other natural noise algorithms, the Berlin noise algorithm, through interpolation and smoothing functions, can generate continuous and smooth noise, avoiding sharp edges, and performs exceptionally well in simulating natural phenomena. This provides a new approach for building geologically realistic models.

[0004] The second challenge is how to quickly generate datasets that closely approximate real-world anomaly fields in batches. Traditional forward modeling methods suffer from large memory footprints and low computational efficiency, primarily due to the storage and solving of large-scale matrix equations. To address this, researchers have proposed methods such as wavelet transform, matrix compression, and fast Fourier transform to reduce the memory consumption of large-scale matrices. However, with increasing data size and finer grid subdivision, computational complexity and memory consumption rise dramatically, making these methods insufficient for fast forward modeling. In recent years, researchers have attempted to use deep learning for approximate fitting of forward modeling, significantly accelerating the process, but accuracy still needs further improvement. Therefore, researching fast forward modeling methods based on deep learning can effectively solve the problems faced by traditional forward modeling methods.

[0005] The third issue is that traditional deep learning methods primarily rely on Convolutional Neural Networks (CNNs), mainly including traditional CNNs and the Unet method. Among these, Unet variants such as Unet3+ and ResUnet are increasingly being used due to their ability to reconstruct density anomalies with clear boundaries and strong robustness to noise. However, when high-resolution spatial information is crucial, traditional networks may unintentionally homogenize key features, leading to inaccurate inversion results due to the loss of detail. Combining CNNs and Transformer architectures can overcome these limitations and leverage the advantages of both models simultaneously. Due to its self-attention mechanism and multilayer perceptron (MLP) structure, it can effectively capture global features. Some researchers have begun to combine local and global features and have achieved good results. However, in the field of deep learning for gravity and magnetic field data inversion, researchers still mainly focus on local features, lacking the simultaneous application of both local and global features. Therefore, simultaneously utilizing local and global features of gravity and magnetic field data is an important direction for current deep learning inversion.

[0006] The above research shows that developing a suitable model for gravity and magnetic anomaly sources and a rapid forward and inverse modeling method can provide technical support for future gravity and magnetic exploration work. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for constructing and rapidly performing forward and inverse modeling of geophysical gravity and magnetic anomaly source models, thereby addressing the shortcomings of the prior art.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a method for constructing and rapidly forwarding and inversely performing geophysical gravity and magnetic anomaly source models, comprising:

[0009] Based on the geological information of the study area, the source body modeling parameters of the gravity and magnetic anomaly field in the study area were set.

[0010] A rectangular region is selected in the study area, Berlin noise is generated, hard thresholding is performed, and the rectangular region is divided into multiple two-dimensional independent blocks through the connection relationship.

[0011] A three-dimensional model matrix is ​​generated for each independent block, and the matrices are merged to obtain the final three-dimensional model matrix M that can describe multiple underground anomalies;

[0012] For the three-dimensional model matrix describing multiple underground anomalies, the kernel function matrix is ​​calculated using forward modeling formulas in combination with topography and observation surfaces, thereby obtaining simulated gravity and magnetic anomaly data. The training set and test set of the simulated gravity and magnetic anomaly forward modeling response are obtained by combining parallel computing methods.

[0013] Construct a fast forward modeling network, establish a nonlinear mapping from the 3D model matrix to simulated anomalous data through the fast forward modeling network, and calculate the forward response of the 3D model matrix;

[0014] The fast forward modeling network with optimal weight parameters was trained using the training and test sets of gravity and magnetic anomaly forward modeling responses.

[0015] Generate a training set and a test set for gravity and magnetic anomalies to train the inversion network;

[0016] An inversion network based on the concurrent module CFTBlock, combining CNN and Transformer, is constructed to establish a nonlinear mapping from gravity and magnetic anomaly data to a three-dimensional model matrix. Using the gravity and magnetic anomaly training set used to train the inversion network, a random noise addition strategy is employed to train the network and obtain an inversion network with optimal weight parameters.

[0017] Furthermore, the specific method for setting the source body modeling parameters for the gravity and magnetic anomaly field in the study area is as follows:

[0018] Based on the geological information of the study area, the lithological distribution and physical property values ​​of different regions are statistically analyzed to obtain the spatial location and range of physical property values ​​of the exploration target. Thus, when simulating the gravity and magnetic anomaly field, the physical property intervals and distribution ranges are set as modeling parameters for the gravity and magnetic anomaly. At the same time, based on the measurement results, the observation plane and topographic information are set.

[0019] Furthermore, the specific method for selecting a rectangular region in the study area, generating Berlin noise, performing hard thresholding, and dividing the rectangular region into multiple independent two-dimensional blocks through connectivity is as follows:

[0020] In the study area, the size of the rectangular region where the Berlin noise is to be generated is set, and the Berlin noise is generated; hard thresholding refers to setting the element values ​​in the two-dimensional noise matrix below a preset minimum physical property threshold. α The elements are set to 0; then, based on whether each non-zero element has a non-zero element in its four adjacent positions (above, below, left, and right), the entire rectangular area is divided into... N A set of independent two-dimensional blocks that are not connected to each other 1 ...A N .

[0021] Furthermore, the specific method for generating a three-dimensional model matrix for each independent block and merging them to obtain the final three-dimensional model matrix M that can describe multiple underground anomalies is as follows:

[0022] The continuous underground geometric space is discretized into a finite number of small units, and a three-dimensional matrix is ​​used to describe the distribution and physical properties of gravity and magnetic anomalies in the underground space, thus obtaining a three-dimensional model matrix.

[0023] For each independent two-dimensional block A l Create a 3D model matrix mt with an initial state of all zeros. l , l = 1, 2,..., N Its dimensions are nx × ny × nz , nx , ny , nz These represent the number of grid cells in the study area along the X, Y, and Z directions, respectively; mt is defined as... l,n Indicates block A l The corresponding three-dimensional model matrix mt l In the depth layer n Two-dimensional slices at the location, n = 1, 2,..., nz Randomly generate a depth layer number. n ,1≤ n ≤ nz ;Transfer two-dimensional block A l Copy the data to its corresponding 3D model matrix mt l Two-dimensional slice mt l,n In this context, it serves as the initial core layer of the block's anomaly; based on this core layer... n Starting from, and simultaneously moving towards n Decrease direction and n Diffusion occurs across all depth layers in the increasing direction; this diffusion process ultimately results in each independent two-dimensional block A. l Generate a three-dimensional model matrix mt l ; Generate a 3D model matrix mt from all independent 2D blocks l By merging, we obtain the final three-dimensional model matrix M that can describe multiple underground anomalies;

[0024] Furthermore, when the initial core layer of the block anomaly diffuses upward or downward, the attenuation value β is first subtracted from each element of the two-dimensional matrix of the core layer, and then the current attenuation coefficient γ is used to reduce the overall result. Finally, the horizontal direction is adjusted according to the shift step size to generate a stepped or tilted model.

[0025] Furthermore, the method employs parallel computing technology to generate in batches three-dimensional model matrices describing gravity and magnetic anomalies with different spatial distributions and physical properties, specifically:

[0026] Each thread independently completes the entire construction process of a single 3D model matrix, including the generation of Berlin noise, independent block division, 3D model matrix diffusion calculation, and finally outputting the final 3D model matrix M, ensuring that the 3D model matrices generated by different threads are independent and diverse in spatial distribution and physical properties.

[0027] Furthermore, the method constructs a boundary reconstruction module and combines it with a CNN to construct a fast forward modeling network;

[0028] The fast forward modeling network uses ResUNet as its backbone architecture and consists of an encoder for feature extraction and data compression and a decoder for forward modeling and data reconstruction. After feature reconstruction at each layer of the decoder, a boundary reconstruction module based on a dual residual mechanism is used to enhance feature reconstruction, thereby further improving the accuracy of feature reconstruction.

[0029] The boundary recognition module first extracts primary boundary features through residual blocks; then it introduces residual connections to fuse the original features with the reconstructed features, preserving key feature information; finally, it performs secondary feature reconstruction through a composite output layer containing 3×3 convolutions and 1×1 convolutions.

[0030] Furthermore, the specific method for generating the gravity and magnetic anomaly training and test sets for training the inversion network is as follows:

[0031] Based on existing geological and geophysical data and prior information of the target exploration area, a 1×10 model is created using a three-dimensional model matrix generation method. 4 ~1×10 5 A set of three-dimensional model matrices conforming to the geological characteristics of the region is generated. Subsequently, using a fast forward modeling network with optimal weight parameters, the gravity and magnetic anomaly data corresponding to these three-dimensional model matrices are calculated, forming the training dataset for the inversion network.

[0032] Furthermore, the inversion network consists of an input layer, a core processing layer, and an output layer. The input layer receives gravity and magnetic anomaly data, and the core processing layer consists of an encoder-decoder containing four sets of encoder-decoder blocks. The inversion network extracts high-dimensional features from the gravity and magnetic anomaly data through three layers of downsampling. Each encoder-decoder block integrates multiple concurrent modules CFTBlock and introduces skip connections to achieve multi-scale feature fusion, effectively preserving the detailed information of the anomaly data. Finally, the feature map output by the decoder is converted into a three-dimensional model matrix through a 1×1 convolution of the output layer, which can reflect the spatial morphology and physical properties of the underground anomaly.

[0033] Furthermore, the concurrent module CFTBlock includes a CNN branch and a Transformer branch. The CNN branch adopts a two-stage design to gradually extract and enhance local features: the first stage consists of two convolutional blocks and residual connections, which capture local features through convolution operations and alleviate the gradient vanishing problem by utilizing residual connections; the second stage, based on convolution and residual connections, fuses the global features of the Transformer branch to achieve the combination of local features and global context.

[0034] The Transformer branch consists of a multi-head self-attention module and an MLP module; it uses a multi-head self-attention mechanism to capture global features; the MLP module performs a non-linear transformation on the output of the self-attention module to further enhance the feature representation capability.

[0035] The beneficial effects of adopting the above technical solution are as follows: A method for constructing and rapidly forwarding and inverting geophysical gravity and magnetic anomaly source models allows for variations in the physical properties of subsurface anomaly models, making them more consistent with reality. It exhibits high physical property values ​​at the center, gradually decreasing towards both sides, and displays a characteristic of being large in the middle and small at both ends in the three-dimensional model morphology, while also allowing for the existence of small anomalies, which is consistent with actual geological models. Using a rapid forward modeling method based on boundary reconstruction modules, batch forward modeling of subsurface anomaly models can be performed quickly with low errors. Using a high-resolution inversion method combined with a random noise addition strategy can achieve better inversion results, improve the accuracy of subsurface physical property recovery, and obtain results rapidly. Verification has shown that the method proposed in this invention has advantages such as more realistic source body modeling, rapid and accurate forward modeling, and high-resolution inversion capabilities, making it suitable for better interpretation of measured gravity and magnetic data. This invention is of great significance for improving the processing and interpretation accuracy of geophysical gravity and magnetic field data. Attached Figure Description

[0036] Figure 1 A flowchart illustrating a method for constructing a geophysical gravity and magnetic anomaly source model and performing rapid forward and inverse modeling, provided in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of a simulated gravity and magnetic anomaly model provided in an embodiment of the present invention, wherein (a) is a three-dimensional model and (b) is a cross-sectional view;

[0038] Figure 3 This is a diagram of the fast forward modeling network architecture for a 3D model of gravity and magnetic anomalies based on CNN and a boundary reconstruction module, provided in an embodiment of the present invention.

[0039] Figure 4 The test dataset E provided for the embodiments of the present invention d Box-shaped diagram;

[0040] Figure 5 A diagram of the inversion network architecture based on the concurrent module CFTBlock, which combines CNN and Transformer, provided for embodiments of the present invention;

[0041] Figure 6 The following is a module architecture diagram of the concurrent module CFTBlock provided in the embodiments of the present invention: (a) feature map implementation of spatial alignment for dimensionality increase and decrease; (b) implementation details of CNN branch, Transformer branch and FCU.

[0042] Figure 7 The residual density model, noisy data, and cross-sectional view provided in the embodiments of the present invention are as follows: (a) is a three-dimensional model (the result of plotting the absolute value of residual density > 0.3), (b) is noisy data, and (c) is a cross-sectional view of the residual density model z=425m.

[0043] Figure 8 The inversion results and cross-sectional views provided in the embodiments of the present invention are as follows: (a) is a three-dimensional result diagram (drawing the result with an absolute value of residual density > 0.3), and (b) is a cross-sectional view of the inversion result z = 425m.

[0044] Figure 9 The following are gravity anomaly data and inversion results of Venton Salt Dome provided in this embodiment of the invention: (a) is the gravity anomaly data of Venton Salt Dome; (b) is the three-dimensional spatial diagram of the prediction model (plotting the part where the absolute value of the remaining density is >0.3); (c) is the profile of the prediction model at z=525m; (d) is the profile of the prediction model at y=3334.25km; and (e) is the profile of the prediction model at x=442.550km. Detailed Implementation

[0045] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0046] In this embodiment, a method for constructing a geophysical gravity and magnetic anomaly source model and performing rapid forward and inverse modeling is described, such as... Figure 1 As shown, it includes the following steps:

[0047] Step 1: Parameter setting: Based on geological information such as drilling, seismic profiles, and geological profiles, the source body modeling parameters for the gravity and magnetic anomaly field in the study area are set.

[0048] Before setting parameters, based on previous work, including geological information such as drilling, seismic profiles, and geological profiles, we statistically analyze the lithological distribution and physical property values ​​in different regions to obtain the approximate spatial location and physical property values ​​of the exploration target. This allows us to set the physical property range and distribution range, among other gravity and magnetic anomaly modeling parameters, when simulating gravity and magnetic anomalies. At the same time, we set the observation plane and topographic information based on the measurement data.

[0049] In this embodiment, parameters are set for the measured data to be used later, with the minimum physical property value set to 0.55 and the maximum burial depth range set to 200m.

[0050] Step 2: Select a rectangular region in the study area, generate Berlin noise, perform hard thresholding, and divide the rectangular region into multiple two-dimensional independent blocks through connectivity;

[0051] In the study area, the size of the rectangular region where the Berlin noise is to be generated is set, and the Berlin noise is generated; hard thresholding refers to setting the element values ​​in the two-dimensional noise matrix below a preset minimum physical property threshold. α The elements are set to 0; then, based on whether each non-zero element has a non-zero element in its four adjacent positions (above, below, left, and right), the entire rectangular area is divided into... N A set of independent two-dimensional blocks that are not connected to each other 1 ...A N ;

[0052] In this embodiment, the study area is divided into a 64×64 grid, and the breadth-first search method is used to divide the independent blocks.

[0053] Step 3: Assign different independent blocks to different depth layers in the underground three-dimensional space of the study area, and perform attenuation diffusion in the depth direction. Each independent block generates a three-dimensional model matrix, and the matrices are merged to obtain the final three-dimensional model matrix M that can describe multiple underground anomalies.

[0054] To facilitate calculation, the continuous underground geometric space is discretized into a finite number of small units, i.e., finite element meshing is used. The three-dimensional underground space of the study area is divided into an nx×ny×nz mesh (nx, ny, and nz represent the number of mesh units in the X, Y, and Z directions of the study area, respectively, and the mesh size is consistent with the study area on the plane). A three-dimensional matrix is ​​used to describe the distribution and physical property values ​​of gravity and magnetic anomalies in the underground space, resulting in a three-dimensional model matrix (i.e., a three-dimensional model matrix describing the spatial distribution and physical property characteristics of gravity and magnetic anomalies). The minimum physical property threshold α is set based on the geological information obtained in step 1 and represents the minimum non-zero value (residual density or magnetization intensity) allowed in the three-dimensional model matrix.

[0055] For each independent two-dimensional block A l Create a 3D model matrix mt with an initial state of all zeros. l , l = 1, 2,..., N Its dimensions are nx × ny × nz ( nx , ny , nz (These represent the number of grid cells in the study area along the X, Y, and Z directions, respectively); define mt l,n Indicates block A l The corresponding three-dimensional model matrix mt l In the depth layer n ( n = 1, 2, ..., nzA 2D slice at location ); a depth layer number is randomly generated. n ,1≤ n ≤ nz ;Transfer two-dimensional block A l Copy the data to its corresponding 3D model matrix mt l Two-dimensional slice mt l,n In this context, it serves as the initial core layer of the block's anomaly; based on this core layer... n Starting from, and simultaneously moving towards n Decrease direction and n Diffusion occurs across all depth layers in the increasing direction; during diffusion, the value of each element in the two-dimensional matrix gradually decreases with increasing distance from the core layer; this diffusion process ultimately results in each independent two-dimensional block A. l Generate a three-dimensional model matrix mt l ; Generate a 3D model matrix mt from all independent 2D blocks l By merging, we obtain the final three-dimensional model matrix M that can describe multiple underground anomalies;

[0056] When the initial core layer of the block anomaly diffuses upward or downward, the attenuation value β is first subtracted from each element of the two-dimensional matrix of the core layer. Then, the current attenuation coefficient γ is used to reduce the overall result. Finally, the horizontal direction is adjusted according to the shift step size in order to generate a stepped or tilted model. The shift step size represents the number of grids that the data moves in one direction when it diffuses upward or downward.

[0057] 3D model matrix mt l The two-dimensional slice at depth n+k is shown in the following formula:

[0058] (1);

[0059] (2);

[0060] (3);

[0061] Among them, mt l,n+k Represents the three-dimensional model matrix mt l In the depth layer n+k Two-dimensional slices at the location, ( p , q () represents an element in the two-dimensional slice. k =±1, ±2…,± MS , MS This represents the preset maximum number of unidirectional diffusion steps. FL The filter function is represented by , the base attenuation coefficient is represented by , and the current attenuation coefficient is represented by . γ Based on diffusion steps stepAttenuation, s x , s y The horizontal shift step size is for different directions, and its positive and negative signs represent the direction.

[0062] If during the diffusion process, the number of diffusion steps step Exceed MS The iteration stops; each independent two-dimensional block undergoes diffusion to obtain its three-dimensional model matrix mt. l The final three-dimensional model matrix M describing multiple underground anomalies is:

[0063] (4);

[0064] For different 3D model matrix generation processes, all variables are random values ​​except for the spatial partitioning method and the threshold α.

[0065] The three-dimensional model matrix generated in this embodiment supports continuous spatial variation of physical properties (such as residual density and magnetic susceptibility), exhibiting a typical distribution pattern of "high at the center and low at the edges"—that is, the absolute value of the physical property is highest in the central region of the three-dimensional model matrix, and gradually decreases along the central axis to both sides. The three-dimensional morphology is characterized by the largest size at the center and gradually decreasing towards both sides. At the same time, it allows the inclusion of independent anomalies of smaller scale. These features can effectively simulate the typical properties of actual geological structures. Figure 2 ).

[0066] Step 4: Repeat steps 2-3 and combine them with parallel computing techniques to quickly generate three-dimensional model matrices describing gravity and magnetic anomalies with different spatial distributions and physical properties in batches.

[0067] Repeat steps 2-3, employing OpenMP multithreaded parallel technology to efficiently generate the 3D model matrix M. Specifically, each thread independently completes the entire construction process of a single model, including Berlin noise generation, independent block partitioning, 3D model matrix diffusion calculation, and finally outputting the final 3D model matrix M. This ensures that the 3D model matrices generated by different threads are independent and diverse in spatial distribution and physical properties. By combining parallel computing technology, rapid generation of 3D model matrices can be achieved, enabling the rapid batch generation of the required number of 3D model matrices within a short time.

[0068] Step 5: Based on the three-dimensional model matrix describing multiple underground anomalies obtained in Step 4, the kernel function matrix is ​​calculated using the forward modeling formula, combined with the terrain and observation surface, to obtain the simulated gravity and magnetic anomaly data. The training set and test set of the simulated gravity and magnetic anomaly forward modeling response are obtained by combining parallel computing methods, which are used for subsequent research on fast forward modeling methods.

[0069] The formula for obtaining the simulated gravity and magnetic anomaly data d using the kernel function matrix is:

[0070] (5);

[0071] Where G is the kernel function matrix calculated by the forward modeling formula, and m is the vectorized result of the three-dimensional model matrix M obtained in step 4.

[0072] The calculation of Equation 5 is accelerated in parallel using the OpenMP method. Specifically, OpenMP's multithreading technology is used to calculate multiple matrix multiplications simultaneously and save the results.

[0073] In this embodiment, the forward modeling used is calculated in a Cartesian coordinate system, with coordinates ranging from 0 to 3200m, 0 to 3200m, and 0 to 1000m along the X, Y, and Z axes, respectively. The distance between the measuring point and the measuring line is 50m. Gravity anomaly data V are calculated for the established three-dimensional model matrix. z The total magnetic anomaly ΔT value was also calculated. A total of 50,000 sets of gravity anomaly data and 50,000 sets of magnetic anomaly data were generated, and divided into training and testing sets in a 4:1 ratio.

[0074] Step 6: Construct a boundary reconstruction module and, in conjunction with a CNN, build a fast forward modeling network. Establish a nonlinear mapping from the 3D model matrix to simulated anomalous data through the fast forward modeling network, and calculate the forward response of the 3D model matrix. Train the fast forward modeling network with optimal weight parameters using the training and test sets created in Step 5, and establish a validation set to verify the training effect.

[0075] The fast forward modeling network uses ResUNet as its backbone architecture and consists of an encoder for feature extraction and data compression and a decoder for forward modeling and data reconstruction. After feature reconstruction at each layer of the decoder, a boundary reconstruction module based on a dual residual mechanism is used to enhance and reconstruct features, which further improves the accuracy of feature reconstruction.

[0076] The boundary recognition module first extracts primary features using residual blocks (ResBlock) to obtain reconstructed features. Then, residual connections are introduced to fuse the primary features with the reconstructed features, preserving key feature information. Finally, secondary feature reconstruction is performed through a composite output layer containing 3×3 convolutions and 1×1 convolutions. This dual residual optimization structure enhances the feature representation capability of the network and improves the computational accuracy of the forward modeling task.

[0077] The fast forward modeling network takes a 3D model matrix M describing multiple underground anomalies as its input. The input data first passes through an encoder. The encoder gradually reduces the data size through multiple downsampling operations and extracts deep features using a residual module (ResBlock) that can effectively train deep networks, resulting in a high-dimensional feature representation containing key information. The high-dimensional features are then input into a decoder, which gradually restores the data size through upsampling operations. In each layer of the decoding process, not only is upsampling performed, but the current layer's features are also reconstructed and enhanced by a boundary reconstruction module, and the corresponding forward modeling result is calculated. To fuse information at different scales, the output of each layer of the decoder is fused step-by-step with the more refined low-dimensional feature information from the corresponding layer of the encoder through nearest neighbor interpolation. This design allows the network to effectively utilize the more original and detailed information of the input data when calculating the result at the current scale.

[0078] The three-dimensional model matrices describing multiple underground anomalies generated in step 5 and their corresponding forward response data are used to form a gravity and magnetic anomaly training set. The fast forward modeling network is trained using this set. During the training process, the training effect is monitored in real time using the test set to prevent the model from over-relying on the training data (overfitting) and to evaluate its generalization ability. Finally, the weight parameters that perform best (highest accuracy) on the test set are selected for the fast calculation of the forward response of the three-dimensional model matrix.

[0079] In this embodiment, the fast forward modeling network architecture is as follows: Figure 3 As shown, the encoder takes a 3D model matrix with a resolution of nz×64×64 as input (nz is the number of subdivisions in the depth direction). First, it performs channel dimension transformation through 1×1 convolution, then inputs it into a multi-level ResBlock for feature extraction. The ResBlock resolutions are 64×64, 32×32, 16×16, 8×8, and 4×4, following the principle of doubling the number of channels and halving the resolution at each level to achieve multi-scale feature compression and extraction. The decoder is symmetrical to the encoder, using upsampling to gradually restore the feature map resolution until it matches the size of the observation plane. Finally, a 1×1 convolution maps the fused features to simulate anomalous data (i.e., the forward response of the 3D model matrix), with the output resolution consistent with the input, ensuring spatial consistency of the forward modeling results.

[0080] The parameters used are: learning rate 3×10 -4 The learning rate decay strategy used was Cosine Annealing, the batch size was 4, the number of iterations was 100, the optimizer was Adam, and the loss function was SmoothL1. 10,000 pairs of geological models with highly random physical property values ​​were generated as a validation set, including 5,000 pairs of gravity anomaly data and 5,000 pairs of magnetic anomaly data. The error of the model predictions was statistically analyzed, such as... Figure 4 As shown. The evaluation index is the relative error of the data, E. d E d The smaller the value, the more accurate the predicted forward modeling result. The formula is as follows:

[0081] (6);

[0082] Where, d true Represents real data, d pre This represents the predicted data.

[0083] like Figure 4 It can be seen that the E values ​​of gravity anomalies and magnetic anomalies... d Mostly in 5×10 -6 Below that, the median is 1.25 × 10 -6 and 1.28×10 -6 Among them, the E of the magnetic anomaly d The distribution range is larger than that of gravity anomalies. This is because the amplitude of magnetic anomalies is larger than that of gravity anomalies, leading to a larger error in the prediction data. However, the gravity anomaly E d The maximum value is 2×10 -5 The following is the magnetic anomaly E d The maximum value is 6×10 -5 the following.

[0084] Step 7: Based on the three-dimensional model matrix generation method in Steps 2-5 and the fast forward modeling network model with optimal weight parameters obtained in Step 6, generate the gravity and magnetic anomaly training set and test set for subsequent training of the inversion network.

[0085] Based on existing geological and geophysical data (such as historical exploration data, geological maps, etc.) and prior information (such as known rock property ranges, etc.) of the target exploration area, a 1×10 model is created using the three-dimensional model matrix generation method (steps 2-5). 4 ~1×10 5 A three-dimensional model matrix that conforms to the geological characteristics of the region is generated. Then, using the fast forward modeling network with the optimal weight parameters obtained in step 6, the gravity and magnetic anomaly data corresponding to these three-dimensional model matrices are efficiently calculated, forming the training dataset of the inversion network. This method of generating gravity and magnetic anomaly data based on prior knowledge of the target region aims to enable the optimal three-dimensional model matrix obtained by finally training the inversion network to more accurately invert the spatial distribution and physical properties of the actual underground gravity and magnetic anomalies in the region.

[0086] In this embodiment, a total of 100,000 datasets were generated to train the inversion network. According to the test, taking the underground density model as an example, it takes 5 minutes and 35 seconds to generate 100,000 three-dimensional model matrix data of different gravity anomalies, and 4 minutes and 21.4 seconds to generate the corresponding simulated gravity and magnetic anomaly data. The simulated gravity and magnetic anomaly data is calculated using a forward modeling network, which uses GPU parallelism. The parallel parameters are set as follows: the number of parallel file reading threads is 8, and the batch size, i.e., the number of forward models calculated at the same time, is 256.

[0087] Step 8: Construct an inversion network based on the concurrent module CFTBlock, which combines CNN and Transformer to establish a nonlinear mapping from gravity and magnetic anomaly data to a three-dimensional model matrix; Using the gravity and magnetic anomaly training set created in Step 7 for training the inversion network, train the network with a random noise addition strategy to obtain an inversion network with optimal weight parameters; randomly select a model from the validation set and combine it with experimental data to verify the inversion network's performance.

[0088] The inversion network consists of an input layer, a core processing layer, and an output layer. The input layer receives gravity and magnetic anomaly data, and the core processing layer consists of an encoder-decoder containing four sets of encoder-decoder blocks. The inversion network extracts high-dimensional features from the gravity and magnetic anomaly data through three layers of downsampling. Each encoder-decoder block integrates 3 to 4 concurrent modules CFTBlock, and skip connections are introduced to achieve multi-scale feature fusion, effectively preserving the detailed information of the anomaly data. Finally, the feature map output by the decoder is converted into a three-dimensional model matrix through a 1×1 convolution of the output layer, which can reflect the spatial morphology and physical properties of the underground anomaly.

[0089] The concurrent module CFTBlock includes a CNN branch and a Transformer branch. The CNN branch employs a two-stage design to progressively extract and enhance local features: the first stage consists of two convolutional blocks and residual connections, capturing local features through convolutional operations and mitigating the vanishing gradient problem by utilizing residual connections; the second stage, based on the convolutions and residual connections, fuses the global features from the Transformer branch, achieving an efficient combination of local features and global context. As network depth increases, the feature map resolution gradually decreases, while the number of channels increases accordingly. The second stage aims to reduce the number of parameters while retaining richer semantic information.

[0090] Correspondingly, the Transformer branch consists of a multi-head self-attention module and an MLP module. The multi-head self-attention mechanism is used to capture global features, improving the model's ability to focus on key features; the MLP module performs a non-linear transformation on the output of the self-attention module, further enhancing the feature representation capability.

[0091] Compared to traditional CNNs, CNN branches use MaxPool for downsampling, which expands the receptive field and reduces the number of parameters, but inevitably leads to the loss of some detailed information. The Transformer branch, through its self-attention mechanism, dynamically allocates feature weights, effectively mitigating the information loss problem of pooling operations, thus generating inversion results with clear boundaries and high resolution. In each training iteration, random noise is added to the input gravity and magnetosurgical anomaly data, rather than adding noise directly during training set creation. This ensures that the same data maintains its difference in each training iteration by adding noise on the fly, thereby improving the method's generality and stability.

[0092] In this embodiment, the inverted network architecture diagram and the concurrency module CFTBlock are shown below. Figure 5 and Figure 6 CFTBlock is a core component of ICT-Inv Net, combining local features from CNNs and global features from Transformers to significantly improve the reconstruction accuracy of predicted anomaly geometry and physical properties. The CNN branch extracts local spatial features, containing spatial location information, while the Transformer branch obtains long-range global encoded features through a self-attention mechanism, enhancing global feature representation capabilities. The FCU module fuses and exchanges features from the CNN and Transformer branches, improving the robustness of feature representation.

[0093] This embodiment establishes gravity anomaly data that differs from the training and test sets, such as... Figure 7 As shown, during each training iteration, 0% to 5% random noise was randomly added to the input data. In the theoretical testing phase, to test the algorithm's robustness to noise, 5% Gaussian noise was added to the theoretical data. The constructed high-resolution inversion method was used for testing, and the results are shown below. Figure 8 As shown.

[0094] To verify the practicality of the method of this invention, this embodiment also used measured geophysical gravity anomaly data from the Vinton Salt Dome region in the United States for testing. We selected a test area of ​​approximately 3200m × 3200m for the data experiment. The underground area was divided into 64 × 64 × 20 small cuboids, each with a size of 50 × 50 × 50m. 3 Kriging interpolation was used to grid and interpolate the data, and a high-pass filter with a spatial wavelength of 5500 μm was used to remove the regional background field. The minimum absolute density value was set to 0.55 g / cm³ based on prior information. 3 The inversion results are as follows: Figure 9As shown in the figure, the method proposed in this invention enables high-resolution inversion of gravity anomaly data in the Venton Salt Dome region. The obtained inversion results have high resolution and are beneficial for the processing and interpretation of gravity and magnetic field data.

[0095] As can be seen, the geophysical gravity and magnetic anomaly source model construction and rapid forward and inverse modeling method proposed in this invention can achieve rapid generation of large datasets with low error, providing technical support for subsequent inversion work. The proposed inversion method combining local and global features can obtain higher resolution results, which is beneficial to improving the application of gravity and magnetic exploration in practical work.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for constructing and quickly inverting a geophysical gravity anomaly source model, characterized in that, The application relates to a method for training a joint CNN and Transformer inversion network based on a concurrent module CFTBlock. The method comprises the following steps: According to the geological information of a research area, source body modeling parameters of a gravity and magnetic anomaly field in the research area are set; A rectangular region is selected in the research area, Berlin noise is generated, hard thresholding is performed, and the rectangular region is divided into multiple two-dimensional independent blocks through a connection relationship; A three-dimensional model matrix is generated for each independent block, and a final three-dimensional model matrix M capable of describing multiple underground anomaly bodies is obtained by merging; For each independent two-dimensional block A l Create a three-dimensional model matrix mt with an initial state of all zeros l , l = 1, 2,..., N, with dimensions nx×ny×nz, where nx, ny, nz represent the number of grid elements in the X, Y, Z directions of the study area; define mt l, n to represent the block A l Corresponding three-dimensional model matrix mt l Two-dimensional slice at depth layer n, n = 1, 2,..., nz; randomly generate a depth layer number n, 1≤n≤nz; copy the data of two-dimensional block A l to the two-dimensional slice mt l of its corresponding three-dimensional model matrix mt l, n , as the initial core layer of the block anomaly body; A continuous underground geometric space is discretized into a limited number of small units, and a three-dimensional model matrix is used to describe the distribution and physical property values of the gravity and magnetic anomaly bodies in the underground space; This diffusion process ultimately results in each independent two-dimensional block A l Generate a three-dimensional model matrix mt l ; Generate a 3D model matrix mt from all independent 2D blocks l By merging, we obtain the final three-dimensional model matrix M that can describe multiple underground anomalies; The core layer n is taken as a starting point, and diffusion is performed in all depth layers in the direction of decreasing n and the direction of increasing n; For the three-dimensional model matrix describing multiple underground anomaly bodies, a kernel function matrix is calculated by using a forward calculation formula in combination with terrain and an observation surface, so that simulated gravity and magnetic anomaly data are obtained, and a training set and a test set of simulated gravity and magnetic anomaly forward response are obtained by using a parallel computing method; A fast forward network is constructed, a nonlinear mapping from the three-dimensional model matrix to the simulated anomaly data is established through the fast forward network, and the forward response of the three-dimensional model matrix is calculated; The training set and the test set of the gravity and magnetic anomaly forward response are used to train the fast forward network, so that a fast forward network with optimal weight parameters is obtained; A training set and a test set of gravity and magnetic anomaly data for training an inversion network are generated; A joint CNN and Transformer inversion network based on a concurrent module CFTBlock is constructed, a nonlinear mapping from the gravity and magnetic anomaly data to the three-dimensional model matrix is established, and the inversion network is trained by using the training set of the gravity and magnetic anomaly data and a random noise adding strategy, so that an inversion network with optimal weight parameters is obtained; The inversion network comprises an input layer, a core processing layer and an output layer; the input layer is used for receiving the gravity and magnetic anomaly data, the core processing layer is composed of an encoder-decoder comprising four encoding-decoding blocks, the inversion network extracts high-dimensional features of the gravity and magnetic anomaly data through three-layer downsampling, each encoding-decoding block integrates multiple concurrent modules CFTBlock, and a jump connection is introduced to realize multi-scale feature fusion and effectively retain the detail information of the anomaly data; finally, the features output by the decoder are mapped and converted into the three-dimensional model matrix through 1*1 convolution of the output layer, and the three-dimensional model matrix can reflect the spatial form and physical property characteristics of the underground anomaly body; The concurrent module CFTBlock comprises a CNN branch and a Transformer branch, the CNN branch adopts a two-stage design, local features are gradually extracted and enhanced, the first stage is composed of two convolution blocks and a residual connection, local features are captured through convolution operation, and the residual connection is used to relieve the gradient vanishing problem; the second stage is based on the convolution and the residual connection, and global features of the Transformer branch are fused, so that the local features and the global context are combined. The branch of the transformer is composed of a multi-head self-attention module and an MLP module; the multi-head self-attention mechanism is used to capture global features; and the MLP module performs nonlinear transformation on the output of the self-attention module, thereby further enhancing the feature expression capability.

2. The method according to claim 1, characterized in that, The specific method for setting the source body modeling parameters of the gravity and magnetic anomaly field in the study area is: According to the geological information of the study area, the lithology distribution and physical property values in different regions are counted to obtain the range of spatial position and physical property values of the exploration target, so as to set the physical property interval and distribution range as the gravity and magnetic anomaly body modeling parameters when simulating the gravity and magnetic anomaly field; meanwhile, the observation plane and terrain information are set according to the measurement conditions.

3. The method according to claim 1, characterized in that, The specific method for selecting a rectangular region in the study area, generating a Berlin noise, performing hard thresholding, and dividing the rectangular region into multiple independent two-dimensional blocks through a connection relationship is: The size of the rectangular region to be generated in the study area is set, and the Berlin noise is generated; hard thresholding refers to setting the elements in the two-dimensional matrix noise whose values are lower than the preset minimum physical property threshold α to 0; Subsequently, according to whether each non-0 element has non-0 elements in its upper, lower, left and right four adjacent positions, the entire rectangular region is divided into N independent two-dimensional blocks A1...A N .

4. The method according to claim 3, characterized in that, When the initial core layer of the block anomaly body diffuses upwards or downwards, first, subtract the attenuation value β from each element in the two-dimensional matrix of the core layer, then use the current attenuation coefficient γ to reduce the result as a whole, and finally adjust the horizontal direction according to the fault step to generate a stepped or inclined model.

5. The method according to claim 4, wherein, The method adopts parallel computing technology to batch generate three-dimensional model matrices of gravity and magnetic anomaly bodies describing different spatial distributions and physical property characteristics, specifically: Each thread independently completes the complete construction process of a single three-dimensional model matrix, including Berlin noise generation, independent block division, three-dimensional model matrix diffusion calculation, and finally outputting the final three-dimensional model matrix M, ensuring that the three-dimensional model matrices generated by different threads are independent and diverse in spatial distribution and physical property characteristics.

6. The method according to claim 5, wherein, The method constructs a boundary reconstruction module and jointly constructs a fast forward network with CNN; The fast forward network takes ResUNet as the backbone architecture, is composed of an encoder for feature extraction and data compression and a decoder for forward calculation and data reconstruction, and after feature reconstruction at each layer of the decoder, adopts a boundary reconstruction module based on a double residual mechanism for feature enhancement reconstruction, further improving the feature reconstruction accuracy; The boundary identification module first extracts primary boundary features through a residual block; then introduces residual connection to fuse the original features and reconstructed features, retaining key feature information; Finally, a composite output layer containing 3x3 convolution and 1x1 convolution is used to implement secondary feature reconstruction.

7. The method according to claim 6, characterized in that, The specific method for generating a gravity and magnetic anomaly training set and test set for training the inversion network is: According to the existing geological and geophysical data and prior information of the target exploration area, a 1×10 4 ~1×10 5 group of three-dimensional model matrices conforming to the geological characteristics of the area are created based on a three-dimensional model matrix generation method. Subsequently, the fast forward network with the optimal weight parameter is used to calculate the gravity and magnetic anomaly data corresponding to the three-dimensional model matrices, thereby constituting the training data set of the inversion network.

Images