A three-dimensional gravity inversion intelligent interpretation method, device and medium

Abstract

The present application relates to the technical field of geophysical inversion, and particularly relates to a three-dimensional gravity inversion intelligent interpretation method, device and medium, the method comprises gravity data correction and regularization, frequency domain forward and inversion without displaying sensitivity matrix construction or storage, GPU acceleration solving, inversion result structured summary, result constrained local knowledge retrieval and traceable report generation, realizing automatic processing from original gravity anomaly data to three-dimensional density structure interpretation text. The method repeatedly completes matrix vector multiplication through frequency domain operator; meanwhile, the inversion output is converted into structured constraints such as density range, anomaly peak value, main control depth, anomaly centroid, residual and RMS convergence history, and the local knowledge base retrieval and report generation are driven accordingly, and the method significantly reduces matrix storage demand and forward calculation time.

Description

Technical Field

[0001] This invention relates to the field of geophysical inversion technology, and in particular to a three-dimensional gravity inversion intelligent interpretation method, device and medium. Background Technology

[0002] Three-dimensional gravity inversion is a crucial technique for reconstructing subsurface density distribution, identifying anomaly locations, characterizing regional tectonic boundaries, and constraining deep geological structures. However, with the increase in the number of observation points, the decrease in grid scale, and the shift from two-dimensional profiles to complete three-dimensional volumetric imaging, traditional inversion methods based on spatial domain integral kernels and explicit sensitivity matrices face challenges such as high computational cost, high memory consumption, and long iteration times. Especially in three-dimensional volumetric meshes with hundreds of thousands to millions of elements, the storage of the sensitivity matrix and matrix-vector multiplication often become the main factors limiting the accuracy and scale of inversion.

[0003] In existing technologies, frequency domain potential field calculation and FFT acceleration methods can reduce the computational complexity of forward modeling, GPU parallel computing can improve the speed of large-scale numerical operations, and iterative algorithms such as conjugate gradients can avoid directly solving large-scale linear equation systems. However, in practice, data correction, anomaly regularization, 3D inversion, result slice display, error evaluation, literature comparison, and report writing are usually completed in different software or scripts, resulting in broken processes, results that are difficult to reproduce, and a lack of traceability of interpretive basis.

[0004] In recent years, large language models and retrieval-enhanced generation technologies have begun to be used for the processing of geoscientific data and text generation. However, without the joint constraints of current inversion results, map evidence, and literature sources, the generated content is prone to deviating from the actual model results and is difficult to meet the requirements of geophysical interpretation for evidence chains, verifiability, and professional consistency.

[0005] Therefore, there is an urgent need for a three-dimensional gravity inversion method that can combine efficient computation with intelligent interpretation of results constraints, so as to solve the efficiency problem in large-scale three-dimensional inversion and improve the automation and traceability of subsequent interpretation reports. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a three-dimensional gravity inversion intelligent interpretation method, device, and medium. By combining gravity data preprocessing, ordinary kriging regularization, frequency domain GaussFFT forward modeling and adjoint operator, conjugate gradient inversion, GPU parallel execution, result structured summarization, local vector retrieval, and large language model report generation, this invention solves the problems of high storage overhead of explicit sensitivity matrix, long forward and adjoint calculation time, fragmented interpretation process of inversion results, and low efficiency of manual report generation in traditional three-dimensional gravity inversion.

[0007] In a first aspect, the present invention provides a three-dimensional gravity inversion intelligent interpretation method, comprising: S1: Acquire gravity observation data of the target area and preprocess the gravity observation data; S2: Correct the preprocessed gravity data to obtain gravity anomaly data; S3: Transform gravity anomaly data into a unified local coordinate system and interpolate discrete gravity anomalies to a regular two-dimensional grid using the ordinary kriging method; S4: Construct a 3D model mesh based on the study area, target depth, and resolution requirements, and determine... , , The number of grids and the grid spacing in three directions are used to discretize the underground space into multiple rectangular grid cells, and the reference density model, data weight matrix and model constraint parameters are initialized. S5: Calculate the horizontal wavenumber grid and the three-dimensional frequency domain forward modeling kernel function in the frequency domain. The three-dimensional frequency domain forward modeling kernel function is obtained by multiplying the gravity spectrum kernel with the depth response term, and the Legendre-Gauss integration points and weights are introduced into the horizontal integration of the grid cells for correction. S6: Construct a matrix-independent frequency domain forward modeling operator and a frequency domain adjoint operator. The frequency domain forward modeling operator is used to predict gravity anomalies by performing a fast Fourier transform, frequency domain multiplication, and inverse transform on any density model vector. The frequency domain adjoint operator is used to calculate gradient correlation terms on any residual vector. S7: Establish a regularized inversion objective function that includes a depth weighting term, and use the conjugate gradient method to iteratively solve the regularized inversion objective function. Calculate the matrix-vector product in each iteration, and output the 3D density inversion result after the iteration converges. S8: The three-dimensional density inversion results are processed in a structured manner to obtain a structured summary, and an automated report containing geophysical interpretation, methodological description and evidence source identification is generated through a large language model.

[0008] Preferably, in S3, the variation function model used in the ordinary Kriging method is one of the following: a spherical model, an exponential model, or a Gaussian model.

[0009] Preferably, in S5, the calculation process of the three-dimensional frequency domain forward modeling kernel function is as follows: Let the horizontal wave number be and , the synthetic wave number ; Gravity spectral nuclei Where G is the gravitational constant, 10 5 This is the conversion factor for gravity units; For the z-th layer model unit, the depth response term ,in, For the elevation of the observation surface, and These represent the top and bottom coordinates of the model layer, respectively. Three-dimensional frequency domain forward modeling kernel The expression is: .

[0010] Preferably, in S7, the expression for the regularized inversion objective function including the depth-weighted term is as follows: ; Where m is the density model vector to be determined. This represents the frequency domain forward modeling operator. This is gravity anomaly data. For the data weight matrix, The model constraint matrix, As a reference density model, For regularization parameters; The model constraint matrix Introducing depth-weighted terms Where z is the center depth of the grid cell, This is the depth offset. It is a decay index used to weaken the dominant effect of shallow sensitivity.

[0011] Preferably, in S7, the normal equation of the regularized inversion objective function is as follows: ; In each iteration of the conjugate gradient method, the matrix in the normal equation is not explicitly computed; instead, the matrix-vector product is calculated through a first frequency domain forward modeling and a first frequency domain adjoint operation. ,in, , This indicates the current search direction.

[0012] Preferably, in S7, each iteration of the conjugate gradient method includes: ; ; ; ; ; in, For the first The residual vector of the next iteration For the first Density model of the next iteration.

[0013] Preferably, in S8, the structured summary includes at least the following fields: model range, mesh size, density range, maximum coordinates, minimum coordinates, master anomaly depth layer, centroid of high anomaly region, initial root mean square error, final root mean square error, and number of iterations. When performing vectorized retrieval on the local knowledge base, the retrieval objects include papers, technical reports, regional geological descriptions, map annotations, and historical project documents; the embedding model used for vectorization is a general embedding model, an embedding model in the earth science field, or a vectorization model deployed locally by the user.

[0014] Preferably, during the execution of S5, S6 and S7, the frequency domain kernel function, depth response term, model array and iteration vector are kept in GPU memory for parallel computation, reducing data transfer between CPU and GPU; when GPU is unavailable, computation automatically falls back to CPU.

[0015] In a second aspect, a computer device includes a memory and a processor; The memory is used to store computer programs that can run on the processor; When the processor executes the computer program, it implements the steps of the three-dimensional gravity inversion intelligent interpretation method as described above.

[0016] In a third aspect, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the three-dimensional gravity inversion intelligent interpretation method described above.

[0017] The application of the technical solution of the present invention has at least the following beneficial effects: This invention provides a three-dimensional gravity inversion intelligent interpretation method that repeatedly performs matrix-vector multiplication using frequency domain operators. Simultaneously, it converts the inversion output into structured constraints such as density range, anomalous peaks, master control depth, anomalous centroids, residuals, and RMS convergence history, thereby driving local knowledge base retrieval and report generation. This invention significantly reduces matrix storage requirements and forward modeling computation time.

[0018] The method of this invention uses conjugate gradient iteration to solve the regularized inversion objective function. Each iteration only requires operator-level forward and adjoint operations, which can maintain good scalability in high-resolution 3D meshes.

[0019] The method of this invention combines GPU parallel computing with frequency domain operators, enabling the frequency domain kernel, model volume and iteration vector to be computed centrally on the GPU, reducing unnecessary data transmission, and is suitable for large-scale three-dimensional gravity inversion.

[0020] The method of this invention automatically organizes the inversion results into a structured summary, and then uses the structured summary to drive the retrieval of local literature and reports, so that subsequent interpretation is no longer a generalized topic search, but an evidence retrieval directly constrained by the abnormal features of the current model.

[0021] The method of this invention restricts the large language model to result constraints and evidence constraints for generating explanatory text, which can improve the efficiency of report generation, while retaining the source of results and evidence, thereby improving the traceability of geophysical interpretation. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the steps of the three-dimensional gravity inversion intelligent interpretation method in a preferred embodiment of the present invention; Figure 2 A comparison chart of CPU and GPU speedup ratios under different grid sizes; Figure 3 The figure shows the comparison of forward modeling accuracy and error under the two-sphere theoretical model, where (a) is the analytical solution; (b) is the forward modeling result of the method in this embodiment; (c) is the forward modeling result of SimPEG; (d) is the error between the method in this embodiment and the analytical solution; and (e) is the error between SimPEG and the analytical solution. Figure 4 The figure shows a comparison of the observed values, predicted values, and residual distributions under the double-sphere theoretical model. In this figure, (a) is the observed data; (b) is the predicted result obtained by the method inversion in this embodiment; (c) is the predicted result obtained by the SimPEG method; (d) is the method error in this embodiment; and (e) is the SimPEG method error. Figure 5 The figures show a comparison of the theoretical model of the two spheres, the 3D reconstruction results, and orthogonal slices. (a) is the actual 3D density model; (b) is the orthogonal slice of the method in this embodiment at X=238.3 m; (c) is the orthogonal slice of SimPEG at X=242.2 m; (d) is the 3D reconstruction of the method in this embodiment; (e) is the orthogonal slice of the method in this embodiment at Y=488.3 m; (f) is the orthogonal slice of SimPEG at Y=492.2 m; (g) is the 3D reconstruction of SimPEG; (h) is the orthogonal slice of the method in this embodiment at Z=488.3 m; and (i) is the orthogonal slice of SimPEG at Z=492.2 m. Figure 6Map showing the distribution of the Bushveld field data study area and the Bouguer gravity anomaly observation points; Figure 7 The images show a comparison of observed anomalies, predicted anomalies, and residual fields in Bushveld field data. (a) shows the anomalies observed on a 128*128 grid; (b) shows the prediction results of the method in this embodiment; (d) shows the SimPEG prediction results; (c) shows the residual plot of the method in this embodiment; and (e) shows the SimPEG residual plot. Figure 8 The image shows the 3D density structure and representative slice results of Bushveld field data. (a), (b), and (c) show the slices obtained by the method in this embodiment, and (d), (e), and (f) show the slices obtained by SimPEG. (a) and (d) are cross sections with fixed longitude, (b) and (e) are cross sections with fixed latitude, and (c) and (f) are depth slices. Figure 9 Flowchart for local knowledge base retrieval and structured report generation. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: like Figure 1 As shown, this embodiment provides a three-dimensional gravity inversion intelligent interpretation method, including: S1: Acquire gravity observation data of the target area and preprocess the gravity observation data.

[0026] In this embodiment, gravity observation data includes measurement point coordinates, elevation, measured gravity value, necessary topographic, density and coordinate information, and preprocessing includes integrity checks, outlier identification and format standardization.

[0027] S2: Correct the preprocessed gravity data to obtain gravity anomaly data.

[0028] In this embodiment, gravity data correction includes normal gravity correction, free air correction, Bouguer correction, terrain correction, and regional field / residual field separation for highlighting target anomalies. It should be noted that, depending on the data source, this embodiment may perform all of the above correction steps, or select only some of them.

[0029] S3: Transform gravity anomaly data into a unified local coordinate system and interpolate discrete gravity anomalies to a regular two-dimensional grid using the ordinary kriging method.

[0030] In this embodiment, the variogram model used in the ordinary Kriging method is one of the spherical model, exponential model, or Gaussian model. Preferably, this embodiment uses a spherical variogram and sets nlags=12 to interpolate irregular measurement point data to a 128×128 regular grid.

[0031] S4: Construct a 3D model mesh based on the study area, target depth, and resolution requirements, and determine... , , The number of grids and the grid spacing in three directions are used to discretize the underground space into multiple rectangular grid cells, and the reference density model, data weight matrix and model constraint parameters are initialized.

[0032] In this embodiment, the 3D model mesh can be set to 32×32×32, 64×64×64, 96×96×48, 128×128×128 or other resolutions; when video memory and memory conditions permit, a higher resolution mesh is preferred to improve the 3D anomaly recovery capability.

[0033] S5: Calculate the horizontal wavenumber grid and the three-dimensional frequency domain forward modeling kernel function in the frequency domain. The three-dimensional frequency domain forward modeling kernel function is obtained by multiplying the gravity spectrum kernel with the depth response term, and the Legendre-Gauss integration points and weights are introduced into the horizontal integral of the grid cell for correction.

[0034] In this embodiment, the Legendre-Gauss integration order can be set to a natural number between 2 and 12; specifically, the integration order is preferably set to 8 in this embodiment to balance forward modeling accuracy and computational speed.

[0035] Furthermore, the calculation process of the three-dimensional frequency domain forward modeling kernel function is as follows: Let the horizontal wave number be and , the synthetic wave number ; Gravity spectral nuclei Where G is the gravitational constant, 10 5 This is the conversion factor for gravity units; For the z-th layer model unit, the depth response term ,in, For the elevation of the observation surface, and These represent the top and bottom coordinates of the model layer, respectively. Three-dimensional frequency domain forward modeling kernel The expression is: .

[0036] S6: Construct a matrix-independent frequency-domain forward modeling operator and a frequency-domain adjoint operator. The frequency-domain forward modeling operator is used to predict gravity anomalies by applying a fast Fourier transform, frequency-domain multiplication, and inverse transform to an arbitrary density model vector. The frequency-domain adjoint operator is used to calculate gradient correlation terms for an arbitrary residual vector.

[0037] S7: Establish a regularized inversion objective function that includes a depth weighting term. Use the conjugate gradient method to iteratively solve the regularized inversion objective function. In each iteration, the matrix-vector product is calculated by calling the frequency domain forward operator once and once. After the iteration converges, the three-dimensional density inversion result is output.

[0038] In this embodiment, the expression for the regularized inversion objective function including the depth-weighted term is as follows: ; Where m is the density model vector to be determined. This represents the frequency domain forward modeling operator. This is gravity anomaly data. For the data weight matrix, The model constraint matrix, As a reference density model, For regularization parameters; The model constraint matrix Introducing depth-weighted terms Where z is the center depth of the grid cell, This is the depth offset. The attenuation index is used to weaken the dominant effect of shallow sensitivity. In this embodiment, the depth-weighted term is used to weaken the dominant effect of shallow sensitivity and improve the recovery ability of deep anomalies.

[0039] Furthermore, the normal equation for the regularized inversion objective function is as follows: ; In each iteration of the conjugate gradient method, the matrix in the normal equation is not explicitly computed; instead, the matrix-vector product is calculated through a first frequency domain forward modeling and a first frequency domain adjoint operation. ,in, , This indicates the current search direction.

[0040] Furthermore, each iteration of the conjugate gradient method includes: ; ; ; ; ; in, For the first The residual vector of the next iteration For the first Density model of the next iteration.

[0041] In this embodiment, the regularization parameter The value is 0.1, which is the depth-weighted index. The value is 4.5, which is the convergence threshold. The value is 1×10 -5 .

[0042] S8: The three-dimensional density inversion results are processed in a structured manner, anomaly feature fields are extracted and a structured summary is generated. The structured summary is used as a retrieval condition to perform a vectorized retrieval of the local knowledge base, recalling evidence fragments related to the current anomaly features, spatial location and depth range. The structured summary, the three-dimensional density inversion results and the evidence fragments are then input into the large language model to generate an automated report containing geophysical interpretation, methodological description and evidence source identification under the constraints of the results.

[0043] In this embodiment, the structured summary includes at least the following fields: model range, mesh size, density range, maximum coordinates, minimum coordinates, master anomaly depth layer, centroid of high anomaly region, initial root mean square error, final root mean square error, and number of iterations. It should be noted that the structured summary in this embodiment can be saved as JSON, CSV, NPZ, or a database record.

[0044] Furthermore, when performing vectorized retrieval on the local knowledge base, the retrieval objects include papers, technical reports, regional geological descriptions, map annotations, and historical project documents; the embedding model used for vectorization is a general embedding model, an embedding model in the field of earth sciences, or a vectorization model deployed locally by the user.

[0045] Furthermore, the results in this embodiment are subject to the following constraints: the generated content must first describe the current inversion results, and then discuss its possible geological significance in conjunction with the search evidence; when the search evidence is insufficient or the results are not unique, the report should clearly provide an uncertainty warning.

[0046] In this embodiment, the GPU computing modes include three types of workloads: forward only, inversion only, and full process. Forward only is used for forward speed testing, inversion only is used for inversion iteration testing, and full process is used for testing the entire process from data input to result output.

[0047] like Figure 2As shown, in the CPU vs. GPU comparison test, only the execution backend was changed while keeping the model size, data input, and inversion parameters consistent. The test workloads included three types: forward modeling only, inversion only, and the complete process. When the mesh size reached 128×128×128, the GPU speedup ratios for the three workloads—forward modeling only, inversion only, and the complete process—were approximately 19, 57, and 61, respectively, indicating that the acceleration effect of this invention becomes more significant as the problem size increases.

[0048] Furthermore, during the execution of S5, S6, and S7, the frequency domain kernel function, depth response term, model array, and iteration vector are kept in GPU memory for parallel computation, reducing data transfer between the CPU and GPU; when the GPU is unavailable, computation automatically falls back to the CPU.

[0049] To verify the effectiveness of the method in this embodiment, a three-dimensional density anomaly model of two spheres was designed for numerical experiments.

[0050] The model region is 1000 m × 1000 m × 1000 m, discretized into a regular 128 × 128 × 128 grid, with a total of 2,097,152 elements. The centers of the two spherical anomalies are located at (250, 500, 500) m and (750, 500, 500) m respectively, both with a radius of 120 m and a density difference of +1000 kg / m³. 3 and -1000 kg / m 3 .

[0051] This embodiment uses GPU-accelerated GaussFFT forward and inverse processes and compares them with the standard open-source geophysical inversion framework SimPEG process.

[0052] like Figure 3 The figure shows a comparison of the forward modeling accuracy and error under the double-sphere theoretical model, where gz represents the first-order component of gravity in the z-direction. (a) is the analytical solution; (b) is the forward modeling result of the method in this embodiment; (c) is the forward modeling result of SimPEG; (d) is the error between the method in this embodiment and the analytical solution; and (e) is the error between SimPEG and the analytical solution. In the forward modeling test, the RMSE of the method in this embodiment relative to the analytical reference solution is 2.783 × 10⁻⁶. -5 mGal, with a maximum absolute error of 4.606 × 10⁻⁶. -5 mGal; the total modeling time was reduced from 114.847 s for the SimPEG process to 0.48 s.

[0053] like Figure 4The figure shows a comparison of the observed values, predicted values, and residual distributions under the dual-sphere theoretical model. (a) represents the observed data; (b) represents the predicted results obtained from the inversion using the method in this embodiment; (c) represents the predicted results obtained using the SimPEG method; (d) represents the method error in this embodiment; and (e) represents the SimPEG method error. In the inversion test, this embodiment ran on a complete 128×128×128 model grid and a 128×128 observed data grid, with a final data RMS of 0.0026. The SimPEG process ran on a 64×64×64 model grid, with a final full data prediction RMS of 0.0027. The total inversion time for this embodiment was 6.975 s, while the SimPEG process took 25.333 s, representing an inversion speedup of approximately 3.632 times.

[0054] like Figure 5 The diagram shows the theoretical model of the two spheres, the 3D reconstruction results, and a comparison of orthogonal slices. (a) is the actual 3D density model; (b) is the orthogonal slice at X=238.3 m using the method of this embodiment; (c) is the orthogonal slice at X=242.2 m using SimPEG; (d) is the 3D reconstruction diagram of the method of this embodiment; (e) is the orthogonal slice at Y=488.3 m using the method of this embodiment; (f) is the orthogonal slice at Y=492.2 m using SimPEG; (g) is the 3D reconstruction diagram of SimPEG; (h) is the orthogonal slice at Z=488.3 m using the method of this embodiment; and (i) is the orthogonal slice at Z=492.2 m using SimPEG. Comparing the 3D density model recovered by the method of this invention with the theoretical model and the SimPEG process results, it can be seen that the method of this invention can better maintain the main spatial location, polarity, and geometric shape of the positive and negative anomalies, and is superior to or no less than the standard SimPEG process in terms of model RMSE and structural correlation.

[0055] like Figure 6 The map shown is of the Bushveld field data study area and the distribution of Bouguer gravity anomaly observation points. In this embodiment, processed Bouguer gravity anomaly data from the Bushveld region was used for testing. The study area ranges from approximately 25.5°E to 31.0°E and 27.0°S to 23.0°S.

[0056] like Figure 7 The above diagram shows a comparison of observed anomalies, predicted anomalies, and residual fields in Bushveld field data. (a) shows anomalies observed on a 128*128 grid; (b) shows the prediction results of the method in this embodiment; (d) shows the SimPEG prediction results; (c) shows the residual plot of the method in this embodiment; and (e) shows the SimPEG residual plot. Figure 7 As shown, the method in this embodiment can recover the trend of first-order gravity anomaly changes in the region and obtain a predicted anomaly field that is similar to the standard SimPEG process but with clearer local contrast.

[0057] like Figure 8 The diagram shows the 3D density structure and representative slice results of the Bushveld field data. (a), (b), and (c) show the slices obtained using the method in this embodiment, while (d), (e), and (f) show the slices obtained using SimPEG. (a) and (d) are sections with fixed longitude, (b) and (e) are sections with fixed latitude, and (c) and (f) are depth slices. The slices obtained using the method in this embodiment were extracted at 29.066°E, -24.161°S, and 11.9 km; the SimPEG slices were extracted at the nearest available coordinates, namely 29.094°E, -24.171°S, and 12.0 km. Figure 8 As shown, in the 3D inversion results of field data, the method of this embodiment can generate a regional-scale density structure image, which can be used to identify deep density anomalies, lateral density variations, and the main anomaly extension directions. It should be noted that this result is not the only geological interpretation, but rather a regional-scale density imaging result obtained by constraining gravity data.

[0058] like Figure 9 As shown, after the inversion is completed, the method in this embodiment can also extract model dimensions, density range, peak voxels, master control depth layer, centroid of high anomaly regions, and RMS convergence features, and use these as retrieval conditions to call the local vector knowledge base, returning evidence fragments related to the current anomaly location, depth, and density features. Subsequently, the report generation module combines the current result image and the retrieved evidence to automatically generate a structured report containing data sources, methodological procedures, result descriptions, anomaly explanations, and uncertainty descriptions.

[0059] In this embodiment, the process of generating the structured report includes: Knowledge base construction involves extracting model dimensions, density range, peak voxels, master control depth layers, centroids of high anomaly regions, and RMS convergence features as retrieval criteria, and calling the local vector knowledge base to construct a searchable vector database. Summarize the inversion results and construct evidence-guided queries; RGA is used to retrieve evidence fragments related to the current anomaly location, depth, and density features from the vector database; The retrieved evidence fragments and inversion results are input into the large model, which then constrains the model to generate a structured report guided by the results.

[0060] In addition, this embodiment also provides a computer device, including a memory and a processor; The memory is used to store computer programs that can run on the processor; When the processor executes the computer program, it implements the steps of the three-dimensional gravity inversion intelligent interpretation method as described above.

[0061] It should be noted that computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0062] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0063] In addition, this embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the three-dimensional gravity inversion intelligent interpretation method described above.

[0064] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-described intelligent interpretation method for three-dimensional gravity inversion. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the intelligent interpretation method for three-dimensional gravity inversion provided in the above embodiments, and will not be repeated here.

[0065] The above description is only a preferred embodiment of the present invention and does not limit the scope of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the protection scope of the present invention.

Claims

1. A three-dimensional gravity inversion intelligent interpretation method, characterized in that, Includes the following steps: S1: Acquire gravity observation data of the target area and preprocess the gravity observation data; S2: Correct the preprocessed gravity data to obtain gravity anomaly data; S3: Transform gravity anomaly data into a unified local coordinate system and interpolate discrete gravity anomalies to a regular two-dimensional grid using the ordinary kriging method; S4: Construct a 3D model mesh based on the study area, target depth, and resolution requirements, and determine... , , The number of grids and the grid spacing in three directions are used to discretize the underground space into multiple rectangular grid cells, and the reference density model, data weight matrix and model constraint parameters are initialized. S5: Calculate the horizontal wavenumber grid and the three-dimensional frequency domain forward modeling kernel function in the frequency domain. The three-dimensional frequency domain forward modeling kernel function is obtained by multiplying the gravity spectrum kernel with the depth response term, and the Legendre-Gauss integration points and weights are introduced into the horizontal integration of the grid cells for correction. S6: Construct a matrix-independent frequency domain forward modeling operator and a frequency domain adjoint operator. The frequency domain forward modeling operator is used to predict gravity anomalies by performing a fast Fourier transform, frequency domain multiplication, and inverse transform on any density model vector. The frequency domain adjoint operator is used to calculate gradient correlation terms on any residual vector. S7: Establish a regularized inversion objective function that includes a depth weighting term, and use the conjugate gradient method to iteratively solve the regularized inversion objective function. Calculate the matrix-vector product in each iteration, and output the 3D density inversion result after the iteration converges. S8: The three-dimensional density inversion results are processed in a structured manner to obtain a structured summary, and an automated report containing geophysical interpretation, methodological description and evidence source identification is generated through a large language model.

2. The intelligent interpretation method for three-dimensional gravity inversion according to claim 1, characterized in that, In S3, the variation function model used in the ordinary Kriging method is one of the following: a spherical model, an exponential model, or a Gaussian model.

3. The intelligent interpretation method for three-dimensional gravity inversion according to claim 1, characterized in that, In S5, the calculation process of the three-dimensional frequency domain forward modeling kernel function is as follows: Let the horizontal wave number be and , the synthetic wave number ; Gravity spectral nuclei Where G is the gravitational constant, 10 5 This is the conversion factor for gravity units; For the z-th layer model unit, the depth response term ,in, For the elevation of the observation surface, and These represent the top and bottom coordinates of the model layer, respectively. Three-dimensional frequency domain forward modeling kernel The expression is: .

4. The intelligent interpretation method for three-dimensional gravity inversion according to claim 1, characterized in that, In S7, the expression for the regularized inversion objective function that includes a depth-weighted term is as follows: ； Where m is the density model vector to be determined. This represents the frequency domain forward modeling operator. This is gravity anomaly data. For the data weight matrix, The model constraint matrix, As a reference density model, For regularization parameters; The model constraint matrix Introducing depth-weighted terms Where z is the center depth of the grid cell, This is the depth offset. It is a decay index used to weaken the dominant effect of shallow sensitivity.

5. The intelligent interpretation method for three-dimensional gravity inversion according to claim 4, characterized in that, In S7, the normal equation for the regularized inversion objective function is as follows: ； In each iteration of the conjugate gradient method, the matrix in the normal equation is not explicitly computed; instead, the matrix-vector product is calculated through a first frequency domain forward modeling and a first frequency domain adjoint operation. ,in, , This indicates the current search direction.

6. The intelligent interpretation method for three-dimensional gravity inversion according to claim 5, characterized in that, In S7, each iteration of the conjugate gradient method includes: ； ； ； ； ； in, For the first The residual vector of the next iteration For the first The density model of the next iteration.

7. The intelligent interpretation method for three-dimensional gravity inversion according to claim 1, characterized in that, In S8, the structured summary includes at least the following fields: model range, mesh size, density range, maximum coordinates, minimum coordinates, master anomaly depth layer, centroid of high anomaly region, initial root mean square error, final root mean square error, and number of iterations. When performing vectorized retrieval on the local knowledge base, the retrieval objects include papers, technical reports, regional geological descriptions, map annotations, and historical project documents; the embedding model used for vectorization is a general embedding model, an embedding model in the earth science field, or a vectorization model deployed locally by the user.

8. The intelligent interpretation method for three-dimensional gravity inversion according to claim 1, characterized in that, During the execution of S5, S6, and S7, the frequency domain kernel function, depth response term, model array, and iteration vector are kept in GPU memory for parallel computation, reducing data transfer between the CPU and GPU; when the GPU is unavailable, computation automatically falls back to the CPU.

9. A computer device, characterized in that, Including memory and processor; The memory is used to store computer programs that can run on the processor; When the processor executes the computer program, it implements the steps of the three-dimensional gravity inversion intelligent interpretation method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the three-dimensional gravity inversion intelligent interpretation method as described in any one of claims 1 to 8.

Images