Deep learning constraint-based joint inversion method of electrical method and seismic, storage medium

Abstract

The application provides a method for jointly inverting electrical method and seismic method based on deep learning constraint, and a storage medium, wherein the method for jointly inverting comprises the following steps: constructing a geological anomaly body model, and constructing a resistivity database and a seismic wave velocity database based on the model; training a formed resistivity / seismic wave velocity joint inversion network by using the constructed resistivity database and seismic wave velocity database, wherein the resistivity / seismic wave velocity joint inversion network is a double-channel neural network FeatIntNet; calculating a target function based on the output result of the resistivity / seismic wave velocity joint inversion network, optimizing the resistivity / seismic wave velocity joint inversion network by using the target function, and obtaining an optimized joint inversion network; and inputting test data into the optimized joint inversion network to obtain the output result of the optimized joint inversion network. The application effectively integrates feature information provided by different physical detection methods, combines the advantages of detection methods with different resolution sensitivities, and realizes a more accurate inversion process.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration technology, specifically relating to a method for joint inversion of electrical and seismic methods based on deep learning constraints, and a storage medium. Background Technology

[0002] Seismic and resistivity methods are two of the most commonly used geophysical exploration methods, widely applied in engineering surveys, hydrogeology, environmental investigations, and resource exploration. Resistivity methods have a wide detection range, are sensitive to low-resistivity bodies, and offer flexible field operations; however, they suffer from low vertical resolution and poor boundary characterization of anomalies. Compared to resistivity methods, seismic methods offer higher accuracy and resolution, and can better characterize fine subsurface structures; however, the lack of low-wavenumber components makes it difficult to recover the wave velocity distribution in deep strata, leading to distorted inversion results. Therefore, there is considerable research on combined seismic and resistivity inversion. Compared to single-method inversion, combined inversion reduces the ambiguity of inversion results and improves accuracy by fully leveraging the advantages of various geophysical methods and compensating for their respective shortcomings.

[0003] According to the inventors, most commonly used joint inversion methods are based on rock property relationships or structural similarity. The former is theoretically based on the idea that under certain geological conditions, physical parameters can be linked through physical expressions or empirical formulas. The latter utilizes the similarity in the boundary positions of the physical parameters to be inverted, establishing connections between them structurally so that the inversion results reflect common boundary changes in different physical parameters. A common method is to use cross gradients as structural constraints. However, traditional joint inversion methods require multiple iterations of inversion and interpretation to obtain a model that conforms to prior information, which is time-consuming and lacks flexibility. In recent years, the outstanding performance of deep learning in computer vision and image processing has made it a very attractive technology, and it has been widely applied to geophysical inversion problems. However, it is currently generally used for inversion problems of single detection methods, and no research has yet used deep learning network constraints to replace traditional cross gradient constraints to achieve joint inversion.

[0004] There are two main challenges in incorporating deep learning methods into the joint inversion process: First, traditional joint inversion uses cross gradient functions as structural constraints to make the two geophysical models structurally consistent. The primary problem in introducing deep learning into traditional joint inversion is how to use deep learning methods to replace cross gradients as structural constraints, guide network parameters to make meaningful updates, thereby modeling the complex mapping relationship between input and output, and ultimately constructing a direct end-to-end mapping relationship.

[0005] Second, since the physical property parameters of different detection methods have different sensitivities in joint inversion, their inversion resolutions are not the same. How to design a network structure to deeply extract the features of each physical model and reasonably integrate the advantages of multiple physical property parameters is an important issue in using deep learning methods to solve the joint inversion problem. Summary of the Invention

[0006] Therefore, the technical problem to be solved by this invention is to provide a method for joint inversion of electrical and seismic methods based on deep learning constraints and a storage medium. The method uses deep learning to provide structural constraints and constructs the correspondence between the independent inversion models of seismic and resistivity and the real models through neural networks. It effectively integrates the feature information provided by different physical detection methods and combines the advantages of detection methods with different resolution sensitivities to achieve a more accurate inversion process.

[0007] To address the aforementioned problems, this invention provides a joint inversion method for electrical and seismic methods based on deep learning constraints, comprising the following steps: S100, a geological anomaly model is constructed through computer numerical simulation, a resistivity database is constructed based on the anomaly model through forward and inverse numerical calculations, and a seismic velocity database is constructed based on the anomaly model through forward and inverse numerical calculations. S200, the resistivity / seismic velocity joint inversion network is trained using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network FeatIntNet. S300, Calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and use the objective function to optimize the resistivity / seismic wave velocity joint inversion network to obtain the optimized joint inversion network; S400, input the test data into the optimized joint inversion network to obtain the output result of the optimized joint inversion network.

[0008] In some implementations... The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. αp represents the true model value, and αp represents the corresponding weight of each physical model.

[0009] In some implementations... The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC mS These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. This is an edge detection operator constructed using the Laplace transform.

[0010] In some implementations... The objective function also includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch.

[0011] In some implementations... The objective function L is: .

[0012] The present invention also provides a deep learning-constrained electrical and seismic joint inversion device, comprising: The database construction unit is used to construct a geological anomaly model through computer numerical simulation, construct a resistivity database by performing forward and inverse numerical calculations of resistivity based on the anomaly model, and construct a seismic velocity database by performing forward and inverse numerical calculations of seismic velocity based on the anomaly model. The neural network training unit is used to train the constructed resistivity / seismic velocity joint inversion network using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network, FeatIntNet. The neural network optimization unit is used to calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and to optimize the resistivity / seismic wave velocity joint inversion network using the objective function to obtain the optimized joint inversion network. The test output unit is used to input test data into the optimized joint inversion network and obtain the output result of the optimized joint inversion network.

[0013] In some implementations... The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. αp represents the true model value, and αp represents the corresponding weight of each physical model.

[0014] In some implementations... The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC m S These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. This is an edge detection operator constructed using the Laplace transform.

[0015] In some implementations... The objective function also includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch.

[0016] The present invention also provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described deep learning-constrained joint inversion method of electrical and seismic methods.

[0017] The present invention provides a method and storage medium for joint inversion of electrical and seismic methods based on deep learning constraints, which has the following advantages: This invention introduces deep learning into the traditional joint inversion method, proposes the idea of ​​using deep learning for structural constraints, utilizes a dual-channel neural network to establish the mapping relationship between models, and integrates the complementary features of different physical parameters through orthogonal constraints during network training. This establishes a joint inversion method of electrical and seismic methods based on deep learning constraints, effectively integrating the feature information provided by different physical detection methods, and combining the advantages of detection methods with different resolution sensitivities, thus achieving a more accurate inversion process.

[0018] Meanwhile, this invention introduces an edge detection operator to further constrain the training process of the neural network. The edge detection operator further extracts high-level edge feature information, and by adding collinear fused edge feature information to the objective function, the final inversion result changes to the common boundary reflected by the two physical property parameters, thereby further strengthening the structural similarity and improving the accuracy.

[0019] Furthermore, by optimizing the neural network through mean squared error and adding adaptive smoothness constraints to the regularization term, the overfitting problem of deep learning methods and the loss of edge information as the number of iterations increases are reduced. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the steps of the deep learning-constrained joint inversion method of electrical and seismic methods according to the present invention. Figure 2This is a flowchart of the steps in a specific embodiment of the deep learning-constrained joint inversion method of electrical and seismic methods according to the present invention; Figure 3 This is a schematic diagram of the joint inversion network FeatIntNet in an embodiment of the present invention; Figure 4 This is a comparison chart of the output results obtained using the joint inversion method of the present invention, the output results obtained using the traditional independent inversion method, and the actual model; Figure 5 This is a schematic diagram of the structure of the electrical and seismic joint inversion device based on deep learning constraints of the present invention. Detailed Implementation

[0021] See also Figures 1 to 4 As shown in the embodiment of the present invention, a joint inversion method for electrical and seismic methods based on deep learning constraints is provided, comprising the following steps: S100, a geological anomaly model is constructed through computer numerical simulation, a resistivity database is constructed based on the anomaly model through forward and inverse numerical calculations, and a seismic velocity database is constructed based on the anomaly model through forward and inverse numerical calculations. Specifically, based on common geological types, anomaly models were constructed using caves, isolated boulders, lithological interfaces, faults, etc., and 14,580 datasets were established. For each sample data, the dataset also has its corresponding inversion results, and is divided into training set, validation set and test set in a 10:1:1 ratio. S200, the resistivity / seismic velocity joint inversion network is trained using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network, FeatIntNet; that is, the resistivity / seismic velocity databases are inverted independently in the training set (i.e.,...). Figure 2 The seismic wave velocity and resistivity models (i.e., the resistivity database and seismic wave velocity database mentioned above) are used as inputs to train the resistivity / seismic wave velocity joint inversion network. In this example, the joint inversion network mainly consists of the FeatIntNet neural network used to find the model mapping relationship and the edge detection operator used to extract edge information.

[0022] Specifically, the dual-channel neural network FeatIntNet is built on the U-net network architecture. FeatIntNet is an end-to-end encoder-decoder convolutional neural network. The encoder consists of four downsampling modules, each including two 3×3 convolutional layers, a non-linear ReLU layer, and a 2×2 max pooling layer. The decoder consists of four upsampling modules. Upsampling allows low-resolution images containing high-level abstract features to be converted to high-resolution images while preserving these features, ultimately yielding a result with the same size as the input data. Each upsampling module includes a deconvolution operation to double the feature map size. After deconvolution, the deconvolution result is concatenated with the corresponding feature map from the encoding step (copy and crop), achieving the fusion of deep and shallow features. Finally, the concatenated feature map undergoes two more 3×3 convolution operations. The specific structure is as follows: Figure 3 As shown.

[0023] S300, Calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and use the objective function to optimize the resistivity / seismic wave velocity joint inversion network to obtain the optimized joint inversion network; S400, input the test data into the optimized joint inversion network to obtain the output result of the optimized joint inversion network.

[0024] In this technical solution, a geological model database is established based on common geological types such as karst caves, isolated boulders, lithological interfaces, and faults. Seismic wave velocity and resistivity data are used for independent inversion. Each inversion begins with the corresponding observation data, initial model, and other prior information, and ends with the inverted model. For each sample data, the dataset also has its corresponding inversion result. Based on the U-net network architecture, a dual-channel neural network, Feature Integration Neural Network (FeatIntNet), is constructed for feature fusion and establishing mapping relationships between models. Seismic wave velocity and resistivity inversion models are used as input to train the network. Through the fusion and extraction of feature information by the network, two physical models with higher structural consistency are output.

[0025] This invention introduces deep learning based on a dual-channel neural network into the traditional joint inversion method, proposes the idea of ​​using deep learning for structural constraints, utilizes a dual-channel neural network to establish the mapping relationship between models, and integrates the complementary features of different physical parameters through orthogonal constraints during network training. This establishes a joint inversion method for electrical and seismic methods based on deep learning constraints, effectively integrating the feature information provided by different physical detection methods, and combining the advantages of detection methods with different resolution sensitivities, thus achieving a more accurate inversion process.

[0026] In some implementations... The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. αp represents the true model value, and αp represents the corresponding weight of each physical model.

[0027] In this technical solution, the aforementioned MSE is used to measure the deviation between the predicted model and the real model, and gradient backpropagation is performed to optimize the neural network.

[0028] In some implementations... The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC m S These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. This is an edge detection operator constructed using the Laplace transform. The Laplace operator is a second-order differential linear operator. In image edge processing, the second-order derivative has stronger edge localization capabilities and better sharpening effects. Therefore, when performing image edge processing, the second-order derivative operator is directly used instead of the first-order derivative, as defined below: .

[0029] Second-order differential operator Defined as gradient ( The divergence of ) ),Right now .in: In this technical solution, the edge information extracted by the edge detection operator is subject to collinearity constraint, and the objective function is added to constrain the training process of the neural network. The edge detection operator further extracts high-level edge feature information, and by adding collinear fused edge feature information to the objective function, the final inversion result changes to the common boundary reflected by the two physical property parameters, thereby further strengthening the structural similarity and improving the accuracy.

[0030] In some implementations, the objective function further includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch.

[0031] In this technical solution, a regularization term is further added to the objective function, which increases the adaptive smoothness constraint, reduces the overfitting problem of deep learning methods, and reduces the loss of edge information as the number of iterations increases.

[0032] Thus, in a preferred implementation, the objective function L is: .

[0033] To clarify the effectiveness of the proposed deep learning-constrained joint inversion method for electrical and seismic methods, the inventors conducted tests and verifications. (See attached document.) Figure 4 As shown, this method has high inversion accuracy and stable performance, especially in its more precise characterization of the boundaries of anomalous bodies.

[0034] See also Figure 5 As shown, according to an embodiment of the present invention, a deep learning-constrained electrical resistivity and seismic inversion device is provided, comprising: The database construction unit is used to construct a geological anomaly model through computer numerical simulation, construct a resistivity database by performing forward and inverse numerical calculations of resistivity based on the anomaly model, and construct a seismic velocity database by performing forward and inverse numerical calculations of seismic velocity based on the anomaly model. Specifically, based on common geological types, anomaly models were constructed using caves, isolated boulders, lithological interfaces, faults, etc., and 14,580 datasets were established. For each sample data, the dataset also has its corresponding inversion results, and is divided into training set, validation set and test set in a 10:1:1 ratio. The neural network training unit is used to train the constructed resistivity / seismic velocity joint inversion network using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network, FeatIntNet; that is, it independently inverts each data point in the training set (i.e.,...). Figure 2The seismic wave velocity and resistivity models (i.e., the resistivity database and seismic wave velocity database mentioned above) are used as inputs to train the resistivity / seismic wave velocity joint inversion network. In this example, the joint inversion network mainly consists of the FeatIntNet neural network used to find the model mapping relationship and the edge detection operator used to extract edge information.

[0035] Specifically, the dual-channel neural network FeatIntNet is built on the U-net network architecture. FeatIntNet is an end-to-end encoder-decoder convolutional neural network. The encoder consists of four downsampling modules, each including two 3×3 convolutional layers, a non-linear ReLU layer, and a 2×2 max pooling layer. The decoder consists of four upsampling modules. Upsampling allows low-resolution images containing high-level abstract features to be converted to high-resolution images while preserving these features, ultimately yielding a result with the same size as the input data. Each upsampling module includes a deconvolution operation to double the feature map size. After deconvolution, the deconvolution result is concatenated with the corresponding feature map from the encoding step (copy and crop), achieving the fusion of deep and shallow features. Finally, the concatenated feature map undergoes two more 3×3 convolution operations. The specific structure is as follows: Figure 3 As shown.

[0036] The neural network optimization unit is used to calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and to optimize the resistivity / seismic wave velocity joint inversion network using the objective function to obtain the optimized joint inversion network. The test output unit is used to input test data into the optimized joint inversion network and obtain the output result of the optimized joint inversion network.

[0037] In this technical solution, a geological model database is established based on common geological types such as karst caves, isolated boulders, lithological interfaces, and faults. Seismic wave velocity and resistivity data are used for independent inversion. Each inversion begins with the corresponding observation data, initial model, and other prior information, and ends with the inverted model. For each sample data, the dataset also has its corresponding inversion results. Based on the U-net network architecture, a dual-channel neural network, Feature Integration Neural Network (FeatIntNet), is constructed for feature fusion and establishing mapping relationships between models. Seismic and resistivity separate inversion models are used as inputs to train the network. Through the fusion and extraction of feature information by the network, two physical models with higher structural consistency are output.

[0038] This invention introduces deep learning based on a dual-channel neural network into the traditional joint inversion method, proposes the idea of ​​using deep learning for structural constraints, utilizes a dual-channel neural network to establish the mapping relationship between models, and integrates the complementary features of different physical parameters through orthogonal constraints during network training. This establishes a joint inversion method for electrical and seismic methods based on deep learning constraints, effectively integrating the feature information provided by different physical detection methods, and combining the advantages of detection methods with different resolution sensitivities, thus achieving a more accurate inversion process.

[0039] In some implementations... The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. αp represents the true model value, and αp represents the corresponding weight of each physical model.

[0040] In this technical solution, the aforementioned MSE is used to measure the deviation between the predicted model and the real model, and gradient backpropagation is performed to optimize the neural network.

[0041] In some implementations... The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC m S These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. This is an edge detection operator constructed using the Laplace transform. The Laplace operator is a second-order differential linear operator. In image edge processing, the second-order derivative has stronger edge localization capabilities and better sharpening effects. Therefore, when performing image edge processing, the second-order derivative operator is directly used instead of the first-order derivative, as defined below: .

[0042] Second-order differential operator Defined as gradient ( The divergence of ) ),Right now .in: In this technical solution, the edge information extracted by the edge detection operator is subject to collinearity constraint, and the objective function is added to constrain the training process of the neural network. The edge detection operator further extracts high-level edge feature information, and by adding collinear fused edge feature information to the objective function, the final inversion result changes to the common boundary reflected by the two physical property parameters, thereby further strengthening the structural similarity and improving the accuracy.

[0043] In some implementations, the objective function further includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch.

[0044] In this technical solution, a regularization term is further added to the objective function, which increases the adaptive smoothness constraint, reduces the overfitting problem of deep learning methods, and reduces the loss of edge information as the number of iterations increases.

[0045] Thus, in a preferred implementation, the objective function L is: .

[0046] The present invention also provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the above-described deep learning-constrained joint inversion method of electrical and seismic methods.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0052] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A joint inversion method for electrical and seismic methods based on deep learning constraints, characterized in that, Includes the following steps: S100, a geological anomaly model is constructed through computer numerical simulation, a resistivity database is constructed based on the anomaly model through forward and inverse numerical calculations, and a seismic velocity database is constructed based on the anomaly model through forward and inverse numerical calculations. S200, the resistivity / seismic velocity joint inversion network is trained using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network FeatIntNet. S300, Calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and use the objective function to optimize the resistivity / seismic wave velocity joint inversion network to obtain the optimized joint inversion network; S400, input the test data into the optimized joint inversion network to obtain the output result of the optimized joint inversion network; The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. The values ​​are the actual model values, and αp is the corresponding weight of each physical model. The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC m S These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. An edge detection operator constructed using Laplace transform; The objective function also includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch; The objective function L is: 。 2. A deep learning-constrained electrical resistivity and seismic inversion device, characterized in that, include: The database construction unit is used to construct a geological anomaly model through computer numerical simulation, construct a resistivity database by performing forward and inverse numerical calculations of resistivity based on the anomaly model, and construct a seismic velocity database by performing forward and inverse numerical calculations of seismic velocity based on the anomaly model. The neural network training unit is used to train the constructed resistivity / seismic velocity joint inversion network using the constructed resistivity database and seismic velocity database. The resistivity / seismic velocity joint inversion network is a dual-channel neural network, FeatIntNet. The neural network optimization unit is used to calculate the objective function based on the output of the resistivity / seismic wave velocity joint inversion network, and to optimize the resistivity / seismic wave velocity joint inversion network using the objective function to obtain the optimized joint inversion network. The test output unit is used to input test data into the optimized joint inversion network and obtain the output result of the optimized joint inversion network. The objective function includes the data mismatch term MSE. Where DC and S represent resistivity and seismic wave velocity, The predicted value output by the neural network. The values ​​are the actual model values, and αp is the corresponding weight of each physical model. The objective function also includes a collinearity constraint term Ledge for edge detection. , where m DC m S These represent the DC resistivity model and the seismic wave velocity model output by the dual-channel neural network, respectively. An edge detection operator constructed using Laplace transform; The objective function also includes a regularization term. , , where λ is the weight factor of the smoothing regularization term, Epoch is the total number of network training iterations, and epoch is the current training iteration, with values ​​of 1, 2, ..., Epoch; The objective function L is: 。 3. A computer-readable storage medium, characterized in that, It stores multiple instructions, which are adapted to be loaded and executed by the processor of the terminal device as described in claim 1, which is a deep learning-constrained joint inversion method for electrical and seismic methods.

Images