MultiResUNet neural network intelligent speed modeling method and device based on physical constraint

Abstract

The invention provides a MultiResUNet neural network intelligent speed modeling method and device based on physical constraints. The method comprises the following steps: constructing a MultiResUNet network model; designing different types of seismic velocity models, and performing forward modeling on the velocity models to obtain a corresponding data set; the data set is preprocessed, the preprocessing comprises normalization and resampling, and the data set is divided into a training set and a test set according to a preset proportion; carrying out training optimization on the network by taking the loss of the seismic record obtained by forward modeling of the network prediction speed model and the original seismic record as physical constraints; and testing the trained network model by using the test set. According to the method, the problem that the network precision of a conventional U-Net network is reduced due to a semantic gap between an encoder and a decoder is solved, so that higher prediction precision is obtained, high-precision seismic velocity modeling is successfully realized, efficient processing of seismic data can be performed, and effective guidance and help are provided for oil-gas exploration.

Description

Technical Field

[0001] The present invention belongs to the field of exploration geophysics, and in particular, relates to a MultiResUNet neural network intelligent velocity modeling method and device based on physical constraints. Background Art

[0002] Currently, velocity model construction is an essential step in seismic exploration because it is used throughout the entire process of seismic exploration, including seismic data acquisition, processing and interpretation. A good velocity model is a prerequisite for migration imaging technology. Many techniques explored by predecessors, such as migration velocity analysis, tomography and full waveform inversion, have the same purpose, which is to build a more accurate velocity model.

[0003] With the development of artificial intelligence, deep learning methods have played a great role in many fields such as natural language understanding and visual target recognition. Geophysicists are also actively exploring the application of deep learning related technologies in seismic data processing such as noise suppression, first arrival picking, velocity spectrum picking, seismic data reconstruction and interpolation, and scattered wave imaging. In recent years, experts and scholars at home and abroad have gradually begun to apply deep learning methods to the field of seismic velocity modeling. At present, the existing neural networks for intelligent velocity modeling are mainly U-Net series networks.

[0004] Conventional U-Net series networks mainly use skip connection layers to directly connect the encoder and decoder. However, there is often a semantic gap between the corresponding levels of the encoder and decoder of U-Net networks. Directly using skip connection layers for connection will affect the network performance. Therefore, a more suitable network structure is needed for speed modeling. Existing intelligent modeling technologies based on deep learning often directly use the loss between the predicted speed model and the true speed model to perform back propagation and update network parameters. There is a lack of physical constraints during the training process and it is strongly dependent on the true speed model. Therefore, introducing physical constraints during the training process is very necessary to enhance network performance and improve network applicability. Summary of the invention

[0005] In view of the shortcomings of the prior art, the embodiments of the present application provide a MultiResUNet neural network intelligent velocity modeling method and device based on physical constraints, which successfully realizes high-precision seismic velocity modeling. First, a MultiResUNet network is constructed, which uses a U-Net network as the main network framework, and replaces the convolutional layer and the skip connection layer in the traditional U-Net network with MultiRes blocks and Res paths, respectively. By replacing the convolutional layer with a MultiRes block, features of different scales in the seismic record can be identified, which is conducive to improving the network's ability to depict details. In addition, by replacing the skip connection layer with the Res path, the performance degradation problem caused by the semantic gap between the corresponding levels of the encoder-decoder in the conventional U-Net series network can be avoided, thereby obtaining higher prediction accuracy. In addition, the present invention introduces physical constraints in the network training process. It takes earthquake records as input, predicts the earthquake velocity model through the network, and then uses the velocity model to perform new earthquake records according to the original observation system. The loss between the model and the input earthquake record is then used to perform backpropagation and optimize the network parameters, thereby enhancing the stability and efficiency of network learning, enhancing the generalization ability of the network, and making the model results more interpretable and with higher prediction accuracy, providing support and assistance for subsequent earthquake data processing.

[0006] In a first aspect, an embodiment of the present application provides a MultiResUNet neural network intelligent velocity modeling method based on physical constraints, including: S1, constructing a MultiResUNet network model; S2, designing different types of seismic velocity models, and forward modeling the velocity models to obtain corresponding data sets; S3, preprocessing the data set, wherein the preprocessing includes normalization and resampling, and dividing the data set into a training set and a test set according to a preset ratio; S4, training and optimizing the network based on physical constraints; S5, using the test set to test the trained network model.

[0007] Among them, in step S1, the MultiResUNet network uses the U-Net network as the main network framework, and replaces the convolutional layer and the cross connection layer in the traditional U-Net network with MultiRes blocks and Res paths respectively.

[0008] Among them, the MultiRes block in the network in step S1 includes a 3×3 convolutional layer, a 5×5 convolutional layer, and a 7×7 convolutional layer arranged in parallel; the Res path in the network in step S1 includes a series of residual structures.

[0009] Among them, in step S1, the MultiResUNet network includes a symmetrical structure of a contraction path for capturing seismic record characteristics and an expansion path for achieving precise positioning. This symmetrical form is an encoder-decoder structure.

[0010] Wherein, in step S2, the seismic velocity model includes a simple velocity model and a complex velocity model; and in step S2, after the velocity model is obtained, seismic records are obtained by finite difference forward simulation.

[0011] In step S3, the data set is divided into a training set and a test set in a ratio of 9:1.

[0012] Among them, in step S4, the earthquake record is used as input, and the earthquake velocity model is predicted through the network. Then, the velocity model is used to perform new earthquake records according to the original observation system. Finally, the loss between the earthquake record and the input earthquake record is used for back propagation and network parameters optimization. The standard mean square error is used to measure the loss between the two earthquake records. The formula is:

[0013]

[0014] In the formula, S RMS is the loss function, is the value of the real shot gather record at the i-th sample point, is the value of the shot gather record at the i-th sample point predicted by the network;

[0015] The specific process of optimizing network parameters in step 4 is:

[0016] (1) Calculate the gradient of the loss function

[0017]

[0018] In the formula, ω is the network parameter to be optimized, N is the number of samples, l is the loss of a single sample, and L is the sum of the losses of all samples;

[0019] (2) Use the gradient descent method to update the network parameters ω:

[0020]

[0021] In the formula, ω n is the network parameter updated after the nth gradient descent, ω n+1 is the network parameter after the n+1th gradient descent update, and r is the learning rate.

[0022] In a second aspect, an embodiment of the present application provides a MultiResUNet neural network intelligent speed modeling device based on physical constraints, comprising:

[0023] Construction unit, used to build MultiResUNet network model;

[0024] The forward modeling unit is used to design different types of seismic velocity models and perform forward modeling on the velocity models to obtain the corresponding data sets;

[0025] A preprocessing unit, used to preprocess the data set, wherein the preprocessing includes normalization and resampling, and dividing the data set into a training set and a test set according to a preset ratio;

[0026] An optimization unit, used to optimize network training based on physical constraints;

[0027] The test unit is used to test the trained network model using the test set.

[0028] In a third aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0029] In a fourth aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0030] The MultiResUNet neural network intelligent speed modeling method and device based on physical constraints in the embodiment of the present application have the following beneficial effects:

[0031] This application intelligently realizes the direct mapping from seismic records to velocity models, solving the problem of reduced network accuracy due to the semantic gap between the encoder and decoder in conventional U-Net networks, thereby achieving higher prediction accuracy; compared with pure data-driven neural network learning, by adding physical constraints to the network loss function, strengthening the stability and efficiency of network learning, and enhancing the generalization ability of the network, the model results are more interpretable, and high-precision seismic velocity modeling can be achieved. Efficient processing of seismic data can be performed, thereby providing effective guidance and assistance for oil and gas exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of the MultiResUNet neural network intelligent speed modeling method based on physical constraints in an embodiment of the present application;

[0033] Figure 2 Another schematic diagram of the process of the MultiResUNet neural network intelligent speed modeling method based on physical constraints in an embodiment of the present application;

[0034] Figure 3 This is a schematic diagram of the structure of the MultiResUNet network in the embodiment of the present application;

[0035] Figure 4This is a schematic diagram of the structure of the MultiRes block in an embodiment of the present application;

[0036] Figure 5 This is a schematic diagram of the structure of the Res path in the embodiment of the present application;

[0037] Figure 6 This is a two-layer simple model for the embodiment of this application;

[0038] Figure 7 Single shot seismic records for the two-layer simple model forward modeling of the embodiment of the present application;

[0039] Figure 8 A complex velocity model constructed by local cutting of the Marmousi model in the embodiment of the present application;

[0040] Fig. 9 A single shot seismic record of the complex model forward modeling described above in the embodiment of the present application;

[0041] Fig.10 This is the network training loss curve of the embodiment of the present application;

[0042] Fig.11 The network performance test results of the embodiment of this application;

[0043] Fig.12 This is a structural schematic diagram of the MultiResUNet neural network intelligent speed modeling device based on physical constraints in an embodiment of the present application. DETAILED DESCRIPTION

[0044] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0045] In the following introduction, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The following introduction provides multiple embodiments of the present invention, and different embodiments can be replaced or combined, so this application can also be considered to include all possible combinations of the same and / or different embodiments recorded. Therefore, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then this application should also be considered to include embodiments containing one or more of all other possible combinations of features A, B, C, and D, even though the embodiment may not be clearly recorded in the following text.

[0046] Embodiment 1

[0047] The present invention provides a MultiResUNet neural network intelligent velocity modeling method based on physical constraints, which successfully realizes high-precision seismic velocity modeling.

[0048] like Figure 1As shown, the MultiResUNet neural network intelligent velocity modeling method based on physical constraints of the present application includes: step S1, constructing a MultiResUNet network model; step S2, designing different types of seismic velocity models, and forward modeling the velocity models to obtain corresponding data sets; step S3, preprocessing the data set, the preprocessing includes normalization, resampling, and dividing the data set into a training set and a test set according to a preset ratio; step S4, training and optimizing the network based on physical constraints; step S5, using the test set to test the trained network model.

[0049] This application develops a new deep learning network based on the MultiResUNet neural network intelligent speed modeling method with physical constraints, which solves the problem of reduced network accuracy due to the semantic gap between the encoder and the decoder of the conventional U-Net network, thereby obtaining a higher-precision speed model. Compared with the use of the real speed model and the predicted speed model to calculate the loss, the physical constraints are added to enhance the stability and efficiency of network learning, enhance the generalization ability of the network, and make the model results more interpretable, thereby obtaining a modeling result with higher predictive accuracy.

[0050] Embodiment 2

[0051] This application uses multi-shot seismic records as the input of the MultiResUNet neural network, predicts the seismic velocity model through the network, and then uses the velocity model to perform new seismic records according to the original observation system, and then uses the loss of the input seismic records to perform back propagation and update the network parameters. The MultiResUNet network uses the U-Net network as the main network framework, and replaces the convolutional layer and the cross-connection layer in the traditional U-Net network with MultiRes blocks and Res paths respectively. By introducing physical constraints in the training process and using the MultiResUNet network instead of the conventional U-Net series network, the present invention achieves high-precision prediction of simple velocity models and complex velocity models. The present invention is described in detail below in conjunction with the accompanying drawings.

[0052] like Figure 2 As shown, the MultiResUNet neural network intelligent speed modeling method based on physical constraints proposed in the present invention has the following specific steps:

[0053] Step S10: Construct MultiResUNet network model. Figure 3As shown in the figure, the MultiResUNet network uses the U-Net network as the main network framework, replacing the convolutional layer and the cross-connection layer in the traditional U-Net network with MultiRes blocks and Res paths respectively. The MultiResUNet network consists of a symmetrical structure of a contraction path for capturing seismic recording characteristics and an expansion path for precise positioning. This symmetric form is an encoder-decoder structure. Figure 4 As shown in the figure, the MultiRes block in the network is composed of 3×3 convolutional layers, 5×5 convolutional layers, and 7×7 convolutional layers arranged in parallel. Through this multi-resolution design, the network can obtain features of different scales in seismic records and better obtain detailed features in seismic records. Figure 5 As shown in the figure, the Res path in the network is composed of a series of residual structures. Through this residual design, the performance degradation problem caused by the semantic gap between the corresponding levels of the encoder and decoder in the conventional U-Net series network can be effectively solved, thereby achieving higher prediction accuracy.

[0054] Step S20: forward modeling to obtain training data. In order to make the MultiResUNet network suitable for various velocity models and have stronger generalization ability, the present invention designs different types of seismic velocity models, including velocity models with 2-6 layers of different layer thicknesses, layer positions, and dip angles, and velocity models composed of Marmousi model blocks. The present invention designs a total of 500 velocity models with 2-6 layers of different layer thicknesses, layer positions, and dip angles, and a complex velocity model composed of 1,000 Marmousi blocks.

[0055] After the velocity model is obtained, the seismic record is obtained by finite difference forward simulation. The observation system of the forward part of the present invention is set as follows: the shot points are set on the surface, the shot spacing is 20m, and they are evenly distributed from 10 meters to 1270 meters in the horizontal direction, with a total of 64 shots; the detection points are also set on the surface, and the track spacing is 10m, and they are evenly distributed from 0 meters to 1270 meters in the horizontal direction, with a total of 128 tracks. In the forward modeling process, a Ricker wavelet with a main frequency of 30Hz is used as the source, and the time sampling interval is set to 8ms. The shot points move in sequence, and 128 detectors are fixed for reception. Finally, seismic shot data with a size of 64*128*256 is obtained.

[0056] like Figure 6 A simple two-layer model is shown. Figure 7 The following is a single shot record obtained through forward modeling. Figure 8 The figure shows a complex velocity model constructed by local cutting of the Marmousi model. Fig. 9 for Figure 8 A single-shot earthquake record being performed by the model.

[0057] Step S30: normalize and resample the earthquake data being performed by the above model, and divide it into a training set and a test set in a ratio of 9:1 to complete the data set production.

[0058] Step S40: Network training based on physical constraints

[0059] The training environment is:

[0060] (1) Processor: 13th Gen Intel(R) Core(TM) i7-13700KF;

[0061] (2) Memory: 64 GB;

[0062] (3) System type: 64-bit operating system;

[0063] (4) Graphics card: NVIDIA GeForce RTX4090, 24GB;

[0064] (5) Platform: PyTorch neural network framework based on PyCharm compiler.

[0065] Based on the above training environment, the MultiResUNet network based on physical constraints in the present invention is trained using the training set. First, the earthquake record is used as input, and the earthquake velocity model is predicted through the network. Then, the velocity model is used to perform new earthquake records according to the original observation system. Finally, the loss between the earthquake record and the input earthquake record is used for back propagation and optimization of network parameters. The standard mean square error (NRMS) is used to measure the loss between the two earthquake records, as shown in formula (1):

[0066]

[0067] Among them, S RMS is the loss function, is the value of the real shot gather record at the i-th sample point, is the value of the shot gather record at the i-th sample point predicted by the network.

[0068] Taking the MultiResUNet network with physical constraints as an example, the specific process of optimizing network parameters is as follows:

[0069] (1) Calculate the gradient of the loss function

[0070]

[0071] Among them, ω is the network parameter to be optimized, N is the number of samples, l is the loss of a single sample, and L is the sum of the losses of all samples.

[0072] (2) Use the gradient descent method to update the network parameters ω:

[0073]

[0074] Among them, ω n is the network parameter updated after the nth gradient descent, ω n+1 is the network parameter after the n+1th gradient descent update, and r is the learning rate.

[0075] The loss function curve of the network of the present invention is compared with that of the conventional U-Net network without physical constraints to verify the advantages of the network of the present invention. The training parameters of the two networks are the same, and the specific settings are as follows: maximum learning times (MAX_EPOCH) = 500, batch size (BATCH_SIZE) = 64, learning rate (LR) = 0.001. The loss reduction curve of the MultiResUNet network based on physical constraints in the present invention is shown in Figure 2. Fig.10 As shown in (a) in the figure, the loss decrease curve of the conventional U-Net network without physical constraints is as follows Fig.10 Comparing the two curves, we can see Fig.10 The loss reduction curve in (b) shows a very drastic cliff-like drop, and the curve oscillates severely, which shows that for a conventional U-Net network without physical constraints, it is difficult for the network parameters to approach the optimal solution in a relatively stable trend, thus affecting the prediction results.

[0076] By comparing with a conventional U-Net network without physical constraints, it can be seen that compared with using the real speed model and the predicted speed model to calculate the loss, the method of the present invention enhances the stability and efficiency of network learning, enhances the generalization ability of the network, and makes the model results more interpretable, thereby achieving higher prediction accuracy.

[0077] Step S50: Use the test set to test the trained network model

[0078] Fig.11 This is a comparative analysis chart of the prediction effect of a certain block model of the Marmousi speed model in the test set, such as Fig.11 As shown, Fig.11 (a) is the true velocity model. Fig.11 (b) is the result predicted by the MultiResUNet network based on physical constraints in the present invention. Fig.11 (c) is the result predicted by the conventional U-Net network without physical constraints. Fig.11 (b) Fig.11In (c), it can be seen that the velocity model predicted by the conventional U-Net network without physical constraints cannot accurately depict the complex velocity interface, and the inverted velocity value has a large error. The velocity model predicted by the MultiResUNet network based on physical constraints in the present invention can accurately depict the details in the velocity model. Fig.11 The real velocity model a in is basically consistent, so the present invention has great guiding significance for practical work.

[0079] The method of the present application includes: constructing a MultiResUNet network model; designing different types of seismic velocity models, and forward modeling the velocity models to obtain corresponding data sets; preprocessing the data sets, wherein the preprocessing includes normalization and resampling, and dividing the data sets into training sets and test sets according to a preset ratio; training and optimizing the network based on the loss of seismic records obtained by forward modeling of the network prediction velocity model and the original seismic records as physical constraints; and testing the trained network model using the test set. The present invention solves the problem that the conventional U-Net network has reduced network accuracy due to the semantic gap between the encoder and the decoder, thereby obtaining higher prediction accuracy, successfully realizing high-precision seismic velocity modeling, and being able to efficiently process seismic data, thereby providing effective guidance and assistance for oil and gas exploration.

[0080] The present invention innovatively applies the MultiResUNet network to the process of intelligent modeling of seismic velocity. This network uses the U-Net network as the main network framework, and replaces the convolutional layer and the jump connection layer in the traditional U-Net network with MultiRes blocks and Res paths respectively. By replacing the convolutional layer with a MultiRes block, features of different scales in seismic records can be identified, which is conducive to improving the network's ability to depict details. In addition, by replacing the jump connection layer with the Res path, the performance degradation problem caused by the semantic gap between the corresponding levels of the encoder-decoder in the conventional U-Net series network can be avoided, thereby obtaining higher prediction accuracy.

[0081] The present invention innovatively introduces physical constraints in the network training process. It takes earthquake records as input, predicts earthquake velocity models through the network, and then uses the velocity model to perform new earthquake records according to the original observation system. Then, it uses the loss of the input earthquake records to perform back propagation and optimize network parameters. By adding physical constraints to the network, the deficiency of the traditional intelligent velocity modeling method lacking physical constraints is overcome, thereby strengthening the stability and efficiency of network learning, enhancing the generalization ability of the network, and making the model results more interpretable, and obtaining modeling results with higher prediction accuracy.

[0082] The present invention has good prediction effects for both simple models and complex models, can realize high-precision seismic velocity modeling, and can perform efficient processing of seismic data, thereby providing effective guidance and assistance for oil and gas exploration.

[0083] Embodiment 3

[0084] like Fig.12 As shown, the present application also provides a MultiResUNet neural network intelligent velocity modeling device based on physical constraints, including: a construction unit 201, used to construct a MultiResUNet network model; a forward modeling unit 202, used to design different types of seismic velocity models, and forward model the velocity model to obtain a corresponding data set; a preprocessing unit 203, used to preprocess the data set, the preprocessing includes normalization, resampling, and dividing the data set into a training set and a test set according to a preset ratio; an optimization unit 204, used to train and optimize the network based on physical constraints; a testing unit 205, used to test the trained network model using a test set.

[0085] In the present application, an embodiment of a MultiResUNet neural network intelligent speed modeling device based on physical constraints is basically similar to an embodiment of a MultiResUNet neural network intelligent speed modeling method based on physical constraints. For relevant matters, please refer to the introduction of the embodiment of a MultiResUNet neural network intelligent speed modeling method based on physical constraints.

[0086] The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned MultiResUNet neural network intelligent speed modeling method steps based on physical constraints. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a microdrive, and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0087] The present application also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A MultiResUNet neural network intelligent speed modeling method based on physical constraints, characterized in that: include: S1, build MultiResUNet network model; S2, design different types of seismic velocity models, and perform forward modeling on the velocity models to obtain corresponding data sets; S3, preprocessing the data set, wherein the preprocessing includes normalization and resampling, and dividing the data set into a training set and a test set according to a preset ratio; S4, training and optimizing the network based on physical constraints; S5, use the test set to test the trained network model.

2. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to claim 1 is characterized in that: In step S1, the MultiResUNet network uses the U-Net network as the main network framework, and replaces the convolutional layer and the cross-connection layer in the traditional U-Net network with MultiRes blocks and Res paths respectively.

3. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to claim 2 is characterized in that: The MultiRes block in the network in step S1 includes a 3×3 convolutional layer, a 5×5 convolutional layer, and a 7×7 convolutional layer arranged in parallel; the Res path in the network in step S1 includes a series of residual structures.

4. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to any one of claims 1 to 3, characterized in that: In step S1, the MultiResUNet network includes a symmetrical structure of a contraction path for capturing seismic recording characteristics and an expansion path for achieving precise positioning. This symmetrical form is an encoder-decoder structure.

5. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to any one of claims 1 to 3, characterized in that: In step S2, the seismic velocity model includes a simple velocity model and a complex velocity model; and in step S2, after the velocity model is obtained, seismic records are obtained by finite difference forward simulation.

6. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to any one of claims 1 to 3, characterized in that: In step S3, the data set is divided into a training set and a test set in a ratio of 9:

1.

7. The MultiResUNet neural network intelligent velocity modeling method based on physical constraints according to any one of claims 1 to 3, characterized in that: In step S4, the earthquake record is used as input, and the earthquake velocity model is predicted through the network. Then, the velocity model is used to perform new earthquake records according to the original observation system. Finally, the loss between the earthquake record and the input earthquake record is used for back propagation and network parameters optimization. The standard mean square error is used to measure the loss between the two earthquake records. The formula is: Among them, S RMS is the loss function, is the value of the real shot gather record at the i-th sample point, is the value of the shot gather record at the i-th sample point predicted by the network; The specific process of optimizing the network parameters in step S4 includes: (1) Calculate the gradient of the loss function Among them, ω is the network parameter to be optimized, N is the number of samples, l is the loss of a single sample, and L is the sum of the losses of all samples; (2) Use the gradient descent method to update the network parameters ω: Among them, ω n is the network parameter updated after the nth gradient descent, ω n+1 is the network parameter after the n+1th gradient descent update, and r is the learning rate.

8. A MultiResUNet neural network intelligent velocity modeling device based on physical constraints, characterized in that: include: Construction unit, used to build MultiResUNet network model; The forward modeling unit is used to design different types of seismic velocity models and perform forward modeling on the velocity models to obtain the corresponding data sets; A preprocessing unit, used to preprocess the data set, wherein the preprocessing includes normalization and resampling, and dividing the data set into a training set and a test set according to a preset ratio; An optimization unit, used to optimize network training based on physical constraints; The test unit is used to test the trained network model using the test set.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images