Neural network inversion method, device, electronic equipment and medium

Abstract

The application discloses a neural network inversion method and device, electronic equipment and medium. The method can include: establishing a basic training set and constructing an initial neural network; updating the neural network parameters of the initial neural network to obtain an optimized neural network; for a target seismic wave gather, inverting an initial subsurface structure model through the optimized neural network; performing full waveform inversion on the initial subsurface structure model, comparing the inversion result with an error of a known full waveform inversion result, adjusting the optimized neural network, and obtaining a final neural network; and inverting through the final neural network. The application can quickly and accurately obtain the subsurface structure of a target region by continuously optimizing the neural network.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration, and more specifically, to a neural network inversion method, device, electronic equipment and medium. Background Art

[0002] The most critical factor in seismic imaging is the accuracy of the velocity model. Currently, there are many approaches to velocity modeling, including interpolation from well data, traveltime-based tomography, and full-waveform inversion (FWI) using seismic waveforms and energy. Full-waveform inversion (FWI) is the most comprehensive in principle. The specific steps are: first, an initial velocity model is given; then, seismic gathers are generated through forward modeling. The model is updated based on the differences between these and actual seismic records, and the final subsurface structure is iteratively determined. While this method theoretically yields more accurate subsurface velocity information, it is computationally intensive and depends on the accuracy of the initial model, implying significant manual computation. To improve computational efficiency, some researchers have proposed using neural networks to derive subsurface velocity models directly from seismic gather data. This approach leverages the principles of machine learning, using deep learning training on the seismic model itself and the forward-modeled seismic gathers to generate an operator that directly converts seismic gathers into velocity models. While this technique works well with modeled data, it often falls short when applied to real-world data. Real-world data often lacks reliable subsurface velocity information as labels, making it difficult to apply to neural network training.

[0003] Therefore, it is necessary to develop a neural network inversion method, device, electronic equipment and medium based on traditional iterative inversion.

[0004] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0005] The present invention proposes a neural network inversion method, device, electronic equipment and medium, which can quickly and accurately obtain the underground structure of the target area by continuously optimizing the neural network.

[0006] In a first aspect, an embodiment of the present disclosure provides a neural network inversion method, comprising:

[0007] Establish a basic training set and build an initial neural network;

[0008] Updating the neural network parameters for the initial neural network to obtain an optimized neural network;

[0009] For the target seismic wave gather, an initial underground structure model is obtained by inversion using the optimized neural network;

[0010] Performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining a final neural network;

[0011] Inversion is performed through the final neural network.

[0012] Preferably, updating the neural network parameters for the initial neural network to obtain an optimized neural network includes:

[0013] Inputting the data of the basic training set into the initial neural network to obtain output data;

[0014] The error between the output data and the underground structure model is calculated, and the neural network parameters are updated to obtain an optimized neural network.

[0015] Preferably, performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining the final neural network comprises:

[0016] Performing full waveform inversion on the initial underground structure model to obtain an initial inversion result;

[0017] The error between the initial inversion result and the known full waveform inversion result is calculated, the neural network parameters are updated, and the optimized neural network is adjusted. If the error is less than a set threshold, the final neural network is obtained.

[0018] Preferably, the neural network parameters are updated using formula (1):

[0019] loss=λ1|OM|1+λ2|OM|2 (1)

[0020] Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, and λ1 and λ2 are weight coefficients.

[0021] Preferably, it also includes:

[0022] The target seismic wave gather is replaced, the final neural network is used as the optimized neural network, and the neural network is repeatedly learned and optimized to improve the inversion accuracy.

[0023] As a specific implementation of the embodiment of the present disclosure,

[0024] In a second aspect, the present disclosure also provides a neural network inversion device, including:

[0025] Initial neural network building module, establish basic training set, and build initial neural network;

[0026] An optimized neural network construction module is used to update the neural network parameters of the initial neural network to obtain an optimized neural network;

[0027] An underground structure model inversion module is configured to obtain an initial underground structure model by inverting the target seismic wave gathers through the optimized neural network;

[0028] A final neural network construction module performs full waveform inversion on the initial underground structure model, compares the error between the inversion result and the known full waveform inversion result, and adjusts the optimized neural network to obtain a final neural network;

[0029] An inversion module performs inversion through the final neural network.

[0030] Preferably, updating the neural network parameters for the initial neural network to obtain an optimized neural network includes:

[0031] Inputting the data of the basic training set into the initial neural network to obtain output data;

[0032] The error between the output data and the underground structure model is calculated, and the neural network parameters are updated to obtain an optimized neural network.

[0033] Preferably, performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining the final neural network comprises:

[0034] Performing full waveform inversion on the initial underground structure model to obtain an initial inversion result;

[0035] The error between the initial inversion result and the known full waveform inversion result is calculated, the neural network parameters are updated, and the optimized neural network is adjusted. If the error is less than a set threshold, the final neural network is obtained.

[0036] Preferably, the neural network parameters are updated using formula (1):

[0037] loss=λ1|OM|1+λ2|OM|2 (1)

[0038] Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, and λ1 and λ2 are weight coefficients.

[0039] Preferably, it also includes:

[0040] The target seismic wave gather is replaced, the final neural network is used as the optimized neural network, and the neural network is repeatedly learned and optimized to improve the inversion accuracy.

[0041] In a third aspect, an embodiment of the present disclosure further provides an electronic device, the electronic device comprising:

[0042] a memory storing executable instructions;

[0043] A processor runs the executable instructions in the memory to implement the neural network inversion method.

[0044] In a fourth aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the neural network inversion method is implemented.

[0045] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be described in detail in the accompanying drawings and subsequent detailed descriptions incorporated herein, which together serve to explain the specific principles of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0047] Figure 1 A flowchart showing the steps of a neural network inversion method according to an embodiment of the present invention is shown.

[0048] Figure 2 A schematic diagram of an actual seismic wave gather according to an embodiment of the present invention is shown.

[0049] Figure 3 A schematic diagram showing underground structure information obtained through manual participation according to an embodiment of the present invention is shown.

[0050] Figure 4 A schematic diagram showing an optimized neural network inversion result according to an embodiment of the present invention is shown.

[0051] Figure 5 A schematic diagram of a target seismic wave gather according to an embodiment of the present invention is shown.

[0052] Figure 6 A schematic diagram showing the target seismic wave gather in the optimized neural network inversion result according to an embodiment of the present invention is shown.

[0053] Figure 7 A schematic diagram showing full waveform inversion results according to an embodiment of the present invention is shown.

[0054] Figure 8 A schematic diagram showing the final neural network inversion result according to one embodiment of the present invention is shown.

[0055] Figure 9 A schematic diagram of a new seismic wave gather according to an embodiment of the present invention is shown.

[0056] Figure 10 A schematic diagram showing the final neural network inversion result of a new seismic wave gather according to an embodiment of the present invention is shown.

[0057] Figure 11 A block diagram of a neural network inversion device according to an embodiment of the present invention is shown.

[0058] Description of reference numerals:

[0059] 201. Initial neural network construction module; 202. Optimized neural network construction module; 203. Underground structure model inversion module; 204. Final neural network construction module; 205. Inversion module. DETAILED DESCRIPTION

[0060] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0061] The present invention provides a neural network inversion method, comprising:

[0062] Establish a basic training set and build an initial neural network.

[0063] Specifically, part of the forward synthetic data or part of the actual data that has been used for initial underground structure analysis in the past is used to make the seismic wave gathers and the initial underground structure model form a one-to-one corresponding data set to establish a basic training set for later training.

[0064] Construct an initial neural network. The structure of the neural network is selected as a convolutional neural network. Depending on the type of data, two-dimensional convolution or three-dimensional convolution can be selected. The characteristic information of the seismic wave gather is continuously obtained through convolution, and then the corresponding underground structure information is obtained through the deconvolution layer. At this time, the parameters in the neural network are initialized to random and have nothing to do with wave propagation. Therefore, the use of synthetic data can also provide the relationship between seismic wave records and underground structures to train the neural network.

[0065] The neural network parameters are updated for the initial neural network to obtain an optimized neural network. In one example, the neural network parameters are updated for the initial neural network to obtain the optimized neural network, including: inputting data from a basic training set into the initial neural network to obtain output data; calculating the error between the output data and the underground structure model, updating the neural network parameters, and obtaining the optimized neural network.

[0066] In one example, the neural network parameters are updated using formula (1):

[0067] loss=λ1|OM|1+λ2|OM|2 (1)

[0068] Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, λ1 and λ2 are weight coefficients, and λ1+λ2=1.

[0069] Specifically, the data from the basic training set is input into the initial neural network to obtain output data. The error between the output data and the underground structure model is calculated, and the neural network parameters are updated using formula (1) to obtain an optimized neural network. When the loss is small enough, the output of the neural network is considered to be close to the underground structure model M, and a preliminarily trained neural network is obtained. This neural network is considered to have the ability to invert simple underground structures, and its inversion result can be used as the initial model for the traditional iterative method.

[0070] For the target seismic wave gathers, the initial underground structure model is obtained by inversion through optimizing the neural network.

[0071] Specifically, an optimized neural network is used to invert the initial underground structure model corresponding to the target seismic wave gather. Here, the initial model of the seismic wave gather is no longer manually generated. Instead, the trained neural network is used to output the seismic wave gather, reducing manual intervention.

[0072] Performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting and optimizing the neural network, and obtaining the final neural network; in one example, performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting and optimizing the neural network, and obtaining the final neural network includes: performing full waveform inversion on the initial underground structure model to obtain an initial inversion result; calculating the error between the initial inversion result and the known full waveform inversion result, updating the neural network parameters, adjusting and optimizing the neural network, and obtaining the final neural network if the error is less than a set threshold.

[0073] In one example, the method further includes: replacing the target seismic wave gather, using the final neural network as the optimized neural network, and repeatedly performing learning and optimization on the neural network to improve the inversion accuracy.

[0074] Specifically, full waveform inversion is performed on the initial underground structure model to obtain the initial inversion result. The error between the initial inversion result and the known full waveform inversion result is calculated, and the neural network parameters are updated using formula (1). The neural network is adjusted and optimized. If the error is less than the set threshold, the final neural network is obtained. Compared with the optimized neural network, the final neural network learns more detailed information about the underground structure from the traditional iterative inversion method, resulting in a more accurate inversion result.

[0075] It is also possible to replace the target seismic wave gather with a new one, use the final neural network as the optimized neural network, and repeatedly perform learning and optimization on the neural network to improve the inversion accuracy.

[0076] Inversion is performed through the final neural network.

[0077] Specifically, the final neural network is used to invert the new seismic wave gathers.

[0078] The present invention also provides a neural network inversion device, comprising:

[0079] Initial neural network building module, establishes basic training set, and builds initial neural network.

[0080] Specifically, part of the forward synthetic data or part of the actual data that has been used for initial underground structure analysis in the past is used to make the seismic wave gathers and the initial underground structure model form a one-to-one corresponding data set to establish a basic training set for later training.

[0081] Construct an initial neural network. The structure of the neural network is selected as a convolutional neural network. Depending on the type of data, two-dimensional convolution or three-dimensional convolution can be selected. The characteristic information of the seismic wave gather is continuously obtained through convolution, and then the corresponding underground structure information is obtained through the deconvolution layer. At this time, the parameters in the neural network are initialized to random and have nothing to do with wave propagation. Therefore, the use of synthetic data can also provide the relationship between seismic wave records and underground structures to train the neural network.

[0082] An optimized neural network construction module is used to update the neural network parameters of the initial neural network to obtain an optimized neural network. In one example, the neural network parameters are updated for the initial neural network to obtain the optimized neural network, including: inputting data from a basic training set into the initial neural network to obtain output data; calculating the error between the output data and the underground structure model, updating the neural network parameters, and obtaining the optimized neural network.

[0083] In one example, the neural network parameters are updated using formula (1):

[0084] loss=λ1|OM|1+λ2|OM|2 (1)

[0085] Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, and λ1 and λ2 are weight coefficients.

[0086] Specifically, the data from the basic training set is input into the initial neural network to obtain output data. The error between the output data and the underground structure model is calculated, and the neural network parameters are updated using formula (1) to obtain an optimized neural network. When the loss is small enough, the output of the neural network is considered to be close to the underground structure model M, and a preliminarily trained neural network is obtained. This neural network is considered to have the ability to invert simple underground structures, and its inversion result can be used as the initial model for the traditional iterative method.

[0087] The underground structure model inversion module obtains the initial underground structure model by inverting the target seismic wave gather through optimizing the neural network.

[0088] Specifically, an optimized neural network is used to invert the initial underground structure model corresponding to the target seismic wave gather. Here, the initial model of the seismic wave gather is no longer manually generated. Instead, the trained neural network is used to output the seismic wave gather, reducing manual intervention.

[0089] The final neural network construction module performs full waveform inversion on the initial underground structure model, compares the error between the inversion result and the known full waveform inversion result, adjusts and optimizes the neural network, and obtains the final neural network. In one example, the full waveform inversion is performed on the initial underground structure model, the error between the inversion result and the known full waveform inversion result is compared, and the neural network is adjusted and optimized to obtain the final neural network, including: performing full waveform inversion on the initial underground structure model to obtain the initial inversion result; calculating the error between the initial inversion result and the known full waveform inversion result, updating the neural network parameters, and adjusting and optimizing the neural network. If the error is less than a set threshold, the final neural network is obtained.

[0090] In one example, the method further includes: replacing the target seismic wave gather, using the final neural network as the optimized neural network, and repeatedly performing learning and optimization on the neural network to improve the inversion accuracy.

[0091] Specifically, full waveform inversion is performed on the initial underground structure model to obtain the initial inversion result. The error between the initial inversion result and the known full waveform inversion result is calculated, and the neural network parameters are updated using formula (1). The neural network is adjusted and optimized. If the error is less than the set threshold, the final neural network is obtained. Compared with the optimized neural network, the final neural network learns more detailed information about the underground structure from the traditional iterative inversion method, resulting in a more accurate inversion result.

[0092] It is also possible to replace the target seismic wave gather with a new one, use the final neural network as the optimized neural network, and repeatedly perform learning and optimization on the neural network to improve the inversion accuracy.

[0093] Inversion module, performs inversion through the final neural network.

[0094] Specifically, the final neural network is used to invert the new seismic wave gathers.

[0095] The present invention also provides an electronic device, which includes: a memory storing executable instructions; and a processor that runs the executable instructions in the memory to implement the above-mentioned neural network inversion method.

[0096] The present invention also provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned neural network inversion method.

[0097] To facilitate understanding of the solutions and effects of the embodiments of the present invention, four specific application examples are given below. Those skilled in the art should understand that these examples are only for facilitating understanding of the present invention, and any specific details thereof are not intended to limit the present invention in any way.

[0098] Example 1

[0099] Figure 1 A flowchart showing the steps of a neural network inversion method according to an embodiment of the present invention is shown.

[0100] like Figure 1 As shown, the neural network inversion method includes: step 101, establishing a basic training set and constructing an initial neural network; step 102, updating the neural network parameters for the initial neural network to obtain an optimized neural network; step 103, inverting the target seismic wave channel set by optimizing the neural network to obtain an initial underground structure model; step 104, performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining a final neural network; step 105, performing inversion through the final neural network.

[0101] Using two-dimensional data as an example, the iterative method is full waveform inversion:

[0102] Figure 2 A schematic diagram of an actual seismic wave gather according to an embodiment of the present invention is shown.

[0103] Figure 3 A schematic diagram showing underground structure information obtained through manual participation according to an embodiment of the present invention is shown.

[0104] Use as Figure 2The actual seismic wave gathers shown are similar to those shown in Figure 3 The underground structure information obtained by manual participation shown in the figure constitutes a corresponding data body, and then multiple such data bodies together constitute the basic training set of the neural network; the neural network is built, the network parameters are randomly initialized, and the initial neural network is obtained.

[0105] Figure 4 A schematic diagram showing an optimized neural network inversion result according to an embodiment of the present invention is shown.

[0106] The initial neural network is trained using the basic training set. When the inversion effect of the neural network is good, the training is stopped to obtain the optimized neural network. The inversion results are as follows: Figure 4 shown.

[0107] Figure 5 A schematic diagram of a target seismic wave gather according to an embodiment of the present invention is shown.

[0108] Figure 6 A schematic diagram showing the target seismic wave gather in the optimized neural network inversion result according to an embodiment of the present invention is shown.

[0109] Figure 7 A schematic diagram showing full waveform inversion results according to an embodiment of the present invention is shown.

[0110] Using optimized neural networks Figure 5 The target seismic wave gathers shown in the figure are inverted to obtain the initial underground structure model, as shown in Figure 6 As shown, the initial underground structure model is used for full waveform inversion. The results of full waveform inversion are shown in Figure 7 shown.

[0111] Figure 8 A schematic diagram showing the final neural network inversion result according to one embodiment of the present invention is shown.

[0112] The neural network is trained and optimized so that its output is similar to the result of full waveform inversion. After training, the final inversion result of the neural network for seismic wave gathers is as follows: Figure 8 As shown, it contains more detailed information and presents the underground structure more perfectly, indicating that the neural network obtained here can be used for rough detailed inversion of underground structures.

[0113] Figure 9 A schematic diagram of a new seismic wave gather according to an embodiment of the present invention is shown.

[0114] Figure 10 A schematic diagram showing the final neural network inversion result of a new seismic wave gather according to an embodiment of the present invention is shown.

[0115] For the final neural network model, we can directly invert the following Figure 9 The underground structure corresponding to the new seismic wave gathers shown in Figure 10 The final neural network can be used for preliminary judgment in production and can also be used as an initial model for full waveform inversion to produce more accurate underground structure.

[0116] Example 2

[0117] Figure 11 A block diagram of a neural network inversion device according to an embodiment of the present invention is shown.

[0118] like Figure 11 As shown, the neural network inversion device includes:

[0119] Initial neural network construction module 201, establishes a basic training set and constructs an initial neural network;

[0120] The optimized neural network construction module 202 updates the neural network parameters for the initial neural network to obtain an optimized neural network;

[0121] The underground structure model inversion module 203 is configured to obtain an initial underground structure model by inverting the target seismic wave gathers through optimizing the neural network;

[0122] The final neural network construction module 204 performs full waveform inversion on the initial underground structure model, compares the error between the inversion result and the known full waveform inversion result, adjusts and optimizes the neural network, and obtains the final neural network;

[0123] The inversion module 205 performs inversion through the final neural network.

[0124] As an optional solution, the neural network parameters are updated for the initial neural network to obtain an optimized neural network including:

[0125] Input the data of the basic training set into the initial neural network to obtain output data;

[0126] The error between the output data and the underground structure model is calculated, and the neural network parameters are updated to obtain the optimized neural network.

[0127] As an optional solution, full waveform inversion is performed on the initial underground structure model, the error between the inversion result and the known full waveform inversion result is compared, and the neural network is adjusted and optimized to obtain the final neural network including:

[0128] Perform full waveform inversion on the initial underground structure model to obtain initial inversion results;

[0129] The error between the initial inversion result and the known full waveform inversion result is calculated, the neural network parameters are updated, and the neural network is adjusted and optimized. If the error is less than the set threshold, the final neural network is obtained.

[0130] As an alternative, the neural network parameters are updated using formula (1):

[0131] loss=λ1|OM|1+λ2|OM|2 (1)

[0132] Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, and λ1 and λ2 are weight coefficients.

[0133] As an option, it also includes:

[0134] The target seismic wave gather is replaced, and the final neural network is used as the optimized neural network. The neural network is repeatedly learned and optimized to improve the inversion accuracy.

[0135] Example 3

[0136] The present disclosure provides an electronic device comprising: a memory storing executable instructions; and a processor executing the executable instructions in the memory to implement the above-mentioned neural network inversion method.

[0137] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0138] The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, flash memory, etc.

[0139] The processor may be a central processing unit (CPU) or other form of processing unit having data processing capability and / or instruction execution capability, and may control other components in the electronic device to perform desired functions. In one embodiment of the present disclosure, the processor is used to execute the computer-readable instructions stored in the memory.

[0140] Those skilled in the art should understand that in order to solve the technical problem of how to obtain a good user experience, this embodiment may also include well-known structures such as a communication bus and an interface, and these well-known structures should also be included in the scope of protection of this disclosure.

[0141] For detailed description of this embodiment, please refer to the corresponding description in the aforementioned embodiments, which will not be repeated here.

[0142] Example 4

[0143] An embodiment of the present disclosure provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the neural network inversion method is implemented.

[0144] According to an embodiment of the present disclosure, a computer-readable storage medium stores non-transitory computer-readable instructions, which, when executed by a processor, execute all or part of the steps of the aforementioned methods of the embodiments of the present disclosure.

[0145] The above-mentioned computer-readable storage media include, but are not limited to, optical storage media (e.g., CD-ROMs and DVDs), magneto-optical storage media (e.g., MOs), magnetic storage media (e.g., magnetic tapes or mobile hard disks), media with built-in rewritable non-volatile memory (e.g., memory cards), and media with built-in ROM (e.g., ROM cartridges).

[0146] Those skilled in the art should understand that the above description of the embodiments of the present invention is only for the purpose of illustrative purposes only to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any given examples.

[0147] While various embodiments of the present invention have been described above, the above description is intended to be illustrative, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A neural network inversion method, characterized in that: include: Establish a basic training set and build an initial neural network; Updating the neural network parameters for the initial neural network to obtain an optimized neural network; For the target seismic wave gather, an initial underground structure model is obtained by inversion using the optimized neural network; Performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining a final neural network; performing inversion through the final neural network; Among them, the neural network parameters are updated by formula (1): loss=λ1|OM|1+λ2|OM|2 (1) Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, λ1 and λ2 are weight coefficients; Among them, also include: The target seismic wave gather is replaced, the final neural network is used as the optimized neural network, and the neural network is repeatedly learned and optimized to improve the inversion accuracy.

2. The neural network inversion method according to claim 1, wherein: Updating the neural network parameters for the initial neural network to obtain an optimized neural network includes: Inputting the data of the basic training set into the initial neural network to obtain output data; The error between the output data and the underground structure model is calculated, and the neural network parameters are updated to obtain an optimized neural network.

3. The neural network inversion method according to claim 1, wherein: Performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining the final neural network includes: Performing full waveform inversion on the initial underground structure model to obtain an initial inversion result; The error between the initial inversion result and the known full waveform inversion result is calculated, the neural network parameters are updated, and the optimized neural network is adjusted. If the error is less than a set threshold, the final neural network is obtained.

4. A neural network inversion device, characterized in that: include: Initial neural network building module, establish basic training set, and build initial neural network; An optimized neural network construction module is used to update the neural network parameters of the initial neural network to obtain an optimized neural network; An underground structure model inversion module is configured to obtain an initial underground structure model by inverting the target seismic wave gathers through the optimized neural network; A final neural network construction module performs full waveform inversion on the initial underground structure model, compares the error between the inversion result and the known full waveform inversion result, and adjusts the optimized neural network to obtain a final neural network; an inversion module, performing inversion through the final neural network; Among them, the neural network parameters are updated by formula (1): loss=λ1|OM|1+λ2|OM|2 (1) Among them, loss is the difference of neural network output, O is the neural network output, M is the underground structure model, λ1 and λ2 are weight coefficients; Among them, also include: The target seismic wave gather is replaced, the final neural network is used as the optimized neural network, and the neural network is repeatedly learned and optimized to improve the inversion accuracy.

5. The neural network inversion device according to claim 4, wherein: Updating the neural network parameters for the initial neural network to obtain an optimized neural network includes: Inputting the data of the basic training set into the initial neural network to obtain output data; The error between the output data and the underground structure model is calculated, and the neural network parameters are updated to obtain an optimized neural network.

6. The neural network inversion device according to claim 4, wherein: Performing full waveform inversion on the initial underground structure model, comparing the error between the inversion result and the known full waveform inversion result, adjusting the optimized neural network, and obtaining the final neural network includes: Performing full waveform inversion on the initial underground structure model to obtain an initial inversion result; The error between the initial inversion result and the known full waveform inversion result is calculated, the neural network parameters are updated, and the optimized neural network is adjusted. If the error is less than a set threshold, the final neural network is obtained.

7. An electronic device, characterized in that: The electronic device comprises: a memory storing executable instructions; A processor that runs the executable instructions in the memory to implement the neural network inversion method according to any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the neural network inversion method according to any one of claims 1 to 3.

Images