Inversion method, device, equipment and medium

Abstract

The application provides an inversion method, device, equipment and medium, constructs a reversible neural network model and a coupling objective function, and randomly generates multiple groups of Gassmann fluid item elastic parameters based on the elastic parameter characteristics of a to-be-observed area; the multiple groups of Gassmann fluid item elastic parameters are input into a forward mapping process of the reversible neural network, and the reversible neural network is iteratively trained based on the coupling objective function until the coupling objective function converges, so that a trained reversible neural network model is obtained; actual seismic data of the to-be-observed area is input into a reverse mapping process of the trained reversible neural network model, and an elastic parameter inversion result of the to-be-observed area is obtained. The method provided in the application can improve the inversion result accuracy.

Description

Technical Field

[0001] This application relates to the field of geophysics, and in particular to an inversion method, apparatus, device and medium. Background Technology

[0002] The Gassmann fluid term is an important elastic parameter in reservoir fluid identification. Inversion is a common method for obtaining the Gassmann fluid term in oil and gas exploration, and it can be divided into indirect and direct methods. Indirect inversion first obtains velocity and density based on seismic data, and then calculates the Gassmann fluid term using empirical formulas. Direct inversion uses the forward modeling equation that characterizes the reflection coefficient using the Gassmann fluid term to directly obtain the Gassmann fluid term from seismic data. Direct inversion can avoid the cumulative errors that exist in the indirect inversion calculation process to some extent.

[0003] The current direct inversion method using amplitude variation with offset (AVO) technology is too dependent on the initial low-frequency parameters and training samples in the seismic data. This can lead to low accuracy of the inversion results when the initial low-frequency parameters or training samples are limited. Summary of the Invention

[0004] This application provides an inversion method, apparatus, equipment, and medium that can improve the accuracy of inversion results.

[0005] In a first aspect, this application provides an inversion method, the method comprising:

[0006] A reversible neural network model and a coupled objective function are constructed, and multiple sets of Gassmann fluid term elastic parameters are randomly generated based on the elastic parameter characteristics of the region to be observed.

[0007] By inputting the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, a trained reversible neural network model is obtained; wherein, the forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the reverse mapping of the reversible neural network model reflects the mapping of seismic data to elastic parameters;

[0008] The actual seismic data of the area to be observed is obtained, and the actual seismic data of the area to be observed is input into the inverse mapping process of the trained reversible neural network model to obtain the elastic parameter inversion result of the area to be observed.

[0009] In one example, constructing the reversible neural network model and coupling the objective function includes:

[0010] A reversible neural network model is constructed, wherein the loss function of the reversible neural network model includes a supervised loss function defined by the error between seismic data, a first unsupervised loss function defined by the error between the latent variable distribution and the Gaussian distribution, and a second unsupervised loss function defined by the error between the elastic parameter distributions;

[0011] The supervised loss function is defined using a zero-delay cross-correlation function, and the first unsupervised loss function and the second unsupervised loss function are defined using a maximum mean difference function. Based on the set weight coefficients of each loss function, the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function are coupled to obtain the coupled objective function.

[0012] In one example, the step of inputting the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, to obtain the trained reversible neural network model, includes:

[0013] The elastic parameters are input into the forward modeling equation, which directly characterizes the reflection coefficient, to obtain the synthetic seismic data corresponding to the elastic parameters.

[0014] The elastic parameters are input into the forward mapping process of the reversible neural network model to obtain the predicted earthquake data and latent variables corresponding to the elastic parameters output by the reversible neural network model under the forward mapping process; and the predicted earthquake data and latent variables corresponding to the elastic parameters are input into the reverse mapping process of the reversible neural network to obtain the predicted elastic parameters corresponding to the elastic parameters output by the reversible neural network model under the reverse mapping process.

[0015] Based on the errors between the synthetic and predicted seismic data corresponding to the elastic parameters, the errors between the latent variable distribution and the Gaussian distribution corresponding to the elastic parameters, and the errors between the data distribution of the elastic parameters and the corresponding predicted elastic parameters, the reversible neural network model is iteratively updated using the backpropagation algorithm based on the coupled objective function until the coupled objective function converges, thus obtaining the trained reversible neural network model.

[0016] In one example, the step of acquiring actual seismic data of the area to be observed, inputting the actual seismic data of the area to be observed into the inverse mapping process of the trained reversible neural network model, and obtaining the elastic parameter inversion result of the area to be observed includes:

[0017] Obtain the actual seismic data and final latent variables of the area to be observed. The final latent variables are the latent variables most recently output by the invertible neural network model in the forward mapping process when the coupled objective function converges.

[0018] The actual seismic data of the area to be observed and the final latent variables are input into the inverse mapping process of the trained reversible neural network model to obtain the elastic parameter inversion results of the area to be observed.

[0019] In one example, the random generation of multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the observed region includes:

[0020] Based on empirical formulas, multiple P-wave velocities are randomly generated, and the corresponding S-wave velocity and density for each P-wave velocity are calculated. Multiple S-wave velocities and multiple densities are obtained by adding random numbers and / or replacing low frequencies.

[0021] The multiple sets of Gassmann fluid term elastic parameters are calculated based on the multiple longitudinal wave velocities, the multiple transverse wave velocities, and the multiple densities.

[0022] Determine whether the multiple sets of Gassmann fluid term elastic parameters meet the set random generation rules, and remove elastic parameters that do not meet the random generation rules from the multiple sets of Gassmann fluid term elastic parameters. The random generation rules include at least one of the following: the value range of the randomly generated elastic parameter is not greater than the value range of the elastic parameter of the area to be observed; the thinnest layer thickness of the randomly generated elastic parameter is less than the thinnest layer thickness identified based on the seismic data of the area to be observed; and the vertical variation trend of the randomly generated elastic parameter conforms to the geological laws of the area to be observed.

[0023] In one example, the reversible neural network model includes a forward reversible block and a reverse reversible block of complementary affine coupling layers;

[0024] The forward invertible block is used to decompose the elastic parameters input during the forward mapping process into a first elastic parameter and a second elastic parameter; process the first elastic parameter based on a first activation function to obtain a first processing result; process the first elastic parameter based on a second activation function to obtain a second processing result; and output predicted earthquake data by using matrix inner product calculation and summation based on the second elastic parameter, the first processing result, and the second processing result; and process the predicted earthquake data based on a third activation function to obtain a third processing result; process the second elastic parameter based on a fourth activation function to obtain a fourth processing result; and output latent variables by using matrix inner product calculation and summation based on the first elastic parameter, the third processing result, and the fourth processing result.

[0025] The reverse reversible block is used in the reverse mapping process to process the input seismic data based on the third activation function to obtain a fifth processing result; to process the input seismic data based on the fourth activation function to obtain a sixth processing result; and to output a first predicted elasticity parameter by using matrix inner product calculation and difference calculation based on the fifth processing result, the sixth processing result, and the input latent variables; to obtain a seventh processing result by processing the first activation function and the first predicted elasticity parameter; to obtain an eighth processing result by processing the second activation function and the first predicted elasticity parameter; and to output a second predicted elasticity parameter by using matrix inner product calculation and difference calculation based on the seventh processing result, the eighth processing result, and the input seismic data; and to synthesize the first predicted elasticity parameter and the second predicted elasticity parameter to output a predicted elasticity parameter.

[0026] On the other hand, this application provides an inversion apparatus, the apparatus comprising:

[0027] The building module is used to construct a reversible neural network model and couple the objective function, and randomly generate multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the region to be observed;

[0028] The input module is used to input the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and to iteratively train the reversible neural network based on the coupling objective function until the coupling objective function converges, thereby obtaining a trained reversible neural network model; wherein, the forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the reverse mapping of the reversible neural network model reflects the mapping of seismic data to elastic parameters;

[0029] The acquisition module is used to acquire the actual seismic data of the area to be observed, input the actual seismic data of the area to be observed into the inverse mapping process of the trained reversible neural network model, and obtain the elastic parameter inversion result of the area to be observed.

[0030] In one example, the building module is specifically used to build a reversible neural network model, the loss function of which includes a supervised loss function defined by the error between seismic data, a first unsupervised loss function defined by the error between the latent variable distribution and the Gaussian distribution, and a second unsupervised loss function defined by the error between the elastic parameter distributions;

[0031] The construction module is further configured to define the supervised loss function using a zero-delay cross-correlation function, define the first unsupervised loss function and the second unsupervised loss function using a maximum mean difference function, and couple the supervised loss function, the first unsupervised loss function and the second unsupervised loss function according to the set weight coefficients of each loss function to obtain the coupled objective function.

[0032] In another aspect, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0033] The memory stores computer-executed instructions;

[0034] The processor executes computer execution instructions stored in the memory to implement the method described above.

[0035] In another aspect, this application provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, are used to implement the method as described in the preceding claim.

[0036] The inversion method, apparatus, equipment, and medium provided in this application first construct a reversible neural network model and a coupling objective function, and randomly generate multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the area to be observed. Then, by inputting the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, a trained reversible neural network model is obtained. The forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the backward mapping of the reversible neural network model reflects the mapping of seismic data to elastic parameters. Finally, the actual seismic data of the area to be observed is obtained, and the actual seismic data of the area to be observed is input into the backward mapping process of the trained reversible neural network model to obtain the elastic parameter inversion result of the area to be observed. This scheme, by constructing a bidirectional reversible neural network model and a coupling objective function, and training the reversible neural network model, obtains the elastic parameter inversion result of the area to be observed based on the trained reversible neural network model, reducing the dependence on initial low-frequency parameters and training samples in the seismic data and improving the accuracy of the inversion result. Attached Figure Description

[0037] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0038] Figure 1This is a schematic diagram illustrating an application scenario for this application;

[0039] Figure 2 A flowchart illustrating an inversion method provided in Embodiment 1 of this application;

[0040] Figure 3 A flowchart illustrating another inversion method provided in Embodiment 1 of this application;

[0041] Figure 4 A flowchart illustrating another inversion method provided in Embodiment 1 of this application;

[0042] Figure 5 This is a schematic diagram of the training process of a reversible neural network.

[0043] Figure 6 A flowchart illustrating another inversion method provided in Embodiment 1 of this application;

[0044] Figure 7 A flowchart illustrating another inversion method provided in Embodiment 1 of this application;

[0045] Figure 8 This is a schematic diagram of the computation process of a reversible neural network.

[0046] Figure 9 An inversion device provided in Embodiment 2 of this application;

[0047] Figure 10 A schematic diagram of the elastic parameters, shear modulus, and density of the Gassmann fluid term in the Marmousi model;

[0048] Figure 11 Example of stacked seismic data;

[0049] Figure 12 The example shows the Gassmann fluid term elastic parameter inversion results corresponding to real seismic data;

[0050] Figure 13 This is a schematic diagram illustrating the extraction of data from wells A and B as an example.

[0051] Figure 14 Another example of post-stack seismic data;

[0052] Figure 15 The example shows the inversion results of the elastic parameters of the Gassmann fluid term corresponding to another real seismic data.

[0053] Figure 16 This is a schematic diagram of another type of extracted data from wells A and B;

[0054] Figure 17This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of this application.

[0055] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0056] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0057] The specific application scenario for this application is in the field of geophysics. Figure 1 This is a schematic diagram of an application scenario for this application. a represents a conventional neural network, b represents a reversible neural network, and the Gassmann fluid term is an important elastic parameter in reservoir fluid identification. Inversion is a common method in the field of oil and gas exploration to obtain the elastic parameter of the Gassmann fluid term. Generally, the elastic parameter of the Gassmann fluid term is obtained by inverting the seismic data of the area to be explored.

[0058] Currently, methods for obtaining the Gassmann fluid term elastic parameters from seismic data based on conventional neural networks are generally divided into indirect and direct methods. Direct inversion can avoid the cumulative errors present in indirect inversion calculations to some extent. However, AVO direct inversion methods can be divided into model-driven inversion and data-driven inversion. Model-driven inversion is dependent on the initial low-frequency parameters in the seismic data, while data-driven inversion requires a large number of representative training samples. Therefore, AVO direct inversion methods are highly dependent on the initial low-frequency parameters and training samples. Without a large number of initial low-frequency parameters or training samples, the accuracy of obtaining the Gassmann fluid term elastic parameters from AVO direct inversion will be low.

[0059] However, the method provided in this application can be based on constructing a reversible neural network model and coupling an objective function. The reversible neural network is a bidirectional mapping, with the forward mapping being the mapping of the Gassmann fluid term elastic parameters to seismic data and latent variables, and the reverse mapping being the mapping of seismic data and latent variables to the Gassmann fluid term elastic parameters. The reversible neural network model is trained, and based on the trained reversible neural network model, the seismic data of the area to be observed is input into the reverse mapping of the trained reversible neural network model to obtain the elastic parameter inversion results of the area to be observed. This reduces the dependence on the initial low-frequency parameters and training samples in the seismic data and improves the accuracy of the inversion results.

[0060] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0061] The technical solutions of this application will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. In the description of this application, unless otherwise expressly specified and limited, the terms should be broadly understood within the art. The embodiments of this application will now be described with reference to the accompanying drawings.

[0062] Example 1

[0063] Figure 2 This is a flowchart illustrating an inversion method provided in Embodiment 1 of this application, as shown below. Figure 2 As shown, the method includes:

[0064] Step 201: Construct a reversible neural network model and couple the objective function, and randomly generate multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the region to be observed;

[0065] Step 202: By inputting the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, a trained reversible neural network model is obtained; wherein, the forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the reverse mapping of the reversible neural network model reflects the mapping of seismic data to elastic parameters;

[0066] Step 203: Obtain the actual seismic data of the area to be observed, input the actual seismic data of the area to be observed into the reverse mapping process of the trained reversible neural network model, and obtain the elastic parameter inversion result of the area to be observed.

[0067] The execution subject of this embodiment is a seepage law characterization device, which can be implemented by a computer program, such as application software; or it can be implemented as a medium storing relevant computer programs, such as a USB flash drive or cloud drive; or it can be implemented by a physical device that integrates or installs relevant computer programs, such as a chip.

[0068] With a scenario example, the reversible neural network model constructed in this application is used to fit the bidirectional mapping relationship between the Gassmann fluid elastic parameter and seismic data. The forward mapping is based on the mapping from the Gassmann fluid elastic parameter to the seismic data, and the reverse mapping is based on the mapping from the seismic data to the Gassmann fluid elastic parameter. Since conventional neural networks with unidirectional mapping typically learn the underdetermined inversion process from seismic data to the Gassmann fluid elastic parameter, their training efficiency is low and their stability is not high. Therefore, using the reversible neural network model based on the bidirectional mapping can improve the efficiency and accuracy of the inversion. A representative region is selected as the observation region in this application. The selection of the observation region can be a region where an earthquake has occurred. Based on the actual earthquake situation in the observation region, the characteristics of the Gassmann fluid elastic parameter of the observation region are determined. Based on these characteristics, multiple sets of Gassmann fluid elastic parameters that conform to the characteristics are randomly generated. Common Gassmann fluid elastic parameters include: Lamé constant (λ, μ), Young's modulus (E), Poisson's ratio (ν), bulk modulus (K), shear modulus (G), density (ρ), and wave velocity (Vp, Vs). For a homogeneous, isotropic, perfectly elastic medium, for the three pairs of elastic moduli λ-μ, E-ν, and KG, knowing only one pair is sufficient to determine the other two pairs. Combined with the formation density, the P-wave and S-wave velocities Vp and Vs of the entire medium or formation can be determined.

[0069] The multiple sets of Gassmann fluid elastic parameters generated above are input as training data into the forward mapping process of the reversible neural network to complete the training of the reversible neural network. Specifically, a coupling objective function is set in the reversible neural network. This coupling objective function characterizes the error between the training data and the training results based on the training data at various stages of the training process. When the error gradually decreases, it indicates that the coupling objective function is gradually converging. When the coupling objective function completes convergence, it indicates that the training of the reversible neural network is complete. After the training of the reversible neural network is completed, the actual seismic data when an earthquake occurs in the area to be observed is obtained. This can be done by recording the seismic data when an earthquake occurs in the area to be observed and querying the recorded seismic data. Because the inverse mapping of the reversible neural network is based on the mapping of seismic data to the elastic parameters of the Gassmann fluid term, the seismic data of the area to be observed can be input into the inverse mapping process of the trained reversible neural network. Based on the inverse mapping of the reversible neural network, the inversion result of the area to be observed is obtained, and the inversion result is the elastic parameters of the Gassmann fluid term of the area to be observed.

[0070] This example first constructs a reversible neural network model and a coupled objective function, and randomly generates multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the region to be observed. Then, by inputting these multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupled objective function until the coupled objective function converges, a trained reversible neural network model is obtained. The forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the backward mapping reflects the mapping of seismic data to elastic parameters. Finally, the actual seismic data of the region to be observed is obtained, and the actual seismic data of the region to be observed is input into the backward mapping process of the trained reversible neural network model to obtain the elastic parameter inversion results of the region to be observed. This example, by constructing a bidirectional reversible neural network model and a coupled objective function, and training the reversible neural network model, obtains the elastic parameter inversion results of the region to be observed based on the trained reversible neural network model, reducing the dependence on initial low-frequency parameters and training samples in the seismic data and improving the accuracy of the inversion results.

[0071] Optional, Figure 3 A flowchart illustrating another inversion method provided in Embodiment 1 of this application is shown below. Figure 3 As shown, in step 201, constructing the reversible neural network model and coupling the objective function includes:

[0072] Step 301: Construct a reversible neural network model. The loss function of the reversible neural network model includes a supervised loss function defined by the error between seismic data, a first unsupervised loss function defined by the error between the latent variable distribution and the Gaussian distribution, and a second unsupervised loss function defined by the error between the elastic parameter distributions.

[0073] Step 302: Define the supervised loss function using the zero-delay cross-correlation function, define the first unsupervised loss function and the second unsupervised loss function using the maximum mean difference function, and couple the supervised loss function, the first unsupervised loss function and the second unsupervised loss function according to the set weight coefficients of each loss function to obtain the coupled objective function.

[0074] In the context of a scenario example, when constructing the reversible neural network, a loss function can be built into the network since it requires training. Because seismic data is crucial for obtaining the inversion results, and because there is a correlation between the actual seismic data and the model-predicted seismic data in the observed area, the error between the seismic data can be defined as a supervised loss function based on the zero-latency cross-correlation function. During the training process of the reversible neural network, it learns and memorizes the seismic data. The loss function in supervised training is often used to reflect the degree of inconsistency between the actual values ​​of the samples and the model's predicted values. Besides defining the error between the seismic data as a supervised loss function, based on the principle of the maximum mean difference function and the characteristics of unsupervised learning, the error between key parameters during the training process of the reversible neural network can be defined as an unsupervised loss function. For example, the error between the latent variable distribution and the Gaussian distribution of the reversible neural network can be used as the first unsupervised loss function, and the error between the elastic parameter distributions can be used as the second unsupervised loss function. The latent variable distribution of the reversible neural network consists of parameters inherent to the reversible neural network. The role of the latent variables is to store important information related to the elastic parameters lost during the forward process. The Gaussian distribution can be any Gaussian distribution. The weights corresponding to the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function are determined based on the quality of the seismic data and the Gaussmann fluid term elastic parameter data. The objective function can be obtained by coupling the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function with their respective weights. For example, if the weights corresponding to the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function are a1, a2, and a3, respectively, and the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function are SL... R USL z With USL mThe objective function can be defined as:

[0075] L=a1·SL R +a2·USL z +a3·USL m

[0076] Where L is the objective function. This example improves the efficiency of training the reversible neural network and ensures its accuracy by setting the error between key parameters involved in training the reversible neural network as a loss function and coupling the loss function to the objective function. The convergence of the objective function determines whether the reversible neural network has been trained successfully.

[0077] Optional, Figure 4 A flowchart illustrating another inversion method provided in Embodiment 1 of this application is shown below. Figure 4 As shown, in step 202, the reversible neural network model is obtained by inputting the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, including:

[0078] Step 401: Input the elastic parameters into the forward modeling equation that directly characterizes the reflection coefficient by the elastic parameters to obtain the synthetic seismic data corresponding to the elastic parameters;

[0079] Step 402: Input the elastic parameters into the forward mapping process of the reversible neural network model to obtain the predicted earthquake data and latent variables corresponding to the elastic parameters output by the reversible neural network model under the forward mapping process; and input the predicted earthquake data and latent variables corresponding to the elastic parameters into the reverse mapping process of the reversible neural network to obtain the predicted elastic parameters corresponding to the elastic parameters output by the reversible neural network model under the reverse mapping process.

[0080] Step 403: Based on the error between the synthetic seismic data and the predicted seismic data corresponding to the elastic parameters, the error between the latent variable distribution and the Gaussian distribution corresponding to the elastic parameters, and the error between the data distribution of the elastic parameters and the corresponding predicted elastic parameters, the reversible neural network model is iteratively updated using the backpropagation algorithm based on the coupling objective function until the coupling objective function converges, thereby obtaining the trained reversible neural network model.

[0081] Combined with scenario examples, Figure 5This diagram illustrates the training process of a reversible neural network. During training, the first set of Gassmann fluid elastic parameters from the multiple sets of Gassmann fluid elastic parameters is first input into the forward modeling equation. This equation synthesizes seismic data from the input elastic parameters. Therefore, inputting the first set of Gassmann fluid elastic parameters into the forward modeling equation yields the first synthesized seismic data corresponding to these parameters. The first set of Gassmann fluid elastic parameters is then input into the forward mapping process of the reversible neural network, causing it to output the first predicted seismic parameter and the first latent variable corresponding to these parameters. The first predicted seismic parameter and the first latent variable are then input into the reversible neural network, causing it to output the first predicted elastic parameter. This process of obtaining the first predicted elastic parameter based on the first set of Gassmann fluid elastic parameters constitutes the first iterative training of the reversible neural network.

[0082] The error between the first synthetic seismic data and the first predicted seismic data is taken as the first error, the error between the first latent variable and the Gaussian distribution is taken as the second error, the Gaussian distribution can be any Gaussian distribution, and the error between the first elastic parameter and the first predicted elastic parameter is taken as the third error. Based on the first error, the second error and the third error, the value of the coupling objective function is obtained in the first iteration. In the first iteration training, the value of the coupling objective function is obtained by summing the first error, the second error and the third error.

[0083] After obtaining the value of the coupling objective function in the first iteration, the reversible neural network is updated based on the backpropagation algorithm. After updating the reversible neural network, the second set of Gassmann fluid elastic parameters are input into the forward mapping between the forward equation and the reversible neural network. The reversible neural network is then trained in a second iteration based on the second set of Gassmann fluid elastic parameters to obtain the value of the coupling objective function under the second iteration of training of the second set of Gassmann fluid elastic parameters. The reversible neural network is then updated again using the backpropagation algorithm.

[0084] Following the method described above, based on the generated multiple sets of Gassmann fluid elastic parameters, the reversible neural network is iteratively trained sequentially to obtain the value of the coupling objective function corresponding to each iteration. After each iteration, the reversible neural network is updated using the backpropagation algorithm until the coupling objective function converges, at which point the training of the reversible neural network stops. It is worth noting that if the coupling objective function does not converge after training the reversible neural network using the multiple sets of Gassmann fluid elastic parameters, the iterative training of the reversible neural network is restarted, starting from the first set of Gassmann fluid elastic parameters, until the multiple sets of Gassmann fluid elastic parameters converge.

[0085] This example trains the reversible neural network sequentially using the elastic parameters of the Gassmann fluid terms, and monitors the training completion status of the reversible neural network by the convergence of the coupled objective function, which can make the obtained reversible neural network more accurate.

[0086] Optional, Figure 6 A flowchart illustrating another inversion method provided in Embodiment 1 of this application is shown below. Figure 6 As shown, step 203 includes:

[0087] Step 601: Obtain the actual seismic data and final latent variables of the area to be observed. The final latent variables are the latent variables most recently output by the invertible neural network model in the forward mapping process when the coupled objective function converges.

[0088] Step 602: Input the actual seismic data of the area to be observed and the final latent variables into the inverse mapping process of the trained reversible neural network model to obtain the elastic parameter inversion results of the area to be observed.

[0089] In a scenario example, if the reversible neural network is trained n times before the coupled objective function converges, the latent variable obtained during the nth training iteration is taken as the final latent variable. Based on seismic survey techniques, actual seismic data generated when an earthquake occurs in the observed area is obtained. This actual seismic data, along with the final latent variable, is used as input data in the inverse mapping process of the reversible neural network. This allows the reversible neural network to output the elastic parameter inversion result, thus obtaining the elastic parameter distribution corresponding to the observed area.

[0090] This example is based on using the final latent variable as part of the input data in the inverse mapping process of the reversible neural network, which allows the reversible neural network to obtain more accurate inversion results during the inversion process.

[0091] Optional, Figure 7 A flowchart illustrating another inversion method provided in Embodiment 1 of this application is shown below. Figure 7 As shown, in step 201, multiple sets of Gassmann fluid term elastic parameters are randomly generated based on the elastic parameter characteristics of the region to be observed, including:

[0092] Step 701: Based on empirical formulas, randomly generate multiple P-wave velocities, calculate the corresponding S-wave velocity and density for each P-wave velocity, and obtain multiple S-wave velocities and multiple densities by adding random numbers and / or replacing low frequencies.

[0093] Step 702: Calculate the multiple sets of Gassmann fluid term elastic parameters based on the multiple longitudinal wave velocities, the multiple transverse wave velocities, and the multiple densities;

[0094] Step 703: Determine whether the multiple sets of Gassmann fluid term elastic parameters meet the set random generation rules, and remove elastic parameters that do not meet the random generation rules from the multiple sets of Gassmann fluid term elastic parameters. The random generation rules include at least one of the following: the value range of the randomly generated elastic parameter is not greater than the value range of the elastic parameter of the area to be observed; the thinnest layer thickness of the randomly generated elastic parameter is less than the thinnest layer thickness identified based on the seismic data of the area to be observed; and the vertical variation trend of the randomly generated elastic parameter conforms to the geological laws of the area to be observed.

[0095] In the process of generating multiple sets of Gassmann fluid term elastic parameters, the generation rules for these elastic parameters must be met. The first rule is that the range of values ​​for randomly generated Gassmann fluid term elastic parameter data must be less than or equal to the range of values ​​for Gassmann fluid term elastic parameters in the area to be observed. This means that the upper and lower limits of the Gassmann fluid term elastic parameters in the area to be observed must be determined first, and the randomly generated Gassmann fluid term elastic parameters must fall between these limits. The second rule is that the thinnest layer thickness that can be identified in the randomly generated Gassmann fluid term elastic parameter data must be less than the thinnest layer thickness that can be identified from the seismic data in the area to be observed. This means that all randomly generated Gassmann fluid term elastic parameter data must be identified before the subsequent training of the reversible neural network can be completed. The third rule is that the vertical variation trend of the randomly generated Gassmann fluid term elastic parameter data must conform to the geological patterns of the area to be observed. Based on the three elastic parameter generation rules mentioned above, multiple P-wave velocities can first be generated using a first empirical formula. To ensure the diversity of the generated Gassmann fluid term elastic parameters, different second empirical formulas can be applied to each of the multiple P-wave velocities to generate the corresponding S-wave velocity. Furthermore, different third empirical formulas can be applied to each P-wave velocity to generate the corresponding density. The Gassmann fluid term elastic parameters are divided into high-frequency and low-frequency components. Besides obtaining multiple P-wave velocities, S-wave velocities, and densities through empirical formulas, multiple sets of random numbers can be added to represent the remaining P-wave velocities, S-wave velocities, and densities. Additionally, some low-frequency values ​​in the parameters can be replaced. This method yields a diverse set of Gassmann fluid term elastic parameters.

[0096] After obtaining multiple sets of Gassmann fluid term elastic parameters, each set of Gassmann fluid term elastic parameters is compared with the generation rule of elastic parameters. If, during the calculation process using the second empirical formula and the third empirical formula, or during the addition of multiple sets of random numbers or the replacement of low-frequency values, the result of one set of Gassmann fluid term elastic parameters does not conform to the generation rule of elastic parameters, then that set of Gassmann fluid term elastic parameters that does not conform to the generation rule of elastic parameters is deleted. Finally, each set of Gassmann fluid term elastic parameters obtained conforms to the generation rule of elastic parameters. Finally, based on the longitudinal wave velocity, transverse wave velocity, and density in each set of Gassmann fluid term elastic parameters, the Gassmann fluid term and shear modulus corresponding to each set of Gassmann fluid term elastic parameters are calculated using the first calculation formula. The Gassmann fluid term and shear modulus are the elastic parameters in each set of Gassmann fluid term elastic parameters that are ultimately used to train the reversible neural network.

[0097] This example restricts the generation of multiple sets of Gassmann fluid term elastic parameters by specifying the generation rules for elastic parameters, resulting in higher quality and greater utilization value of the generated multiple sets of Gassmann fluid term elastic parameters. Furthermore, by generating multiple sets of Gassmann fluid term elastic parameters through various empirical formulas, adding multiple sets of random numbers, or replacing low-frequency values, the generated multiple sets of Gassmann fluid term elastic parameters can be made more diverse.

[0098] Optionally, the reversible neural network model includes a forward reversible block and a reverse reversible block of complementary affine coupling layers;

[0099] The forward invertible block is used to decompose the elastic parameters input during the forward mapping process into a first elastic parameter and a second elastic parameter; process the first elastic parameter based on a first activation function to obtain a first processing result; process the first elastic parameter based on a second activation function to obtain a second processing result; and output predicted earthquake data by using matrix inner product calculation and summation based on the second elastic parameter, the first processing result, and the second processing result; and process the predicted earthquake data based on a third activation function to obtain a third processing result; process the second elastic parameter based on a fourth activation function to obtain a fourth processing result; and output latent variables by using matrix inner product calculation and summation based on the first elastic parameter, the third processing result, and the fourth processing result.

[0100] The reverse reversible block is used in the reverse mapping process to process the input seismic data based on the third activation function to obtain a fifth processing result; to process the input seismic data based on the fourth activation function to obtain a sixth processing result; and to output a first predicted elasticity parameter by using matrix inner product calculation and difference calculation based on the fifth processing result, the sixth processing result, and the input latent variables; to obtain a seventh processing result by processing the first activation function and the first predicted elasticity parameter; to obtain an eighth processing result by processing the second activation function and the first predicted elasticity parameter; and to output a second predicted elasticity parameter by using matrix inner product calculation and difference calculation based on the seventh processing result, the eighth processing result, and the input seismic data; and to synthesize the first predicted elasticity parameter and the second predicted elasticity parameter to output a predicted elasticity parameter.

[0101] Based on a scenario example, since the reversible neural network can perform three functions—forward mapping, backward mapping, and calculation of the coupled objective function—it can be divided into three modules: a forward reversible block, a backward reversible block, and a complementary affine coupling layer. The forward reversible block implements the forward mapping process, the backward reversible block implements the backward mapping process, and the complementary affine coupling layer calculates the coupled objective function. Because the Gassmann fluid term elastic parameters have high-frequency and low-frequency variations, the Gassmann fluid term and shear modulus in each group of Gassmann fluid term elastic parameters are respectively divided into high-frequency Gassmann fluid term, low-frequency Gassmann fluid term, high-frequency shear modulus, and low-frequency shear modulus. The high-frequency Gassmann fluid term and the high-frequency shear modulus are used as the first elastic parameter, and the low-frequency Gassmann fluid term and the low-frequency shear modulus are used as the second elastic parameter. The first and second elastic parameters are then simultaneously input into the forward reversible block. Figure 8 This is a schematic diagram of the computation process of a reversible neural network, as shown below. Figure 8 As shown, the predicted earthquake data and latent variables for the forward mapping output are obtained based on the first and second formulas below.

[0102] First formula: y1=x1⊙exp[s2(x2)]+t2(x2)

[0103] Second formula: y2=x2⊙exp[s1(x1)]+t1(y1)

[0104] Where: y1 is the predicted earthquake data, y2 is the latent variable, x1 is the second elastic parameter, and x2 is the first elastic parameter.

[0105] Specifically, according to the first formula, the process of obtaining the predicted earthquake data through the first elastic parameter and the second elastic parameter can be as follows: first, the first elastic parameter is processed by the first activation function s2 to obtain the first processing result s2(x2); then, the first elastic parameter is processed by the second activation function t2 to obtain the second processing result t2(x2); then, the second elastic parameter and the first processing result are matrix inner product calculation to obtain the processing result x1⊙exp[s2(x2)]; finally, the result is added to the second processing result to obtain the predicted earthquake data y1.

[0106] According to the second formula, the third processing result t1(y1) can be obtained by processing the predicted earthquake data y1 using the third activation function t1, and the fourth processing result s1(x1) can be obtained by processing the second elastic parameter based on the fourth activation function s1. Then, the first elastic parameter and the fourth processing result are matrix inner product calculation to obtain x2⊙exp[s1(x1)], and finally, the third processing result is added to obtain the latent variable y2.

[0107] After obtaining the predicted earthquake data and latent variables, the predicted earthquake data and latent variables are input into the inverse reversible block of the reversible neural network, and the first predicted elastic parameter and the second predicted elastic parameter are obtained based on the third and fourth formulas below, respectively.

[0108] Third formula: x2'=[y2-t1(y1)]⊙exp[-s1(y1)]

[0109] Fourth formula: x1'=[y1-t2(x2)]⊙exp[-s2(x2)]

[0110] Where: x1' is the second elastic parameter, and x2' is the first elastic parameter.

[0111] Specifically, the predicted earthquake data y1 is processed using the third activation function t1 to obtain the fifth processing result t1(y1). The predicted earthquake data y1 is processed using the fourth activation function s1 to obtain the sixth processing result s1(y1). The negative of the sixth processing result is obtained as -s1(y1). The difference between the latent variable y2 and the fifth processing result is obtained as y2-t1(y1). Based on -s1(y1) and y2-t1(y1), the first elastic parameter x2' is calculated using the matrix inner product.

[0112] The first elastic parameter is processed using a first activation function s2 to obtain a seventh processing result s2(x2). The first elastic parameter is then processed using a second activation function t2 to obtain an eighth processing result t2(x2). The seventh processing result is the same as the first processing result, and the eighth processing result is the same as the second processing result. The difference between the predicted earthquake data y1 and the eighth processing result is y2-t1(y1). The negative of the seventh processing result is calculated to obtain -s1(y1). The second elastic parameter x1' is calculated using the matrix inner product of y2-t1(y1) and -s1(y1). It is worth noting that the first, second, third, and fourth activation functions can be the same or different activation functions.

[0113] If the reverse mapping process is an iteration during the training of the reversible neural network, the output first elastic parameter x2' and second elastic parameter x1' are the elastic parameter inversion results output during the iteration. If the reverse mapping process is when the reversible neural network has been trained and is being used to invert and obtain elastic parameters, the predicted earthquake data y1 is replaced with the actual earthquake data of the area to be observed, and the latent variable y2 is replaced with the final latent variable. Then, the first elastic parameter x2' obtained by the reverse mapping represents the high-frequency Gassmann fluid term and high-frequency shear modulus of the elastic parameter inversion result for the area to be observed; the second elastic parameter x1' obtained by the reverse mapping represents the low-frequency Gassmann fluid term and low-frequency shear modulus of the elastic parameter inversion result for the area to be observed. This example uses multiple formulas applied to the forward and reverse mapping processes, enabling the reversible neural network to quickly and accurately obtain inversion results during training and use.

[0114] This embodiment first constructs a reversible neural network model and a coupling objective function, and randomly generates multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the region to be observed. Then, by inputting these multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, a trained reversible neural network model is obtained. The forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the backward mapping reflects the mapping of seismic data to elastic parameters. Finally, the actual seismic data of the region to be observed is obtained, and the actual seismic data of the region to be observed is input into the backward mapping process of the trained reversible neural network model to obtain the elastic parameter inversion results of the region to be observed. This embodiment, by constructing a bidirectional reversible neural network model and a coupling objective function, and training the reversible neural network model, obtains the elastic parameter inversion results of the region to be observed based on the trained reversible neural network model, reducing the dependence on initial low-frequency parameters and training samples in the seismic data and improving the accuracy of the inversion results.

[0115] Example 2

[0116] Figure 9 An inversion device provided in Embodiment 2 of this application, such as Figure 9 As shown, the device includes:

[0117] Module 91 is used to construct a reversible neural network model and couple the objective function, and randomly generate multiple sets of Gassmann fluid term elastic parameters based on the elastic parameter characteristics of the region to be observed.

[0118] Input module 92 is used to input the multiple sets of Gassmann fluid term elastic parameters into the forward mapping process of the reversible neural network, and to iteratively train the reversible neural network based on the coupling objective function until the coupling objective function converges, thereby obtaining a trained reversible neural network model; wherein, the forward mapping of the reversible neural network model reflects the mapping of elastic parameters to seismic data, and the reverse mapping of the reversible neural network model reflects the mapping of seismic data to elastic parameters;

[0119] The acquisition module 93 is used to acquire the actual seismic data of the area to be observed, input the actual seismic data of the area to be observed into the reverse mapping process of the trained reversible neural network model, and obtain the elastic parameter inversion result of the area to be observed.

[0120] In a scenario example, input module 92 inputs the multiple sets of Gassmann fluid term elastic parameters generated above as training data into the forward mapping process of the reversible neural network, completing the training of the reversible neural network. Specifically, construction module 91 sets a coupling objective function in the reversible neural network. The coupling objective function characterizes the error between the training data and the training results based on the training data at various stages of the training process. When the error gradually decreases, it indicates that the coupling objective function is gradually converging. When the coupling objective function completes convergence, it indicates that the training of the reversible neural network is complete. After the training of the reversible neural network is completed, acquisition module 93 acquires the actual seismic data when an earthquake occurs in the area to be observed. This can be done by recording the seismic data when an earthquake occurs in the area to be observed and querying the recorded seismic data. Because the inverse mapping of the reversible neural network is based on the mapping of seismic data to the elastic parameters of the Gassmann fluid term, the seismic data of the area to be observed can be input into the inverse mapping process of the trained reversible neural network. Based on the inverse mapping of the reversible neural network, the inversion result of the area to be observed is obtained, and the inversion result is the elastic parameters of the Gassmann fluid term of the area to be observed.

[0121] This example constructs a bidirectional reversible neural network model and a coupling objective function, and trains the reversible neural network model. Based on the trained reversible neural network model, the elastic parameter inversion results of the observed area are obtained, reducing the dependence on initial low-frequency parameters and training samples in the seismic data and improving the accuracy of the inversion results. Optionally, the construction module 91 is specifically used to construct the reversible neural network model. The loss function of the reversible neural network model includes a supervised loss function defined by the error between seismic data, a first unsupervised loss function defined by the error between the latent variable distribution and the Gaussian distribution, and a second unsupervised loss function defined by the error between the elastic parameter distributions. The construction module 91 is further used to define the supervised loss function using a zero-delay cross-correlation function, define the first unsupervised loss function and the second unsupervised loss function using a maximum mean difference function, and couple the supervised loss function, the first unsupervised loss function and the second unsupervised loss function according to the set weight coefficients of each loss function to obtain the coupling objective function. In the context of a scenario example, when constructing the reversible neural network, since the network requires training, a loss function can be built within it. Because seismic data is crucial for obtaining the inversion results, and because there is a correlation between the actual seismic data and the model-predicted seismic data in the observed area, the error between the seismic data can be defined as a supervised loss function based on the zero-latency cross-correlation function. During the training process of the reversible neural network, it learns and memorizes the seismic data. The loss function in supervised training is often used to reflect the degree of inconsistency between the true values ​​of the samples and the model's predicted values. Besides defining the error between the seismic data as a supervised loss function, based on the principle of the maximum mean difference function and the characteristics of unsupervised learning, the error between key parameters during the training process of the reversible neural network is defined as an unsupervised loss function. The weights corresponding to the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function are determined according to the quality of the seismic data and the Gassmann fluid term elastic parameter data. The objective function can be obtained by coupling the supervised loss function, the first unsupervised loss function, and the second unsupervised loss function with their respective weights. This example improves the efficiency of training the reversible neural network and ensures its accuracy by setting the error between key parameters involved in training the reversible neural network as a loss function and coupling the loss function to an objective function. The convergence of the objective function determines whether the reversible neural network has been trained successfully.In the training process of the reversible neural network, the first set of Gassmann fluid elastic parameters from the multiple sets of Gassmann fluid elastic parameters is first input into the forward modeling equation. This forward modeling equation synthesizes seismic data from the input elastic parameters. Therefore, inputting the first set of Gassmann fluid elastic parameters into the forward modeling equation yields the first synthesized seismic data corresponding to the first set of Gassmann fluid elastic parameters. Then, the first set of Gassmann fluid elastic parameters is input into the forward mapping process of the reversible neural network, causing the reversible neural network to output the first predicted seismic parameter and the first latent variable corresponding to the first set of Gassmann fluid elastic parameters. The first predicted seismic parameter and the first latent variable are then input into the reversible neural network, causing the reversible neural network to output the first predicted elastic parameter corresponding to the first set of Gassmann fluid elastic parameters. This process of obtaining the first predicted elastic parameter based on the first set of Gassmann fluid elastic parameters is considered the first iterative training of the reversible neural network. The error between the first synthetic seismic data and the first predicted seismic data is taken as the first error; the error between the first latent variable and the Gaussian distribution is taken as the second error (the Gaussian distribution can be any type); and the error between the first elastic parameter and the first predicted elastic parameter is taken as the third error. Based on the first, second, and third errors, the value of the coupling objective function in the first iteration is obtained. In the first iteration training, the value of the coupling objective function is obtained by summing the first, second, and third errors. After obtaining the value of the coupling objective function in the first iteration, the reversible neural network is updated based on the backpropagation algorithm. After updating the reversible neural network, the second set of Gassmann fluid elastic parameters are input into the forward mapping between the forward equation and the reversible neural network. Based on the second set of Gassmann fluid elastic parameters, the reversible neural network is trained in a second iteration to obtain the value of the coupling objective function in the second iteration training of the second set of Gassmann fluid elastic parameters. The reversible neural network is then updated again using the backpropagation algorithm. According to the above method, based on the generated multiple sets of Gassmann fluid term elastic parameters, the reversible neural network is iteratively trained sequentially to obtain the value of the coupling objective function corresponding to each iteration. After each iteration is completed, the reversible neural network is updated based on the backpropagation algorithm until the coupling objective function converges, at which point the training of the reversible neural network is stopped.It is worth mentioning that if the coupled objective function does not converge after training the reversible neural network with the multiple sets of Gassmann fluid elastic parameters sequentially, then the iterative training of the reversible neural network is re-executed, starting from the first set of Gassmann fluid elastic parameters, until the multiple sets of Gassmann fluid elastic parameters converge. In a scenario example, if the coupled objective function converges after n iterations of training the reversible neural network, the latent variable obtained during the nth iteration is taken as the final latent variable. Based on seismic data survey techniques, actual seismic data generated when an earthquake occurs in the observed area is obtained. This actual seismic data, along with the final latent variable, is used as input data in the inverse mapping process of the reversible neural network, so that the reversible neural network outputs the elastic parameter inversion result, i.e., the elastic parameter distribution corresponding to the observed area. This example uses the final latent variable as part of the input data in the inverse mapping process of the reversible neural network, enabling the reversible neural network to obtain more accurate inversion results during the inversion process. In the scenario example, when generating multiple sets of Gassmann fluid elastic parameters, the generation rules for elastic parameters must be satisfied. Based on these rules, multiple P-wave velocities can be generated using a first empirical formula. To ensure the diversity of the generated Gassmann fluid elastic parameters, a different second empirical formula can be applied to each P-wave velocity to generate the corresponding S-wave velocity, and a different third empirical formula can be applied to each P-wave velocity to generate the corresponding density. The Gassmann fluid elastic parameters are divided into high-frequency and low-frequency components. In addition to obtaining multiple P-wave velocities, S-wave velocities, and densities through empirical formulas, multiple sets of random numbers can be added to represent the remaining P-wave velocities, S-wave velocities, and densities. Low-frequency values ​​in some parameters can also be replaced. This method yields multiple sets of Gassmann fluid elastic parameters with diversity.After obtaining multiple sets of Gassmann fluid term elastic parameters, each set of Gassmann fluid term elastic parameters is compared with the generation rule of elastic parameters. If, during the calculation process using the second empirical formula and the third empirical formula, or during the addition of multiple sets of random numbers or the replacement of low-frequency values, the result of one set of Gassmann fluid term elastic parameters does not conform to the generation rule of elastic parameters, then that set of Gassmann fluid term elastic parameters that does not conform to the generation rule of elastic parameters is deleted. Finally, each set of Gassmann fluid term elastic parameters obtained conforms to the generation rule of elastic parameters. Finally, based on the longitudinal wave velocity, transverse wave velocity, and density in each set of Gassmann fluid term elastic parameters, the Gassmann fluid term and shear modulus corresponding to each set of Gassmann fluid term elastic parameters are calculated using the first calculation formula. The Gassmann fluid term and shear modulus are the elastic parameters in each set of Gassmann fluid term elastic parameters that are ultimately used to train the reversible neural network. This example restricts the generation of multiple sets of Gassmann fluid term elastic parameters by specifying the generation rules for the elastic parameters. This results in higher quality and greater utilization of the generated multiple sets of Gassmann fluid term elastic parameters. Furthermore, by generating multiple sets of Gassmann fluid term elastic parameters through various empirical formulas, adding multiple sets of random numbers, or replacing low-frequency values, the generated multiple sets of Gassmann fluid term elastic parameters can be made more diverse. In the scenario example, since the reversible neural network can perform three functions—forward mapping, backward mapping, and calculation of the coupled objective function—it can be divided into three modules according to function: a forward reversible block, a backward reversible block, and a complementary affine coupling layer. The forward reversible block is used to implement the forward mapping process, the backward reversible block is used to implement the backward mapping process, and the complementary affine coupling layer is used to calculate the coupled objective function. Since the Gassmann fluid term elastic parameters are divided into high-frequency and low-frequency types, the Gassmann fluid term and shear modulus in each group of Gassmann fluid term elastic parameters are respectively divided into high-frequency Gassmann fluid term, low-frequency Gassmann fluid term, high-frequency shear modulus, and low-frequency shear modulus. The high-frequency Gassmann fluid term and the high-frequency shear modulus are used as the first elastic parameter, and the low-frequency Gassmann fluid term and the low-frequency shear modulus are used as the second elastic parameter. The first elastic parameter and the second elastic parameter are simultaneously input into the forward reversible block. After obtaining the predicted earthquake data and latent variables, the predicted earthquake data and the latent variables are input into the reverse reversible block of the reversible neural network.If the reverse mapping process is an iteration during the training of the reversible neural network, the output first elastic parameter x2' and second elastic parameter x1' are the elastic parameter inversion results output during the iteration. If the reverse mapping process is when the reversible neural network has been trained and is being used to invert and obtain elastic parameters, the predicted earthquake data y1 is replaced with the actual earthquake data of the area to be observed, and the latent variable y2 is replaced with the final latent variable. Then, the first elastic parameter x2' obtained by the reverse mapping represents the high-frequency Gassmann fluid term and high-frequency shear modulus of the elastic parameter inversion result for the area to be observed; the second elastic parameter x1' obtained by the reverse mapping represents the low-frequency Gassmann fluid term and low-frequency shear modulus of the elastic parameter inversion result for the area to be observed. This example uses multiple formulas applied to the forward and reverse mapping processes, enabling the reversible neural network to quickly and accurately obtain inversion results during training and use. This embodiment constructs a bidirectional reversible neural network model and a coupled objective function, and trains the reversible neural network model. Based on the trained reversible neural network model, the elastic parameter inversion results of the area to be observed are obtained, which reduces the dependence on the initial low-frequency parameters and training samples in the seismic data and improves the accuracy of the inversion results.

[0122] Example 3

[0123] Using a scenario example, we first tested the model data using a portion of the Marmousi model data. The model data includes 3001 Common Depth Points (CDPs), each CDP consisting of 751 sampling points. Figure 10 This diagram illustrates the elastic parameters, shear modulus, and density of the Gassmann fluid term in the Marmousi model, where a represents the Gassmann fluid term; b represents the shear modulus; and c represents the density. The reflection coefficient was calculated using a forward modeling equation that directly characterizes the reflection coefficient based on the elastic parameters of the Gassmann fluid term. This result was then convolved with a Ricker wavelet (dominant frequency 30 Hz, length 128 ms) to synthesize pre-stack seismic data, which was then used as real seismic data. Figure 11The example uses stacked seismic data. Data from the 1001st and 2001st CDPs are extracted as pseudo-well data, named Well 1 and Well 2, to verify the accuracy of the inversion results obtained by the method of this invention. For example, 1000 sets of Gassmann fluid term elastic parameter data are randomly generated based on the Gassmann fluid term elastic parameter data generation rules, and input into the forward mapping of the constructed reversible neural network model for training. After 300 epochs, the objective function converges. The activation function is the Sigmoid function with a learning rate of 0.01, and the weight coefficients in the objective function are 0.6, 0.3, and 0.1, respectively. Real seismic data is input into the backward mapping of the trained reversible neural network model to obtain the Gassmann fluid term elastic parameter inversion results. Figure 12 The example shows the inversion results of the elastic parameters of the Gassmann fluid term corresponding to real seismic data, where a is the Gassmann fluid term; b is the shear modulus; and c is the density. Figure 13 This is a schematic diagram illustrating the extraction of data from wells A and B as an example. Figure 13 It can be seen that a is well 1 and b is well 2. The elastic parameter inversion results obtained by the method of this application are consistent with... Figure 10 The data shown are generally consistent and agree well with the pseudo-well data. For Well 1, the relative errors between the Gassmann fluid term, shear modulus, and density inversion results and the true values ​​are 0.0342, 0.0351, and 0.0347, respectively; for Well 2, the values ​​are 0.0349, 0.353, and 0.0356, respectively. The above analysis demonstrates that the method of this invention can obtain high-precision Gassmann fluid term elastic parameters without relying on initial low-frequency parameters and training samples. This reduces the dependence on initial low-frequency parameters and training samples in seismic data and improves the accuracy of the inversion results.

[0124] To further verify the effectiveness of the method proposed in this application, this specification uses some F3 open-source actual data for trial calculations. The actual data comes from two wells (Well A and Well B), including 691 CDPs, each CDP comprising 400 sampling points. Figure 14 For another example of post-stack seismic data, the recorded seismic wavelet is extracted. 1500 sets of Gassmann fluid term elastic parameter data are randomly generated based on the Gassmann fluid term elastic parameter data generation rules. These are input into the forward mapping of a pre-constructed reversible neural network model for training. After 400 epochs, the objective function converges. The activation function is the Sigmoid function with a learning rate of 0.01, and the weight coefficients in the objective function are 0.5, 0.4, and 0.1, respectively. Real seismic data is then input into the backward mapping of the trained reversible neural network model to invert the Gassmann fluid term elastic parameters. Figure 15This example shows the Gassmann fluid term elastic parameter inversion result corresponding to another real seismic data. Figure 16 This is a schematic diagram of another extraction of data from wells A and B, where a represents well 1 and b represents well 2. Figure 16 As can be seen, the inversion results obtained by the method of this invention show good agreement with the well logging data. For well A, the relative errors of the inverted Gassmann fluid term, shear modulus, and density compared with the well logging data are 0.0875, 0.0879, and 0.0881, respectively; for well B, the relative errors are 0.893, 0.891, and 0.898, respectively. The above analysis shows that the method of this application can obtain highly accurate Gassmann fluid term elastic parameters using only randomly generated elastic parameter data. This reduces the dependence on initial low-frequency parameters and training samples in the seismic data and improves the accuracy of the inversion results.

[0125] Example 4

[0126] Figure 17 This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of this application, as shown below. Figure 17 As shown, the electronic device includes:

[0127] The server includes a processor 291 and a memory 292; it may also include a communication interface 293 and a bus 294. The processor 291, memory 292, and communication interface 293 can communicate with each other via the bus 294. The communication interface 293 can be used for information transmission. The processor 291 can call logical instructions stored in the memory 292 to execute the method described in Embodiment 1.

[0128] Furthermore, the logic instructions in the aforementioned memory 292 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0129] The memory 292, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this application. The processor 291 executes functional applications and data processing by running the software programs, instructions, and modules stored in the memory 292, that is, implementing the method of Embodiment 1 described above.

[0130] The memory 292 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 292 may include high-speed random access memory and may also include non-volatile memory.

[0131] This application provides a non-transitory computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods described in the foregoing embodiments.

[0132] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention filed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not claimed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0133] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An inversion method characterized by, The method comprises: constructing a reversible neural network model comprising a forward reversible block and a reverse reversible block of complementary affine coupling layers; a loss function of the reversible neural network model comprises a supervised loss function defined by an error between seismic data, a first unsupervised loss function defined by an error between a latent variable distribution corresponding to an elastic parameter and a Gaussian distribution, and a second unsupervised loss function defined by an error between elastic parameter distributions; According to the weight coefficients of each loss function set, the supervised loss function, the first unsupervised loss function and the second unsupervised loss function are coupled to obtain a coupled objective function; Based on the elastic parameter characteristics of the to-be-observed region, a plurality of groups of Gassmann fluid item elastic parameters are randomly generated; By inputting the plurality of groups of Gassmann fluid item elastic parameters into the forward mapping process of the reversible neural network, and based on the coupled objective function, the reversible neural network is iteratively trained until the coupled objective function converges, thereby obtaining a trained reversible neural network model; wherein the forward mapping of the reversible neural network model reflects the mapping of the elastic parameter to the seismic data, and the reverse mapping of the reversible neural network model reflects the mapping of the seismic data to the elastic parameter; The actual seismic data of the to-be-observed area is obtained, the actual seismic data of the to-be-observed area is input into a reverse mapping process of the trained reversible neural network model, and an elastic parameter inversion result of the to-be-observed area is obtained; in the forward mapping process of the reversible neural network model, the forward reversible block is used to split the input elastic parameter into a first elastic parameter and a second elastic parameter; the first elastic parameter is processed based on a first activation function to obtain a first processing result; the first elastic parameter is processed based on a second activation function to obtain a second processing result, and predicted seismic data is output by using matrix inner product calculation and summation calculation according to the second elastic parameter, the first processing result and the second processing result; and the predicted seismic data is processed based on a third activation function to obtain a third processing result; the second elastic parameter is processed based on a fourth activation function to obtain a fourth processing result, and latent variables are output by using matrix inner product calculation and summation calculation according to the first elastic parameter, the third processing result and the fourth processing result; in the reverse mapping process of the reversible neural network model, the reverse reversible block is used to process the predicted seismic data based on the third activation function to obtain a fifth processing result; the predicted seismic data is processed based on the fourth activation function to obtain a sixth processing result, and the first elastic parameter is output by using matrix inner product calculation and difference calculation according to the fifth processing result, the sixth processing result and the input latent variables; and the seventh processing result is obtained by processing the first elastic parameter based on the first activation function; the eighth processing result is obtained by processing the first elastic parameter based on the second activation function, and the second elastic parameter is output by using matrix inner product calculation and difference calculation according to the seventh processing result, the eighth processing result and the predicted seismic data, and the first elastic parameter and the second elastic parameter are combined and output as predicted elastic parameters.

2. The method of claim 1, wherein, Also includes: The supervised loss function is defined by using a zero-delay cross-correlation function, and the first unsupervised loss function and the second unsupervised loss function are defined by using a maximum mean difference function.

3. The method of claim 2, wherein, The reversible neural network model is obtained by inputting the multiple groups of Gassmann fluid item elastic parameters into the forward mapping process of the reversible neural network and iteratively training the reversible neural network based on the coupling objective function until the coupling objective function converges, including: The elastic parameter is input into a forward equation directly represented by the elastic parameter, and synthetic seismic data corresponding to the elastic parameter is obtained; inputting the elastic parameters into a forward mapping process of the reversible neural network model to obtain predicted seismic data and latent variables corresponding to the elastic parameters output by the reversible neural network model in the forward mapping process; and inputting the predicted seismic data and the latent variables corresponding to the elastic parameters into a reverse mapping process of the reversible neural network to obtain predicted elastic parameters corresponding to the elastic parameters output by the reversible neural network model in the reverse mapping process; updating the reversible neural network model based on the coupling objective function in a reverse propagation algorithm until the coupling objective function converges, to obtain the trained reversible neural network model.

4. The method of claim 3, wherein, The actual seismic data of the to-be-observed region is obtained, and the actual seismic data of the to-be-observed region is input into a reverse mapping process of the trained reversible neural network model to obtain an elastic parameter inversion result of the to-be-observed region, including: The actual seismic data of the to-be-observed region and a final latent variable are obtained, the final latent variable being a latent variable output by the reversible neural network model in a forward mapping process for the last time when the coupling objective function converges; The actual seismic data of the to-be-observed region and the final latent variable are input into a reverse mapping process of the trained reversible neural network model to obtain an elastic parameter inversion result of the to-be-observed region.

5. The method of claim 1, wherein, The multiple groups of Gassmann fluid term elastic parameters are randomly generated based on elastic parameter characteristics of the to-be-observed region, including: a plurality of P-wave velocities are randomly generated according to an empirical formula, and a S-wave velocity and a density corresponding to each P-wave velocity are calculated, and a plurality of S-wave velocities and a plurality of densities are obtained by adding random numbers and / or replacing low frequencies; the multiple groups of Gassmann fluid term elastic parameters are calculated according to the plurality of P-wave velocities, the plurality of S-wave velocities and the plurality of densities; whether the multiple groups of Gassmann fluid term elastic parameters meet a set random generation rule is determined, and elastic parameters that do not meet the random generation rule are removed from the multiple groups of Gassmann fluid term elastic parameters, the random generation rule including at least one of the following: a value range of the randomly generated elastic parameters is not greater than an elastic parameter value range of the to-be-observed region, a thinnest layer thickness of the randomly generated elastic parameters is less than a thinnest layer thickness identified according to seismic data of the to-be-observed region, and a vertical variation trend of the randomly generated elastic parameters conforms to a geological law of the to-be-observed region.

6. An inverting device, characterized by The device includes: The constructing module is configured to construct a reversible neural network model, the reversible neural network model comprising a forward reversible block and a reverse reversible block of a complementary affine coupling layer; a loss function of the reversible neural network model comprises a supervised loss function defined based on an error between seismic data, a first unsupervised loss function defined based on an error between a latent variable distribution corresponding to an elastic parameter and a Gaussian distribution, and a second unsupervised loss function defined based on an error between elastic parameter distributions The constructing module is further configured to couple the supervised loss function, the first unsupervised loss function and the second unsupervised loss function according to a weight coefficient of each loss function, to obtain a coupled target function; The constructing module is further configured to randomly generate a plurality of sets of Gassmann fluid item elastic parameters based on an elastic parameter feature of a region to be observed; The input module is configured to input the plurality of sets of Gassmann fluid item elastic parameters to a forward mapping process of the reversible neural network, and iteratively train the reversible neural network based on the coupled target function until the coupled target function converges, to obtain a trained reversible neural network model; wherein the forward mapping of the reversible neural network model reflects mapping of elastic parameters to seismic data, and the reverse mapping of the reversible neural network model reflects mapping of seismic data to elastic parameters; The acquisition module is configured to acquire actual seismic data of the to-be-observed area, and input the actual seismic data of the to-be-observed area into a reverse mapping process of the trained reversible neural network model to obtain an elastic parameter inversion result of the to-be-observed area. In the forward mapping process of the reversible neural network model, the forward reversible block is configured to split the input elastic parameter into a first elastic parameter and a second elastic parameter, process the first elastic parameter based on a first activation function to obtain a first processing result, process the first elastic parameter based on a second activation function to obtain a second processing result, and output predicted seismic data by using matrix inner product calculation and summation calculation according to the second elastic parameter, the first processing result and the second processing result. The predicted seismic data is processed based on a third activation function to obtain a third processing result, and the second elastic parameter is processed based on a fourth activation function to obtain a fourth processing result. The first elastic parameter, the third processing result and the fourth processing result are used to output a latent variable by using matrix inner product calculation and summation calculation. In the reverse mapping process of the reversible neural network model, the reverse reversible block is configured to process the predicted seismic data based on the third activation function to obtain a fifth processing result, process the predicted seismic data based on the fourth activation function to obtain a sixth processing result, and output the first elastic parameter by using matrix inner product calculation and difference calculation according to the fifth processing result, the sixth processing result and the input latent variable. The first elastic parameter is processed based on the first activation function to obtain a seventh processing result, and the first elastic parameter is processed based on the second activation function to obtain an eighth processing result. The seventh processing result, the eighth processing result and the predicted seismic data are used to output the second elastic parameter by using matrix inner product calculation and difference calculation, and the first elastic parameter and the second elastic parameter are combined to output the predicted elastic parameter.

7. The apparatus of claim 6, wherein: the construction module is further configured to define the supervised loss function using a zero-lag cross-correlation function, and define the first unsupervised loss function and the second unsupervised loss function using a maximum mean discrepancy function.

8. An electronic device, comprising: comprising: a processor, and a memory connected to the processor in communication; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method of any one of claims 1-5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are executed by the processor to implement the method of any one of claims 1-5.

Images