Deep network based full waveform inversion acceleration method and related device

Abstract

The application provides a full waveform inversion acceleration method based on a deep network and a related device, and relates to the technical field of seismic exploration. The method comprises the following steps: obtaining first migration data of a first velocity model; inputting the first velocity model and the first migration data into a pre-trained deep learning network to output a second velocity model; the deep learning network is obtained based on iterative data and migration data, the iterative data is iterative data generated when full waveform inversion (FWI) is performed on forward data, the forward data is obtained by performing forward calculation on the iterative data, and the migration data is obtained by performing migration on the forward data based on the iterative data; and performing FWI based on the second velocity model to obtain a target model. The target model obtained by using the above method can accelerate the convergence of the FWI iteration process and improve the divergence problem in the FWI iteration process.

Description

Technical Field

[0001] This application relates to the field of seismic exploration technology, and in particular to a full waveform inversion acceleration method and related apparatus based on deep networks. Background Technology

[0002] Full waveform inversion (FWI) is a high-resolution seismic exploration imaging method that iteratively inverts subsurface medium parameters based on the full waveform information of seismic data. It is currently widely used in the field of oil and gas geophysics and in regional, plate, and global scale seismological research.

[0003] In mathematics, FWI is a strongly nonlinear ill-conditioned inverse problem. From a computational efficiency perspective, it relies on iterative solutions using local gradients, and is therefore limited by the "period jump" phenomenon. "Period jump" refers to a time difference between the observed and predicted signals that is greater than half a period, which makes FWI prone to getting trapped in local minima, and also causes large-scale computation due to its dimensionality.

[0004] Therefore, existing FWI methods are prone to divergence during the iteration process, and FWI iteration is time-consuming and has a slow convergence speed. Summary of the Invention

[0005] This application provides a deep network-based full waveform inversion acceleration method and related apparatus to solve the problems of existing FWI methods being prone to divergence during iteration, and the FWI iteration taking a long time and having a slow convergence speed.

[0006] According to a first aspect of this application, a method for accelerating full waveform inversion based on deep networks is provided, comprising:

[0007] Obtain the first offset data of the first velocity model; input the first velocity model and the first offset data into a pre-trained deep learning network to output a second velocity model; the deep learning network is trained based on iterative data and offset data, the iterative data is generated when performing full waveform inversion (FWI) on forward modeling data, the forward modeling data is obtained by performing forward modeling on the iterative data, and the offset data is obtained by offsetting the forward modeling data based on the iterative data; perform FWI based on the second velocity model to obtain the target model.

[0008] Optionally, before obtaining the first migration data of the first velocity model, the method further includes: obtaining a training set, the training set including M sample velocity models; performing forward modeling on the M sample velocity models respectively to obtain M pre-stack seismic data; performing full waveform inversion on the M pre-stack seismic data respectively to obtain M*N iterative velocity models; wherein, when performing full waveform inversion on any of the pre-stack seismic data, N iterative velocity models are obtained; for any of the iterative velocity models, performing depth migration on the pre-stack seismic data corresponding to the iterative velocity model to obtain the migration data of the iterative velocity model; using the M*N iterative velocity models and the migration data of the M*N iterative velocity models as dual-channel inputs to the deep learning network to be trained, training the deep learning network to obtain the deep learning network; wherein, M and N are both natural numbers.

[0009] Optionally, when performing FWI on any one of the M pre-stack seismic data, N iterations are performed sequentially; wherein, in the N iterations, during the i-th iteration, the model is updated by using the difference between the pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration, so as to obtain the iterative velocity model generated in the i-th iteration.

[0010] Optionally, the loss function of the deep network to be trained satisfies the following formula:

[0011]

[0012] Where G represents the deep learning network of the deep network to be trained, and V i P represents the iterative speed model. i V represents the offset data of the iterative velocity model. T Let represent the sample velocity model, j represent the j-th sample velocity model, x represent the surface spatial coordinates, and z represent the imaging depth.

[0013] Optionally, before performing FWI based on the second velocity model to obtain the target model, the method further includes: determining whether the difference between the first velocity model and the data in the training set is greater than a difference threshold; the step of inputting the first velocity model and the first offset data into a pre-trained deep learning network includes: when the difference between the first velocity model and the data in the training set is less than or equal to the difference threshold, inputting the first velocity model and the first offset data into the pre-trained deep learning network.

[0014] Optionally, performing FWI based on the second velocity model to obtain the target model includes: when the difference between the first velocity model and the training set data is greater than the difference threshold, using the imaging result of the first velocity model as a guide image and the imaging result of the second velocity model as an input image, performing guided filtering to obtain a filtered velocity model; calculating the residual between the second velocity model and the filtered velocity model; superimposing the residual with the first velocity model to obtain a third velocity model; and performing FWI on the third velocity model to obtain the target model.

[0015] According to a second aspect of this application, a deep network-based full waveform inversion acceleration device is provided, comprising:

[0016] The acquisition module is used to acquire the first offset data of the first velocity model;

[0017] The input / output module is used to input the first velocity model and the first offset data into a pre-trained deep learning network and output a second velocity model; the deep learning network is trained based on iterative data and offset data, the iterative data is generated when performing full waveform inversion (FWI) on forward data, the forward data is obtained by performing forward modeling on the iterative data, and the offset data is obtained by offsetting the forward data based on the iterative data;

[0018] The inversion module is used to perform FWI based on the second velocity model to obtain the target model.

[0019] Optionally, the acquisition module is further configured to acquire a training set, which includes M sample velocity models; perform forward modeling on the M sample velocity models to obtain M pre-stack seismic data; the inversion module is further configured to perform full waveform inversion on the M pre-stack seismic data to obtain M*N iterative velocity models; wherein, when performing full waveform inversion on any one of the pre-stack seismic data, N iterative velocity models are obtained; for any one of the iterative velocity models, depth migration is performed on the pre-stack seismic data corresponding to the iterative velocity model to obtain the migration data of the iterative velocity model; the input / output module is further configured to use the M*N iterative velocity models and the migration data of the M*N iterative velocity models as dual-channel inputs to the deep network to be trained, train the deep network to be trained, and obtain the deep learning network; wherein, M and N are both natural numbers.

[0020] Optionally, the inversion module is also used to perform N iterations sequentially when performing FWI on any of the M pre-stack seismic data; wherein, in the N iterations, when performing the i-th iteration, the model is updated by using the difference between the pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration, so as to obtain the iterative velocity model generated in the i-th iteration.

[0021] Optionally, the loss function of the deep network to be trained satisfies the following formula:

[0022]

[0023] Where G represents the deep learning network of the deep network to be trained, and V i P represents the iterative speed model. i V represents the offset data of the iterative velocity model. T Let represent the sample velocity model, j represent the j-th sample velocity model, x represent the surface spatial coordinates, and z represent the imaging depth.

[0024] Optionally, the input / output module is further configured to determine whether the difference between the first velocity model and the training set data is greater than a difference threshold; the step of inputting the first velocity model and the first offset data into the pre-trained deep learning network includes: when the difference between the first velocity model and the training set data is less than or equal to the difference threshold, inputting the first velocity model and the first offset data into the pre-trained deep learning network.

[0025] Optionally, the input / output module is further configured to, when the difference between the first velocity model and the training set data is greater than the difference threshold, use the imaging result of the first velocity model as a guide image and the imaging result of the second velocity model as an input image to perform guided filtering to obtain a filtered velocity model; calculate the residual between the second velocity model and the filtered velocity model; superimpose the residual with the first velocity model to obtain a third velocity model; the inversion module is further configured to perform FWI on the third velocity model to obtain the target model.

[0026] According to a third aspect of this application, an electronic device is provided, comprising:

[0027] At least one processor and memory;

[0028] The memory stores computer-executed instructions;

[0029] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the deep network-based full waveform inversion acceleration method as described in the first aspect above.

[0030] According to a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, which, when executed by a processor, are used to implement the deep network-based full waveform inversion acceleration method described in the first aspect above.

[0031] According to a fourth aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the deep network-based full waveform inversion acceleration method as described in the first aspect above.

[0032] This application provides a method and related apparatus for accelerating full waveform inversion based on deep networks. The method uses the iterative velocity model obtained from FWI iteration and the corresponding offset data as input to a deep learning network for training. When the velocity model and its offset data are input into the trained deep learning network, the network can output a new velocity model that approximates the true velocity model. Using this new velocity model for FWI can accelerate the convergence of subsequent iterations and improve the divergence problem during FWI iteration.

[0033] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0034] 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.

[0035] Figure 1 This is a schematic diagram illustrating the application scenarios involved in the embodiments of this application;

[0036] Figure 2 A flowchart illustrating the full waveform inversion acceleration method based on deep networks provided in this application embodiment;

[0037] Figure 3 A schematic diagram illustrating the process of training a deep learning network as provided in an embodiment of this application;

[0038] Figure 4 This is a schematic diagram of the process of the guided filter-based migration algorithm provided in the embodiments of this application;

[0039] Figure 5 This is a schematic diagram illustrating the process of training a deep learning network as provided in an embodiment of this application.

[0040] Figure 6This is a schematic diagram illustrating the application of this embodiment to a certain FWI iterative convergence speed model;

[0041] Figure 7 This is a schematic diagram illustrating the application of this embodiment of the application to a velocity model where an FWI iteration does not converge.

[0042] Figure 8 A schematic diagram showing the changes in PCC and MAE with the number of iteration steps for the velocity model provided in the embodiments of this application;

[0043] Figure 9 A schematic diagram of the local PCC and MAE corresponding to a certain FWI iterative convergence velocity model provided in an embodiment of this application;

[0044] Figure 10 A schematic diagram of the local PCC and MAE corresponding to a certain FWI iteration non-convergence velocity model provided in an embodiment of this application;

[0045] Figure 11 A comparison chart of the average PCC and MAE metrics of three deep network models provided in the embodiments of this application;

[0046] Figure 12 This is a schematic diagram showing the updated result of the initial velocity model of a velocity model that has converged in an iterative convergence according to an embodiment of this application.

[0047] Figure 13 A schematic diagram of the PCC and MAE corresponding to the updated initial velocity model of a velocity model for a certain FWI iterative convergence provided in an embodiment of this application;

[0048] Figure 14 This is a schematic diagram showing the updated initial velocity model of a velocity model that does not converge in an FWI iteration, as provided in an embodiment of this application.

[0049] Figure 15 This is a schematic diagram of the PCC and MAE corresponding to the updated initial velocity model of a velocity model that does not converge in a certain FWI iteration provided in an embodiment of this application.

[0050] Figure 16 This is a schematic diagram showing the results of applying the embodiments of this application to the Marmousi model;

[0051] Figure 17 A schematic diagram illustrating the processing results of the guided filter migration algorithm applied to the Marmousi initial velocity model;

[0052] Figure 18 This is a schematic diagram of PCC and MAE corresponding to FWI iteration after applying the Marmousi initial velocity model in an embodiment of this application;

[0053] Figure 19 A schematic diagram of the structure of a deep network-based full waveform inversion acceleration device provided in an embodiment of this application;

[0054] Figure 20 This is a schematic diagram of the structure of an electronic device provided in an embodiment 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 represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0057] With the development of computer technology and the improvement of computer performance, FWI has begun to be successfully applied in marine seismic data. In 2008, BP used FWI in the Valhall oil field in Norway to obtain high-precision imaging results. However, due to the lack of low-frequency information and other problems, land seismic data still faces great difficulties in practical applications.

[0058] In 1989, Mora P pointed out that FWI (Focused Inversion Weighing) is equivalent to tomography plus depth migration. Tomography updates the background velocity, while depth migration repositions the reflecting surfaces, ultimately matching the velocity model with the reflecting horizons on the seismic depth migration profile. However, obtaining seismic horizon information and seamlessly integrating it into the FWI process to guide velocity updates has always been a challenging problem. Currently, deep learning technology has been widely applied in seismic inversion and velocity modeling.

[0059] Das et al. first proposed in 2018 to use a two-layer convolutional neural network (CNN) to learn the inversion mapping from seismic data to wave impedance, with the training dataset generated by continuous Gaussian simulations. Alfarraj proposed a semi-supervised learning model that models both the forward and inversion processes simultaneously, increasing the amount of training samples. Guo utilized the memory capacity of RNNs to propose a method combining CNN and bidirectional LSTM to handle the one-dimensional wave impedance inversion problem. The lateral continuity of the inversion results is better than that of 1DCNN, but it requires a large amount of labeled data and has a weak ability to invert high-frequency information. Ge combined geostatistics with a closed-loop network structure. Li et al. proposed using a Generative Adversarial Network (GAN) to generate a distribution space consistent with well logging data, which can then be used as training data for subsequent applications. Wang et al. proposed using a one-dimensional Cycle-GAN to perform velocity inversion on one-dimensional (D) seismic data, leveraging the ability of closed-loop networks such as Cycle-GAN to simultaneously model and learn the forward and inversion processes, thus expanding the diversity of the training sample set. In subsequent work, Wang added bilateral filtering constraints to the one-dimensional closed-loop network, further improving the inversion effect. Later, Q Wang et al. used 2D CNN for inversion, incorporating geological constraints into the inversion process to improve lateral continuity. In terms of velocity modeling, Han Mingliang et al. jointly used reflection waveform data and velocity spectrum as input to a fully convolutional neural network to achieve the mapping from seismic data to a velocity model. Bo Mao from Jilin University proposed a velocity modeling convolutional neural network that takes self-excited and self-harvested seismic data as input and outputs the velocity distribution of the target area.

[0060] Applying deep learning to full waveform inversion is currently a research hotspot. Sun proposed using a recurrent neural network (RNN) to model the inversion process, achieving unsupervised full waveform inversion. Liu Sitong from Harbin Institute of Technology proved that automatic gradient differentiation based on deep learning is equivalent to the adjoint state method commonly used in full waveform inversion. However, under the L2 norm objective function, the problem of FWI being prone to getting trapped in local minima due to the time difference between the observed and predicted signals being greater than half a cycle (called "cycle jump") caused by inaccurate background velocity remains unresolved.

[0061] Currently, the mainstream solutions for addressing "cycle jumps" and aiding FWI convergence can be summarized into three aspects: objective function optimization, model space expansion, and low-frequency signal reconstruction. These solutions are essentially error functional estimations constrained by prior information. However, in practical applications, the prior information used in these methods may not conform to the complex and ever-changing realities, easily leading to inaccurate inversion results.

[0062] To facilitate understanding, the application scenarios of the embodiments of this application will be introduced first.

[0063] Figure 1 This is a schematic diagram illustrating an application scenario involved in an embodiment of this application. For example... Figure 1 As shown, the application scenario of this application embodiment involves shot point 101, detector 102 and server 103.

[0064] A shot was fired at shot point 101 to simulate an earthquake. Actual seismic data was obtained using geophone 101, and a full waveform inversion was performed on the server using the actual seismic data and the initial velocity model to obtain the subsurface velocity model. The initial velocity model is typically obtained using ray tomography, migration velocity analysis, or sonic logging interpolation.

[0065] However, since FWI is a highly nonlinear and ill-posed problem, it may be affected by "period jumps" or the initial velocity model, and the iteration process requires several seismic wave forward modeling calculations. Therefore, improving the divergence problem in the FWI iteration process and improving computational efficiency are the two major challenges of FWI.

[0066] Based on this, this application proposes a deep network-based method for accelerating full waveform inversion (FWI) to address the aforementioned problems of divergence and low computational efficiency in the FWI iteration process. In this application, the iterative velocity model obtained from FWI iteration and the corresponding offset data are used as input to a deep learning network for training. When the velocity model and its offset data are input into the trained deep learning network, the network can output a new velocity model that approximates the true velocity model. Using this new velocity model for FWI accelerates the convergence of subsequent iterations and improves the divergence problem during FWI iteration.

[0067] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0068] Figure 2This is a flowchart illustrating the full waveform inversion acceleration method based on deep networks provided in an embodiment of this application. Figure 2 As shown, the method of this application embodiment, executed by a server, includes the following steps:

[0069] S201, Obtain the first offset data of the first velocity model.

[0070] In this embodiment, the first velocity model can be any iterative velocity model in the FWI iteration process, and the first velocity model can be V. I (x, z) represents the spatial coordinates of the Earth's surface, z represents the imaging depth, and I represents the number of iteration steps.

[0071] The first offset data is the offset data corresponding to the first velocity model. This data can be represented by P. I (x, z) represents the first migration data. In this embodiment, based on the first velocity model, depth migration (DM) technology can be used to migrate pre-stack seismic data to obtain the first migration data. This process can be expressed by the following formula:

[0072] P I (x, z) = DM(V) I (x, z), S(x) s x r ,t)) (1)

[0073] Where P represents the offset data; S(x s x r (t) represents pre-stack seismic data, also known as actual observed seismic data. In practical applications, pre-stack seismic data can be directly measured by detectors. s The x-axis represents the Earth's surface spatial coordinates of the earthquake source. r The surface spatial coordinates of the receiver point are represented; t represents time; DM represents the depth migration operator. In this embodiment, the first migration data can be obtained using the reverse time migration (RTM) technique in depth migration technology.

[0074] In one possible implementation, the server obtains the first offset data corresponding to the first velocity model through depth offset technology.

[0075] S202. Input the first velocity model and the first offset data into a pre-trained deep learning network and output the second velocity model. The deep learning network is trained based on the iterative data and the offset data. The iterative data is the iterative data generated when performing full waveform inversion FWI on the forward modeling data. The forward modeling data is obtained by performing forward modeling on the iterative data. The offset data is obtained by offsetting the forward modeling data based on the iterative data.

[0076] In this embodiment, the deep learning network can be a generative adversarial network (GAN) or a convolutional neural network (CNN). The second velocity model is obtained by inputting the first velocity model and the first offset data into a pre-trained deep learning network. The second velocity model can be... The second velocity model represents a significant improvement in quality compared to the first velocity model, more closely approximating the true velocity model. Taking a GAN (Generative Adversarial Network) deep learning network as an example, the process of obtaining the second velocity model can be represented by the following formula:

[0077]

[0078] Deep learning networks are trained using iterative and offset data. The training process for deep learning networks can be found by referring to... Figure 3 The description in the text will not be repeated here.

[0079] In one possible implementation, the server inputs a first velocity model and first offset data into a pre-trained deep learning network and outputs a second velocity model that approximates the real velocity model.

[0080] S203. Perform FWI based on the second velocity model to obtain the target model.

[0081] In this embodiment, the target model is a velocity model obtained through full waveform inversion. In practical applications, since the specific velocity model of the underground structure is unknown, existing methods mainly obtain the specific velocity model of the underground structure, i.e., the target model, through FWI (Full Waveform Inversion).

[0082] In one possible implementation, after replacing the first velocity model with the second velocity model, the server performs subsequent FWI iterations based on the second velocity model to obtain the target model.

[0083] Therefore, in this embodiment of the application, a second velocity model that approximates the real velocity model can be obtained by inputting the first velocity model and the corresponding first offset data obtained during the FWI iteration process into a pre-trained deep learning network. The second velocity model can replace the first velocity model and be applied to the subsequent FWI iteration process, which can accelerate the convergence of the subsequent FWI iteration process and improve the divergence problem in the FWI iteration process.

[0084] Based on the above embodiments, this application also provides a method for training deep learning networks. For example... Figure 3 As shown, the method includes the following steps:

[0085] S301. Obtain the training set, which includes M sample velocity models.

[0086] In this embodiment, sample velocity models in the training set can be directly generated via a server. The training set includes M sample velocity models used to train the deep learning network. The velocity model can also be called the true velocity model, referring to the velocity parameters of the underground structure. The velocity model uses... Let T be an abbreviation for truth, representing the true velocity model, and j represent the j-th velocity model.

[0087] S302. Perform forward modeling on the velocity models of the M samples respectively to obtain M pre-stack seismic data.

[0088] In this embodiment of the application, forward modeling refers to obtaining the seismic waveform by using forward modeling with known velocity parameters of the underground structure.

[0089] In one possible implementation, forward modeling is performed on the M sample velocity models on the server to obtain the corresponding M pre-stack seismic data. For example, for the j-th sample velocity model... Perform forward modeling to obtain the corresponding pre-stack seismic data S. j (x s x r ,t).

[0090] S303. Perform full waveform inversion on M pre-stack seismic data points to obtain M*N iterative velocity models; wherein, when performing full waveform inversion on any one pre-stack seismic data point, N iterative velocity models are obtained. Here, M and N are both natural numbers.

[0091] In the N iterations, during the i-th iteration, the model is updated by using the difference between any pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration, thus obtaining the iterative velocity model generated in the i-th iteration.

[0092] In this embodiment, the iterative velocity model refers to any velocity model obtained during the iterative inversion process of pre-stack seismic data obtained by forward modeling of the velocity model in the training set. The iterative velocity model can be used... This means that each iteration yields one iteration rate model, and N iterations yield N iteration rate models.

[0093] In the FWI iteration process, the i-th iteration of FWI updates the model by using the difference between any pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration. For example, using the j-th velocity model... For example, The corresponding pre-stack seismic data is S j (x s x r ,t), using S j (x s x r ,t) is obtained by full waveform inversion in, It is to utilize With S j (x s x r The difference between t and t is obtained. This process can be expressed by the following formula:

[0094]

[0095] FWI is the full waveform inversion operator.

[0096] It is understandable that when i = 1, for The corresponding forward modeling data, This is the initial velocity model.

[0097] S304. For any given iterative velocity model, perform depth migration on the pre-stack seismic data corresponding to that iterative velocity model to obtain the migration data for that iterative velocity model.

[0098] A detailed description of the depth migration method can be found in step S201. For example, the velocity model of the i-th iteration of FWI. Corresponding pre-stack seismic data S j (x s x r Offset the data by t), and obtain the offset data as follows:

[0099] S305. Use the offset data of M*N iterative velocity models and M*N iterative velocity models as dual-channel inputs to train the deep network to obtain the deep learning network.

[0100] The loss function of the deep network to be trained satisfies the following formula:

[0101]

[0102] Where G represents a pre-trained deep learning network.

[0103] In one possible implementation, the offset data of M*N iterative velocity models and M*N iterative velocity models are used as dual-channel inputs to train the deep network. For example, in the case of a GAN deep learning network, data pairs are constructed. As a dual-channel input to a deep network, training a deep learning network GAN makes its output approximate the corresponding real speed model. During training, the L1 loss function can be used to measure the output of a deep learning network. The differences between them are used to adjust the parameters of the deep learning network and obtain a well-trained deep learning network.

[0104] In this way, the trained deep learning network can output a target model that approximates the real speed model based on the input data, satisfying a preset relationship.

[0105] Based on the above embodiments, this application also provides a transfer learning method based on guided filtering. For example... Figure 4 As shown, before obtaining the target model by performing FWI based on the second velocity model, the method includes the following steps:

[0106] S401. The server determines whether the difference between the first speed model and the training set data is greater than the difference threshold.

[0107] In this embodiment of the application, the difference between the first velocity model and the training set data can refer to a significant difference between the two velocity structures or a significant difference in the velocity value range.

[0108] S402, Inputting the first velocity model and the first offset data into a pre-trained deep learning network, including: when the difference between the first velocity model and the data in the training set is less than or equal to a difference threshold, inputting the first velocity model and the first offset data into the pre-trained deep learning network.

[0109] In one possible implementation, when the server determines that the difference between the first velocity model and the training set data is less than or equal to a difference threshold, it inputs the first velocity model and the first offset data into a pre-trained deep learning network to obtain a second velocity model. Based on the second velocity model, FWI can be performed to obtain the target model.

[0110] S403. When the difference between the first velocity model and the training set data exceeds a difference threshold, the imaging result of the first velocity model is used as the guide image, and the imaging result of the second velocity model is used as the input image. Guided filtering is performed to obtain the filtered velocity model. The residual between the second velocity model and the filtered velocity model is calculated. The residual is superimposed with the first velocity model to obtain the third velocity model. FWI is performed on the third velocity model to obtain the target model.

[0111] In one possible implementation, V is used. new(x, z) represents the first velocity model whose difference from the training set data is greater than the difference threshold. The corresponding offset data can be represented by P. new Let (x, z) represent V. new (x, z) and P new Inputting (x, z) into a pre-trained deep learning network yields the second velocity model. V new The velocity imaging results of (x, z) are used as the guiding image. The velocity imaging result is used as the input image, and guided filtering is performed to obtain V. Guided (x, z). Calculate With V Guided The residuals between (x, z) are used as stratigraphic information, and the residuals are compared with... Add them together to get use Replace V new (x, z) are used for subsequent FWI iterations.

[0112] This process can be represented by the following formula:

[0113]

[0114]

[0115]

[0116] Here, GF represents the guided filter operator. The diameter of the guided filter kernel can be 11, and the normalization parameter ∈ can be set to 0.01.

[0117] In this way, when the difference between the first velocity model and the velocity model in the training set is greater than the difference threshold, the guided filter transfer algorithm can improve the inaccuracy of the deep learning network output caused by the large difference between the two velocity models, and thus improve the subsequent FWI iteration process.

[0118] Based on the above embodiments, this application also provides a more specific method for accelerating full waveform inversion based on deep networks, which includes the following steps:

[0119] S501, the server randomly generates 190 speed models.

[0120] S502, for each velocity model Shot data was generated using a Rayleigh wavelet with a main frequency of 20Hz. The horizontal and vertical sampling intervals were dx = dz = 10.0m, with 788 horizontal and 266 vertical sampling points, respectively. The starting coordinate of the shot point was 6, with an interval of 8, for a total of 98 shots.

[0121] Among them, the Ricker wavelet is a type of seismic wavelet, which is a short pulse vibration. Shot data refers to actual observed seismic data. In practical applications, shot data is the data collected by the geophone during shot firing, reflecting the stratigraphic data corresponding to the coordinates of the geophone and the shot point. Shot point refers to the location where the shot is fired.

[0122] S503, for each velocity model Perform 300 FWI iterations. The corresponding one-dimensional initial velocity during the FWI iteration process is: After every 10 iterations, the obtained velocity model is used as a basis. Reverse time migration is performed to obtain migrated seismic data.

[0123] Among them, the data corresponding to the first 150 velocity models were selected. The data corresponding to the 20 velocity models serves as the training set. The remaining 20 velocity models serve as the validation set. As a test set, all velocity models and RTM seismic data are 266*788 pixels in size.

[0124] S504, will As the dual-channel input to a deep neural network, the process of training a deep neural network can be described as follows: Figure 5 As shown.

[0125] The deep neural network can use the Pix2Pix network model. The Pix2Pix network model is a generative adversarial network model with a generator (G) and a discriminator (D). The Pix2Pix network model in this embodiment uses the U-Net network structure.

[0126] In one possible implementation, the velocity model and corresponding depth migration seismic data After standardization, the data are merged into two input channels. Since the geological velocity range on land is generally between 1000 m / s and 8000 m / s, while the value range of depth-migrated seismic data (signal amplitude is generally between -0.2 and 0.2) is different, a large difference between the two could lead to numerical instability in the model. Therefore, the velocity model needs to be optimized. and corresponding depth migration seismic data Standardize them separately.

[0127] The standardization process is as follows: The data is processed by subtracting the mean and dividing the variance, transforming it into data with a mean of 0 and a variance of 1; To perform a linear transformation, we can transform it to the range (-11) using the following formula:

[0128]

[0129] S505, The updated velocity image is output by generator G. The generated velocity model is then compared with the actual velocity model by the discriminator D. The degree of similarity.

[0130]

[0131]

[0132]

[0133]

[0134] Where n represents noise, implemented through a Dropout layer to prevent model overfitting; G * This represents the loss function of GAN.

[0135] In this embodiment of the application, when training the deep learning network, the input data is randomly cropped into a region of size 256*256, the initial learning rate is 0.0002, the Adam optimizer is used, and the network is trained for 200 epochs.

[0136] Pix2Pix, which uses the U-Net network structure, is a fully convolutional network model that can be directly applied to velocity model updates of any size. When using a pre-trained deep learning network, the first velocity model and the corresponding RTM seismic data are directly input into the network to obtain the second velocity model.

[0137] When evaluating the quality of a speed model output by a pre-trained deep learning network, the Pearson correlation coefficient (PCC) and mean absolute error (MAE) can be used.

[0138] PCC is used to measure the similarity between two variables. For two images X and Y with length H and width W, the PCC index can be calculated as follows:

[0139]

[0140] MAE is the average of absolute errors, calculated as follows:

[0141]

[0142] The embodiments of this application can be implemented based on PyTorch and can be deployed and run on a machine with an Intel(R) Core(TM) i9-10900X CPU@3.70GHz, 64GB RAM, and an RTX 2080Ti graphics card.

[0143] Figure 6 This application demonstrates a speed model for the iterative convergence of a certain FWI process. A schematic diagram. Where 'c' represents the velocity model number. Figure 6 (a) in the figure is the velocity model obtained after the 50th FWI iteration. Figure 6 (b) in the figure is the RTM migration seismic data obtained using (a). Figure 6 (c) in the middle is using and The resulting updated velocity model Figure 6 (d) in the model represents the actual velocity. Figure 6 (e) in the figure represents the velocity model. and The MAE and PCC metrics.

[0144] according to Figure 6 As can be seen, when the FWI iteration converges, the method provided in this application embodiment can input the velocity model and corresponding offset data obtained based on the FWI iteration into a pre-trained deep learning network to obtain a velocity model updated by the deep learning network. This velocity model has a higher approximation to the real velocity model and can improve the output results of FWI in subsequent FWI iterations.

[0145] Figure 7 This application describes a velocity model that does not converge in a certain FWI iteration. Schematic diagram. Where n represents the speed model number. Figure 7 (a) is the velocity model obtained after the 50th FWI iteration. Figure 7 (b) is the RTM imaging seismic data obtained using (a). Figure 7 (c) is the method that utilizes and Updated velocity model Figure 7 (d) is the true velocity model. Figure 7 (e) is the velocity model. and The MAE and PCC metrics.

[0146] according to Figure 7 As can be seen, when the FWI iteration does not converge, the method provided in this application embodiment can input the velocity model and corresponding offset data obtained based on the FWI iteration into a pre-trained deep learning network to obtain a velocity model updated by the deep learning network. This velocity model has a higher approximation to the real velocity model and can improve the output results of FWI in subsequent FWI iterations.

[0147] Figure 8 This is a schematic diagram showing the changes in PCC and MAE with the number of iteration steps for the velocity model provided in this embodiment. (Regarding the velocity model...) and Using a one-dimensional initial velocity model, perform 300 FWI iterations. Starting from the 10th iteration, adjust the obtained values ​​every 10 iterations. and The update is performed using the method provided in the embodiments of this application, resulting in... and The iteration rounds i = 10, 20, ... 290, 300.

[0148] Figure 8 (a) is and The curves showing the changes in PCC and MAE with the number of iterations i. Figure 8 (b) is and The curves showing the changes in PCC and MAE with iteration number i. From Figure 8 As can be seen from (a), when the FWI iteration converges, the method provided in this application embodiment can improve the speed model quality of the FWI output, and its use in subsequent FWI iterations can improve the FWI output results. Figure 8 As can be seen from (b), when the FWI iteration diverges, the method provided in this application embodiment can correct the divergent velocity model results, which is beneficial to the subsequent FWI iterations.

[0149] Figure 9 This is a schematic diagram of the local PCC and MAE corresponding to a certain FWI iterative convergence speed model provided in an embodiment of this application. Figure 9 In the middle, take a 5*5 sliding window and calculate the velocity model respectively. and The local PCC and MAE indices. Figure 9 (a) is the velocity model. Figure 9 (b) is Local PCC index, Figure 9 (c) is The local MAE index, Figure 9 (d) is the velocity model Figure 9 (e) is Local PCC index, Figure 9 (f) is The local MAE indicator. From Figure 9 It can be seen that the method provided in this application embodiment can effectively improve the local errors in the velocity model obtained by FWI iteration, and can effectively correct obvious inaccuracies in the deeper parts of the model.

[0150] Figure 10 This is a schematic diagram of the local PCC and MAE corresponding to a certain FWI iterative non-convergent velocity model provided in an embodiment of this application. Figure 10 In the middle, take a 5*5 sliding window and calculate the velocity model respectively. and The local PCC and MAE indices. Figure 10 (a) is the velocity model. Figure 10 (b) is Local PCC index, Figure 10 (c) is The local MAE index, Figure 10 (d) is the velocity model Figure 10(e) is Local PCC index, Figure 10 (f) is The local MAE indicator. From Figure 10 It can be seen that, in the case of FWI iterative divergence, the method provided in this application embodiment can effectively correct the large-area errors in the velocity model.

[0151] Figure 11 A comparison chart of the average PCC and MAE metrics for three deep network models provided in this application embodiment. The three deep network models are: Pix2Pix, U-Net, and M-RUDSR. Figure 11 To compare the average PCC and MAE metrics of the three deep network models on the test set provided in step S503, from... Figure 11 It can be seen that using the Pix2Pix model can achieve relatively better results.

[0152] Figure 12 A certain FWI iterative convergence speed model provided in the embodiments of this application One-dimensional initial velocity model The updated results are shown in the diagram. Figure 12(a) is a one-dimensional initial velocity model. Figure 12 (b) is using The obtained RTM seismic data imaging results, Figure 12 (c) is the true velocity model. Figure 12 (d) is the updated velocity model. Figure 12 Figure (e) compares the PCC and MAE indices before and after the update. As can be seen from Figure 12, the initial velocity model updated using the method provided in this embodiment shows significant improvements in both PCC and MAE, which is beneficial for subsequent FWI iterations.

[0153] Figure 13 This diagram illustrates the PCC and MAE corresponding to the updated initial velocity model of a velocity model for iterative convergence of a certain FWI provided in an embodiment of this application. Figure 12 The updated velocity model shown And the original one-dimensional velocity model Using each velocity as an initial velocity, perform FWI iterations and compare the PCC and MAE indices of the velocity model obtained at each step. Figure 13 As can be seen, compared with the results of FWI using the original one-dimensional velocity model, FWI using the updated initial velocity model provided by the method in this application embodiment can converge faster and achieve better results.

[0154] Figure 14 A velocity model for a certain FWI iteration that does not converge, provided in an embodiment of this application. One-dimensional initial velocity model The updated results are shown in the diagram. Figure 14 (a) is a one-dimensional initial velocity model. Figure 14 (b) is using The obtained RTM seismic data imaging results, Figure 14 (c) is the true velocity model. Figure 14 (d) is the updated velocity model. Figure 14 (e) is a comparison of the PCC and MAE indicators before and after the update. From Figure 14 It can be seen that the quality of the initial velocity model updated using the method provided in the embodiments of this application is improved, which is beneficial to the subsequent iterations of FWI.

[0155] Figure 15 This diagram illustrates the PCC and MAE corresponding to the updated initial velocity model of a non-convergent FWI velocity model provided in an embodiment of this application. Figure 14 The updated velocity model shown FWI was performed using the initial velocity. Figure 15 (a) shows the use of and The initial velocities were used for FWI iterations, and the PCC and MAE indices of the velocity model obtained at each step were compared. Figure 15 (b) shows the use of Using the initial velocity as the basis for FWI iteration, the velocity model obtained in step 10 is updated again using the method provided in the embodiments of this application, and then the PCC and MAE index results obtained at each step of the FWI iteration are continued. Figure 15 As can be seen from (a), when using In cases where FWI fails to converge, using the initial velocity updated by the method provided in this application embodiment for FWI can effectively correct the divergent iterative process. Figure 15 As can be seen in (b), the use Using the initial velocity model for FWI iteration, and updating the velocity model again using the method provided in this application embodiment after a certain number of steps, can further improve subsequent FWI iterations and adjust them to the convergence direction.

[0156] Figure 16 This is a schematic diagram showing the results of applying the embodiments of this application to the Marmousi model. Figure 16 (a) is Marmousi's initial velocity model V marm (x, z), Figure 16 (b) is using V marm RTM seismic data P obtained from (x, z) marm (x, z), Figure 16 (c) is the Marmousi real velocity model. Figure 16 (d) is the speed model obtained directly using network updates. Figure 16 (e) is The residual V of the guided filtering result res (x, z), Figure (d) shows the result obtained using the transfer algorithm. Figure 17 A schematic diagram showing the processing results of the guided filter migration algorithm applied to the Marmousi initial velocity model.

[0157] Combination Figure 16 and Figure 17It can be seen that the PCC and MAE metrics of the speed model obtained using the guided filter transfer algorithm are significantly improved compared to the speed model obtained by directly applying deep learning network updates. Furthermore, although the PCC and MAE metrics of the speed model obtained using the guided filter transfer algorithm are not significantly improved compared to the initial speed, according to... Figure 16 As can be seen from (f), the velocity model obtained by the guided filter migration algorithm has richer stratigraphic structure information, which is beneficial to subsequent FWI iterations.

[0158] Figure 18 This diagram illustrates the PCC and MAE corresponding to the FWI iteration performed after applying the Marmousi initial velocity model in this embodiment of the application. The main frequency used for FWI here is 10Hz. Figure 18 (a) shows the use of V respectively marm (x, z) and The curves showing the changes in PCC and MAE with the number of iterations during FWI iterations, using the initial velocity as the basis. Figure 18 (b) demonstrates the use of After five rounds of FWI iteration using the initial velocity as the starting point, a deep learning network and transfer algorithm are used for further updates. The results are then used to analyze the changes in PCC and MAE during subsequent FWI iterations. Figure 18 As can be seen in (a), when using the initial velocity model V marm When the FWI iteration of (x, z) fails to converge, the result updated using the network plus transfer algorithm is used. Performing FWI iteration can effectively alleviate the divergence process of FWI iteration, as can be seen from Figure 18(b). Using the initial velocity model for FWI iteration, and updating the velocity model again using the method provided in this application embodiment after a small number of iterations, can further improve subsequent FWI iterations.

[0159] Figure 19 This is a schematic diagram of a deep network-based full waveform inversion acceleration device 500 provided in an embodiment of this application. The device in this embodiment can be in software and / or hardware form. As shown in Figure 19, the deep network-based full waveform inversion acceleration device provided in this embodiment includes: an acquisition module 501, an input / output module 502, and an inversion module 503. Wherein:

[0160] The acquisition module 501 is used to acquire the first offset data of the first velocity model.

[0161] The input / output module 502 is used to input the first velocity model and the first offset data into a pre-trained deep learning network and output the second velocity model. The deep learning network is trained based on iterative data and offset data. The iterative data is generated when performing full waveform inversion (FWI) on the forward modeling data. The forward modeling data is obtained by performing forward modeling on the iterative data. The offset data is obtained by offsetting the forward modeling data based on the iterative data.

[0162] Inversion module 503 is used to perform FWI based on the second velocity model to obtain the target model.

[0163] Optionally, the acquisition module 501 is further used to acquire a training set, which includes M sample velocity models; perform forward modeling on the M sample velocity models respectively to obtain M pre-stack seismic data; the inversion module 503 is further used to perform full waveform inversion on the M pre-stack seismic data respectively to obtain M*N iterative velocity models; wherein, when performing full waveform inversion on any pre-stack seismic data, N iterative velocity models are obtained; for any iterative velocity model, depth migration is performed on the pre-stack seismic data corresponding to any iterative velocity model to obtain the migration data of any iterative velocity model; the input-output module 502 is further used to use the M*N iterative velocity models and the migration data of the M*N iterative velocity models as dual-channel inputs to the deep network to be trained, and train the deep network to be trained to obtain a deep learning network; wherein, M and N are both natural numbers.

[0164] Optionally, the inversion module 503 is also used to perform N iterations sequentially when performing FWI on any one of the M pre-stack seismic data; wherein, in the N iterations, when performing the i-th iteration, the model is updated by using the difference between any pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration, so as to obtain the iterative velocity model generated in the i-th iteration.

[0165] Optionally, the loss function of the deep network to be trained satisfies the following formula:

[0166]

[0167] Where G represents the deep learning network to be trained, and V i Represents the iterative speed model, P i V represents the offset data of the iterative velocity model. T Let represent the sample velocity model, j represent the j-th sample velocity model, x represent the surface spatial coordinates, and z represent the imaging depth.

[0168] Optionally, the input / output module 502 is further configured to determine whether the difference between the first velocity model and the training set data is greater than a difference threshold; and to input the first velocity model and the first offset data into a pre-trained deep learning network, including: when the difference between the first velocity model and the training set data is less than or equal to the difference threshold, inputting the first velocity model and the first offset data into the pre-trained deep learning network.

[0169] Optionally, the input / output module 502 is further configured to, when the difference between the first velocity model and the training set data is greater than the difference threshold, use the imaging result of the first velocity model as the guide image and the imaging result of the second velocity model as the input image, perform guided filtering to obtain the filtered velocity model; calculate the residual between the second velocity model and the filtered velocity model; superimpose the residual with the first velocity model to obtain the third velocity model; the inversion module 503 is further configured to perform FWI on the third velocity model to obtain the target model.

[0170] The deep network-based full waveform inversion acceleration device provided in this application embodiment can be used to execute the deep network-based full waveform inversion acceleration method provided in any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0171] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0172] Figure 20 This is a schematic diagram of an electronic device structure provided in an embodiment of the present invention. Figure 20 As shown, the electronic device 600 may include at least one processor 601 and a memory 602.

[0173] The memory 602 is used to store programs. Specifically, the program may include program code, which includes computer operation instructions.

[0174] The memory 602 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0175] The processor 601 executes computer execution instructions stored in the memory 602 to implement the device control method described in the foregoing method embodiments. The processor 601 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. Specifically, when implementing the device control method described in the foregoing method embodiments, the electronic device may be, for example, a terminal, a server, or other electronic device with processing capabilities.

[0176] Optionally, the electronic device 600 may also include a communication interface 603. In specific implementations, if the communication interface 603, memory 602, and processor 601 are implemented independently, they can be interconnected via a bus to complete communication. The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, or an extended industry standard architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not imply that there is only one bus or one type of bus.

[0177] Optionally, in a specific implementation, if the communication interface 603, memory 602, and processor 601 are integrated on a single chip, then the communication interface 603, memory 602, and processor 601 can communicate through an internal interface.

[0178] This application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps of the methods described above.

[0179] This application also provides a computer program product, including computer instructions that, when executed by a processor, implement the various steps in the methods described above.

[0180] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0181] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for accelerating full-waveform inversion based on deep networks, characterized in that, The method includes: Obtain the first offset data of the first velocity model; The first offset data is processed by subtracting the mean and dividing the variance to transform it into data with a mean of 0 and a variance of 1; the first velocity model is linearly transformed to the range (-1, 1). The transformed first velocity model and the transformed first offset data are input into a pre-trained deep learning network to output a second velocity model; the deep learning network is trained based on iterative data and offset data, the iterative data is generated when performing full waveform inversion (FWI) on forward modeling data, the forward modeling data is obtained by performing forward modeling on the iterative data, and the offset data is obtained by offsetting the forward modeling data based on the iterative data; Determine whether the difference between the first speed model and the training set data is greater than a difference threshold; When the difference between the first velocity model and the training set data is less than or equal to the difference threshold, FWI is performed based on the second velocity model to obtain the target model. When the difference between the first velocity model and the training set data is greater than the difference threshold, the imaging result of the first velocity model is used as the guide image, and the imaging result of the second velocity model is used as the input image. Guided filtering is performed to obtain the filtered velocity model. The residual between the second velocity model and the filtered velocity model is calculated. The residual is superimposed with the first velocity model to obtain the third velocity model. The third velocity model is subjected to FWI to obtain the target model. The training process of the deep learning network includes: acquiring a training set, which comprises M sample velocity models; performing forward modeling on each of the M sample velocity models to obtain M pre-stack seismic data; and performing full waveform inversion on each of the M pre-stack seismic data to obtain M... N iterative velocity models are obtained; wherein, when performing full waveform inversion on any of the pre-stack seismic data, N iterative velocity models are obtained; for any of the iterative velocity models, depth migration is performed on the pre-stack seismic data corresponding to the iterative velocity model to obtain the migration data of the iterative velocity model; the M... N iterative rate models and the M The offset data of N iterative velocity models are used as dual-channel inputs to train the deep network to obtain the deep learning network; wherein, M and N are both natural numbers, and during the training of the deep network to be trained, the L1 loss function is used to measure the difference between the output of the deep learning network and the sample velocity model.

2. The method according to claim 1, characterized in that, When performing FWI on any one of the M pre-stack seismic data, N iterations are performed sequentially. In the N iterations, during the i-th iteration, the model is updated by using the difference between any pre-stack seismic data and the forward model corresponding to the iterative velocity model generated in the (i-1)-th iteration, thus obtaining the iterative velocity model generated in the i-th iteration.

3. The method according to claim 2, characterized in that, The loss function of the deep network to be trained satisfies the following formula: in, This represents the deep learning network to be trained. This represents the iterative speed model. This represents the offset data of the iterative velocity model. Let j represent the sample velocity model, x represent the surface spatial coordinates, and z represent the imaging depth.

4. A full-waveform inversion acceleration device based on deep networks, characterized in that, include: The acquisition module is used to acquire the first offset data of the first velocity model; The transformation module is used to perform a subtraction of mean and division of variance on the first offset data, transforming it into data with a mean of 0 and a variance of 1; and to perform a linear transformation on the first velocity model, transforming it to the range (-1, 1). The input / output module is used to input the transformed first velocity model and the transformed first offset data into a pre-trained deep learning network and output a second velocity model; the deep learning network is trained based on iterative data and offset data, the iterative data is generated when performing full waveform inversion (FWI) on forward data, the forward data is obtained by performing forward modeling on the iterative data, and the offset data is obtained by offsetting the forward data based on the iterative data; The input / output module is further configured to determine whether the difference between the first velocity model and the training set data is greater than a difference threshold; when the difference between the first velocity model and the training set data is less than or equal to the difference threshold, the inversion module is configured to perform FWI based on the second velocity model to obtain the target model; The input / output module is further configured to, when the difference between the first velocity model and the training set data is greater than the difference threshold, use the imaging result of the first velocity model as a guide image and the imaging result of the second velocity model as an input image to perform guided filtering to obtain a filtered velocity model; calculate the residual between the second velocity model and the filtered velocity model; and superimpose the residual with the first velocity model to obtain a third velocity model; the inversion module is further configured to perform FWI on the third velocity model to obtain a target model; The acquisition module is also used to acquire a training set, which includes M sample velocity models; and to perform forward modeling on the M sample velocity models to obtain M pre-stack seismic data. The inversion module is also used to perform full waveform inversion on the M pre-stack seismic data respectively to obtain M N iterative velocity models are obtained; wherein, when performing full waveform inversion on any of the pre-stack seismic data, N iterative velocity models are obtained; for any of the iterative velocity models, depth migration is performed on the pre-stack seismic data corresponding to the iterative velocity model to obtain the migration data of the iterative velocity model. The input / output module is also used to convert the M N iterative rate models and the M The offset data of N iterative velocity models are used as dual-channel inputs to train the deep network to obtain the deep learning network; wherein, M and N are both natural numbers, and during the training of the deep network to be trained, the L1 loss function is used to measure the difference between the output of the deep learning network and the sample velocity model.

5. An electronic device, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the deep network-based full waveform inversion acceleration method as described in any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the deep network-based full waveform inversion acceleration method as described in any one of claims 1-3.

7. A computer program product comprising a computer program that, when executed by a processor, implements the full waveform inversion acceleration method based on deep networks as described in any one of claims 1-3.

Images