Deep learning based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs

Abstract

The application discloses a method for pre-stack seismic velocity inversion of natural gas hydrate reservoir based on deep learning. According to the occurrence characteristics of natural gas hydrate and free gas under actual geological conditions, combined with the real seabed layered sedimentary area and the public ocean velocity model, a synthetic velocity model of thin layer associated structure containing high and low speed alternating hydrate and free gas is constructed; the model is forward by solving the wave equation to generate forward seismic records; the forward seismic records and the synthetic velocity model are registered, and the training, testing and verification sets are divided; the deep neural network is constructed, and the training environment and hyperparameters are configured; the perception loss is introduced, the mixed loss function is designed, the network training is multi-dimensionally constrained, the network is optimized and the optimal parameters are saved; the optimal parameters are called to batch inversion of seismic records, and high-precision velocity inversion and thin layer identification are realized. Through the above method, the inversion efficiency and reservoir characterization ability are improved, and effective technical means are provided for natural gas hydrate exploration.

Description

Technical Field

[0001] This invention relates to the field of natural gas hydrate exploration technology, and in particular to a deep learning-based method for pre-stack seismic velocity inversion of natural gas hydrate reservoirs. Background Technology

[0002] Natural gas hydrates are cage-like crystalline compounds formed from natural gas and water molecules. Because they can only exist stably under high pressure and low temperature conditions, they are mainly found in deep-sea sedimentary layers and polar permafrost regions, exhibiting complex geological conditions and high exploration difficulty. Under standard temperature and pressure conditions, 1 cubic meter of hydrate can release approximately 164 cubic meters of natural gas; meanwhile, its main component, methane, is a strong greenhouse gas, and its impact on global warming is far greater than that of carbon dioxide. Therefore, from both the perspective of marine resource exploration and sustainable development, the accurate identification and monitoring of natural gas hydrates has significant scientific and practical implications.

[0003] In natural gas hydrate reservoirs, the hydrate-bearing layer typically overlies the free gas layer, with a significant velocity difference between the two, forming a thin-layered structure with alternating high and low velocities. This complex structure makes the seismic wave propagation process more variable, easily leading to energy attenuation and interface wave interference, thus increasing the difficulty of reservoir boundary identification and weakening the ability to distinguish subtle stratigraphic structures and changes in physical properties.

[0004] In existing natural gas hydrate reservoir exploration technologies, grid-based tomographic imaging methods have been implemented to construct velocity models for natural gas hydrates; or pre-stack inversion techniques and rock physical analysis are used in conjunction with well logging data and seismic data to distinguish reservoirs. However, these methods generally suffer from problems such as long model construction cycles, high computational resource consumption, and limited inversion resolution, making it difficult to accurately characterize the thin-layer boundary between hydrate layers and free gas layers.

[0005] Furthermore, in complex seafloor geological environments, its sensitivity to initial conditions and insufficient noise resistance further limit its application in the fine-grained inversion and quantitative characterization of natural gas hydrate reservoirs. In natural gas hydrate exploration, seismic inversion is a core technology for constructing high-resolution velocity models. Among these, full waveform inversion (FWI), as a physics-based high-precision inversion method, is widely used for subsurface medium velocity reconstruction. This method leverages the high sensitivity of pre-stack seismic data to velocity changes, iteratively minimizing the difference between the synthetic and measured wavefields, thereby significantly improving the identification accuracy of hydrate-bearing strata.

[0006] However, FWI still suffers from problems such as large computational load, strong non-uniqueness of solutions, and high dependence on initial velocity models, making it difficult to achieve high-precision characterization and stable identification of natural gas hydrate reservoir boundaries in complex seabed environments.

[0007] To overcome these shortcomings, deep learning technology has been introduced into the field of seismic exploration in recent years. However, its inversion performance is still constrained by many factors, making it difficult to fully capture the sensitive features of thin-layer boundaries and deep semantic information, and thus failing to meet the requirements for high-precision inversion of natural gas hydrate and free gas reservoirs.

[0008] In summary, it is necessary to propose an inversion method that is computationally efficient, numerically stable, and can accurately characterize the boundaries of natural gas hydrate reservoirs, so as to improve the inversion accuracy and reservoir characterization capabilities. Summary of the Invention

[0009] The purpose of this invention is to provide a deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs. This method addresses the alternating high and low velocity characteristics of marine natural gas hydrates and free gas under actual geological conditions by proposing a synthetic velocity modeling method for thin-layer associated structures. This method is used to construct a velocity model that conforms to the actual seafloor sedimentary environment, providing a high-quality training sample matching foundation with real geological scenes for deep learning inversion. Based on the U-Net network structure, this invention constructs a deep neural network architecture suitable for pre-stack seismic velocity inversion. This architecture can deeply mine the complex wavefield response features in pre-stack seismic records, achieving high-precision reconstruction of the velocity structure of thin layers in natural gas hydrate and free gas reservoirs. Furthermore, this invention constructs a hybrid loss function composed of mean square error, structural similarity loss, and perception loss, realizing multi-dimensional joint constraints on velocity deviation, structural continuity, and semantic consistency. This loss function not only effectively suppresses network misjudgment of the seafloor background and improves the deep semantic consistency of velocity estimation, but also enhances the network's ability to identify thin-layer structures of hydrates and free gas, thereby significantly improving the accuracy and reliability of pre-stack seismic velocity inversion.

[0010] To achieve the above objectives, this invention provides a deep learning-based method for pre-stack seismic velocity inversion in natural gas hydrate reservoirs, comprising the following steps:

[0011] S1. Based on the occurrence characteristics of natural gas hydrates and free gas under actual geological conditions, and taking real seabed sedimentary areas and publicly available ocean velocity models as the geological background, a synthetic velocity model containing thin-layered associated structures of high-speed hydrates and low-speed free gas is constructed.

[0012] S2. Based on the synthetic velocity model established in S1, the propagation process of seismic waves in complex media is simulated by forward modeling the wave equation to generate corresponding forward modeling seismic records, which are used to construct training samples for neural networks.

[0013] S3. Register the synthetic velocity model established in S1 with the forward seismic records generated in S2, and divide the training, testing and validation sets according to the data source, data size and characteristic distribution.

[0014] S4. Based on U-Net as the backbone, construct the SC-UNeXt deep neural network, and configure the training environment and hyperparameters.

[0015] S5. Introduce perceptual loss to construct a hybrid loss function, constrain the network training process, and save the optimal network parameters according to the evaluation index.

[0016] S6. Call the optimal network parameters saved in S5, and perform batch inversion of the data in test mode to achieve high-precision inversion and identification of natural gas hydrate and free gas reservoirs.

[0017] The preferred S1 modeling method specifically includes:

[0018] (1) Stochastic geological reservoir modeling method: Based on the occurrence parameters of natural gas hydrates and free gas in the target sea area, including but not limited to reservoir depth, planar distribution range and layer thickness, a synthetic velocity model containing thin-layer associated structures of high-velocity hydrates and low-velocity free gas is constructed under the real seabed sedimentary background.

[0019] (2) Reservoir velocity modeling method based on public datasets: Using publicly available ocean velocity models and geological parameter data, a synthetic velocity model with potential occurrence characteristics is constructed under complex geological background, which includes thin-layer associated structures of high-speed hydrates and low-speed free gas.

[0020] Preferably, in S2, the wave equation is a second-order acoustic wave equation, as shown below:

[0021] ;

[0022] in, Represents spatial coordinates; Indicates the time of transmission; Functional representation of velocity model; Represents the sound wave field; The Ricker wavelet is specifically represented as:

[0023] ;

[0024] in, Main frequency.

[0025] Preferably, S4 specifically includes:

[0026] The specific structure of the S41, SC-UNeXt deep neural network is as follows:

[0027] (1) The ConvNeXt residual block is introduced into the encoder structure to replace the original two-dimensional convolutional downsampling module;

[0028] (2) The scSE attention mechanism is introduced into the skip connection structure. The scSE attention mechanism adaptively adjusts the weights of the wave field features extracted by the encoder in the spatial dimension and the channel dimension.

[0029] (3) A Pixel-Shuffle upsampling module is introduced into the decoder structure to replace the original two-dimensional deconvolution upsampling module. The Pixel-Shuffle upsampling module efficiently reconstructs and restores the velocity features, reducing the influence of the chessboard artifact.

[0030] S42. Configure network training hyperparameters, including but not limited to learning rate, batch size, regularization coefficient, and maximum number of iterations;

[0031] S43. Set the loss function change judgment condition. When the loss function does not show a downward trend in several consecutive iterations, the dynamic decay mechanism of the learning rate will be automatically triggered.

[0032] S44. Determine the convergence state of the network based on the changing trend of the multidimensional evaluation index. If the evaluation index does not show an optimization trend in several consecutive iterations, the early stopping mechanism is automatically triggered to prevent overfitting.

[0033] Preferably, S5 specifically includes:

[0034] S51. Based on the root mean square error (MSE) and the multi-scale structural similarity index (MSSIM), a hybrid loss function is constructed by introducing perceptual loss. By setting the weights of the root mean square error (MSE), the multi-scale structural similarity index (MSSIM), and the perceptual loss, multi-dimensional constraints are implemented on the velocity deviation of natural gas hydrate reservoirs, the continuity of multi-scale geological structures, and the semantic consistency of deep layers.

[0035] S52. Based on the hybrid loss function described in S51, an optimization algorithm is used to iteratively update the parameters of the deep neural network in order to achieve continuous optimization of model performance.

[0036] S53. Set a preset loss threshold and comprehensive evaluation index, monitor the model performance during network training, and save the network weight parameters with the best performance when the set conditions are met.

[0037] Preferably, the hybrid loss function is as follows:

[0038] ;

[0039] in, The weighting factor represents the relative contribution of the control perception loss; The pixel-structure fidelity loss of MSE and MSSIM is specifically expressed as follows:

[0040] ;

[0041] in, The mean squared error loss function is expressed as follows:

[0042] ;

[0043] in, Indicates the total number of samples; and These represent the actual velocity value and the velocity value obtained through inversion, respectively. express and The loss metric for multi-scale structural similarity between them is specifically expressed as:

[0044] ;

[0045] in, , )and These represent the velocity model at a scale of [scale value missing]. Differences in brightness, contrast, and structural components at that time; , , Representing the scale respectively The corresponding weights are assigned to the luminance, contrast, and structure components. A set representing scale; The perceptual loss function is expressed as follows:

[0046] ;

[0047] in, This represents the set of selected feature layers; , and Corresponding to the first The number of channels, height, and width of the layer feature map; This represents a weighting factor that controls the relative contribution of each layer to the total perceptual loss. This represents a feature extractor used to extract perceptual feature differences.

[0048] Preferably, in S53, the comprehensive evaluation index includes, but is not limited to, structural similarity index, peak signal-to-noise ratio, and relative velocity error;

[0049] Among them, the structural similarity index is used to measure the degree of similarity between the inverted model and the real model in terms of spatial structural features;

[0050] Peak signal-to-noise ratio (PSNR) is used to evaluate the similarity between the inverted model and the true model in terms of overall amplitude distribution.

[0051] Relative velocity error is used to quantify the velocity deviation of the inverted model relative to the true model.

[0052] Preferably, the monitoring of model performance during network training as described in S53 includes the following specific details:

[0053] (1) Let the first The evaluation index is in the first The value at the next iteration is... The historical best value is ;in, , This represents the set of selected comprehensive evaluation indicators;

[0054] (2) Define the criteria for determining the optimal value, as follows:

[0055] If the first The higher the value of each indicator, the better. At that time, the judgment Superior ;

[0056] If the first The smaller the value of each indicator, the better. At that time, the judgment Superior ;

[0057] (3) When any indicator meets the above-mentioned criteria for determining a better value, the historical best value is updated. And record the corresponding iteration number. ;

[0058] When continuous If no evaluation metric achieves a better value within the next iteration, the network performance is considered to have stabilized, and the early stopping mechanism is automatically triggered; the optimal network weight parameters are then saved; among which, continuous No evaluation metric achieved a better value in the next iteration, which is quantified as follows:

[0059] ;

[0060] in, Indicates the first The iteration number corresponding to the most recent better value for each evaluation metric. This is a preset threshold for the number of consecutive iterations.

[0061] Preferably, S6 specifically includes:

[0062] S61. Call the optimal network weights saved in S5 and set the SC-UNeXt deep neural network to test mode. The test mode is configured to freeze network parameters and turn off gradient backpropagation to ensure stability and computational efficiency during the inference phase.

[0063] S62. Using the validation set divided by S3 or the measured pre-stack seismic data of the target sea area as input, perform batch velocity inversion to obtain the P-wave velocity distribution of natural gas hydrates and free gas reservoirs, thereby achieving high-precision inversion and thin-layer structure identification of marine natural gas hydrates and free gas reservoirs.

[0064] Therefore, the present invention employs the aforementioned deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs, and the beneficial effects are as follows:

[0065] (1) This invention can utilize publicly available marine datasets and actual data to artificially embed hydrates and free atmospheres, quickly construct high-quality velocity model datasets, broaden the sources of model training data, and effectively improve the model's generalization ability and adaptability.

[0066] (2) The deep learning method of the present invention does not rely on a high-quality initial model. Through data-driven approach, it can reduce the local extremum problem caused by the inaccuracy of the initial model, effectively avoid inversion distortion and instability, and enhance the robustness of the inversion process.

[0067] (3) The multi-dimensional hybrid loss function described in this invention, in addition to focusing on the pixel-level fitting of the model, also ensures that the inversion results are not only more realistic in terms of visual appearance, but also in terms of geological structure by constraining deeper semantic consistency and structural similarity. This method overcomes the problem that a single loss function in the prior art cannot fully optimize the model, and improves the comprehensiveness and accuracy of the inversion results.

[0068] (4) The inversion method proposed in this invention can effectively overcome the problems of structure loss and oversmoothing under low signal-to-noise ratio conditions. Compared with traditional methods, this invention has stronger noise resistance and can accurately reconstruct the thin-layer structure of hydrates and free gas layers under complex seafloor sedimentation and low signal-to-noise ratio conditions, thereby enhancing the stability and robustness of the inversion results.

[0069] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0070] Figure 1 This is a flowchart of the pre-stack seismic velocity inversion method for natural gas hydrate reservoirs based on deep learning, as described in this invention.

[0071] Figure 2 This is a schematic diagram of the neural network architecture and loss function according to an embodiment of the present invention;

[0072] Figure 3These are partial synthetic velocity models in the synthetic dataset of the embodiments of the present invention. Among them, (a), (b), (c), (d), (e), and (f) are schematic diagrams of synthetic velocity models simulated in a layered sedimentary background using a stochastic geological reservoir modeling method based on the occurrence parameters of natural gas hydrates and free gas in the target sea area.

[0073] Figure 4 This is a specific implementation process of embedding hydrates and free gas layers into the public dataset 3DMO in the embodiments of the present invention. Among them, (a) is a schematic diagram of the 3DMO model obtained by interpolation; (b) is a schematic diagram of the 3DMO model with the range of the theoretical stable region of natural gas hydrate marked; and (c) is a schematic diagram of the workflow of slicing the 3DMO model using a fixed sliding window.

[0074] Figure 5 This is a partial synthetic velocity model of the public dataset 3DMO after embedding hydrates and free gas layers in an embodiment of the present invention. Among them, (a), (b), (c), (d), (e), and (f) are schematic diagrams of synthetic velocity models with potential reservoir characteristics constructed by using the 3DMO model and geological parameter data through the reservoir velocity modeling method based on the public dataset.

[0075] Figure 6 This is a structural diagram of the SC-UNeXt network architecture in an embodiment of the present invention; wherein, (a) is an overall structural diagram of SC-UNeXt; (b) is an architectural diagram of the ConvNeXt residual block (CRB) of the SC-UNeXt encoder structure; and (c) is an architectural diagram of the Pixel-Shuffle upsampling (PSUS) of the SC-UNeXt decoder structure.

[0076] Figure 7 It is the scSE attention module in the SC-UNeXt skip connection structure in this embodiment of the invention;

[0077] Figure 8 These are comparison diagrams of the inversion method of this invention and the results of FWI; where (a) and (b) are synthetic velocity models generated by the stochastic geological reservoir modeling method, serving as velocity labels; (c) and (d) are initial model diagrams used for full waveform inversion; (e) and (f) are velocity result diagrams obtained by inversion using the full waveform inversion method; and (g) and (h) are velocity result diagrams obtained by inversion using the method proposed in this invention.

[0078] Figure 9The images show a comparison and residual plot of the SC-UNeXt network proposed in this invention with representative networks in the semantic segmentation field. (a) and (b) are synthetic velocity models generated using a stochastic geological reservoir modeling method, serving as velocity labels. (c) and (d) are the velocity result plot, evaluation index, and residual plot generated using the DeepLabv3+ network, respectively. (e) and (f) are the velocity result plot, evaluation index, and residual plot generated using the ResUnet network, respectively. (g) and (h) are the velocity result plot, evaluation index, and residual plot generated using the SC-UNeXt network proposed in this invention, respectively.

[0079] Figure 10 The figures show comparison charts and residual plots of the inversion results using different loss function strategies in this invention; (a) and (b) are synthetic velocity models generated by the reservoir velocity modeling method of public datasets, serving as velocity labels; (c) and (d) are the velocity result plot, evaluation index, and residual plot respectively using loss function 1; (e) and (f) are the velocity result plot, evaluation index, and residual plot respectively using loss function 2; (g) and (h) are the velocity result plot, evaluation index, and residual plot respectively using loss function 3; (i) and (j) are the velocity result plot, evaluation index, and residual plot respectively using the loss function proposed in this invention.

[0080] Figure 11 The images show the inversion results of the proposed method under different levels of noise influence. Among them, (a) and (b) are synthetic velocity models generated by the random geological reservoir modeling method, which serve as velocity labels; (c) and (d) are the velocity result diagram and evaluation index diagram under noise-free conditions, respectively; (e) and (f) are the velocity result diagram and evaluation index diagram under 15 dB noise conditions, respectively; (g) and (h) are the velocity result diagram and evaluation index diagram under 10 dB noise conditions, respectively; and (i) and (j) are the velocity result diagram and evaluation index diagram under 5 dB noise conditions, respectively. Detailed Implementation

[0081] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0082] This invention presents a deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs, comprising the following steps:

[0083] S1. Based on the occurrence characteristics of natural gas hydrates and free gas under actual geological conditions, and taking real seabed sedimentary areas and publicly available ocean velocity models as the geological background, a synthetic velocity model containing thin-layered associated structures of high-speed hydrates and low-speed free gas is constructed.

[0084] S2. Based on the synthetic velocity model established in S1, the propagation process of seismic waves in complex media is simulated by forward modeling the wave equation to generate corresponding forward modeling seismic records, which are used to construct training samples for neural networks.

[0085] S3. Register the synthetic velocity model established in S1 with the forward seismic records generated in S2, and divide the training, testing and validation sets according to the data source, data size and characteristic distribution.

[0086] S4. Based on U-Net as the backbone, construct the SC-UNeXt deep neural network, and configure the training environment and hyperparameters.

[0087] S5. Introduce perceptual loss to construct a hybrid loss function, constrain the network training process, and save the optimal network parameters according to the evaluation index.

[0088] S6. Call the optimal network parameters saved in S5, and perform batch inversion of the data in test mode to achieve high-precision inversion and identification of natural gas hydrate and free gas reservoirs.

[0089] Example

[0090] like Figure 1 As shown, the present invention provides a deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs, comprising the following steps:

[0091] S1. Based on the occurrence characteristics of natural gas hydrates and free gas under actual geological conditions, and taking real seabed sedimentary areas and publicly available ocean velocity models as the geological background, a synthetic velocity model containing a thin layer of associated structures of high-speed hydrates and low-speed free gas is constructed.

[0092] This embodiment uses two modeling schemes to generate a synthetic velocity model, which are adapted to two scenarios: one with limited measured data in the target sea area and the other relying on a public dataset.

[0093] Option 1: Employ a stochastic geological reservoir modeling method to simulate the P-wave velocity structure of hydrate and free gas reservoirs in layered sedimentary layers. This method constructs a synthetic velocity model that includes a thin layer of associated high-velocity hydrates and low-velocity free gas, based on natural gas hydrate and free gas parameters consistent with the target sea area.

[0094] This synthetic velocity model contains 8 to 10 horizontal strata with smooth transitions between each stratum and is generated through Gaussian perturbation.

[0095] The velocity of the upper seawater layer was randomly selected within the range of 1500 to 1530 m / s, and the thickness was 1000 to 1300 meters, which is consistent with the measurement data of the target sea area.

[0096] The velocity of the sedimentary layer below the seawater layer gradually increases with depth, with the increment randomly selected between 100 and 120 m / s, and the total thickness is 400 to 500 meters.

[0097] The hydrate units have a fixed velocity of 2300 m / s, are located at a depth of approximately 200 to 500 meters below the seabed, have a thickness that varies between 60 and 200 meters, and their orientation is basically parallel to the seabed surface.

[0098] The free gas layer imparts a fixed velocity of 1600 m / s, located 20 to 60 meters below the hydrate reservoir, with a thickness randomly distributed within the range of 40 to 100 meters.

[0099] The synthetic velocity model established using the above method is discretized on a 200×200 point two-dimensional grid, with a spatial sampling interval of 20 meters. Figure 3 As shown.

[0100] Option 2: Employing a reservoir velocity modeling method based on public datasets, a synthetic velocity model with potential occurrence characteristics, encompassing thin-layered associated structures of high-velocity hydrates and low-velocity free gas, is constructed under complex geological conditions. Details are as follows:

[0101] The modeling was based on the 3D SEG / EAGE Marine Overthrust (3DMO) dataset, which approximates the wave velocity of North Sea sediments through velocity scaling. A three-dimensional P-wave velocity model containing a 200-meter-thick seawater layer was built based on this dataset.

[0102] Hydrate layer; within the theoretical gas hydrate stable zone (GHSZ) of the target area, multiple thin hydrate layer regions are randomly labeled from the velocity model in the above-mentioned synthetic dataset using an image annotation program, and the labeled hydrate layers are projected onto the three-dimensional P-wave velocity model in the 3DMO dataset, with the velocity set to 2300 m / s and the thickness randomly selected within the range of 100 to 250 meters.

[0103] Free gas thin layer: In multiple directions below the hydrate layer, multiple free gas thin layer regions are randomly labeled from the velocity model in the above synthetic dataset using an image annotation program, and the labeled free gas layers are projected onto the model in the 3DMO dataset. Specifically, 1 to 3 free gas layers are set 10 to 20 meters below the hydrate layer, with a velocity of 1600 m / s and a thickness of 50 to 100 meters.

[0104] Since the original 3DMO data size is 185×801×801 with a sampling interval of 25 meters, this method interpolates along the x and z axes to improve resolution, resulting in a model with a size of 200×800×801 and a sampling interval of 20 meters. Subsequently, slicing is performed along the y-axis to generate 801 two-dimensional profiles. Then, based on the vertical range of the z-axis, a 200×200 velocity model is extracted along the x-axis with a step size of 50. The overall workflow for preparing the above dataset is as follows: Figure 4 As shown.

[0105] The final synthetic velocity model was discretized on a 200×200 point two-dimensional grid with a spatial sampling interval of 20 meters to ensure consistency with the training parameters of the synthetic dataset. The discretization result is as follows: Figure 5 As shown.

[0106] The two velocity modeling methods proposed in this step can simulate relatively simple layered sedimentary thin-layer reservoirs; and can also combine publicly available marine datasets to artificially embed hydrate layers and free gas layers in areas where natural gas hydrates may be present, thereby constructing a task-oriented hydrate-free gas velocity model dataset, providing high-quality data support for training and evaluation. This provides high-quality data support for the subsequent training and evaluation of deep neural networks, solving the problems of limited training data sources and insufficient adaptability to actual geological scenarios in existing technologies, and laying a data foundation for high-precision reservoir inversion.

[0107] S2. Based on the synthetic velocity model established in S1, the propagation process of seismic waves in complex media is simulated through forward modeling of the wave equation, generating corresponding forward-modeled seismic records for constructing neural network training samples; wherein, the wave equation is a second-order acoustic wave equation, as shown below:

[0108] ;

[0109] in, Represents spatial coordinates; Indicates the time of transmission; Functional representation of velocity model; Represents the sound wave field; The Ricker wavelet is specifically represented as:

[0110] ;

[0111] in, Main frequency.

[0112] In this embodiment, the finite difference method (FDM) is used for forward modeling: during each excitation process, all surface points are set as receivers to achieve full coverage acquisition of seismic signals. The total number of receivers is 200, and the maximum offset is 4 kilometers. To effectively suppress boundary reflections, a perfectly matched layer (PML) is used as the absorbing boundary. The specific parameter settings for the forward modeling are shown in Table 1.

[0113] Table 1 Parameter settings for forward simulation

[0114]

[0115] S3. Register the synthetic velocity model established in S1 with the forward seismic records generated in S2, and divide the training, testing and validation sets according to the data source, data size and characteristic distribution.

[0116] In this embodiment, the forward modeling seismic records generated by S2 are registered and matched with the synthetic velocity model established by S1, and the training set, test set and validation set are set in a ratio of 7:2:1.

[0117] S4. Based on U-Net as the backbone, construct the SC-UNeXt deep neural network, and configure the training environment and hyperparameters.

[0118] S41, SC-UNeXt deep neural networks, such as Figure 6 As shown in (a), the details are as follows:

[0119] (1) The encoder structure introduces ConvNeXt residual blocks to replace the original two-dimensional convolutional downsampling module, such as Figure 6 As shown in (b).

[0120] (2) Introduce scSE attention mechanism into the skip connection structure, such as Figure 7 As shown, the scSE attention mechanism can adaptively adjust the weights of the wavefield features extracted by the encoder in both spatial and channel dimensions, thereby improving the fusion effect of key geological features, enhancing feature representation capabilities, and further providing accurate feature support for the decoder to reconstruct the synthesis speed model.

[0121] (3) A pixel-shuffle upsampling module is introduced into the decoder structure to replace the original two-dimensional deconvolution upsampling module, such as Figure 6 As shown in (c), the Pixel-Shuffle upsampling module efficiently reconstructs and restores velocity features, effectively reducing the impact of checkerboard artifacts.

[0122] S42. Configure network training hyperparameters, including learning rate, batch size, regularization coefficient, and maximum number of iterations.

[0123] S43. Set the loss function change judgment condition. When the loss function does not show a downward trend in several consecutive iterations, the dynamic decay mechanism of the learning rate will be automatically triggered.

[0124] S44. Determine the convergence state of the network based on the changing trend of the multidimensional evaluation index. If the evaluation index does not show an optimization trend in several consecutive iterations, the early stopping mechanism is automatically triggered to prevent overfitting.

[0125] Compared with traditional physics-driven inversion methods, the SC-UNeXt deep neural network constructed in this step improves efficiency while reducing computational overhead and can handle multiple models, greatly improving the feasibility of engineering applications. Compared with data-driven inversion methods, this invention improves the feature extraction and reconstruction capabilities of the network through structural optimizations such as improved convolutional neural networks (such as CRB, PSUS and scSE attention mechanisms), significantly reducing computational load and training time.

[0126] like Figure 2 As shown, the synergistic effect of seismic record input, SC-UNeXt network, and hybrid loss function on the model training process is clearly presented, providing visual support for subsequent parameter optimization.

[0127] S5. Introduce perceptual loss to construct a hybrid loss function, constrain the network training process, and save the optimal network parameters according to the evaluation index.

[0128] S51. Based on the root mean square error (MSE) and the multi-scale structural similarity index (MSSIM), a perceptual loss is introduced to construct a hybrid loss function. By setting the weights of each loss term, multi-dimensional constraints are applied to the velocity deviation of natural gas hydrate reservoirs, the continuity of multi-scale geological structures, and the semantic consistency of deep layers. The hybrid loss function is shown below:

[0129] ;

[0130] in, It is a weighting factor that controls the relative contribution of perceptual loss; The pixel-structure fidelity loss of MSE and MSSIM is specifically expressed as follows:

[0131] ;

[0132] in, The mean squared error loss function is expressed as follows:

[0133] ;

[0134] in, Indicates the total number of samples; and These represent the actual velocity value and the velocity value obtained through inversion, respectively. express and The loss metric for multi-scale structural similarity between them is specifically expressed as:

[0135] ;

[0136] in, , )and These represent the velocity model at a scale of [scale value missing]. Differences in brightness, contrast, and structural components at that time; , , Representing the scale respectively The corresponding weights are assigned to the luminance, contrast, and structure components. A set representing scale; The perceptual loss function is expressed as follows:

[0137] ;

[0138] in, This represents the set of selected feature layers; , and Corresponding to the first The number of channels, height, and width of the layer feature map; This represents a weighting factor that controls the relative contribution of each layer to the total perceptual loss. This represents a feature extractor used to extract perceptual feature differences.

[0139] S52. Based on the hybrid loss function described in S51, an optimization algorithm is used to iteratively update the parameters of the deep neural network to achieve continuous optimization of model performance.

[0140] S53. Set a preset loss threshold and comprehensive evaluation index, monitor the model performance during network training, and save the optimal network parameters when the set conditions are met.

[0141] In this embodiment, in addition to directly comparing MSE, various comprehensive evaluation metrics are used to assess the quality of the inverted velocity model, such as the Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Relative Velocity Error (RVE). The Structural Similarity Index (SSIM) measures the degree of similarity between the inverted model and the real model in spatial structural features. The calculation formula is as follows:

[0142] ;

[0143] in, and These represent local means; and Indicates the corresponding variance; express and The covariance between the two; and This represents a constant, used to prevent numerical instability when the denominator is small.

[0144] Peak signal-to-noise ratio (PSNR) is used to evaluate the similarity between the inverted model and the true model in terms of overall amplitude distribution. The calculation formula is as follows:

[0145] ;

[0146] in, and These represent the actual velocity value and the velocity value obtained through inversion, respectively. This represents the maximum speed value among the actual speed values; This represents the total number of grid points.

[0147] Relative velocity error (RVE) is used to quantify the velocity deviation of the inverted model relative to the true model. The calculation formula is shown below:

[0148] .

[0149] SSIM ranges from 0 to 1, with values ​​closer to 1 indicating higher structural similarity between the two models; PSNR typically ranges from negative to positive infinity, with higher values ​​indicating higher similarity between the reconstructed model and the true value; RVE is used to quantify the velocity error between the inverted model and the true model, and it is calculated as the relative difference between the model's predicted velocity and the true velocity. The smaller the value, the smaller the velocity error and the higher the inversion accuracy.

[0150] The hybrid loss function designed in this step specifically addresses the shortcomings of single / few index constraints in existing deep learning inversion methods. Existing methods tend to focus on pixel-level fitting (such as using only MSE or SSIM), neglecting deep semantic consistency and multi-scale structural detail constraints, resulting in insufficient boundary and detail reconstruction capabilities.

[0151] Monitoring model performance during network training includes the following:

[0152] (1) Let the first The evaluation index is in the first The value at the next iteration is... The historical best value is ;in, , This represents the set of selected comprehensive evaluation indicators;

[0153] (2) Define the criteria for determining the optimal value, as follows:

[0154] If the first The higher the value of each indicator, the better. At that time, the judgment Superior ;

[0155] If the first The smaller the value of each indicator, the better. At that time, the judgment Superior ;

[0156] (3) When any indicator meets the above-mentioned criteria for determining a better value, its historical best value is updated. And record the corresponding iteration number. ;

[0157] When continuous If no evaluation metric achieves a better value within the next iteration, the network performance is considered to have stabilized, and the early stopping mechanism is automatically triggered; the optimal network weight parameters are then saved; among which, continuous No evaluation metric achieved a better value in the next iteration, which is quantified as follows:

[0158] ;

[0159] in, Indicates the first The iteration number corresponding to the most recent better value for each evaluation metric. This is a preset threshold for the number of consecutive iterations.

[0160] S6. Call the optimal network parameters saved in S5, and perform batch inversion of the data in test mode to achieve high-precision inversion and identification of natural gas hydrate and free gas reservoirs.

[0161] S61. Call the optimal network parameters saved in S5 and set the deep neural network to test mode. The test mode is configured to freeze network parameters and turn off gradient backpropagation to ensure stability and computational efficiency during the inference phase.

[0162] S62. Using the validation set divided by S3 or the measured pre-stack seismic data of the target sea area as input, perform batch velocity inversion to obtain the P-wave velocity distribution of natural gas hydrates and free gas reservoirs, thereby achieving high-precision inversion and thin-layer structure identification of marine natural gas hydrates and free gas reservoirs.

[0163] In this embodiment, to ensure the efficiency and stability of training the deep neural network SC-UNeXt, the hyperparameters for neural network training are shown in Table 2:

[0164] Table 2 Hyperparameters for Neural Network Training

[0165]

[0166] The following experimental results verify the advanced nature and reliability of the method of the present invention.

[0167] 1. Comparison and verification with traditional full-wave inversion FWI.

[0168] 1) Comparison of initial model dependencies, such as Figure 8 As shown, the full waveform inversion FWI relies on an initial model for inversion calculations and obtains the final velocity model through iterative optimization; however, the training of the deep neural network in this invention does not rely on an initial model and directly performs velocity model inversion using only the input seismic records.

[0169] 2) Accuracy of inversion results: By comparison, it can be found that the velocity model obtained by using FWI is unclear, especially in the reconstruction of the boundaries of the hydrate layer and the free gas layer; while the inversion by SC-UNeXt can not only accurately characterize the location and boundary of the hydrate layer and the free gas layer, but also reconstruct the marine layered sedimentary background with high fidelity, showing a clearer and more accurate inversion effect.

[0170] 2. Comparative verification with networks widely used in the field of semantic segmentation.

[0171] Using the synthetic velocity model constructed in S1 and the forward-modeled synthetic seismic records generated in S2 as unified comparison data, the inversion results and residuals of the synthetic data of the network widely used in the semantic segmentation field and the SC-UNeXt proposed in this invention are compared; the comparison results are as follows: Figure 9 As shown, existing methods still have shortcomings in terms of boundary clarity, structural continuity, and background layer reconstruction, making it difficult to meet the high-precision inversion requirements of hydrate-free gas thin layers.

[0172] 3. Results of ablation experiments using hybrid loss functions.

[0173] Ablation experiments were conducted on the hybrid loss function proposed in this invention. The specific loss function strategies used in the experiments are shown in Table 3.

[0174] Table 3 Loss Function Strategy

[0175]

[0176] Among them, MSE, MSSIM and These represent mean squared error, multi-scale structural similarity, and perceptual loss, respectively. The specific results of this ablation experiment are as follows: Figure 10 As shown.

[0177] 4. Input the noise interference level of the seismic record.

[0178] The experiment constructs inversion scenarios with varying noise interference intensity by adding different levels of Gaussian noise to the input seismic records; the specific verification process and results of this experiment are as follows: Figure 11 As shown, the results demonstrate the robustness of the SC-UNeXt proposed in this invention, enabling accurate inversion of hydrate and free gas reservoirs even under low signal-to-noise ratio conditions.

[0179] Therefore, this invention employs the aforementioned deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs, effectively addressing the common problems of low computational efficiency, high ambiguity, and high sensitivity to the initial model in traditional inversion methods. By introducing a high-performance inversion network and a hybrid loss function, the thin-layer structure of hydrate layers and free gas layers can be accurately identified and precisely located in complex seafloor environments. This framework not only improves inversion accuracy but also significantly enhances the robustness and generalization of the inversion, enabling efficient exploration and resource assessment of marine natural gas hydrate reservoirs.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based method for pre-stack seismic velocity inversion in natural gas hydrate reservoirs, characterized in that, Includes the following steps: S1. Based on the occurrence characteristics of natural gas hydrates and free gas under actual geological conditions, and taking real seabed sedimentary areas and publicly available ocean velocity models as the geological background, a synthetic velocity model containing thin-layered associated structures of high-speed hydrates and low-speed free gas is constructed. S2. Based on the synthetic velocity model established in S1, the propagation process of seismic waves in complex media is simulated by forward modeling the wave equation to generate corresponding forward modeling seismic records, which are used to construct training samples for neural networks. S3. Register the synthetic velocity model established in S1 with the forward seismic records generated in S2, and divide the training, testing and validation sets according to the data source, data size and characteristic distribution. S4. Based on U-Net as the backbone, construct the SC-UNeXt deep neural network, and configure the training environment and hyperparameters. S5. Introduce perceptual loss to construct a hybrid loss function, constrain the network training process, and save the optimal network parameters according to the evaluation index. S6. Call the optimal network parameters saved in S5, and perform batch inversion of the data in test mode to achieve high-precision inversion and identification of natural gas hydrate and free gas reservoirs.

2. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 1, characterized in that, S1 modeling method, specifically including: (1) Stochastic geological reservoir modeling method: Based on the occurrence parameters of natural gas hydrates and free gas in the target sea area, including but not limited to reservoir depth, planar distribution range and layer thickness, a synthetic velocity model containing thin-layer associated structures of high-velocity hydrates and low-velocity free gas is constructed under the real seabed sedimentary background. (2) Reservoir velocity modeling method based on public datasets: Using publicly available ocean velocity models and geological parameter data, a synthetic velocity model with potential occurrence characteristics is constructed under complex geological background, which includes thin-layer associated structures of high-speed hydrates and low-speed free gas.

3. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 1, characterized in that, In S2, the wave equation is a second-order acoustic wave equation, as shown below: ； in, Represents spatial coordinates; Indicates the time of transmission; Functional representation of velocity model; Represents the sound wave field; The Ricker wavelet is specifically represented as: ； in, Main frequency.

4. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 1, characterized in that, S4 specifically includes: The specific structure of the S41, SC-UNeXt deep neural network is as follows: (1) The ConvNeXt residual block is introduced into the encoder structure to replace the original two-dimensional convolutional downsampling module; (2) The scSE attention mechanism is introduced into the skip connection structure. The scSE attention mechanism adaptively adjusts the weights of the wave field features extracted by the encoder in the spatial dimension and the channel dimension. (3) A Pixel-Shuffle upsampling module is introduced into the decoder structure to replace the original two-dimensional deconvolution upsampling module. The Pixel-Shuffle upsampling module efficiently reconstructs and restores the velocity features, reducing the influence of the chessboard artifact. S42. Configure network training hyperparameters, including but not limited to learning rate, batch size, regularization coefficient, and maximum number of iterations; S43. Set the loss function change judgment condition. When the loss function does not show a downward trend in several consecutive iterations, the dynamic decay mechanism of the learning rate will be automatically triggered. S44. Determine the convergence state of the network based on the changing trend of the multidimensional evaluation index. If the evaluation index does not show an optimization trend in several consecutive iterations, the early stopping mechanism is automatically triggered to prevent overfitting.

5. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 1, characterized in that, S5 specifically includes: S51. Based on the root mean square error (MSE) and the multi-scale structural similarity index (MSSIM), a hybrid loss function is constructed by introducing perceptual loss. By setting the weights of the root mean square error (MSE), the multi-scale structural similarity index (MSSIM), and the perceptual loss, multi-dimensional constraints are implemented on the velocity deviation of natural gas hydrate reservoirs, the continuity of multi-scale geological structures, and the semantic consistency of deep layers. S52. Based on the hybrid loss function described in S51, an optimization algorithm is used to iteratively update the parameters of the deep neural network in order to achieve continuous optimization of model performance. S53. Set a preset loss threshold and comprehensive evaluation index, monitor model performance during network training, and save the network weight parameters with the best performance when the set conditions are met.

6. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 5, characterized in that, The hybrid loss function is as follows: ； in, The weighting factor represents the relative contribution of the control perception loss; The pixel-structure fidelity loss of MSE and MSSIM is specifically expressed as follows: ； in, The mean squared error loss function is expressed as follows: ； in, Indicates the total number of samples; and These represent the actual velocity value and the velocity value obtained through inversion, respectively. express and The loss metric for multi-scale structural similarity between the two structures is specifically expressed as: ； in, , )and These represent the velocity model at a scale of [scale value missing]. Differences in brightness, contrast, and structural components at that time; , , Representing the scale respectively The corresponding weights are assigned to the luminance, contrast, and structure components. A set representing scale; The perceptual loss function is expressed as follows: ； in, This represents the set of selected feature layers; , and They represent the first The number of channels, height, and width of the layer feature map; This represents a weighting factor that controls the relative contribution of each layer to the total perceptual loss. This represents a feature extractor used to extract perceptual feature differences.

7. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 5, characterized in that, In S53, the comprehensive evaluation indicators include, but are not limited to, structural similarity index, peak signal-to-noise ratio, and relative velocity error; Among them, the structural similarity index is used to measure the degree of similarity between the inverted model and the real model in terms of spatial structural features; Peak signal-to-noise ratio (PSNR) is used to evaluate the similarity between the inverted model and the true model in terms of overall amplitude distribution. Relative velocity error is used to quantify the velocity deviation of the inverted model relative to the true model.

8. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 5, characterized in that, S53 describes the monitoring of model performance during network training, specifically as follows: (1) Let the first The evaluation index is in the first The value at the next iteration is... The historical best value is ;in, , This represents the set of selected comprehensive evaluation indicators; (2) Define the criteria for determining the optimal value, as follows: If the first The higher the value of each indicator, the better. At that time, the judgment Superior ; If the first The smaller the indicator, the better. At that time, the judgment Superior ; (3) When any indicator meets the above-mentioned criteria for determining a better value, the historical best value is updated. And record the corresponding iteration number. ; When continuous If no evaluation metric achieves a better value within the next iteration, the network performance is considered to have stabilized, and the early stopping mechanism is automatically triggered; the optimal network weight parameters are then saved; among which, continuous No evaluation metric achieved a better value in the next iteration, which is quantified as follows: ； in, Indicates the first The iteration number corresponding to the most recent better value for each evaluation metric. This is a preset threshold for the number of consecutive iterations.

9. The deep learning-based pre-stack seismic velocity inversion method for natural gas hydrate reservoirs according to claim 1, characterized in that, S6 specifically includes; S61. Call the optimal network weights saved in S5 and set the SC-UNeXt deep neural network to test mode. The test mode is configured to freeze network parameters and turn off gradient backpropagation to ensure stability and computational efficiency during the inference phase. S62. Using the validation set divided by S3 or the measured pre-stack seismic data of the target sea area as input, perform batch velocity inversion to obtain the P-wave velocity distribution of natural gas hydrates and free gas reservoirs, thereby achieving high-precision inversion and thin-layer structure identification of marine natural gas hydrates and free gas reservoirs.

Images