A Model-Data Dual-Driven Seismic Inversion Method and System for Fractured Attributes

By employing a model-data dual-driven seismic inversion method for fracture attributes, combined with semi-supervised learning and the HTI medium forward model, the problem of low accuracy in fracture attribute inversion in traditional methods is solved, and high-precision fracture attribute prediction is achieved in complex geological contexts.

CN120214890BActive Publication Date: 2026-11-17CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510335614.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-03-20
Publication Date
2026-11-17
Estimated Expiration
2045-03-20

AI Technical Summary

Technical Problem

Traditional methods struggle to accurately characterize the nonlinear relationship between seismic data and fracture properties in complex geological contexts. Existing deep learning methods heavily rely on labeled data and lack geophysical constraints, resulting in low accuracy in fracture property inversion.

Method used

A model-data dual-driven fracture attribute seismic inversion method is adopted. By designing a semi-supervised learning framework and introducing an initial model, a forward model of fracture weakness in HTI medium is constructed to achieve deep integration and synergistic driving of model and data. The inversion network is optimized by using pre-stack seismic data and initial model feature extraction, combined with consistency loss function.

Benefits of technology

It significantly improves the accuracy and computational efficiency of fracture attribute inversion, enhances the robustness and stability of the network, ensures that the inversion process follows geophysical laws, and improves the lateral resolution and reliability of the inversion results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120214890B_ABST
    Figure CN120214890B_ABST
Patent Text Reader

Abstract

This disclosure provides a model-data dual-driven fracture attribute seismic inversion method and system, relating to the field of seismic reservoir identification technology, including: acquiring pre-stack seismic data; inputting the pre-stack seismic data into a model-data dual-driven fracture attribute inversion network to perform fracture weakness parameter inversion and obtain predicted results of fracture weakness parameters; the model-data dual-driven fracture attribute inversion network uses pre-stack seismic data as input to the inversion network and introduces an initial model as inversion constraints; obtaining seismic features using the pre-stack seismic data; obtaining initial model features using the initial model, fusing the seismic features and the initial model features to obtain predicted values ​​of fracture parameters; inputting the predicted values ​​into a forward model for forward simulation to obtain synthetic azimuth seismic record differences, introducing the synthetic azimuth seismic record differences into a consistency loss function to update the parameters of the inversion network, and obtaining the final model-data dual-driven fracture attribute inversion network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of seismic reservoir identification technology, specifically to a model-data dual-driven method and system for seismic inversion of fracture attributes. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Fractures are a key structural element of oil and gas reservoirs, and accurate prediction of fracture properties is of great significance for oil and gas extraction.

[0004] In complex geological contexts, traditional methods predict fracture properties by constructing analytical functions, but these methods struggle to characterize the complex nonlinear relationship between seismic data and fracture properties, limiting prediction accuracy. The development of deep learning has brought new opportunities to the field of geophysical inversion, with deep neural networks, typically represented by Convolutional Neural Networks (CNNs), being widely applied to fracture parameter prediction.

[0005] In crack parameter prediction and crack attribute inversion, convolutional neural networks have shown strong feature extraction capabilities. However, this method is highly dependent on labeled data. The limited and insufficient labeled data seriously affects the prediction accuracy. In addition, all existing methods lack constraints from geophysical principles and cannot ensure that the inversion process follows geophysical laws during the prediction process, resulting in low accuracy of crack attribute inversion. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes a model-data dual-driven method and system for seismic inversion of fracture attributes. It proposes a model-data dual-driven network integrating semi-supervised learning and geophysical constraints. By designing a semi-supervised learning framework that includes a consistency loss function for unlabeled data, the problem of scarce labeled data is alleviated. Furthermore, an initial model is introduced into the inversion network to provide geological prior information for the inversion. A forward model of fracture weakness is constructed based on the HTI medium and embedded into the semi-supervised learning framework, achieving deep integration and synergistic driving of the model and data.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] A model-data dual-driven method for seismic inversion of fracture attributes includes:

[0009] Acquire pre-stack seismic data;

[0010] Prestack seismic data is input into a model-data dual-driven fracture attribute inversion network to perform fracture weakness parameter inversion and obtain the predicted results of fracture weakness parameters.

[0011] The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation, obtaining the difference between synthetic azimuth seismic records; introducing the difference between synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] A model-data dual-driven seismic inversion system for fracture attributes includes:

[0014] The data acquisition module is used to acquire pre-stack seismic data;

[0015] The prediction module is used to input pre-stack seismic data into the model-data dual-driven fracture attribute inversion network, perform fracture weakness parameter inversion, and obtain the prediction results of fracture weakness parameters.

[0016] The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation, obtaining the difference between synthetic azimuth seismic records; introducing the difference between synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network.

[0017] According to some embodiments, the present disclosure adopts the following technical solutions:

[0018] A computer program product includes a computer program that, when executed by a processor, implements the model-data dual-driven fracture attribute seismic inversion method.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned model-data dual-driven fracture attribute seismic inversion method.

[0021] According to some embodiments, the present disclosure adopts the following technical solutions:

[0022] An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the model-data dual-driven fracture attribute seismic inversion method.

[0023] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0024] This disclosed model-data dual-driven seismic inversion method for fracture attributes introduces a semi-supervised learning architecture based on data augmentation technology. This fully exploits unlabeled data information, expands the diversity of the training set, and thus enhances the robustness and stability of the network learning. Simultaneously, to ensure the inversion process follows geophysical laws, initial model constraints are integrated into the inversion network, allowing the inversion work to proceed from a physically sound starting point. Then, based on the horizontally isotropic (HTI) medium theory, a forward model that clearly describes the relationship between fracture attributes and azimuth seismic data is constructed. This model is implemented through the convolution operation of the azimuth PP wave reflection coefficient equation and the seismic wavelet, generating seismic response characteristics of fracture weakness parameters. This lays the theoretical foundation for subsequent inversion work and significantly improves inversion accuracy and computational efficiency.

[0025] This disclosed model-data dual-driven fracture attribute seismic inversion method, based on the principle of Physical Information Guided Network (PINN), organically integrates the constructed forward model and inversion network into a semi-supervised framework, creating a high-precision model-data dual-driven fracture attribute inversion network system. It achieves deep fusion and synergistic driving of model and data. The introduction of initial model constraints significantly improves the lateral resolution, making the characterization of lateral changes in fracture reservoirs more detailed, and providing reliable technical support for fracture weakness inversion under complex geological backgrounds. Attached Figure Description

[0026] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0027] Figure 1 This is a schematic diagram of the structure of the model-data dual-driven crack attribute inversion network in an embodiment of this disclosure;

[0028] Figure 2 This is a schematic diagram of the working process of the model-data dual-driven crack attribute inversion network in an embodiment of this disclosure;

[0029] Figure 3 This is a data-driven forward modeling network structure according to an embodiment of the present disclosure;

[0030] Figure 4 The true crack weakness of the three-dimensional inverse mask model in this embodiment of the disclosure; wherein, Figure 4 (a) in the text is Δ N , Figure 4 (b) in the equation represents Δ T ;

[0031] Figure 5 The normalized seismic data difference of the three-dimensional inverse mask model in this embodiment of the present disclosure;

[0032] in, Figure 5 In the diagram, (a) represents an incident angle of 20°. Figure 5 In (b), the incident angle is 10°.

[0033] Figure 6 The predicted crack weakness of different networks in the inverse masking model according to embodiments of this disclosure;

[0034] in, Figure 6 In the diagram, (a) represents the Δ predicted by Data-Driven-net. N , Figure 6 (b) in the figure represents the Δ predicted by Data-Driven-net. T ; Figure 6 In the figure, (c) represents the Δ predicted by Data-Driven-IM-net. N , Figure 6 In the figure, (d) represents the Δ predicted by Data-Driven-IM-net. T ; Figure 6 In the figure, (d) represents the Δ predicted by Model-Data-Driven-net. N , Figure 6 In this context, (f) represents the Δ predicted by Model-Data-Driven-Net. T .

[0035] Figure 7 The absolute error in predicting crack weakness using different networks in embodiments of this disclosure;

[0036] in, Figure 7 In the figure, (a) and (b) represent the Δ values ​​predicted by Data-Driven-net, respectively. N Δ T ; Figure 7In the figure, (c) and (d) represent the Δ values ​​predicted by Data-Driven-IM-net, respectively. N Δ T ; Figure 7 In the figure, (e) and (f) represent the Δ values ​​predicted by Model-Data-Driven-Net, respectively. N Δ T .

[0037] Figure 8 This involves comparing the network predictions with the actual values ​​of the test set in this embodiment of the disclosure.

[0038] Figure 9 This disclosure provides different networks for predicting crack weakness based on actual data.

[0039] in, Figure 9 In the figure, (a) and (b) represent the Δ values ​​predicted by Data-Driven-IM-net, respectively. N Δ T ; Figure 9 In the figure, (c) and (d) represent the Δ predicted by Model-Data-Driven-Net, respectively. N Δ T .

[0040] Figure 10 In the figure, (a) represents the difference in predicted azimuth seismic records of the Model-Data-Driven-net in this embodiment of the present disclosure;

[0041] Figure 10 In the figure, (a) represents the absolute error of the predicted azimuth earthquake in the Model-Data-Driven-net embodiment of this disclosure;

[0042] Figure 11 This is a comparison between the network prediction values ​​and logging data of the test wells in this embodiment of the disclosure. Detailed Implementation

[0043] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0045] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0046] Example 1

[0047] One embodiment of this disclosure provides a model-data dual-driven seismic inversion method for fracture attributes. It proposes a model-data dual-driven fracture attribute inversion network that integrates semi-supervised learning and geophysical constraints. First, a semi-supervised framework incorporating a consistency loss function for unlabeled data is designed to alleviate the problem of scarce labeled data. Second, an initial model is introduced into the inversion network to provide geological prior information for the inversion. Then, a fracture weakness forward model is constructed based on the HTI medium and embedded into the semi-supervised learning framework, achieving deep fusion and synergistic driving of the model and data. The implementation steps include:

[0048] Step 1: Acquire pre-stack seismic data;

[0049] Step 2: Input the pre-stack seismic data into the model-data dual-driven fracture attribute inversion network to perform fracture weakness parameter inversion and obtain the predicted results of fracture weakness parameters;

[0050] The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation, obtaining the difference between synthetic azimuth seismic records; introducing the difference between synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network.

[0051] As one embodiment, the construction steps of the model-data dual-driven fracture attribute inversion network in the model-data dual-driven fracture attribute seismic inversion method disclosed herein are as follows:

[0052] Step 1: Construct a consistency loss function for unlabeled data, and use a weighted combination strategy to organically integrate it with the supervised learning loss function, thereby successfully constructing a loss function suitable for semi-supervised learning mechanisms, thus building a semi-supervised learning framework.

[0053] Step 2: Based on the principles of convolutional neural networks and recurrent neural networks, an inversion network is constructed to predict fracture weakness from azimuth seismic data. An initial model is introduced as an inversion constraint through a loss function, thereby providing geological and low-frequency information for the inversion.

[0054] Step 3: Based on the PP wave reflection coefficient equation of HTI medium, derive the AVAZ approximate function under the assumption that the upper interface is an isotropic medium, and construct the forward model of crack weakness based on HTI medium using convolution theory.

[0055] Step 4: Based on the principle of physical information-guided neural networks, the completed AVAZ (Amplitude Variation with Azimuth) forward model is integrated into a semi-supervised framework to construct a model-data dual-driven fracture attribute inversion network that uses azimuth seismic records to predict fracture weakness, and a workflow for implementing specific applications is formed.

[0056] As one example, the specific process of constructing and training the model-data dual-driven fracture attribute inversion network in the model-data dual-driven fracture attribute seismic inversion method is as follows:

[0057] S1: A consistency loss function is constructed for unlabeled data, and a weighted combination strategy is used to organically integrate it with the supervised learning loss function, thus successfully constructing a loss function suitable for semi-supervised learning mechanisms;

[0058] Specifically, in supervised learning, the mean squared error (MSE) is often used as the loss function L. sup (·), its expression is:

[0059]

[0060] Where Θ represents the parameters of the inversion network, and N l To label the number of samples, x i (i = 1, 2, ..., N) l ) represents the labeled sample, y i The known sample labels, F Θ (x i ) is the network prediction value of the sample.

[0061] For unlabeled sample x j (j = 1, 2, ..., N) uThe network will provide a corresponding predicted value F. Θ (x j To extract information from unlabeled samples, this disclosure assumes a predicted value F. Θ (x j After being transformed by a certain function f, the transformed value f(F) is obtained. Θ (x j And define the consistency loss function:

[0062]

[0063] Based on the above formula, this disclosure constructs a loss function for a semi-supervised learning mechanism by weighting and combining the supervised learning loss function and the consistency loss function:

[0064] L semi (Θ,f)=αL sup (Θ)+βL cons (Θ,f) (3)

[0065] Among them, L semi (Θ) is the total loss function of the semi-supervised learning network, α,β∈[0,1] are the weight coefficients used to balance the weights of labeled and unlabeled data, and α+β=1.

[0066] S2: Based on the principles of convolutional neural networks and recurrent neural networks, an inversion network for predicting fracture weakness from azimuth seismic data is constructed, and an initial model is introduced as an inversion constraint through a loss function, thereby providing geological and low-frequency information for the inversion.

[0067] Specifically, based on the operational characteristics of convolutional neural networks and recurrent neural networks (RNNs), this disclosure constructs an inversion network consisting of two parts: a seismic feature extraction module and initial model constraints, as follows: Figure 1 As shown.

[0068] Furthermore, the inversion network consists of two parts: a seismic feature extraction module and an initial model constraint. Pre-stack seismic data and an initial model are input into the inversion network, with the initial model serving as the inversion constraint. The seismic feature extraction module includes a global feature extraction module and a local feature extraction module. The global feature extraction module is composed of a series of gated recurrent units (GRUs), while the local feature extraction module is composed of parallel one-dimensional convolutional Conv1d blocks with different dilation factors. These modules are used to extract global and local features from the pre-stack seismic data, respectively, and the global and local features are fused to obtain the seismic features.

[0069] Specifically, the inversion network uses pre-stack seismic data as input. First, the pre-stack seismic data is simultaneously input into two feature extraction modules: a global feature extraction module and a local feature extraction module. The global feature extraction module consists of a series of gated recurrent units (GRUs). The gating mechanism in GRUs can effectively handle the information flow within the sequence, avoiding the gradient vanishing or exploding problems common in traditional RNNs, thus enabling efficient and accurate extraction of global features from complex time-series data such as seismic data. The local feature extraction module consists of parallel one-dimensional convolutional (Conv1d) blocks with different dilation factors, used to capture local features at different scales. Each Conv1d block consists of a one-dimensional convolutional layer, a batch normalization layer (BN), and a hyperbolic tangent activation function (Tanh) in sequence. The extracted local features are integrated through a fully connected layer, and then output through three convolutional blocks. Finally, the output of the feature extraction is the seismic feature, which is a fusion of the local and global features of the pre-stack seismic data. However, during this process, the dimension of the input data changes due to the convolution operation. To ensure that the sampling rates of seismic and well logging data match, this disclosure employs deconvolution blocks to upsample high-dimensional features, with a structure identical to that of convolution blocks. The end of the inversion network is a regression module, consisting of a GRU and a fully connected layer, used to normalize the output data, achieving the transformation from feature representations to fracture parameter estimates.

[0070] Furthermore, an initial model is constructed based on well logging data, seismic attributes, and stratigraphic information, representing the initial distribution of fracture parameters. This initial model includes geological structural information and low-frequency information, which serve as important constraints in fracture parameter inversion. The initial model is processed by the input network of the initial model constraint part. In the inversion network, the input network of the initial model constraint part consists of three convolutional blocks. The first two convolutional blocks are used for feature extraction, and the last convolutional block is used for regression prediction of the parameter domain. The output yields the initial model features.

[0071] Furthermore, the seismic features and the initial model features are fused to output predicted values ​​of fracture weakness parameters. The loss function of this inversion network is:

[0072]

[0073] Where, Θ I1 and Θ I2 These represent the input network parameters for the seismic data and the initial model, respectively, m label and s label These represent the labeled samples and their corresponding labels, m0 is the initial model, and λ is the weighting coefficient. When the initial model does not match the actual fracture parameters due to complex underground structures, the weighting coefficient can be adjusted to prevent the network from overlearning erroneous information from the initial model.

[0074] S3: Based on the reflection coefficient equation of PP wave in HTI medium, the AVAZ approximate function is derived under the assumption that the upper interface is an isotropic medium, and the forward model of crack weakness based on HTI medium is constructed using convolution theory.

[0075] Specifically, in the forward model of fracture weakness in HTI medium, a PP wave reflection coefficient equation based on HTI medium is constructed. This includes: strata with vertical fractures developed in isotropic background strata are characterized as transversely isotropic media with a horizontal axis of symmetry, i.e., HTI medium. Based on the Born approximation and the steady-phase method, the PP wave reflection coefficient equation in HTI medium is expressed as the sum of two components: the azimuth-independent reflection coefficient of the isotropic background medium and the azimuth-dependent reflection coefficient caused by fractures. The specific implementation is as follows:

[0076] Based on the definition of fractured media by HSU et al., ignoring the specific shape of the fracture and assuming it to be an infinitely thin plane, the concept of fracture weakness is proposed, with the specific expression as follows:

[0077]

[0078] Among them, M b and μ b Z represents the P-wave and S-wave moduli of an isotropic background, respectively. N and Z T Δ represents the normal compliance and tangential compliance of the crack, respectively. N Δ represents the normal fracture weakness, used to indicate the fluid composition within the fracture. T This represents the tangential crack weakness, used to indicate crack density.

[0079] Strata with vertical fractures in isotropic background strata can be characterized as transversely isotropic media with a horizontal axis of symmetry, i.e., HTI media. Based on the Born approximation and the steady-phase method (2006), the reflection coefficient equation of PP waves in HTI media can be expressed as the sum of two components, namely, the position-independent reflection coefficient of the isotropic background medium. Reflection coefficient related to the orientation caused by the crack sum:

[0080]

[0081] in:

[0082]

[0083] Where θ represents the longitudinal wave incident angle, φ represents the azimuth angle, ρ represents the density, and ΔM b , Δμ bΔρ and Δρ represent the differences in longitudinal wave modulus and transverse wave density between the reflecting interface, respectively. and This represents the average value of the reflective interface. and This represents the difference in crack strength between the upper and lower layers of the interface. g = μ b / M b It represents the ratio of longitudinal wave modulus to transverse wave modulus.

[0084] Furthermore, in order to focus on the seismic response characteristics of crack weakness and thus eliminate the influence of PP wave reflection under an isotropic background medium, this disclosure, assuming the upper layer is a homogeneous isotropic medium, uses two azimuth angles φ1 and φ2 as examples to calculate the difference in reflection coefficients:

[0085]

[0086] In actual seismic data acquisition, seismic waves undergo multiple reflections and transmissions with the subsurface medium. This disclosure, based on a convolution model, can approximately simulate this forward modeling process, i.e., the seismic data s(θ,φ) is composed of reflection coefficients. The result is obtained by convolution with the seismic wavelet w, taking into account the influence of random noise e:

[0087]

[0088] Since seismic data response is more sensitive to elastic parameters, in order to ensure the stability and accuracy of fracture weakness inversion, this disclosure uses the method of subtracting seismic data from different azimuths, and the following formula can be derived based on equation (9):

[0089]

[0090] Where Δs(θ,φ1,φ2) represents the difference in earthquake amplitude corresponding to different orientations. * =[Δ N ,Δ T ] T F(·) represents the crack weakness to be inverted, and F(·) represents the forward simulation process of crack weakness.

[0091] S4: Based on the principle of physical information-guided neural networks, the constructed AVAZ forward model is integrated into a semi-supervised framework to build and complete a model-data dual-driven fracture attribute inversion network that uses azimuth seismic records to predict fracture weakness. This includes: introducing the HTI medium fracture weakness forward model into the inversion network; the residual between the difference between the generated synthetic azimuth seismic record and the actual azimuth seismic record is introduced into the consistency loss function to evaluate the consistency of the prediction results, and can also participate in the parameter update process of the inversion network to complete the training of the model-data dual-driven fracture attribute inversion network.

[0092] Specifically, to organically integrate geophysical constraints related to fracture parameters into the inversion process, this disclosure draws on the core principles of the Physics-informed Neural Network (PINN) proposed by Raissi. Based on a semi-supervised inversion framework, and with an initial model constraining the network, the transformation is defined as an AVAZ forward model, forming a workflow for a dual-driven inversion method that simultaneously incorporates a physical model and a data-driven model. Figure 2 As shown.

[0093] In this disclosure, during network training, due to the successful introduction of the forward model, the residual between the generated synthetic azimuth seismic record difference and the actual azimuth seismic record difference is incorporated into the consistency loss function as a crucial component. This not only reflects the consistency of the network's prediction results but also participates in the parameter update process of the inversion network, thereby improving the accuracy and reliability of the inversion.

[0094]

[0095] S5: Using the trained model-data dual-driven fracture attribute inversion network, fracture attribute seismic inversion is performed to obtain fracture weakness prediction results, including: acquiring pre-stack seismic data; inputting the pre-stack seismic data into the model-data dual-driven fracture attribute inversion network to perform fracture weakness parameter inversion and obtain the predicted results of fracture weakness parameters.

[0096] Simulation Experiment

[0097] To demonstrate the significant superiority of the proposed method, this disclosure employs comparative tests using networks: (1) Data-Driven Network, where the forward modeling network is as follows: Figure 3 As shown, there are no initial model constraints in the inversion network; (2) Data-Driven-IM-net with initial model constraints; (3) Model-Data-Driven-net.

[0098] This disclosure uses a 3D inverse masking model with an 801×801 watch set. To reduce computational complexity, 100×100 watch sets are extracted at equal intervals for subsequent processing. Figure 4 The model demonstrates realistic fracture weakness, with 187 sampling points per trace and a sampling interval of 2ms. The synthetic seismic data is generated by convolution of the reflection coefficient equations of the Ricker wavelet and PP wave at a dominant frequency of 20Hz, with incident angles of 10° and 20°, and azimuth angles of 45° and 60°. To simulate a realistic geological exploration environment, random noise with a signal-to-noise ratio (SNR) of 5 is injected into the synthetic data, such as... Figure 5 As shown. Subsequently, 0.2% of the data in the inverse masking model was randomly sampled as the training set. The network was trained using the Adam optimization algorithm with an initial learning rate of 0.001, weight coefficients α and β set to 1 and 0.1 respectively, and the number of training iterations was set to 500 rounds.

[0099] Based on the prediction results and residual analysis for different network architectures, see [link / reference]. Figure 6 and Figure 7 The results show that, compared to the network without initial model constraints, the other two networks have better-fitted predicted parameters to the true values ​​and significantly improved lateral resolution. To further verify the network performance, this disclosure selects random single channels on the test set for comparison, such as... Figure 8 As shown in the comparison results, in areas of data mutation, the prediction results of the pure data-driven network deviate significantly and cannot accurately capture the characteristics of rapid data changes, while the model-data dual-driven network can accurately characterize this change.

[0100] To quantitatively evaluate the network inversion performance, this disclosure introduces two indicators for analysis: mean squared error (MSE) and Pearson correlation coefficient (PCC), as shown in Table 1.

[0101] Table 1. Analysis of Mean Squared Error (MSE) and Pearson Correlation Coefficient (PCC)

[0102]

[0103] Table 1 shows that Model-Data-Driven-net has the smallest MSE value compared to other networks, indicating that the predicted value deviates less from the true value; at the same time, it has the largest PCC, which indicates that the predicted value has a stronger linear correlation with the true value, strongly proving that the present invention has higher accuracy and reliability in crack weakness prediction.

[0104] Furthermore, this disclosure also compares the training time of different networks. The training times of Data-Driven-net, Data-Driven-IM-net and Model-Data-Driven-net are 45, 44 and 28 minutes, respectively. Therefore, it is verified that this disclosure can significantly reduce the network training time cost and effectively improve the efficiency of crack weakness prediction.

[0105] To further verify the applicability of the proposed method, this disclosure applies a model-data dual-driven network to an oil and gas field in the Sichuan Basin. The target reservoir in the study area is mainly composed of sandstone, with widespread and well-continuous fractures. Seismic data with azimuth angles of 20° and 125° were selected as inputs and processed for noise reduction and amplitude preservation. This disclosure uses two wells as the training set and interpolates the logging data using stratigraphic information combined with filtering methods to generate an initial model. An additional well was selected for testing.

[0106] Figure 9 and Figure 10 The prediction results of the model-data dual-driven model, the difference between the synthetic azimuth seismic data generated by the forward modeling, and their errors are presented separately. The results show that the fracture weakness predicted by the network peaks in the time interval of 2.35 to 2.4 seconds, indicating a high degree of fracture development in this area. The fracture weakness predicted by the model-data dual-driven network shows a high degree of consistency with the actual well logging curves in terms of overall trend. Furthermore, the difference between the synthetic seismic data predicted by the forward modeling and the actual data matches well.

[0107] To further demonstrate the effectiveness of incorporating the forward model, this disclosure compares the inversion results of test wells using a data-driven network (without a forward model) and a model-data dual-driven network with actual datasets, such as... Figure 11 As shown, it can be observed that in the target reservoir region after 2.4 seconds, the data-driven initial model network failed to effectively invert the peak value of fracture weakness. In contrast, the model-data dual-driven network accurately captured this crucial information, further highlighting its advantages and reliability in inverting fracture weakness in complex reservoirs.

[0108] To verify the effectiveness of the method, this disclosure employs three different network structures—Data-Driven-net, Data-Driven-IM-net, and Model-Data-Driven-net—for comparative analysis on synthetic data. The results show that even with limited labeled data, the semi-supervised framework can still achieve ideal prediction results. The introduction of the initial model significantly improves the lateral resolution, enabling a more detailed characterization of lateral variations in fractured reservoirs. In the comparison of prediction performance metrics such as MSE, PCC, and training time, Model-Data-Driven-net exhibits lower MSE, higher PCC, and optimal training efficiency, highlighting the importance of imposing strong geophysical constraints on the network during the forward modeling process, which can effectively improve prediction accuracy and inversion efficiency. Furthermore, this disclosure applies the proposed method to real-world data. Model-Data-Driven-net not only achieves the highest prediction accuracy but also most closely approximates the differences between synthetic and real seismic data, significantly improving the reliability of the inversion results. Therefore, this demonstrates the improvement in both inversion accuracy and efficiency of this disclosure, providing reliable technical support for weak fracture inversion in complex geological backgrounds.

[0109] Example 2

[0110] One embodiment of this disclosure provides a model-data dual-driven fracture attribute seismic inversion system, comprising:

[0111] The data acquisition module is used to acquire pre-stack seismic data;

[0112] The prediction module is used to input pre-stack seismic data into the model-data dual-driven fracture attribute inversion network, perform fracture weakness parameter inversion, and obtain the prediction results of fracture weakness parameters.

[0113] The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation, obtaining the difference between synthetic azimuth seismic records; introducing the difference between synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network.

[0114] Example 3

[0115] One embodiment of this disclosure provides a computer program product, including a computer program that, when executed by a processor, implements the model-data dual-driven fracture attribute seismic inversion method.

[0116] Example 4

[0117] One embodiment of this disclosure provides a non-transitory computer-readable storage medium for storing computer instructions. When these computer instructions are executed by a processor, they implement the model-data dual-driven fracture attribute seismic inversion method.

[0118] Example 5

[0119] One embodiment of this disclosure provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the model-data dual-driven fracture attribute seismic inversion method.

[0120] This disclosure is made with reference to methods, apparatus (systems), and computer program products according to embodiments of this disclosure.

[0121] The flowcharts and / or block diagrams are used to describe the process. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A model-data dual-driven fracture attribute seismic inversion method, characterized in that, include: Acquire pre-stack seismic data; Prestack seismic data is input into a model-data dual-driven fracture attribute inversion network to perform fracture weakness parameter inversion and obtain the predicted results of fracture weakness parameters. The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation to obtain the difference in synthetic azimuth seismic records; introducing the difference in synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network. In the forward model of fracture weakness in HTI medium, the PP wave reflection coefficient equation based on HTI medium is constructed, including: in an isotropic background stratum with vertical fractures, the stratum is characterized as a transversely isotropic medium with a horizontal axis of symmetry, i.e., HTI medium. Based on the Born approximation and the steady phase method, the PP wave reflection coefficient equation in HTI medium is expressed as the sum of two components, namely the azimuth-independent reflection coefficient of the isotropic background medium and the azimuth-dependent reflection coefficient caused by the fracture.

2. The model-data dual-driven fracture attribute seismic inversion method as described in claim 1, characterized in that, An inversion network is constructed, which consists of two parts: a seismic feature extraction module and an initial model constraint. Pre-stack seismic data and an initial model are input into the inversion network, with the initial model serving as the inversion constraint. The seismic feature extraction module includes a global feature extraction module and a local feature extraction module. The global feature extraction module consists of a series of gated recurrent units (GRUs), while the local feature extraction module consists of parallel one-dimensional convolutional Conv1d blocks with different dilation factors. These modules are used to extract global and local features from pre-stack seismic data, respectively, and the global and local features are then fused to obtain the seismic features.

3. The model-data dual-driven fracture attribute seismic inversion method as described in claim 1, characterized in that, The initial model contains geological structure information and low-frequency information. In the fracture parameter inversion, the initial model is processed by the initial model constraint part. In the inversion network, the initial model constraint part consists of three convolutional blocks. The first two convolutional blocks are used for feature extraction, and the last convolutional block is used for regression prediction of the parameter domain. The output is the initial model features. The seismic features and the initial model features are fused to output the predicted value of fracture weakness parameters.

4. The model-data dual-driven fracture attribute seismic inversion method as described in claim 1, characterized in that, Focusing on the seismic response characteristics of fracture weakness, the influence of PP wave reflection under isotropic background medium is eliminated. Under the assumption that the upper layer is a homogeneous isotropic medium, the difference of reflection coefficients is calculated using two azimuth angles. The forward modeling process is approximated based on the convolution model, that is, the seismic data is obtained by convolving the reflection coefficients with the seismic wavelet. The influence of random noise is also considered. The forward modeling process of deriving fracture weakness is realized by subtracting the seismic data from different azimuth angles.

5. The model-data dual-driven fracture attribute seismic inversion method as described in claim 1, characterized in that, By introducing the HTI medium fracture weakness forward model into the inversion network, the residual between the difference between the generated synthetic azimuth seismic record and the actual azimuth seismic record is introduced into the consistency loss function to evaluate the consistency of the prediction results. At the same time, it can also participate in the parameter update process of the inversion network to complete the training of the model-data dual-driven fracture attribute inversion network.

6. A model-data dual-driven seismic inversion system for fracture attributes, characterized in that, include: The data acquisition module is used to acquire pre-stack seismic data; The prediction module is used to input pre-stack seismic data into the model-data dual-driven fracture attribute inversion network, perform fracture weakness parameter inversion, and obtain the prediction results of fracture weakness parameters. The training process of the model-data dual-driven fracture attribute inversion network includes: using pre-stack seismic data as input to the inversion network and introducing an initial model as an inversion constraint to construct a loss function for a semi-supervised learning mechanism; firstly, extracting features from the pre-stack seismic data to obtain seismic features; extracting features from the geological structural information in the initial model to obtain initial model features; fusing the seismic features and the initial model features to output predicted fracture weakness parameters; inputting the predicted fracture weakness parameters into a forward model based on HTI medium fracture weakness to perform forward simulation to obtain the difference in synthetic azimuth seismic records; introducing the difference in synthetic azimuth seismic records into the consistency loss function of the semi-supervised learning mechanism to update the parameters of the inversion network, thus obtaining the final model-data dual-driven fracture attribute inversion network. In the forward model of fracture weakness in HTI medium, the PP wave reflection coefficient equation based on HTI medium is constructed, including: in an isotropic background stratum with vertical fractures, the stratum is characterized as a transversely isotropic medium with a horizontal axis of symmetry, i.e., HTI medium. Based on the Born approximation and the steady phase method, the PP wave reflection coefficient equation in HTI medium is expressed as the sum of two components, namely the azimuth-independent reflection coefficient of the isotropic background medium and the azimuth-dependent reflection coefficient caused by the fracture.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the model-data dual-driven fracture attribute seismic inversion method as described in any one of claims 1-5.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the model-data dual-driven fracture attribute seismic inversion method as described in any one of claims 1-5.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the model-data dual-driven fracture attribute seismic inversion method as described in any one of claims 1-5.