Speed and anisotropy parameter simultaneous modeling method based on multi-expert learning network
By learning the contribution of dynamic equilibrium parameters through multi-expert learning networks, the problem of insufficient accuracy in velocity and anisotropic parameter modeling is solved, realizing high-precision and automated multi-parameter modeling, which is suitable for oil and gas reservoir exploration and drilling location under complex geological conditions.
Patent Information
- Application Number
- CN202511796332.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-03
AI Technical Summary
Existing technologies struggle to achieve high-precision collaborative modeling of velocity and anisotropic parameters. Traditional methods ignore the intrinsic correlation between multiple parameters and rely on a large amount of well logging data, resulting in high modeling errors and difficulty in adapting to complex geological conditions.
A multi-expert learning network, including a parameter sharing network, a weight network, and a parameter-specific network, is adopted. By dynamically balancing the parameter contributions, the multi-expert learning network is constructed to achieve synchronous modeling of speed and anisotropic parameters.
It achieves high-precision synchronous modeling of velocity and anisotropic parameters, reduces manpower and time costs, adapts to scenarios where logging data is scarce, improves modeling accuracy and automation, and is suitable for complex geological conditions.
Smart Images
Figure CN121598786A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of earthquake multi-parameter modeling technology, and in particular relates to a method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network. Background Technology
[0002] In the field of seismic multi-parameter modeling technology, accurate modeling of velocity and anisotropy parameters is the core foundation for realizing the exploration, prediction, and drilling location of concealed and complex oil and gas reservoirs. The accuracy of this modeling directly affects the reliability of subsequent seismic inversion and geological interpretation. Among current mainstream modeling methods, traditional single-parameter modeling techniques struggle to capture high-order implicit correlations between different parameters, resulting in limited accuracy in multi-parameter collaborative modeling. While existing machine learning-driven modeling schemes attempt to achieve simultaneous multi-parameter modeling, the former ignores the intrinsic correlations between multiple parameters by treating each parameter as an independent variable, while the latter lacks a dynamic weight allocation mechanism to balance the contributions of different parameters. Both suffer from high modeling errors and insufficient exploration of multi-parameter correlations. Furthermore, traditional methods often rely on large amounts of multi-parameter well logging data for modeling and require manual data preprocessing, consuming significant manpower and time costs and being unsuitable for exploration scenarios where well logging data is scarce, further limiting their application effectiveness under complex geological conditions. Summary of the Invention
[0003] The purpose of this invention is to address the aforementioned technical problems by providing a method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network.
[0004] In view of this, the present invention provides a method for simultaneously modeling speed and anisotropic parameters based on multi-expert learning networks, comprising the following steps: Step 1: Obtain a low-precision velocity model and a small amount of multi-parameter logging data. The multi-parameter logging data includes velocity parameters, anisotropy parameters Eta and Delta. Use the low-precision velocity model as network input and the multi-parameter logging data as network labels. Step 2: Construct a multi-expert learning network consisting of a parameter sharing network, a weight network, and a parameter-specific network. The parameter sharing network is used to extract the common features of the velocity parameter, anisotropy parameters Eta and Delta. The parameter-specific network is used to capture the unique features of the velocity parameter, Eta and Delta. The weight network dynamically balances the contribution weights of different parameters. Step 3: Train the network by minimizing the weighted root mean square difference between the input data and the label data, iteratively update the network parameters until the difference between the output and the label data is minimized, and save the trained network model; Step 4: Apply the trained network model to all model parameters to achieve simultaneous modeling of speed and anisotropic parameters.
[0005] Preferably, the parameter sharing network in step two adopts a convolutional neural network structure, which extracts common features of different parameters through multi-layer convolution operations.
[0006] Preferably, the weight network in step two adopts an adaptive weight allocation mechanism, which dynamically adjusts the weight coefficients according to the contribution of different parameters in the modeling process.
[0007] Preferably, the objective function for network training in step three is: ; in This is the total number of input parameters. The weights of each parameter, Indicate each parameter The root mean square error of the corresponding logging parameters is calculated using the following formula: ; in For vectors Total quantity This indicates a multi-expert learning network.
[0008] Preferably, in step three, the network parameter update adopts the Adam (adaptive moment estimation) method, which adjusts the parameter values by calculating the derivative of the objective function with respect to the network parameters, thereby reducing the difference between the network output and the label data.
[0009] Preferably, the number of iterations for network training in step three is no less than 100, until the error between the network output and the multi-parameter logging data is lower than a preset threshold.
[0010] Preferably, the formula for calculating the error is: ; Preferably, the method is suitable for earthquake multi-parameter modeling scenarios and can be applied to multi-parameter modeling of Marmousi models and geological models with complex anisotropic characteristics.
[0011] Preferably, the low-precision velocity model mentioned in step one is low-precision velocity data conventionally obtained in the field of seismic exploration, and the multi-parameter logging data is high-precision logging data containing velocity parameters, Eta and Delta parameters measured in drilling engineering.
[0012] The beneficial effects of this invention are: This invention achieves high-precision synchronous modeling of velocity and anisotropic parameters through a three-layer network architecture: parameter sharing, parameter dedication, and weight allocation. The parameter sharing network effectively extracts common features from multiple parameters, the parameter dedication network accurately captures the unique characteristics of velocity and anisotropic parameters, and the weight network dynamically balances the contributions of different parameters. The synergistic effect of these three layers not only fully explores the higher-order implicit relationships among multiple parameters but also avoids modeling defects caused by neglecting the intrinsic correlations or imbalances in contributions between parameters. The modeling accuracy is significantly superior to traditional single-parameter modeling techniques and existing machine learning-driven modeling schemes.
[0013] Meanwhile, this invention boasts advantages of high automation and strong applicability. It eliminates the need for manual preprocessing of logging data and initial velocity models, enabling fully automated operation of the entire modeling process, significantly reducing manpower and time investment. Furthermore, it requires only a minimal amount of multi-parameter logging data to complete high-precision modeling, flexibly adapting to exploration scenarios where logging data is scarce. Applying the multi-parameter model constructed using this method to full-waveform inversion yields inversion results that highly match the actual geological model. Its forward-modeled seismic records show a high degree of consistency with the actual model's seismic records, providing reliable support for the accurate discovery of concealed and complex oil and gas reservoirs, subsequent oil and gas development planning, and drilling location. Attached Figure Description
[0014] Figure 1 This is a diagram of the multi-expert learning network architecture of the present invention; Figure 2 This is a flowchart illustrating the training process of the multi-expert learning network of this invention. Figure 3 This is a diagram of the actual and initial multi-parameter model of the present invention; Figure 4 This is a diagram showing the results of multi-parameter modeling in this invention; Figure 5 This is a diagram showing the inversion results of the present invention; Figure 6 This is a comparison graph of the multi-parameter single-track curves of the present invention; Figure 7 This is a comparison chart of earthquake records from the present invention; Figure 8 This is a diagram showing the results of multi-parameter modeling in this invention. Detailed Implementation
[0015] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0016] It should be noted that all directional and positional terms used in this invention, such as "up," "down," "left," "right," "front," "back," "vertical," "horizontal," "inner," "outer," "top," "lower," "lateral," "longitudinal," and "center," are only used to explain the relative positional relationships and connections between components in a specific state (as shown in the accompanying drawings). They are merely for the convenience of describing the invention and do not require the invention to be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on the invention. Furthermore, descriptions involving "first," "second," etc., are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.
[0017] In the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0018] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0019] This invention proposes a seismic multi-parameter modeling technique based on a multi-expert learning network. The multi-expert learning network represents the high-order and implicitly complex statistical relationships between multi-parameter well logging data and existing single-parameter models, thereby enabling the construction of a high-precision multi-parameter model. The technical solution of this invention consists of three stages: data preparation, multi-expert learning network construction, and multi-expert learning network training. The data required for this invention includes low-precision velocity models. and a small amount of multi-parameter logging data , Represents the anisotropy parameter Eta, Let Delta represent the anisotropy parameter. The low-precision velocity model serves as the input to the multi-expert learning network. Multi-parameter logging data are used as tags for the network. The multi-expert learning network architecture constructed in this invention is as follows: Figure 1As shown, the system consists of a parameter-sharing network, a weight network, and a parameter-specific network. The parameter-sharing network extracts common features of the velocity parameter, anisotropic parameter Eta, and anisotropic parameter Delta. The parameter-specific network captures the unique anisotropic features of Eta / Delta. The weight network dynamically balances the contribution weights of different parameters to learn diverse parameter combination patterns, thereby more fully capturing the correlations between various parameters.
[0020] After preparing the input data and constructing the multi-expert learning network, training the network begins. The multi-expert learning network minimizes the input data... and tag data The training is performed using the weighted root mean square difference between the two values, and its objective function is: (1); in This is the total number of input parameters. The weights of each parameter, Indicate each parameter The root mean square error of the corresponding logging parameters is calculated using the following formula: (2); in For vectors Total quantity This indicates a multi-expert learning network. Figure 2 This is a flowchart of the training process for a multi-expert learning network. It mainly includes three steps: (1) Use a low-precision velocity model as input data Input into a multi-expert learning network Obtain the corresponding network output. ; (2) Calculate the network output and tag data The objective function of the multi-expert learning network ; (3) Based on the objective function of the multi-expert learning network Update network parameters using the derivatives of the network parameters to reduce network output. and tag data The differences between them.
[0021] Repeat steps (1)-(3) until the network output is complete. and multi-parameter logging data The differences between them are minimized. After network training is complete, the network parameters and network model are saved. Finally, the trained multi-expert learning network model is applied to all model parameters, thereby achieving the construction of a high-precision multi-parameter model.
[0022] To verify the effectiveness of the proposed method, this invention uses the Marmousi model, which has complex anisotropic characteristics, for testing. Figure 3 a, Figure 3 b and Figure 3 c represents the true velocity model, the true Eta model, and the true Delta model, respectively; this invention randomly selects data from three wells as label data. Figure 3 d, Figure 3 e and Figure 3 f represents the initial velocity model, the initial Eta model, and the initial Delta model.
[0023] This invention sets 100 iterations in the training of a multi-expert learning network. Figure 4 a, Figure 4 b and Figure 4 c represents the speed model, Eta model, and Delta model, respectively, built based on a multi-expert learning network. Figure 4 a, Figure 4 b and Figure 4 As can be seen from c, the velocity model, Eta model and Delta model constructed in this invention all have high accuracy and are consistent with the background of the real model.
[0024] To further verify the technical differences and advantages of the multi-expert neural network constructed in this invention compared with existing mainstream multi-task learning networks, the modeling results of this invention and the hard parameter sharing network are compared. Figure 4 d, Figure 4 e and Figure 4 f) and the modeling results of soft parameter sharing networks (unweighted networks) Figure 4 g, Figure 4 h and Figure 4 i) Conduct comparative analysis. For modeling parameter errors, this invention uses the normalized root mean square error (RMSE) as the evaluation metric, i.e. (3); in The modeling results for the neural network, For realistic models, The initial model was used. Quantitative calculations showed that the RMSE of the hard parameter sharing network was 54.58%, the RMSE of the unweighted soft parameter sharing network was 48.05%, and the RMSE of this invention was the lowest, at only 40.34%. Comparative analysis shows that, compared to the hard parameter sharing network, this invention, through a separate design of parameter sharing and dedicated parameters, achieves simultaneous preservation of common and unique characteristics of multiple parameters, resulting in a significant RMSE reduction of 14.24%. Compared to the unweighted soft parameter sharing network, the dynamic weighted network of this invention can reduce the RMSE from 48.05% to 40.34%. This result fully demonstrates that the original three-layer architecture of parameter sharing + dedicated parameters + weight allocation can not only effectively explore the potential correlations between multiple parameters but also achieve a precise balance between parameter correlation and unique characteristics, ultimately achieving high-precision multi-parameter modeling results.
[0025] To further verify the effectiveness of the present invention, a multi-parameter full waveform inversion was performed on the multi-parameter initial model constructed in this invention. Figure 5 a, Figure 5 b and Figure 5 c is the corresponding multi-parameter inversion result. From Figure 5 As can be seen, multi-parameter full-waveform inversion based on the multi-parameter initial model constructed in this invention yields not only a high-precision velocity model but also highly accurate anisotropic Eta and Delta parameter models. This has significant implications for the discovery of concealed and complex oil and gas reservoirs and provides a reliable scientific basis for subsequent oil and gas prediction and well location.
[0026] To further estimate the quality of the multi-parameter inversion results, single-track curves at a horizontal distance of 1.5 km were selected for comparison. For example... Figure 6 As shown, compared with the initial multi-parameter model, the multi-parameter inversion results of this invention have a higher degree of fit with the real multi-parameter model.
[0027] Figure 7 This demonstrates the seismic records obtained through forward modeling based on the real model, the initial model, and multi-parameter inversion results. For example... Figure 6 As shown, the seismic records obtained from the multi-parameter inversion results are in high agreement with the seismic records of the actual model, which further verifies the reliability of the present invention.
[0028] This invention constructs a multi-expert learning neural network to directly build a multi-parameter model from well logging data, overcoming the shortcomings of traditional methods and existing machine learning methods. This invention has the following significant advantages: (1) High accuracy in multi-parameter modeling, because the multi-expert learning neural network constructed in this invention can directly represent the high-order and implicitly complex statistical relationships between existing single-parameter models and multi-parameter high-dimensional well logging data. (2) High degree of automation, because this invention does not require any manual preprocessing of high-dimensional multi-parameter well logging data and existing single-parameter models; the entire modeling process is fully automated, greatly saving manpower and time costs. (3) High efficiency in parameter modeling, because the multi-expert learning neural network constructed in this invention can simultaneously complete the modeling of multiple parameters. (4) Wide applicability, because this invention only requires a very small amount of well logging data to complete the construction of a high-precision multi-parameter model. Based on the multi-parameter model constructed in this invention, high-precision multi-parameter inversion results can be obtained. This not only helps improve the discovery of hidden and complex oil and gas reservoirs, but also provides key guidance for subsequent oil and gas reservoir development and drilling location.
[0029] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.
Claims
1. A method for simultaneously modeling speed and anisotropic parameters based on multi-expert learning networks, characterized in that: Includes the following steps: Step 1: Obtain a low-precision velocity model and a small amount of multi-parameter logging data. The multi-parameter logging data includes velocity parameters, anisotropy parameters Eta and Delta. Use the low-precision velocity model as network input and the multi-parameter logging data as network labels. Step 2: Construct a multi-expert learning network consisting of a parameter sharing network, a weight network, and a parameter-specific network. The parameter sharing network is used to extract the common features of the velocity parameter, anisotropy parameters Eta and Delta. The parameter-specific network is used to capture the unique features of the velocity parameter, Eta and Delta. The weight network dynamically balances the contribution weights of different parameters. Step 3: Train the network by minimizing the weighted root mean square difference between the input data and the label data, iteratively update the network parameters until the difference between the network output and the label data is minimized, and save the trained network model; Step 4: Apply the trained network model to all model parameters to achieve simultaneous modeling of speed and anisotropic parameters.
2. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The parameter sharing network described in step two adopts a convolutional neural network structure, which extracts common features of different parameters through multi-layer convolution operations.
3. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The weight network described in step two adopts an adaptive weight allocation mechanism, which dynamically adjusts the weight coefficients according to the contribution of different parameters in the modeling process.
4. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The objective function for network training in step three is: ； in This is the total number of input parameters. The weights of each parameter, Indicate each parameter The root mean square error of the corresponding logging parameters is calculated using the following formula: ； in For vectors Total quantity This indicates a multi-expert learning network.
5. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: In step three, the network parameters are updated using the Adam (adaptive moment estimation) method. The parameter values are adjusted by calculating the derivative of the objective function with respect to the network parameters, thereby reducing the difference between the network output and the label data.
6. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: In step three, the number of iterations for network training shall not be less than 100, until the difference between the network output and the multi-parameter logging data is lower than the preset threshold.
7. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The formula for calculating the difference is: 。 8. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The method is applicable to multi-parameter seismic modeling scenarios and can be applied to multi-parameter modeling of Marmousi models and geological models with complex anisotropic characteristics.
9. The method for simultaneously modeling speed and anisotropic parameters based on a multi-expert learning network according to claim 1, characterized in that: The low-precision velocity model mentioned in step one refers to low-precision velocity data conventionally obtained in the field of seismic exploration, while the multi-parameter logging data refers to high-precision logging data containing velocity parameters, Eta, and Delta parameters measured in drilling operations.