A deep learning and wave equation jointly driven microseismic velocity inversion method
The microseismic velocity inversion method driven by deep learning and wave equations solves the problem of large errors in inversion results in traditional methods by utilizing Unet neural networks and wave equation constraints. It achieves high-precision and efficient velocity model reconstruction and is suitable for real-time monitoring under low signal-to-noise ratio conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2023-12-14
- Publication Date
- 2026-05-08
AI Technical Summary
Traditional microseismic velocity model inversion methods are limited by low signal-to-noise ratio and the accuracy of initial velocity models, resulting in large errors and low efficiency in the inversion results. Pure data-driven deep learning methods have poor generalization ability and poor interpretability.
A microseismic velocity inversion method jointly driven by deep learning and wave equations is adopted. Nonlinear fitting is performed through Unet neural network, and wave equation constraints are introduced to construct a joint loss function driven by data and physics. Dynamic weights are used to adjust the training focus to improve the network's generalization and interpretability.
It improves the accuracy and robustness of microseismic velocity inversion, enables accurate velocity model reconstruction under low signal-to-noise ratio conditions, and has real-time monitoring capabilities.
Smart Images
Figure CN117706623B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysical microseismic monitoring technology, and in particular to a microseismic velocity inversion method jointly driven by deep learning and wave equations. Background Technology
[0002] Hydraulic fracturing is one of the key technologies for the economic development of unconventional oil and gas reservoirs. Microseismic monitoring technology is one of the most effective methods for real-time monitoring of the effects of hydraulic fracturing. By monitoring the weak earthquakes induced by hydraulic fracturing and other engineering operations, the spatial size and extension morphology of the fracturing fractures can be revealed, which helps to optimize the construction design and effectiveness of fracturing.
[0003] To ensure safe and efficient reservoir development, accurate imaging of fracturing fractures is crucial. Source location directly impacts fracture imaging quality, and velocity, an indispensable parameter in microseismic data processing, directly affects location accuracy. Traditional microseismic velocity model inversion methods primarily involve travel-time inversion and full-waveform inversion. However, limited by the low signal-to-noise ratio of actual data and the accuracy of the initial velocity model, inversion results often get trapped in local minima, resulting in large errors and low efficiency. In recent years, the development of deep learning has brought new prospects to the field of geophysics, especially seismic exploration. Deep learning neural networks can effectively extract the potential mapping relationship between seismic data and velocity models. However, purely data-driven deep learning methods rely heavily on data quality and do not incorporate physical information, leading to weak generalization ability and poor interpretability. Summary of the Invention
[0004] To address the issue that conventional travel time and full waveform velocity inversion methods may be affected by noise and initial velocity in microseismic data processing, this invention provides a microseismic velocity inversion method jointly driven by deep learning and wave equations. By leveraging the nonlinear fitting capability of neural networks and incorporating physical information constrained by wave equations, the accuracy of microseismic velocity inversion is improved, enhancing the network's generalization and interpretability, and providing a reliable velocity model for source location.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:
[0006] A microseismic velocity inversion method jointly driven by deep learning and wave equations includes the following steps:
[0007] S1. Discretize the underground model using a grid and set n detectors in the underground model, where n is a natural number;
[0008] S2. Establish a subsurface velocity model, set the area where the random seismic source is located, and use the finite difference algorithm to calculate the waveform information of the random seismic source at each geophone location;
[0009] S3. Input the waveform records, velocity models and source coordinates into the Unet neural network in batches;
[0010] S4. First, perform pure data-driven calculations to obtain the preliminary velocity structure, then perform physical-driven calculations, use finite difference to solve the wave equation, and obtain simulated seismic waveform records.
[0011] S5. The loss function consists of two parts: the velocity error calculated purely by data and the waveform error calculated by physics.
[0012] The loss function is as follows:
[0013]
[0014] Where L d For the data-driven loss function, x i This is the actual speed value. For predicting velocity values, N is the total number of pixels in a single velocity map;
[0015]
[0016] Where L p Let be the physical-driven loss function, j be the number of sampling points, and r be the number of detectors. For the input waveform data, Waveform data synthesized for a physical driving process;
[0017]
[0018] Where L is the loss function of the entire neural network. When epoch < epochs_syn, pure data-driven operation is performed; when epoch ≥ epochs_syn, physical-driven operation is performed, where w1 and w2 are dynamic weights.
[0019] S6. Calculate the structural similarity between the predicted velocity and the actual velocity, and set their average value as the dynamic weight of the physics-driven loss value.
[0020] The formula for calculating the dynamic weight is:
[0021]
[0022]
[0023] w1 + w2 = 1
[0024] Where SSIM represents structural similarity, SSIM P SSIM represents the structural similarity of P-wave velocities. SSBatchsize represents the structural similarity of the S-wave velocity.
[0025] S7. Use neural networks to extract waveform features and invert and predict underground velocity structure.
[0026] The formula for the neural network to extract waveform features and invert the velocity model is as follows:
[0027] v = Net(d; θ)
[0028] Where, v≡[v p ,v s ] represents the predicted velocity value, Net represents the Unet neural network model, d represents the microseismic waveform data, and θ represents the neural network weights.
[0029] This invention provides a microseismic velocity inversion method jointly driven by deep learning and wave equations. This method is independent of the accuracy of the initial velocity model and possesses strong robustness. It constructs a data-driven and physics-driven loss function based on the wave equation constraining the velocity model inversion. Dynamic weights are added during the loss function setting process to synchronously monitor the velocity reconstruction degree, thus facilitating the adjustment of hyperparameters and influencing the training focus. By introducing the physical information of the wave equation, the generalization and interpretability of the network are effectively improved, further enhancing the accuracy of velocity inversion. Although the neural network training time is relatively long, once trained, it can predict velocity models in real time and has broad application prospects for real-time microseismic monitoring with low signal-to-noise ratios. Attached Figure Description
[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 A flowchart of the method provided in the embodiments of the present invention;
[0032] Figure 2 This is a velocity model and a schematic diagram of the source-detector arrangement provided in an embodiment of the present invention;
[0033] Figure 3 The Unet neural network framework provided in this embodiment of the invention;
[0034] Figure 4 A schematic diagram of the method provided in an embodiment of the present invention;
[0035] Figure 5The loss function curve provided for the embodiments of the present invention;
[0036] Figure 6 The dynamic weighting curve provided in the embodiments of the present invention;
[0037] Figure 7(a) shows the actual P-wave velocity structure;
[0038] Figure 7(b) shows the predicted P-wave velocity structure;
[0039] Figure 7(c) shows the actual S-wave velocity structure;
[0040] Figure 7(d) shows the predicted S-wave velocity structure. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0042] Example 1:
[0043] like Figure 1 As shown in the figure, in this embodiment of the invention, a flowchart of a microseismic velocity inversion method jointly driven by deep learning and wave equation is proposed, which specifically includes the following steps:
[0044] Step S1: Discretize the underground model using a grid of a certain size, and set n detectors in the underground model.
[0045] Specifically, based on the actual conditions of the underground target monitoring area, an underground model is established around the area to be monitored by hydraulic fracturing. This model is then discretized using a grid of a certain size. Furthermore, n geophones are placed at different locations within the underground model according to a specific pattern. The grid size is set according to actual needs; a smaller grid results in higher velocity inversion accuracy, but also a longer computation time.
[0046] Step S2: Establish a subsurface velocity model, set the region where the random seismic source is located, and use the finite difference algorithm to calculate the waveform information of the random seismic source at each detector location.
[0047] Specifically, an underground velocity model is established, and based on the underground velocity model, the waveform information of the random source in the underground model at each detector location is calculated by solving the elastic wave equation using finite difference.
[0048] See Figure 2The diagram shows the velocity model and source-detector arrangement provided in the embodiment. The layered velocity model is 64m*200m in size with a grid spacing of 2m. Fifteen detectors with a spacing of 4m are arranged in the simulated well, with 1201 sampling points. The source wavelet is the Ricker wavelet with a dominant frequency of 60Hz and a sampling interval of 0.2ms. The source is located in the region of the third-layer velocity structure.
[0049] Step S3: Input the waveform records, velocity model and source coordinates into the Unet neural network in batches.
[0050] Specifically, the batch size is set to 4, and the waveform records, corresponding velocity models, and source locations are all input into the Unet neural network according to this batch size. The source coordinates are only loaded and used in solving the wave equation when the neural network executes the physics-driven operation.
[0051] See Figure 3 The image shows the Unet neural network framework. The Unet neural network consists of an encoder and a decoder, and achieves feature fusion by connecting features with the same number of channels through skip-connections.
[0052] Step S4: First, perform pure data-driven calculations to obtain the preliminary velocity structure, then perform physical-driven calculations to solve the wave equation using finite difference methods to obtain simulated seismic waveform records.
[0053] Specifically, the learning rate was set to 0.0001, and a hybrid training strategy was adopted. First, data-driven training was used to find the mapping relationship between waveform and velocity, which has the ability to invert simple layered velocity models, but performs poorly in detail reconstruction. Then, the source coordinates were loaded, the velocity model generated by data-driven training was loaded, the same forward modeling parameters as in step 2 were set, and then the wave equation was solved using finite difference to obtain simulated seismic waveform records.
[0054] See Figure 4 The diagram shown is a schematic representation of a microseismic velocity inversion method jointly driven by deep learning and wave equations, provided in an embodiment. The dashed box in the diagram represents the purely data-driven part, while the area outside the dashed box represents the physics-driven part.
[0055] Step S5: The loss function consists of two parts: the speed error calculated by pure data-driven calculation and the waveform error calculated by physical driving calculation.
[0056] Specifically, the loss function consists of two parts. The root mean square error between the actual speed and the predicted speed in purely data-driven calculations can be expressed by the following formula:
[0057]
[0058] Where Ld For the data-driven loss function, x i This is the actual speed value. For predicting velocity values, N is the total number of pixels in a single velocity map.
[0059] The error between the input waveform and the synthesized waveform in the physics-driven computation can be expressed by the following formula:
[0060]
[0061] Where L p Let be the physical-driven loss function, j be the number of sampling points, and r be the number of detectors. For the input waveform data, Waveform data synthesized for a physical driving process.
[0062] The overall loss function expression is shown below, and consists of two parts:
[0063]
[0064] Where L is the loss function of the entire neural network. When epoch < epochs_syn, pure data-driven operation is performed; when epoch ≥ epochs_syn, physical-driven operation is performed, where w1 and w2 are dynamic weights.
[0065] See Figure 5 The figure shows the variation curve of the loss function in this embodiment. In this example, epochs_syn is set to 50. The curve shows that the first part is purely data-driven, with the curve tending towards a fit, possessing the ability to invert a simple layered velocity model. However, it lacks the ability to reconstruct details, resulting in a low-resolution velocity model. With a total of 80 epochs, the curve shows that the latter part is a physics-driven training process. As the number of epochs increases, the loss function converges more smoothly, and the details of the velocity model are reconstructed better.
[0066] Step S6: Calculate the structural similarity between the predicted velocity and the actual velocity, and set their average value as the dynamic weight of the physics-driven loss value.
[0067] Specifically, the structural similarity between the predicted velocity and the actual velocity is first calculated, and then divided by the batch size to obtain the average structural similarity, which is used as the dynamic weight of the physics-driven loss value. The difference between the average structural similarity and 1 is the weight of the data-driven part.
[0068]
[0069]
[0070] w1 + w2 = 1
[0071] Where SSIM represents structural similarity, SSIM P SSIM is used to calculate the structural similarity of P-waves. S S is the structural similarity calculated for S-waves, and Batchsize is the batch size.
[0072] See Figure 6 As shown, the dynamic curves of w1 and w2 are presented. Their changes depend on the structural similarity of the data-driven computation. The better the velocity model is reconstructed, the larger w2 is, and the smaller w1 is, thus achieving the goal of focusing on physics-driven training.
[0073] Step S7: Based on the microseismic velocity inversion method jointly driven by deep learning and wave equation, the waveform features are extracted using a neural network to invert and predict the underground velocity structure.
[0074] Specifically, by utilizing the strong nonlinear expressive power and feature extraction capabilities of neural networks, a loss function for waveform error and velocity error is fitted to obtain a trained model. When the model is called again, only the waveform record needs to be input to invert and predict the underground velocity structure in real time.
[0075] See Figures 7(a) to 7(d) The figures show the actual and predicted velocity structures of the embodiment. Figures 7(a) and 7(c) show the actual P-wave and S-wave velocity structures, respectively, while Figures 7(b) and 7(d) show the predicted P-wave and S-wave velocity structures, respectively. Both in terms of overall performance and details, this invention can reconstruct the velocity model well and effectively characterize abrupt changes in formation velocity, possessing the ability to recover high-precision inter-layer velocity structures.
[0076] The method proposed in this embodiment is based on deep learning and wave equation constraints. By introducing physical driving after data-driven approaches, physical laws constrain the weights of the neural network, revealing the potential relationship between waveform and velocity. Setting dynamic weights is beneficial for monitoring the fitting degree of the physical driving process and for focusing on the physical driving process, thus constraining the data-driven stage and limiting the velocity model inversion. Comprehensive numerical examples and real-world data examples verify the robustness and interpretability of the proposed method.
[0077] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.
Claims
1. A microseismic velocity inversion method jointly driven by deep learning and wave equations, characterized in that, Includes the following steps: S1. Discretize the underground model using a grid and set n detectors in the underground model, where n is a natural number; S2. Establish a subsurface velocity model, set the area where the random seismic source is located, and use the finite difference algorithm to calculate the waveform information of the random seismic source at each geophone location; S3. Input the waveform records, velocity models and source coordinates into the Unet neural network in batches; S4. First, perform pure data-driven calculations to obtain the preliminary velocity structure, then perform physical-driven calculations, use finite difference to solve the wave equation, and obtain simulated seismic waveform records. S5. The loss function consists of two parts: the velocity error calculated purely by data and the waveform error calculated by physics. S6. Calculate the structural similarity between the predicted velocity and the actual velocity, and set their average value as the dynamic weight of the physics-driven loss value. S7. Use neural networks to extract waveform features and invert and predict underground velocity structure.
2. The microseismic velocity inversion method jointly driven by deep learning and wave equations according to claim 1, characterized in that, In S5, the loss function is as follows: Where L d For the data-driven loss function, x i This is the actual speed value. For predicting velocity values, N is the total number of pixels in a single velocity map; Where L p Let be the physical-driven loss function, j be the number of sampling points, and r be the number of detectors. For the input waveform data, Waveform data synthesized for a physical driving process; Where L is the loss function of the entire neural network. When epoch < epochs_syn, pure data-driven operation is performed; when epoch ≥ epochs_syn, physical-driven operation is performed, where w1 and w2 are dynamic weights.
3. The microseismic velocity inversion method jointly driven by deep learning and wave equations according to claim 1, characterized in that, The formula for calculating the dynamic weights described in S6 is as follows: w1 + w2 = 1 Where SSIM represents structural similarity, SSIM P SSIM represents the structural similarity of P-wave velocities. S SBatchsize represents the structural similarity of the S-wave velocity.
4. The microseismic velocity inversion method jointly driven by deep learning and wave equations according to claim 1, characterized in that, The formula for extracting waveform features from the neural network described in S7 to invert the velocity model is as follows: v = Net(d; θ) Where, v≡[v p ,v s ] represents the predicted velocity value, Net represents the Unet neural network model, d represents the microseismic waveform data, and θ represents the neural network weights.