A data and physics alternating driven semi-supervised seismic wave impedance inversion method
By alternately generating physical pseudo-labels and combining deep learning networks with geophysical laws, the problem of insufficient physical constraints in semi-supervised inversion is solved, achieving high-precision and stable wave impedance inversion, and the output results conform to actual geological laws.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2026-05-15
- Publication Date
- 2026-07-14
AI Technical Summary
Existing semi-supervised seismic impedance inversion methods lack direct and effective physical constraints under conditions with very few well logging labels, resulting in unstable inversion results that are prone to getting trapped in local optima, making it difficult to achieve high-precision impedance inversion.
By alternately generating physical pseudo-labels, and combining deep learning networks with geophysical laws, joint training is performed. The physical pseudo-labels are used to directly correct and constrain the network, forming a powerful physical guidance that enables the network to achieve stable convergence.
Under limited logging labels, high-precision and stable wave impedance inversion was achieved. The output inversion results conform to actual geological laws, and the detail characterization is superior to traditional methods.
Smart Images
Figure CN122386389A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of petroleum geophysical exploration, specifically a semi-supervised seismic impedance inversion method driven by alternating data and physical constraints. Based on a deep learning architecture, this method drives network parameter updates by alternately generating physical pseudo-labels constrained by geophysical laws. By introducing strong physical constraints on top of data-driven approaches, it achieves high-precision impedance inversion under limited logging labels. Background Technology
[0002] Seismic impedance inversion is a core technology for quantitative reservoir prediction and fluid identification. In recent years, deep learning has made significant progress in the field of seismic inversion due to its powerful nonlinear mapping capabilities. However, most existing deep learning inversion methods are supervised learning, and their model inversion heavily relies on a large amount of high-quality well logging label data. In actual oil and gas exploration, drilling costs are high, and the amount of well logging data is often limited, leading to overfitting and insufficient generalization ability in supervised deep learning models.
[0003] To address the label scarcity problem, semi-supervised learning has been increasingly introduced into seismic inversion. Compared to supervised learning, which relies solely on a limited number of labels, semi-supervised inversion methods introduce forward modeling operators on top of supervised learning. Impedance prediction results from unlabeled seismic records are input into the forward modeling operator to obtain synthetic seismic records, and the difference between the synthetic and original unlabeled seismic records is used as an additional loss function. This indirect physical constraint mechanism can effectively utilize massive amounts of unlabeled seismic data, alleviating the data scarcity problem to some extent, and its inversion performance is superior to supervised methods. However, the strength of this indirect constraint relying solely on forward mapping errors is relatively weak. With extremely limited labeled data, it is difficult for the network to establish accurate mapping relationships in the early stages, and indirect residual losses alone are insufficient to effectively guide network convergence, leading to instability in the inversion results and a tendency to get trapped in local optima.
[0004] In summary, while existing semi-supervised inversion methods utilize unlabeled data, they lack direct and effective physical intervention in the impedance space of the prediction results. Overcoming the bottleneck of insufficient indirect constraints in semi-supervised methods, and directly introducing explicit geophysical evolution laws into the deep learning loop with minimal logging labels to form a strong physical constraint error correction and guidance mechanism, thereby achieving high-precision and high-stability impedance inversion, is a pressing technical challenge in the field of geophysical exploration. Summary of the Invention
[0005] To address the aforementioned technical problems, the purpose of this invention is to provide a semi-supervised seismic impedance inversion method driven by alternating data and physical processes. This invention deeply integrates the physical inversion mechanism with a deep learning network. Through an alternating cycle of predicting the initial model, iteratively optimizing the physical model to generate pseudo-labels, and then using these pseudo-labels to feed back into the network for joint training, the invention continuously corrects and constrains the pseudo-labels predicted by the network itself using prior geophysical laws. This overcomes the deficiency in semi-supervised inversion methods that rely solely on forward modeling residuals for insufficient indirect physical constraints, providing direct and powerful physical guidance to the network model, thereby guiding the inversion network towards stable convergence in a direction consistent with actual geological laws.
[0006] To achieve the above objectives, the technical solution adopted in this invention is a semi-supervised seismic wave impedance inversion method driven by alternating data and physical processes, which includes the following steps:
[0007] (1) Obtain post-stack seismic records and well logging impedance curves of the study area to construct labeled and unlabeled datasets;
[0008] (2) Establish a deep learning architecture that includes an inversion network and a forward network, and perform initial joint training using labeled and unlabeled datasets to obtain the initial inversion network and the initial forward network;
[0009] (3) Freeze the inversion network parameters, input the unlabeled dataset into the inversion network to obtain the predicted wave impedance, and use the predicted wave impedance as the initial model for physical inversion;
[0010] (4) Construct a physical forward modeling operator by combining seismic wavelets, and perform physical iterative inversion on the initial model by optimizing the objective function containing physical constraint terms to generate physical pseudo-labels;
[0011] (5) Unfreeze the inversion network parameters, use the physical pseudo-labels as supervision signals for the unlabeled dataset, and update and train the inversion network and the forward network in conjunction with the labeled dataset;
[0012] (6) Repeatedly execute steps (3) to (5) above, and adaptively reduce the number of iterations of the physical iterative inversion during the alternation process until the change of the validation set loss function in N consecutive iterations is less than the preset threshold or reaches the preset maximum number of alternations, then the inversion network is considered to be trained.
[0013] (7) Finally, the post-stack seismic records of the study area are input into the trained inversion network to output high-precision wave impedance inversion results.
[0014] The inversion network and the forward network mentioned in step (2) are independent neural network models; the basic network of the deep learning architecture includes any one of fully connected neural networks, convolutional neural networks, recurrent neural networks, Transformer networks, or hybrid architectures thereof.
[0015] Further, the initial joint training in step (2) includes: calculating the inversion supervision loss and forward mapping loss of labeled data, and calculating the cycle consistency loss of unlabeled data after passing through the closed loop of the inversion network and the forward network, and performing preliminary weight optimization of the inversion network by combining the above losses.
[0016] Furthermore, the objective function containing physical constraints in step (4) is composed of a weighted sum of the data residuals between the synthetic seismic record and the actual observed seismic data, the L1 regularization term constraining the sparsity of the reflection coefficient, and the Tikhonov regularization term constraining the smoothness of the model.
[0017] Further, the update training in step (5) includes: constructing a joint loss function that includes inversion supervision loss, forward mapping loss, cycle consistency loss and pseudo-label inversion loss; wherein, the pseudo-label inversion loss is the error between the current predicted impedance obtained by inputting the unlabeled dataset into the preheating inversion network and the physical pseudo-label; and using the joint loss function to jointly optimize the parameters of the preheating inversion network and the forward modeling network.
[0018] The beneficial effects of this invention are as follows: the physical pseudo-labels are label data generated by iterative optimization of the wave impedance predicted by the inversion network through physical forward modeling operators and physical constraint terms. The alternating driving mechanism uses geophysical prior models to perform physical spatial constraints and corrections on the network prediction results of the unlabeled dataset. By introducing the corrected physical pseudo-labels into the semi-supervised joint update training of the network, the shortcomings of traditional semi-supervised inversion, which relies solely on forward modeling residuals for indirect physical constraint strength, are overcome. This provides direct and powerful physical guidance for the network model, thereby achieving high-precision wave impedance inversion under limited logging label conditions. Attached Figure Description
[0019] Figure 1 A flowchart of a semi-supervised inversion method driven by alternating data and physics provided in an embodiment of the present invention;
[0020] Figure 2 This invention relates to supervised learning, semi-supervised learning, and the final inversion results and errors of the wave impedance in the Marmousi2 geological model.
[0021] Figure 2 a represents the true wave impedance profile;
[0022] Figure 2b is the seismic record profile;
[0023] Figure 2 c represents the wave impedance profile obtained by the supervised learning method inversion;
[0024] Figure 2 d represents the error profile between the inversion result of the supervised learning method and the true wave impedance;
[0025] Figure 2 e represents the wave impedance profile obtained by the semi-supervised learning method.
[0026] Figure 2 f is the error profile between the inversion result of the semi-supervised learning method and the true wave impedance;
[0027] Figure 2 g is the wave impedance profile obtained by the inversion method of the present invention;
[0028] Figure 2 h represents the error profile between the inversion result of the method of this invention and the actual wave impedance. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0030] This invention discloses a semi-supervised seismic impedance inversion method driven by alternating data and physical models. In practical applications, the network is first warmed up using a small amount of labeled well logging data and unlabeled seismic data. Then, the network parameters are frozen, and the prediction results of the preheated inversion network are used as the initial model. This is combined with a seismic wavelet and an optimization algorithm incorporating sparsity and smoothing regularization to perform iterative physical inversion, generating physical pseudo-labels constrained by geophysical laws. Next, the network is unfrozen, and these physical pseudo-labels are used as direct supervision signals from the unlabeled data to feed back into the network for joint update training. By continuously alternating the above process, a two-way closed loop between data-driven and physical model-driven approaches is achieved.
[0031] Example: Taking Marmousi2 theoretical geological model data as an example. This example involves a two-dimensional simulated profile containing a large number of seismic traces. To simulate the scenario of scarce labels in actual exploration, only a small portion of the simulated well-by-well traces are extracted as labeled datasets. Based on the characteristics of this model data, this invention constructs the following alternating-driven inversion workflow:
[0032] Step (1): Data Acquisition and Segmentation. Obtain the post-stack seismic record matrix and corresponding true impedance values of the Marmousi2 model, extract the seismic records and their time-domain impedance curves corresponding to the simulated well locations, and construct a labeled dataset; the remaining non-wellside seismic records are used as an unlabeled dataset. Normalize and preprocess all data.
[0033] Step (2): Establish the architecture and preheating training of the inversion and forward modeling networks. In this example, both the inversion and forward modeling networks use gated recurrent neural networks. The inversion network is used to establish the nonlinear mapping from seismic records to wave impedance, and the forward modeling network is used to fit the forward physical mapping from wave impedance to seismic records. Preheating training is performed using the labeled and unlabeled datasets, and the preheating loss function is constructed as follows: , in, Impedance prediction error for tagged data, For positive mapping error, The loss function represents the cyclic consistency error of unlabeled seismic records after a closed-loop inversion-forward modeling process. By optimizing this loss function, the preheated inversion network is obtained.
[0034] Step (3): Network Freezing and Initial Model Prediction. Freeze the parameters of the preheating inversion network and the forward modeling network. Input the unlabeled dataset into the preheating inversion network, and the network outputs the predicted wave impedance results of the full profile, which are used as the initial model for subsequent physical inversion.
[0035] Step (4): Physical inversion to generate pseudo-labels. A physical forward modeling operator is constructed by combining the seismic wavelet of the study area, and a gradient-based physical inversion objective function is established: , in, For observing earthquake data, W represents the seismic wavelet, and r represents the reflection coefficient. and These are the weighting coefficients. The initial model predicted in step (3) is used. The function is solved by physical iteration optimization, and the converged result is defined as the physical pseudo-label.
[0036] Step (5): Network Unfreezing and Joint Update Training. Unfreeze and invert network parameters. Use physical pseudo-labels as supervision signals to construct the joint loss function: , in, This represents the direct error between the network's current predicted impedance and the physical pseudo-label. This function is used for semi-supervised update training of the inversion network.
[0037] Step (6): Alternate driving and adaptive adjustment. Repeat steps (3) to (5) alternately. During the iteration process, the number of physical inversion iterations is adaptively reduced as the network accuracy improves. When the validation set loss function does not decrease significantly, training is stopped and the best model is saved.
[0038] Step (7): Final Inversion and Output. Finally, the post-stack seismic records of the study area are input into the trained inversion network, and high-precision wave impedance inversion results that conform to physical laws are output.
[0039] To verify the practical technical effect of this invention, inversion comparison experiments were conducted on the same dataset using pure supervised learning, conventional semi-supervised learning, and the semi-supervised seismic impedance inversion method with alternating data and physics driving provided by this invention. The final inversion results and error profiles of the three methods are shown below. Figure 2 As shown in the figure, the comparison of error profiles reveals that the inversion results of the method of this invention significantly reduce the overall inversion error compared to conventional semi-supervised inversion methods. Particularly in the complex fault regions in the middle of the model and the deep anticline structures, this invention demonstrates superior error suppression capabilities. Furthermore, the wave impedance profile output by this method exhibits stronger lateral continuity along the phase axis, and its detail characterization better conforms to geophysical and geological sedimentary patterns.
[0040] The alternating drive mechanism of this invention utilizes a geophysical prior model to directly correct the network prediction results in physical space. By introducing physical pseudo-labels into the training loop, it overcomes the deficiency of traditional semi-supervised inversion methods that rely solely on forward residuals for insufficient indirect physical constraint strength, providing powerful guidance for the network model. The inversion profile output by this method significantly outperforms conventional semi-supervised methods in terms of detail, effectively improving the stability and accuracy of the inversion.
[0041] The above description is merely a summary of the basic principles and implementation methods of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and various improvements should fall within the protection scope of the present invention without departing from its spirit.
Claims
1. A semi-supervised seismic impedance inversion method driven by alternating data and physics, characterized in that... The method includes the following steps: (1) Obtain post-stack seismic records and well logging impedance curves of the study area to construct labeled and unlabeled datasets; (2) Establish a deep learning architecture that includes an inversion network and a forward network, and perform initial joint training using labeled and unlabeled datasets to obtain the initial inversion network and the initial forward network; (3) Freeze the inversion network parameters, input the unlabeled dataset into the inversion network to obtain the predicted wave impedance, and use the predicted wave impedance as the initial model for physical inversion; (4) Construct a physical forward modeling operator by combining seismic wavelets, and perform physical iterative inversion on the initial model by optimizing the objective function containing physical constraint terms to generate physical pseudo-labels; (5) Unfreeze the inversion network parameters, use the physical pseudo-labels as supervision signals for the unlabeled dataset, and update and train the inversion network and the forward network in conjunction with the labeled dataset; (6) Repeatedly execute steps (3) to (5) above, and adaptively reduce the number of iterations of the physical iterative inversion during the alternation process until the change of the validation set loss function in N consecutive iterations is less than the preset threshold or reaches the preset maximum number of alternations, then the inversion network is considered to be trained. (7) Finally, the post-stack seismic records of the study area are input into the trained inversion network to output high-precision wave impedance inversion results.
2. The semi-supervised seismic impedance inversion method driven by alternating data and physics as described in claim 1, characterized in that, The inversion network and the forward network mentioned in step (2) are independent neural network models; the basic network of the deep learning architecture includes any one of fully connected neural networks, convolutional neural networks, recurrent neural networks, Transformer networks, or hybrid architectures thereof.
3. The semi-supervised seismic impedance inversion method driven by alternating data and physics as described in claim 1, characterized in that, The initial joint training in step (2) includes: calculating the inversion supervision loss and forward mapping loss of labeled data, and calculating the cycle consistency loss of unlabeled data after passing through the closed loop of the inversion network and the forward network, and performing preliminary weight optimization of the inversion network by combining the above losses.
4. The semi-supervised seismic impedance inversion method driven by alternating data and physics as described in claim 1, characterized in that, The objective function containing physical constraints in step (4) is composed of a weighted sum of the data residuals between the synthetic seismic record and the actual observed seismic data, the L1 regularization term constraining the sparsity of the reflection coefficient, and the Tikhonov regularization term constraining the smoothness of the model.
5. The semi-supervised seismic impedance inversion method driven by alternating data and physics as described in claim 1, characterized in that, The update training in step (5) includes: constructing a joint loss function that includes inversion supervision loss, forward mapping loss, cycle consistency loss and pseudo-label inversion loss; wherein, the pseudo-label inversion loss is the error between the predicted impedance obtained by inputting the unlabeled dataset into the inversion network and the physical pseudo-label; and using the joint loss function to jointly optimize the parameters of the preheating inversion network and the forward network.