Construction method of full-waveform inversion neural network based on physical perception collaborative fusion

Abstract

The invention belongs to the technical field of seismic exploration and artificial intelligence, and relates to a physical perception collaborative fusion-based full waveform inversion neural network construction method, which comprises the steps of required model data construction, full waveform inversion neural network construction, loss function construction, network pre-training, and network fine tuning and application. Physical feature perception and collaborative fusion are carried out through a U-Net network and an adaptive residual learning module, so that the problem that traditional full-waveform inversion depends on synthetic data training and is not matched with an actually measured data field is solved. The designed pre-training strategy not only can get rid of dependence on a large-scale synthetic data set, but also can effectively relieve the domain offset problem through fine adjustment of measured data, thereby improving the generalization ability of the model. And the adaptive residual learning module enables the neural network to adaptively and dynamically optimize feature expression, and improves inversion precision through collaborative fusion of a physical perception mechanism. According to the method, the physical consistency and imaging reliability of speed model prediction can be remarkably improved while the labor cost is reduced.

Description

Technical Field

[0001] This invention belongs to the fields of seismic exploration and artificial intelligence technology, specifically relating to a method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion. Background Technology

[0002] Oil and gas resources, as one of the most crucial basic energy sources in modern society, play an irreplaceable role in the global energy structure. Due to the significant regional differences in the distribution of oil and gas resources, their exploration and extraction often face complex geological conditions and technological challenges. With the continuous growth of global energy consumption, how to efficiently and cost-effectively develop oil and gas resources in deep or complex geological formations has become a key issue in the energy technology development of various countries.

[0003] Full Waveform Inversion (FWI), a high-precision inversion technique based on the wave equation, has become an indispensable tool in modern geophysical exploration. It deduces the velocity structure of the subsurface medium by minimizing the difference between simulated and measured waveforms, offering advantages such as high imaging accuracy and resolution. However, FWI still faces many challenges in practical applications, especially in scenarios with sparse observational data or low signal-to-noise ratios, where the inversion process is prone to getting trapped in local optima, leading to significant deviations in imaging results. Furthermore, limitations of the acquisition system, the complexity of surface conditions, and high data acquisition costs can all result in insufficient data coverage, thus reducing inversion accuracy. These problems are further exacerbated when dealing with complex geological structures such as deep oil and gas reservoirs or fault-prone areas, significantly restricting the practical application of FWI. Therefore, improving the robustness and computational efficiency of inversion algorithms has become an important direction in current full waveform inversion research.

[0004] Over the past few decades, researchers have proposed various optimization strategies to improve the stability and imaging quality of FWI. Early improvements included multi-scale strategies, frequency progression techniques, weighted objective functions, and the introduction of regularization terms. While these methods alleviated the ill-posedness of inversion to some extent, these physically-based modeling methods are typically heavily reliant on the initial model and perform poorly with highly nonlinear, non-Gaussian noise data. Furthermore, these methods often require extensive parameter tuning, lack automation capabilities, and are computationally expensive, resulting in low efficiency when applied to large-scale 3D models.

[0005] In recent years, with the rapid development of deep learning technology, neural network-based full waveform inversion methods have gradually become an important research direction in the field of geophysics. These methods, by constructing an end-to-end learning framework, can automatically extract complex feature information from seismic observation data, demonstrating significant advantages in improving modeling accuracy and enhancing nonlinear expression capabilities. Early studies mainly employed convolutional neural network (CNN) structures, such as U-Net and ResNet, to achieve data-driven waveform inversion. However, existing data-driven full waveform inversion methods face two core challenges: firstly, the large-scale synthetic datasets required for training the model typically rely on detailed numerical simulations, a process with high computational cost, long modeling cycles, and a high degree of dependence on geological priors; secondly, networks trained on synthetic data are prone to "domain inconsistency" or "distribution shift" problems when applied to actual observation data, leading to a decline in model generalization ability and severely affecting the stability and accuracy of the inversion results.

[0006] To overcome the aforementioned problems, researchers have recently begun to explore the introduction of geological prior information or physical constraints to enhance the rationality and physical consistency of the inversion process. These methods embed information such as wave equations, velocity smoothness constraints, layered structure assumptions, or geostatistical priors into the network training process, enabling the model to follow physical laws and limit meaningless solution spaces during learning. While introducing prior information or physical laws helps alleviate overfitting and enhance model interpretability, it inevitably introduces new problems such as increased model design complexity, difficulty in selecting physical term weights, and limitations on network expressive power due to excessively strong physical constraints. Therefore, how to effectively integrate the expressive power of neural networks with geophysical prior knowledge, improving inversion accuracy while avoiding the inhibition of model flexibility by physical constraints, has become a key technical challenge that current deep learning-based full-waveform inversion methods urgently need to address. Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion, so as to solve the problem of reducing the cost of constructing synthetic data and alleviating the domain mismatch between training data and measured data.

[0008] This invention is achieved through the following technical solution:

[0009] A method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion includes the following steps:

[0010] S1. Construct the initial model of the full waveform inversion neural network for pre-training and fine-tuning the required model data;

[0011] S2. Construct a full-waveform inversion neural network based on physical perception collaborative fusion for high-precision underground velocity model inversion;

[0012] S3. Construct a loss function to measure the performance of the full waveform inversion neural network based on physical perception collaborative fusion;

[0013] S4. Neural Network Pre-training:

[0014] Based on the physical perception-based collaborative fusion full-waveform inversion neural network constructed in step S2, the neural network included in the method is pre-trained using the seismic data obtained from the forward modeling of the initial model to establish an effective mapping relationship between the seismic response and the target velocity model. The results obtained through this mapping are further used to optimize the loss function designed in step S3. Within the set training rounds, the neural network parameters are iteratively updated and adjusted until they converge to the optimal value, completing the pre-training process, and the neural network training weights are saved.

[0015] S5, Network Fine-tuning and Applications:

[0016] After completing the pre-training of the neural network, the obtained full-waveform inversion neural network containing the physical perception collaborative mechanism is fine-tuned and applied in practice. Seismic data from the real model is fed into the neural network to generate the corresponding velocity model prediction results, and its inversion performance is evaluated. If the fine-tuned neural network reaches the expected standard on the set loss function threshold standard, the full-waveform inversion neural network based on physical perception collaborative fusion updated in step S5 is confirmed as the final optimization scheme. If the threshold standard is not met, the training is re-fine-tuned, the neural network parameters are adjusted and optimized, and the retraining operation is performed until the optimization conditions are met and the fine-tuning training effect reaches the threshold standard.

[0017] Furthermore, step S1 specifically includes the following steps:

[0018] S11. Select typical geological models as real velocity models to reflect the complex velocity structure of different underground geological layers. Through the velocity parameters set in these models, simulate the characteristics of underground media that are closer to the real geological environment.

[0019] S12. Perform linear interpolation within the maximum and minimum velocity ranges and extract the overall median for the real velocity model used in step S11 to generate a linear initial model and a uniform initial model, so as to simulate the inversion environment where the initial model and the real model differ under real geological conditions.

[0020] S13. Using the forward modeling method, select the Ricker wavelet as the excitation source, and perform seismic wave propagation simulation on the real velocity model described in step S11 and the initial velocity model described in step S12 respectively, so as to obtain the corresponding real seismic data and initial model seismic data.

[0021] S14. The initial model seismic data and the real model seismic data obtained in step S13 are used for the pre-training and fine-tuning stages, respectively, thus completing the entire process of constructing the dataset.

[0022] Further, step S2 specifically involves the following: The neural network consists of an encoder module, a decoder module, and a designed adaptive residual learning module within the U-Net structure; during the training phase, the acquired seismic data is simultaneously fed into the encoder part of the U-Net and the adaptive residual module for feature extraction; the feature information extracted by the encoder module is then passed to the bottleneck module located in the middle of the network, and subsequently upsampled and restored layer by layer by the U-Net decoder module; the feature map output by the decoder module then undergoes a 1×1 convolution operation to generate intermediate prediction results, and is fused and summed with the output results from the adaptive residual module to generate the final predicted velocity model.

[0023] Furthermore, the encoder module consists of four encoding units, four max-pooling layers, eight 3×3 convolutional layers, activation functions, and normalization operations. Each encoding unit comprises two consecutive 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. The input feature map first enters the first encoding unit, where initial features are extracted through two convolutional layers, followed by downsampling through a 2×2 max-pooling layer. Subsequently, the three sets of encoding units are connected sequentially with the same structure, each set followed by a max-pooling operation after the convolutional module. In each layer, the convolutional operation increases the number of channels, while the max-pooling operation reduces the size of the feature map. The final output of the encoder module is the feature map obtained after four encoding and downsampling operations.

[0024] Furthermore, the bottleneck module consists of a set of two-layer convolutional structures and a residual connection. The input feature map first enters a 3×3 convolutional layer, equipped with a normalization layer and a ReLU activation function. Subsequently, the feature map passes through a second 3×3 convolutional layer, and again undergoes a nonlinear transformation by combining a normalization layer and a ReLU activation function. This module expands the channel dimension while maintaining the original size of the spatial dimension, and introduces a residual connection with an identity mapping, so that the input features are summed with the output of the main branch after matching in the channel dimension, thereby obtaining the output feature map of the bottleneck module.

[0025] Furthermore, the decoder module can perform sampling operations, skip connection concatenation, and convolutional decoding respectively. First, the input feature map is upsampled using transposed convolution to expand its spatial size to the size of the output of the previous encoder layer. Then, the upsampled feature map is concatenated with the feature map of the corresponding layer of the encoder in the channel dimension. The concatenated feature map is then input into a convolutional decoding unit consisting of two 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. The above three steps are repeated four times in the decoder module to restore the spatial resolution of the feature map in turn, thereby realizing the gradual reconstruction from high-dimensional low-resolution features to low-dimensional high-resolution images. Finally, an upsampling operation is used to uniformly adjust the feature map to the target output size.

[0026] Furthermore, the adaptive residual learning module can perform shape transformation, zeroing, and parameter initialization respectively. First, the input data is mapped along the time axis through a linear layer to reduce its dimensionality and extract temporal features. Then, redundant dimensions in the data are removed through dimensionality compression, and the data is copied and expanded on a specific axis to make its spatial shape consistent with the target velocity model, thus achieving structural alignment. Based on this, the processed data is fed into a zeroing layer for initial value definition, and finally defined as learnable parameters. During training, the gradient descent algorithm can be used to jointly optimize the learnable parameters of the U-Net network and the adaptive residual learning module using gradients containing physical information. After the two modules generate intermediate outputs, they are added element-wise in the spatial dimension to achieve collaborative feature fusion.

[0027] Furthermore, in step S3, the introduced loss function is an L2 norm form error metric function, i.e., the root mean square error loss function. This loss function is applied in both the pre-training and parameter fine-tuning stages to jointly guide the optimal convergence of the network; its corresponding mathematical expression is:

[0028]

[0029] In the pre-training stage, x(i,j) represents the predicted velocity model output by the neural network, y(i,j) is the initial velocity model constructed in step S1, and m and n represent the horizontal scale and depth range of the velocity model, respectively. In the parameter fine-tuning stage, x(i,j) represents the seismic data generated by forward modeling from the predicted velocity model, y(i,j) is the seismic data generated by forward modeling from the real velocity model, and m and n here represent the number of time sampling points and the number of receiver channels of the seismic data, respectively.

[0030] Further, step S4 specifically includes the following steps:

[0031] S41. Input the seismic data generated by the initial model through forward modeling into the full waveform inversion neural network based on physical perception collaborative fusion, and perform pre-training to optimize the neural network parameters;

[0032] S42. If the loss function threshold obtained during training does not reach the preset threshold standard, return to step S4, optimize and adjust the network structure or parameter configuration, and retrain.

[0033] S43. When the evaluation index of the training result meets the preset loss function threshold condition, terminate the pre-training process and apply the trained full waveform inversion neural network based on physical perception collaborative fusion to the real speed model.

[0034] Further, step S5 specifically includes the following steps:

[0035] S51. Input the seismic data corresponding to the real velocity model into the pre-trained full-waveform inversion neural network based on physical perception collaborative fusion to perform parameter fine-tuning and model application operations.

[0036] S52. If the performance evaluation index of the loss function obtained during the fine-tuning process does not reach the preset threshold standard, return to step S5, further optimize and adjust the network parameters, and re-execute the training process to continuously improve the model performance.

[0037] S53. When the evaluation index of the results obtained in the fine-tuning and application stages meets the preset loss function threshold requirements, the training process is terminated, and the final inversion speed model that meets the accuracy requirements is output.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. This invention combines the U-Net model to design the network structure, which enables the network to extract multi-scale seismic features while having stronger feature reconstruction ability and spatial information preservation ability compared with traditional convolutional neural network structures;

[0040] 2. This invention designs a phased training strategy of "pre-training + fine-tuning". This strategy, without relying on the construction of additional large-scale synthetic datasets, uses simulated data generated by existing initial models for initial training and performs parameter fine-tuning on measured observation data. This enables the neural network to balance inversion accuracy and generalization ability, effectively reducing data preparation costs and significantly alleviating the domain mismatch problem between synthetic data and measured data.

[0041] 3. This invention designs an adaptive residual learning module, which enhances the adaptability to observed data while maintaining the consistency of the model structure by introducing a learnable temporal feature mapping and dimension alignment mechanism, thereby effectively improving the prediction accuracy of the velocity model.

[0042] 4. This invention designs a physical perception collaborative fusion strategy, which embeds the seismic wave equation and geological prior information into the neural network training process, making the inversion results more consistent with the physical laws of the underground medium and improving the physical consistency and imaging reliability of the full waveform inversion. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A schematic diagram of a full-waveform inversion neural network structure based on physical perception collaborative fusion;

[0045] Figure 2 Schematic diagram of the adaptive residual learning module structure;

[0046] Figure 3 Flowchart of a full-waveform inversion neural network based on physical perception collaborative fusion;

[0047] Figure 4 Comparison of inversion results of the Marmousi model under different observation conditions;

[0048] Figure 5 Comparison of inversion results of the Overthrust model under different observation conditions. Detailed Implementation

[0049] The present invention will be further described below with reference to embodiments:

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0051] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0052] This invention was implemented on the Anaconda platform using the Python compiler, with the operating system Ubuntu 22.04LTS, a single NVIDIA GeForce RTX 4090 graphics card, and the PyTorch 2.5.1 deep learning framework. Its core is a proposed method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion. Due to the limitations of complex geological structures and the nonlinear characteristics of seismic wavefields, traditional full-waveform inversion methods often face local minima traps. Existing deep learning-based full-waveform inversion methods struggle to effectively integrate physical equation constraints with data-driven features. This invention designs a full-waveform inversion neural network based on physical perception collaborative fusion. By pre-training the neural network on an initial model, it retains the advantages of data-driven approaches while avoiding the dependence on synthetic datasets found in traditional methods. This method achieves collaborative optimization of the network architecture by introducing physical information gradients to synchronously update the parameters of the U-net network and the adaptive residual learning module, thereby effectively combining the advantages of physical constraints and data-driven approaches and significantly improving inversion performance. The invented network was tested using the Marmousi and Overthrust standard geological models. When the evaluation index of the test results meets the set index threshold, the trained full waveform inversion network based on physical perception collaborative fusion is taken as the optimal inversion network.

[0053] The present invention provides a method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion, comprising the following steps:

[0054] S1. Construct a full waveform inversion neural network for pre-training and fine-tune the required model data;

[0055] S11. Typical geological models such as Marmousi and Overthrust are selected as real velocity models to reflect the complex velocity structure of different underground geological layers. Through the velocity parameters set in these models, the characteristics of underground media that are closer to the real geological environment are simulated.

[0056] S12. Perform linear interpolation within the maximum and minimum velocity ranges and extract the overall median for the real velocity model used in step S11 to generate a linear initial model and a uniform initial model, so as to simulate the inversion environment where the initial model and the real model differ under real geological conditions.

[0057] S13. Using the forward modeling method, select the Ricker wavelet as the excitation source, and perform seismic wave propagation simulation on the real velocity model described in step S11 and the initial velocity model described in step S12 respectively, so as to obtain the corresponding real seismic data and initial model seismic data.

[0058] S14. The initial model seismic data and real seismic data obtained in step S13 are used as pre-training and fine-tuning stages, respectively, thus completing the entire process of constructing the dataset.

[0059] S2. Construct a full-waveform inversion neural network based on physical perception collaborative fusion for high-precision underground velocity model inversion.

[0060] The neural network consists of an encoder module, a decoder module, and a designed adaptive residual learning module within the U-Net structure. During the training phase, the acquired seismic data is simultaneously fed into both the U-Net encoder module and the adaptive residual module for feature extraction. The feature information extracted by the encoder module is then passed to the bottleneck module located in the middle of the network, and subsequently, it is upsampled and reconstructed layer by layer by the U-Net decoder module. The feature map output by the decoder module is then subjected to a 1×1 convolution operation to generate intermediate prediction results, which are then fused and summed with the output results from the adaptive residual module to generate the final predicted velocity model.

[0061] The encoder module consists of four encoding units, four max-pooling layers, eight 3×3 convolutional layers, activation functions, and normalization operations. Each encoding unit comprises two consecutive 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. The input feature map first enters the first encoding unit, where it undergoes two convolutional layers to extract initial features, followed by downsampling through a 2×2 max-pooling layer. Subsequently, the three encoding units are connected sequentially with the same structure, each group followed by a max-pooling operation after the convolutional module. In each layer, the convolutional operation increases the number of channels, while the max-pooling operation reduces the size of the feature map. The final output of the encoder module is the feature map obtained after four encoding and downsampling operations.

[0062] The bottleneck module consists of a set of two-layer convolutional structures and a residual connection. The input feature map first enters a 3×3 convolutional layer, which is equipped with a normalization layer and a ReLU activation function. Then, the feature map passes through a second 3×3 convolutional layer, which again combines a normalization layer and a ReLU activation function for nonlinear transformation. This module expands the channel dimension while keeping the original size of the spatial dimension unchanged, and introduces a residual connection with an identity mapping, so that the input features are added to the main branch output after matching in the channel dimension, thereby obtaining the output feature map of the bottleneck module.

[0063] The decoder module can perform upsampling, skip connection concatenation, and convolutional decoding. First, the input feature map is upsampled using transposed convolution, expanding its spatial size to match the output size of the previous encoder layer. Then, the upsampled feature map is concatenated with the feature map of the corresponding encoder layer along the channel dimension. The concatenated feature map is then input into a convolutional decoding unit consisting of two 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. These three steps are repeated four times in the decoder module to restore the spatial resolution of the feature map in sequence, thereby achieving a gradual reconstruction from high-dimensional, low-resolution features to a low-dimensional, high-resolution image. Finally, an upsampling operation is used to uniformly adjust the feature map to the target output size.

[0064] The adaptive residual learning module can perform shape transformation, zeroing, and parameter initialization. First, the input data is mapped along the time axis through a linear layer to reduce its dimensionality and extract temporal features. Then, redundant dimensions in the data are removed through dimensionality compression, and the data is copied and expanded along a specific axis to ensure its spatial shape is consistent with the target velocity model, achieving structural alignment. Based on this, the processed data is fed into a zeroing layer for initial value definition, and finally defined as learnable parameters. During training, these parameters can be jointly optimized by the U-Net network and the adaptive residual learning module using gradients containing physical information through gradient descent. After the two modules generate intermediate outputs, they are added element-wise along the spatial dimension to achieve collaborative feature fusion.

[0065] S3. Construct a loss function to measure the performance of the full waveform inversion neural network based on physical perception collaborative fusion;

[0066] Specifically, the loss function is an error metric function in the form of the L2 norm, i.e., the root mean square error loss function. This loss function is applied in both the pre-training and parameter fine-tuning stages to jointly guide the optimal convergence of the network; its corresponding mathematical expression is:

[0067]

[0068] In the pre-training phase, x(i,j) represents the predicted velocity model output by the neural network, y(i,j) is the initial velocity model constructed in step S1, and m and n represent the horizontal scale and depth range of the velocity model, respectively. In the parameter fine-tuning phase, x(i,j) represents the seismic data generated by forward modeling from the predicted velocity model, y(i,j) is the seismic data generated by forward modeling from the real velocity model, and m and n here correspondingly represent the number of time sampling points and the number of receiver channels for the seismic data.

[0069] S4. Neural Network Pre-training:

[0070] Based on the physical perception-based collaborative fusion full-waveform inversion neural network constructed in step S2, the neural network included in the method is pre-trained using seismic data obtained from the initial model through forward modeling. This aims to establish an effective mapping relationship between the seismic response and the target velocity model. The results obtained through this mapping are further used to optimize the loss function designed in step S3. Within the set training rounds, the neural network parameters are iteratively updated and adjusted until convergence to the optimal value, completing the pre-training process and saving the neural network training weights.

[0071] Specifically, it includes the following steps:

[0072] S41. Input the seismic data generated by the initial model through forward modeling into the full waveform inversion neural network based on physical perception collaborative fusion, and perform pre-training to optimize the neural network parameters;

[0073] S42. If the loss function threshold obtained during training does not reach the preset threshold standard, return to step S4, optimize and adjust the network structure or parameter configuration, and retrain.

[0074] S43. When the evaluation index of the training result meets the preset loss function threshold condition, terminate the pre-training process and apply the trained full waveform inversion neural network based on physical perception collaborative fusion to the real speed model.

[0075] S5, Network Fine-tuning and Applications:

[0076] After completing the pre-training of the neural network, the obtained full-waveform inversion neural network containing the physical sensing collaborative mechanism is fine-tuned and applied in practice. Seismic data from real models are fed into the neural network to generate corresponding velocity model prediction results, and its inversion performance is evaluated. If the fine-tuned neural network meets the expected standard on the set loss function threshold, the full-waveform inversion neural network based on physical sensing collaborative fusion updated in step S5 is confirmed as the final optimized scheme; if the threshold standard is not met, the training is re-fine-tuned, the neural network parameters are adjusted and optimized, and the retraining operation is performed until the optimization conditions are met and the fine-tuning training effect reaches the threshold standard.

[0077] Specifically, it includes the following steps:

[0078] S51. Input the seismic data corresponding to the real velocity model into the pre-trained full-waveform inversion neural network based on physical perception collaborative fusion to perform parameter fine-tuning and model application operations.

[0079] S52. If the performance evaluation index of the loss function obtained during the fine-tuning process does not reach the preset threshold standard, return to step S5, further optimize and adjust the network parameters, and re-execute the training process to continuously improve the model performance.

[0080] S53. When the evaluation index of the results obtained in the fine-tuning and application stages meets the preset loss function threshold requirements, the training process is terminated, and the final inversion speed model that meets the accuracy requirements is output.

[0081] Example 1

[0082] This embodiment provides an application of a high-precision velocity modeling and waveform inversion using a full-waveform inversion neural network based on physical perception collaborative fusion, as detailed below:

[0083] In this embodiment, real seismic observation data is input into a full-waveform inversion neural network based on physical sensing collaborative fusion, and the output velocity model is as follows: Figure 4 and Figure 5 As shown. Figure 4 Specifically, it includes: 1. Linear initial model and observation conditions containing Gaussian noise; 2. Linear initial model and observation conditions with missing low-frequency data; 3. Uniform initial model conditions. Figure 5 The conditions include: 1. Linear initial model and observation conditions containing Gaussian noise; 2. Linear initial model and observation conditions with missing low-frequency data; 3. Uniform initial model conditions.

[0084] Experimental results show that the velocity model obtained by this invention has higher accuracy and better conforms to the actual geological structure characteristics. Furthermore, under complex observation conditions such as Gaussian noise, missing low-frequency data, and a uniform initial model, the full-waveform inversion neural network based on physical perception fusion constructed in this invention still maintains robust inversion performance, verifying its reliability and potential in practical applications.

[0085] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion, characterized in that, Includes the following steps: S1. Construct the initial model of the full waveform inversion neural network for pre-training and fine-tuning the required model data; S2. Construct a full-waveform inversion neural network based on physical perception collaborative fusion for high-precision underground velocity model inversion; S3. Construct a loss function to measure the performance of the full waveform inversion neural network based on physical perception collaborative fusion; S4. Neural Network Pre-training: Based on the physical perception-based collaborative fusion full-waveform inversion neural network constructed in step S2, the neural network included in the method is pre-trained using the seismic data obtained from the forward modeling of the initial model to establish an effective mapping relationship between the seismic response and the target velocity model. The results obtained through this mapping are further used to optimize the loss function designed in step S3. Within the set training rounds, the neural network parameters are iteratively updated and adjusted until they converge to the optimal value, completing the pre-training process, and the neural network training weights are saved. S5, Network Fine-tuning and Applications: After completing the pre-training of the neural network, the parameters of the obtained full waveform inversion neural network containing the physical perception collaborative mechanism are fine-tuned and applied in practice. Seismic data from the real model are fed into the neural network to generate the corresponding velocity model prediction results and evaluate its inversion performance. If the fine-tuned neural network reaches the expected standard on the set loss function threshold standard, the full waveform inversion neural network based on physical perception collaborative fusion updated in step S5 is confirmed as the final optimization scheme. If the threshold standard is not met, the training is readjusted, the neural network parameters are adjusted and optimized, and the retraining operation is performed until the optimization conditions are met and the fine-tuning training effect reaches the threshold standard.

2. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Select typical geological models as real velocity models to reflect the complex velocity structure of different underground geological layers. Through the velocity parameters set in these models, simulate the characteristics of underground media that are closer to the real geological environment. S12. Perform linear interpolation within the maximum and minimum velocity ranges and extract the overall median for the real velocity model used in step S11 to generate a linear initial model and a uniform initial model, so as to simulate the inversion environment where the initial model and the real model differ under real geological conditions. S13. Using the forward modeling method, select the Ricker wavelet as the excitation source, and perform seismic wave propagation simulation on the real velocity model described in step S11 and the initial velocity model described in step S12 respectively, so as to obtain the corresponding real seismic data and initial model seismic data. S14. The initial model seismic data and the real model seismic data obtained in step S13 are used for the pre-training and fine-tuning stages, respectively, thus completing the entire process of constructing the dataset.

3. The method for constructing a full-waveform inversion neural network according to claim 1, characterized in that, Step S2 is as follows: The neural network consists of an encoder module, a decoder module, and a designed adaptive residual learning module in the U-Net structure. During the training phase, the collected seismic data is simultaneously fed into the encoder part of the U-Net and the adaptive residual module for feature extraction. The feature information extracted by the encoder module is then passed to the bottleneck module located in the middle of the network, and then upsampled and restored layer by layer through the U-Net decoder module. The feature map output by the decoder module is then subjected to a 1×1 convolution operation to generate an intermediate prediction result, which is then fused and summed with the output result from the adaptive residual module to generate the final predicted speed model.

4. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that: The encoder module consists of four encoding units, four max-pooling layers, eight 3×3 convolutional layers, activation functions, and normalization operations. Each encoding unit comprises two consecutive 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. The input feature map first enters the first encoding unit, where it undergoes two convolutional layers to extract initial features, followed by downsampling through a 2×2 max-pooling layer. Subsequently, the three encoding units are connected sequentially with the same structure, each group followed by a max-pooling operation after the convolutional module. In each layer, the convolutional operation increases the number of channels, while the max-pooling operation reduces the size of the feature map. The final output of the encoder module is the feature map obtained after four encoding and downsampling operations.

5. The method for constructing a full-waveform inversion neural network according to claim 1, characterized in that: The bottleneck module consists of a set of two-layer convolutional structures and a residual connection. The input feature map first enters a 3×3 convolutional layer, which is equipped with a normalization layer and a ReLU activation function. Then, the feature map passes through a second 3×3 convolutional layer, which again combines a normalization layer and a ReLU activation function for nonlinear transformation. This module expands the channel dimension while keeping the original size of the spatial dimension unchanged, and introduces a residual connection with an identity mapping, so that the input features are added to the main branch output after matching in the channel dimension, thereby obtaining the output feature map of the bottleneck module.

6. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that: The decoder module can perform sampling operations, skip connection concatenation, and convolutional decoding. First, the input feature map is upsampled using transposed convolution to expand its spatial size to the size of the output of the previous encoder layer. Then, the upsampled feature map is concatenated with the feature map of the corresponding layer of the encoder along the channel dimension. The concatenated feature map is then input into a convolutional decoding unit consisting of two 3×3 convolutional layers, two BatchNorm normalization layers, and two ReLU activation functions. The above three steps are repeated four times in the decoder module to restore the spatial resolution of the feature map in turn, thereby realizing the gradual reconstruction from high-dimensional low-resolution features to low-dimensional high-resolution images. Finally, an upsampling operation is used to uniformly adjust the feature map to the target output size.

7. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that: The adaptive residual learning module can perform shape transformation, zeroing, and parameter initialization respectively; first, the input data is mapped along the time axis through a linear layer to reduce its dimensionality and extract time features; Subsequently, redundant dimensions in the data are removed through dimensionality compression, and the data is copied and expanded along a specific axis to ensure its spatial shape aligns with the target velocity model, achieving structural alignment. Based on this, the processed data is fed into a zeroing layer for initial value definition, and finally defined as learnable parameters. During training, these parameters can be jointly optimized using gradients containing physical information via gradient descent to optimize the learnable parameters of the U-Net network and the adaptive residual learning module. After the two modules generate intermediate outputs, they are added element-wise along the spatial dimension to achieve collaborative feature fusion.

8. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that: In step S3, the introduced loss function is an L2 norm form error metric function, i.e., the root mean square error loss function. This loss function is applied in both the pre-training and parameter fine-tuning stages to jointly guide the optimal convergence of the network; its corresponding mathematical expression is: In the pre-training stage, x(i,j) represents the predicted velocity model output by the neural network, y(i,j) is the initial velocity model constructed in step S1, and m and n represent the horizontal scale and depth range of the velocity model, respectively. In the parameter fine-tuning stage, x(i,j) represents the seismic data generated by forward modeling from the predicted velocity model, y(i,j) is the seismic data generated by forward modeling from the real velocity model, and m and n here represent the number of time sampling points and the number of receiver channels of the seismic data, respectively.

9. The method for constructing a full-waveform inversion neural network according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41. Input the seismic data generated by the initial model through forward modeling into the full waveform inversion neural network based on physical perception collaborative fusion, and perform pre-training to optimize the neural network parameters; S42. If the loss function threshold obtained during the training process does not reach the preset threshold standard, return to step S4, optimize and adjust the network structure or parameter configuration, and retrain. S43. When the evaluation index of the training result meets the preset loss function threshold condition, terminate the pre-training process and apply the trained full waveform inversion neural network based on physical perception collaborative fusion to the real speed model.

10. The method for constructing a full-waveform inversion neural network based on physical perception collaborative fusion according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51. Input the seismic data corresponding to the real velocity model into the pre-trained full-waveform inversion neural network based on physical perception collaborative fusion to perform parameter fine-tuning and model application operations. S52. If the performance evaluation index of the loss function obtained during the fine-tuning process does not reach the preset threshold standard, return to step S5, further optimize and adjust the network parameters, and re-execute the training process to continuously improve the model performance. S53. When the evaluation index of the results obtained in the fine-tuning and application stages meets the preset loss function threshold requirements, the training process is terminated, and the final inversion speed model that meets the accuracy requirements is output.

Images