A forward modeling prediction method for seismic wavefields based on a lightweight TransUNet network

Abstract

This invention discloses a seismic wavefield forward modeling prediction method based on a lightweight TransUNet network, relating to the fields of geophysical exploration and artificial intelligence. The method includes: acquiring a subsurface medium velocity model; inputting the velocity model into a lightweight TransUNet network, where the encoder extracts local geological structural features, a lightweight Transformer bottleneck layer models the global correlation features during the wavefield evolution process, and the decoder recovers the spatial distribution of the seismic wavefield at the target time; training the network using a composite loss function to establish a nonlinear mapping relationship between the subsurface medium velocity model and a two-dimensional seismic wavefield snapshot at the target time; and inputting the velocity model to be predicted into the trained model to obtain the corresponding seismic wavefield prediction result. This invention combines high prediction accuracy with computational efficiency, effectively reducing the computational overhead of traditional numerical forward modeling methods.

Description

Technical Field

[0001] This invention relates to the fields of geophysical exploration and artificial intelligence, and particularly to a seismic wavefield forward modeling prediction method based on a lightweight TransUNet network. This invention falls under the category of seismic numerical simulation and intelligent agent modeling technology, and is suitable for rapid prediction of seismic wavefield snapshots at target times corresponding to subsurface medium velocity models. It can be used for rapid wavefield estimation in scenarios such as seismic forward modeling, seismic data processing, complex structural imaging, and full-waveform inversion preprocessing. Background Technology

[0002] Forward modeling of seismic wavefields is a fundamental problem in geophysical exploration. Its core lies in solving the dynamic response process of seismic waves propagating underground based on subsurface medium parameters. Through forward modeling, the reflection, refraction, diffraction, and scattering behavior of seismic waves in layered media, fault structures, velocity anomalies, and complex stratigraphic interfaces can be analyzed, providing theoretical and data support for tasks such as migration imaging, reservoir prediction, velocity modeling, and full waveform inversion.

[0003] Traditional forward modeling methods for seismic wavefields mainly include numerical methods such as the finite difference method, finite element method, spectral element method, and pseudospectral method. These methods are typically based on the discrete solution of the wave equation, obtaining a full-time-history wavefield snapshot or synthetic record by progressively advancing the wavefield in the time and spatial domains. While these numerical methods have a clear physical basis and high solution accuracy, they incur significant computational overhead under complex seismic areas, fine-grid conditions, and long propagation times. Especially in large-scale seismic areas, multi-shot-point scenarios, or inversion iterative scenarios, traditional forward modeling methods often become the main computational bottleneck in the overall process.

[0004] In practical applications, the computational complexity of traditional numerical methods is typically influenced by factors such as grid scale, time step, boundary conditions, and medium complexity. To meet the requirements of numerical stability and simulation accuracy, smaller spatial and temporal sampling intervals are often required, which significantly increases the number of wavefield update steps, thus significantly increasing memory consumption and computation time. For tasks that require repeated forward modeling calls, such as parameter scanning, inversion optimization, or rapid field evaluation, traditional numerical simulation methods struggle to balance accuracy and efficiency.

[0005] In recent years, with the development of deep learning technology, establishing a nonlinear mapping relationship between subsurface medium models and seismic responses using neural networks has gradually become an important direction for accelerating seismic forward modeling research. This type of method typically represents the subsurface medium velocity model as a two-dimensional or three-dimensional mesh image and uses a target-time wavefield snapshot or multi-time wavefield sequence as the network output. It learns the mapping operator from medium parameters to wavefield response through a data-driven approach, thereby avoiding the repeated execution of the complete numerical discrete solution process during the inference stage.

[0006] Existing seismic wavefield prediction methods based on convolutional neural networks (CNNs) have shown good performance in modeling local features. Convolutional operations can effectively extract layer boundaries, anomaly contours, and local structural changes from velocity models, and have good expressive power for local features such as reflection interfaces. However, the receptive field expansion of pure convolutional networks usually relies on the stacking of convolutional layers and pooling operations, which is insufficient when dealing with long-distance propagation paths, complex wavefront phase relationships, and long-distance wave coupling, and makes it difficult to fully capture the globally dependent features in wavefield evolution.

[0007] The Transformer architecture has achieved widespread success in natural language processing and vision tasks due to its self-attention mechanism, which enables it to directly model long-range dependencies between different spatial locations. Introducing the Transformer into seismic wavefield prediction tasks helps enhance the network's ability to express global wavefront structure, propagation phase relationships, and cross-regional correlation features. However, the standard Transformer typically has high computational complexity and memory overhead on high-resolution feature maps, especially in geophysical tasks where input data often contains high-resolution spatial structures and complex texture features. Directly using the standard Transformer can easily lead to excessively large parameter counts, training difficulties, and decreased inference efficiency.

[0008] Therefore, how to construct a seismic wavefield forward modeling prediction method that can retain the efficient representation of local geological structures by convolutional networks, capture the wavelength propagation correlation of seismic waves using the Transformer mechanism, and simultaneously take into account model lightweighting, training stability, and inference efficiency has become an urgent technical problem to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of high computational cost, insufficient global dependency modeling capability of traditional convolutional networks, and heavy computational burden of standard Transformers in existing seismic wavefield forward modeling simulations. It provides a seismic wavefield forward prediction method based on a lightweight TransUNet network. This method extracts local structural features through a convolutional encoder, models global propagation relationships through a lightweight Transformer module, and then restores spatial resolution by combining a decoder and skip connections, so as to achieve efficient mapping from the subsurface velocity model to the wavefield snapshot at the target time.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A seismic wavefield forward modeling prediction method based on a lightweight TransUNet network includes the following steps: Step S1: Obtain the subsurface medium velocity model and construct a seismic wavefield training sample corresponding to the subsurface medium velocity model; the training sample includes the subsurface medium velocity model and a target time seismic wavefield snapshot.

[0012] Step S2 involves preprocessing the underground medium velocity model; the preprocessing includes standardization, size adaptation, and optional data augmentation to ensure that the input data meets the requirements for network training and inference.

[0013] Step S3: Input the preprocessed subsurface medium velocity model into a lightweight TransUNet network. The lightweight TransUNet network includes an encoder, a bottleneck layer, and a decoder. The encoder is used to extract local geological structure features, the bottleneck layer is used to extract global correlation features during wavefield propagation, and the decoder is used to restore spatial resolution step by step and output a wavefield snapshot at the target time.

[0014] Step S4: The lightweight TransUNet network is trained using training samples. The amplitude error and structural similarity between the predicted wavefield and the real wavefield are constrained by the composite loss function to obtain the trained seismic wavefield forward modeling prediction model.

[0015] Step S5: Input the subsurface medium velocity model to be predicted into the trained seismic wavefield forward model, and obtain the seismic wavefield prediction result at the corresponding target time through a single forward propagation.

[0016] Furthermore, the training samples in step S1 can be generated using a numerical forward modeling method, preferably based on discrete numerical solutions of the two-dimensional acoustic wave equation, to obtain wave field snapshots of different underground medium models at different time steps, thereby establishing a one-to-one correspondence between the input velocity model and the output wave field image.

[0017] The preprocessing in step S2 includes standardization, size adaptation, and data augmentation. Standardization reduces the impact of different velocity ranges on network training; this can be achieved by mapping the velocity model to a uniform numerical range or by standardization based on the mean and standard deviation. Size adaptation ensures the input size meets the requirements of multiple downsampling and upsampling operations in the network; this can be achieved through zero-padding, interpolation, or boundary expansion. Data augmentation can employ random horizontal flipping, random velocity perturbations, or subtle structural perturbations consistent with physical laws to improve the model's generalization ability to complex geological conditions.

[0018] The encoder consists of multiple convolutional modules, preferably residual convolutional blocks, to enhance feature extraction capabilities and improve gradient propagation performance during deep network training. During the downsampling stage, convolutional layers with a stride of 2 are preferred for downsampling the feature maps, thereby reducing spatial information loss caused by traditional max pooling and enhancing the model's ability to perceive wavefield reflection interfaces, velocity abrupt boundaries, and local anomalies.

[0019] The bottleneck layer employs a lightweight Transformer module. This module first flattens and linearly projects the deep feature map output by the encoder, and introduces positional encoding to preserve spatial location information. Then, it utilizes a multi-head self-attention mechanism to model long-range dependencies between different spatial locations, and extracts global high-level features through a feedforward neural network. Compared to the standard Transformer, this invention achieves lightweighting by reducing the number of encoding layers, lowering the embedding dimension, and reducing the number of attention heads, thereby reducing the number of model parameters and computational overhead. Furthermore, a regularization layer can be added after the sub-layers to suppress overfitting.

[0020] The decoder employs a combination of stepwise upsampling and convolutional reconstruction to restore the wavefront image resolution, with bilinear interpolation and convolutional fusion structure being the preferred approach. To preserve shallow details from the encoding stage, skip connections corresponding to the encoder layers are introduced in each stage of the decoder, fusing shallow features with deep semantic features to improve the recovery of wavefront details, weak reflection events, and local structural edges.

[0021] In step S4, a composite loss function is used to optimize the network parameters. This composite loss function includes at least the mean absolute error loss and the structural similarity loss. The mean absolute error loss is used to constrain global amplitude deviation, while the structural similarity loss is used to constrain wavefront morphology, phase continuity, and structural consistency. By weighting and combining these two types of losses, the fidelity of the prediction results in terms of energy distribution and spatial structure can be improved simultaneously.

[0022] The AdamW optimization algorithm is preferred for updating network parameters to improve training stability and parameter convergence efficiency. In practice, it can also be combined with learning rate adjustment strategies, model weight preservation strategies, and validation set evaluation mechanisms to obtain a better-performing model after training.

[0023] When making predictions for large-scale seismic activity areas, a sliding window block prediction method can be used. Specifically, the large-scale velocity model is divided into multiple overlapping sub-regions, each of which is input into a trained seismic wavefield forward model for inference. The prediction results of each sub-region are then smoothly stitched together to obtain the wavefield prediction results for the complete activity area. Preferably, cosine weighting or other smooth weighting methods are used in the overlapping regions to reduce stitching artifacts and discontinuities at block boundaries.

[0024] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention adopts a hybrid architecture that combines convolutional neural networks with lightweight Transformers. This retains the advantages of convolutional networks in extracting local geological structures while enhancing the model's ability to express long-range dependencies in wavefield propagation, thereby improving the ability to predict complex wavefront structures.

[0025] Secondly, by designing the Transformer module in a lightweight manner, this invention effectively controls the number of model parameters and memory usage, enabling the model to maintain high prediction accuracy while having better training and inference efficiency, making it more suitable for engineering applications.

[0026] Third, by constructing a composite loss function that includes mean absolute error loss and structural similarity loss, this invention enables the model to focus not only on amplitude error, but also on the consistency of wave field morphology and phase structure, thereby improving the physical rationality and structural fidelity of wave field prediction results.

[0027] Fourth, by introducing a sliding window block prediction and overlapping region smooth stitching strategy, the present invention makes the method applicable not only to fixed-size samples, but also to wave field prediction tasks in a larger area, thus having strong engineering scalability.

[0028] Fifth, the method proposed in this invention can serve as an efficient supplement to traditional numerical forward modeling methods, and has good application value in scenarios where rapid prediction of the wave field response at a target time is required. Attached Figure Description

[0029] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0030] Figure 1 This is a schematic diagram of the overall framework of the seismic wavefield forward modeling prediction method based on the lightweight TransUNet network in an embodiment of the present invention; Figure 2 This is a flowchart of a seismic wavefield forward modeling prediction method based on a lightweight TransUNet network, as described in an embodiment of the present invention. Figure 3 This is a schematic diagram of the lightweight TransUNet network structure in an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of the lightweight Transformer bottleneck layer structure in an embodiment of the present invention.

[0032] Figure 5 This is a schematic diagram of large-scale work area sliding window prediction and splicing in an embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] like Figure 1 , Figure 2 As shown, this embodiment provides a seismic wavefield forward modeling prediction method based on a lightweight TransUNet network, specifically including the following steps: Step S1: Obtain the subsurface medium velocity model and construct a seismic wavefield training sample corresponding to the subsurface medium velocity model; the training sample includes the subsurface medium velocity model and a target time seismic wavefield snapshot; wherein, the subsurface medium velocity model can be generated by artificially constructing a layered model, fault model, velocity anomaly model or on the basis of an existing seismic velocity model, covering different dip angles, velocity contrasts and structural complexities, so as to enhance the diversity of the sample.

[0036] In this embodiment, the training samples are constructed based on the numerical forward modeling results of the two-dimensional acoustic wave equation. Preferably, the two-dimensional acoustic wave equation is discretized and solved on a regular grid to obtain wavefield snapshots at different time steps. To reduce the influence of boundary reflections on the samples, absorbing boundary conditions are introduced at the boundary of the computational domain, preferably using perfectly matched layer absorbing boundary conditions. The seismic source can be a Ricker wavelet to simulate common source forms in exploration.

[0037] In this embodiment, the input data includes two parts: a velocity model and corresponding wavefield data. The original size of the velocity model can be set to 70×70, and the wavefield data can consist of two-dimensional snapshots at multiple time steps, such as wavefield samples with 98 time steps, each time step being 70×70 in size. The data type can be uniformly set to 32-bit floating-point numbers. All data can be organized according to sample index, time step index, and spatial grid position to facilitate subsequent training calls and label matching.

[0038] Step S2: Preprocess the underground medium velocity model; To meet the input requirements of the lightweight TransUNet network, the original velocity model and target wavefield samples need to be preprocessed. First, the velocity model is standardized. Standardization aims to reduce the dimensional influence caused by different velocity ranges and improve training stability. Specifically, minimum-maximum normalization can be used to map velocity values ​​to the [0,1] interval; or mean-variance normalization can be used to map samples to a zero-mean, unit-standard-deviation distribution. In this embodiment, interval normalization is preferred to simplify the engineering implementation process.

[0039] Secondly, the sample size is adapted. Since the network contains multiple downsampling operations, the input size is preferably divisible by a power of 2. Based on the implementation method of this embodiment, the original 70×70 velocity model can be expanded to 80×80 using a zero-padding method to meet the size requirements of the encoder's multi-layer downsampling structure; the corresponding wave field snapshot is also simultaneously adjusted to the same size to ensure consistent input and output spatial dimensions.

[0040] Furthermore, data augmentation can be applied to the training samples to improve the model's adaptability to complex geological structures and velocity perturbations. Preferred data augmentation methods include random horizontal flipping, slight velocity perturbations, local smoothing perturbations, or geometric perturbations consistent with the rationality of the geological structure. In this embodiment, random horizontal flipping and small-amplitude random velocity perturbations can be used for augmentation. For dataset partitioning, the training set, validation set, and test set can be divided in an 8:1:1 ratio to balance the number of training samples and the stability of model evaluation.

[0041] For wavefield labels, optional amplitude transformations can be performed based on amplitude attenuation characteristics. For example, logarithmic gain or other amplitude compression methods can be used to reduce training imbalance caused by excessively small amplitudes in deep wavefields. However, this step is not a necessary feature of this invention, and in some embodiments, the original wavefield amplitude can be learned directly.

[0042] Step S3: Input the preprocessed underground medium velocity model into the lightweight TransUNet network; like Figure 3As shown, this embodiment uses a lightweight TransUNet network as the backbone network for wavefield prediction. The network consists of three parts: an encoder, a lightweight Transformer bottleneck layer, and a decoder.

[0043] The encoder's function is to extract multi-scale local features from the input velocity model, including layer boundaries, velocity abrupt change surfaces, fault edges, lens contours, and local velocity anomaly information.

[0044] In this embodiment, the input layer receives a single-channel tensor of shape (N, 1, 80, 80), where N is the batch size. The encoder can consist of four convolutional modules, with the number of feature channels increasing progressively at each level. Preferably, the first level has 64 channels, and subsequent levels increase to 128, 256, and 512 channels respectively. Each convolutional module preferably uses a residual convolutional block structure to enhance feature representation and improve gradient propagation during deep training.

[0045] Each residual convolutional block can contain two 3×3 convolutional layers with a stride of 1 and padding of 1, along with a batch normalization layer and a non-linear activation function. The residual branch directly adds the input features to the convolutional transformation result, and then outputs the result after passing it through the activation function, thereby mitigating the gradient degradation problem in deep network training.

[0046] During the downsampling stage, it is preferable to use convolutional layers with a stride of 2 to downsample the feature map, rather than traditional max pooling layers. The advantage of using convolutional downsampling is that while reducing the spatial size, it can retain more information sensitive to reflection events, refraction events, and structural boundaries in a learnable manner, avoiding excessive loss of local structural features caused by pooling operations.

[0047] After multiple levels of convolution and downsampling, the input velocity model is encoded into a deep feature map with stronger semantic information. In this embodiment, the encoder can output a high-dimensional feature tensor of size 10×10×512 for subsequent processing by the lightweight Transformer module.

[0048] like Figure 4 As shown, the bottleneck layer, located between the encoder and decoder, is a key component of this invention for enhancing global modeling capabilities. Its purpose is to overcome the limitations of traditional convolutional receptive fields, explicitly modeling long-distance dependencies between different spatial locations, thereby more accurately representing global phase correlations and cross-regional coupling characteristics in wavefield propagation.

[0049] In this embodiment, the deep feature map output by the encoder is first flattened and linearly projected to generate a sequence vector suitable for Transformer processing, and positional encoding is added to preserve spatial location information. Subsequently, the sequence is input into multiple lightweight Transformer encoding layers, each of which includes a multi-head self-attention sublayer and a feedforward neural network sublayer, along with residual connections and layer normalization.

[0050] Considering that the standard Transformer has a high number of layers, embedding dimension, and heads, which leads to a significant increase in the number of parameters and computational complexity, this embodiment is designed to be lightweight. Preferably, the number of Transformer encoding layers can be set to 4, the embedding dimension can be set to 512, and the number of multi-head attention heads can be set to 8, so as to balance feature representation capability and computational complexity.

[0051] In some implementations, a Dropout layer can be introduced after the Transformer sublayer to suppress overfitting. The Dropout probability can be set to 0.1 to improve the model's generalization performance on small to medium-sized training datasets.

[0052] The function of the decoder is to restore the high-level semantic features output by the bottleneck layer to the target spatial resolution step by step, and finally output the wave field snapshot at the target time.

[0053] In this embodiment, the decoder can perform upsampling by combining bilinear interpolation and convolutional reconstruction. Bilinear interpolation is simple to implement, has low computational overhead, and provides smooth spatial recovery, making it suitable for lightweight model deployment. After each level of upsampling, it is preferable to further fuse features and restore local details through convolutional layers.

[0054] To enhance wavefront detail recovery, this invention introduces a skip connection structure symmetrical to the encoder in the decoder. Specifically, the shallow high-resolution features output from the corresponding layer of the encoder are concatenated with the features of the current layer of the decoder along the channel dimension, and then fused through a convolutional layer. This approach supplements shallow spatial edge information during the recovery process, helping to preserve the detailed representation of wavefront boundaries, layer reflections, and local weak events.

[0055] The decoder outputs a single-channel wavefield prediction image, corresponding to the two-dimensional wavefield amplitude matrix at the target time. The output activation method can be selected according to the range of wavefield data values. In some implementations, linear output can be used, or an appropriate normalized activation function can be used to adapt to the training label range.

[0056] Step S4: Train the lightweight TransUNet network using training samples; The lightweight TransUNet network was trained using the aforementioned training samples. The training objective was to enable the network to learn the nonlinear mapping relationship from the underground velocity model to the target time-time wavefield snapshot.

[0057] In this embodiment, training employs a composite loss function for optimization. This composite loss function includes at least mean absolute error (MAE) loss and structural similarity loss. The MAE loss constrains the deviation between the predicted and true wavefields in point-by-point amplitude; the structural similarity loss constrains the consistency between the predicted results and the true wavefields in local texture, wavefront morphology, and phase continuity at the image structure level. The weighted combination of these two losses helps to simultaneously improve prediction accuracy and structural fidelity.

[0058] Regarding specific parameter settings, the weight coefficient λ can be adjusted according to the training task. In this embodiment, λ can be set to around 0.7 to achieve a balance between amplitude accuracy and structural continuity in the model.

[0059] Regarding the optimization algorithm, this embodiment preferably uses the AdamW optimizer to update network parameters. AdamW combines adaptive learning rate and explicit weight decay mechanisms, which helps improve training stability and convergence speed. Based on the parameter settings of this embodiment, the initial learning rate can be set to 1×10^-4, the number of training epochs to 150, and the batch size to 2.

[0060] In some implementations, a learning rate adjustment strategy may also be used. For example, cosine annealing learning rate scheduling may be employed, or linear warm-up may be introduced in the early stages of training to improve early training stability. Considering that these training strategies are more of an engineering optimization technique than a necessary limitation of the method of this invention, they are described in the specification as implementation methods and not as essential features of this invention.

[0061] During training, the combined loss can be calculated using the validation set after each training cycle, and the weights of the models with better performance on the validation set can be saved. If the validation set loss no longer decreases within several consecutive rounds, an early stopping strategy can be used to terminate training to avoid overfitting.

[0062] Step S5: Input the subsurface medium velocity model to be predicted into the trained seismic wavefield forward model, and obtain the seismic wavefield prediction result at the corresponding target time through a single forward propagation. After model training is complete, the subsurface medium velocity model to be predicted is preprocessed in the same manner as during the training phase and then input into the trained lightweight TransUNet model. The model can output the two-dimensional wavefield prediction result corresponding to the target time through a single forward propagation.

[0063] Unlike traditional finite difference methods that require multi-time-step discrete propagation, this invention does not require solving the wave equation step by step during the inference stage. Instead, it directly uses the learned mapping operator to complete the target wave field snapshot prediction, thereby significantly shortening the inference time.

[0064] In actual testing, the lightweight TransUNet model can output a snapshot of the target wavefield relatively quickly, and the prediction results have high consistency with the real wavefield in terms of wavefront position, energy intensity and propagation pattern, thus demonstrating that the method can balance prediction efficiency and structural fidelity.

[0065] like Figure 5 As shown, when the actual velocity model size of the work area is larger than the fixed input size of the network, this invention can use a sliding window method for block prediction. Specifically, the large-size velocity model is divided into multiple overlapping sub-regions, each sub-region having the same or compatible size as the network input size; wave field prediction is performed on each sub-region separately; finally, the prediction results corresponding to all sub-regions are reassembled according to their spatial positions to obtain the target time wave field prediction result for the complete work area.

[0066] To reduce stitching artifacts that may appear at block boundaries, this embodiment sets a certain overlap ratio between adjacent windows, for example, around 25%; within the overlapping area, cosine smoothing weighting or other smoothing weighting methods are used for fusion. This method enables adjacent prediction blocks to transition smoothly near the boundary, reduces artificial edge effects, and improves the continuity of the entire wavefield prediction image.

[0067] This embodiment achieves rapid mapping from the subsurface medium velocity model to a target-time wavefield snapshot through the aforementioned data construction, network design, training, and inference processes. Verification shows that the lightweight TransUNet model can output the target wavefield snapshot quickly and maintains high consistency with numerical forward modeling results in terms of wavefront position, energy distribution, and propagation morphology.

[0068] The proposed method maintains good structural recovery capabilities for layered models, models with tilt variations, models containing faults, and models with velocity anomalies. Especially in scenarios requiring simultaneous modeling of local boundary effects and global propagation relationships, the synergistic effect of the convolutional encoder and the lightweight Transformer bottleneck layer enhances the representation of complex wavefield events.

[0069] It should be noted that the parameters involved in the above embodiments, such as input size, number of channels, number of Transformer layers, embedding dimension, number of attention heads, loss function weights, learning rate, number of training rounds, batch size, and window overlap ratio, can all be adjusted according to different data scales, hardware conditions, and task requirements, and should not be construed as limiting the scope of protection of this invention.

[0070] Without altering the basic technical concept of this invention, the above embodiments can be further extended. For example, in some embodiments, training strategies such as learning rate preheating and cosine annealing can be further introduced during the training phase to improve model convergence speed and training stability.

[0071] For example, in some implementations, richer data augmentation strategies can be applied to the velocity model, including but not limited to local smoothing perturbations, slight velocity gradient perturbations, and random fine-tuning of geological boundaries, to further improve the model's generalization ability under different tectonic scenarios.

[0072] For example, in some embodiments, the model can also be used to provide approximate wavefield prediction results for subsequent seismic inversion, rapid initial value assessment, or real-time response analysis. However, these applications are extensions of the method of the present invention and are not necessary limitations on the basic technical solution of the present invention.

[0073] In summary, this invention provides a seismic wavefield forward modeling prediction method that combines good prediction accuracy with high computational efficiency through the collaborative design of a convolutional encoder, a lightweight Transformer bottleneck layer, and a decoder, combined with a composite loss training mechanism suitable for wavefield images. This method can improve inference speed while maintaining structural representation capabilities, and has promising prospects for engineering applications.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A seismic wavefield forward modeling prediction method based on a lightweight TransUNet network, characterized in that, Includes the following steps: Obtain a subsurface medium velocity model and construct a seismic wavefield training sample corresponding to the subsurface medium velocity model; the training sample includes the subsurface medium velocity model and a target time snapshot of the seismic wavefield. A seismic wavefield forward modeling prediction model based on a lightweight TransUNet network was constructed and trained using training samples. The seismic wavefield forward modeling prediction model takes a preprocessed subsurface medium velocity model and the corresponding wavefield snapshot as input, and extracts the local geological structure features and global wavefield evolution features of the subsurface medium velocity model. A nonlinear mapping relationship is established between the subsurface medium velocity model and the target time seismic wavefield snapshot, outputting the corresponding two-dimensional seismic wavefield amplitude matrix at the target time. The lightweight TransUNet network includes an encoder layer, a bottleneck layer, and a decoder. The encoder consists of multi-layer convolutional modules used to extract local geological structural features. The bottleneck layer is a lightweight Transformer module used to capture global correlation features in the long-distance evolution of the wavefield. The lightweight Transformer module includes a multi-head self-attention sublayer and a feedforward neural network sublayer. By reducing the number of Transformer encoder layers and / or optimizing the multi-head attention dimension, the number of network parameters and computational overhead are reduced. The decoder is used to progressively restore the spatial resolution and output the target time wavefield snapshot. The subsurface medium velocity model to be predicted is input into the trained seismic wavefield forward model, and the corresponding seismic wavefield prediction result is obtained through a single forward propagation.

2. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 1, characterized in that, Constructing seismic wavefield training samples corresponding to the underground medium velocity model includes: performing discrete numerical solutions based on the two-dimensional acoustic wave equation to obtain wavefield snapshots of different underground medium models at different time steps, thereby establishing a one-to-one correspondence between the input velocity model and the output wavefield image.

3. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 1, characterized in that, Also includes: The underground medium velocity model is preprocessed to obtain network input data; The preprocessing of the underground medium velocity model includes: The obtained underground medium velocity model is standardized. The size of the underground medium velocity model is adjusted to match the input scale of the lightweight TransUNet network; The training data is augmented, and the augmentation process includes at least one of horizontal flipping and random perturbation.

4. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 1, characterized in that, The seismic wavefield forward modeling prediction model takes the preprocessed subsurface medium velocity model as input and extracts the local geological structural features and global wavefield evolution features of the subsurface medium velocity model, including: An encoder is used to extract features from the input underground medium velocity model; Input the deep features output by the encoder into the bottleneck layer; A decoder is used to upsample and reconstruct the output features of the bottleneck layer step by step to restore the spatial resolution of the seismic wavefield. In the decoder feature recovery process, skip connections are introduced between the encoder and decoder to fuse shallow detail features and deep semantic features.

5. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 4, characterized in that, The encoder is composed of residual convolutional blocks, and in the downsampling stage, a convolutional layer with a stride of 2 is used to downsample the feature map.

6. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 4, characterized in that, A nonlinear mapping relationship is established between the subsurface medium velocity model and the target time seismic wavefield snapshot, outputting the corresponding two-dimensional seismic wavefield amplitude matrix at the target time, including: Linear projection is performed on the feature map output by the encoder, and position encoding is introduced; Input the feature sequence after linear projection into the lightweight Transformer module; A regularization layer is set after the sub-layer of the lightweight Transformer module to suppress model overfitting, and the processed features are passed to the decoder for feature reconstruction.

7. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 1, characterized in that, A seismic wavefield forward modeling prediction model based on a lightweight TransUNet network was trained using training samples, including: Construct a training sample set, which includes a subsurface medium velocity model and its corresponding seismic wave field snapshot at the target time; A composite loss function is used to train the lightweight TransUNet network. The composite loss function includes at least the mean absolute error loss and the structural similarity loss. The composite loss function is a weighted combination of the mean absolute error loss and the structural similarity loss, which is used to simultaneously constrain the amplitude error of the predicted wavefield and the wavefront structural similarity. Iteratively update the network parameters; The trained seismic wavefield forward model is obtained based on the changes in the validation set loss.

8. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 7, characterized in that, The network parameters are iteratively updated, including using the AdamW optimization algorithm to update the network parameters and using a learning rate adjustment strategy to complete network training.

9. The seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 1, characterized in that, For the seismic wavefield prediction task of large-scale work area, the underground medium velocity model is divided into multiple overlapping sub-regions, which are then input into the trained seismic wavefield forward model for prediction. The prediction results of each sub-region are then stitched together to obtain the seismic wavefield prediction results of the complete work area.

10. A seismic wavefield forward modeling prediction method based on a lightweight TransUNet network according to claim 9, characterized in that, When stitching together the prediction results of sub-regions, a smooth weighted fusion strategy is adopted for overlapping areas to reduce prediction discontinuities at the stitching edges.

Images