Three-dimensional geologic structure modeling method based on attention enhancement SinGAN

Abstract

The invention discloses a three-dimensional geologic structure modeling method based on attention enhancement SinGAN, and relates to the technical field of geologic modeling, and the three-dimensional geologic structure modeling method based on attention enhancement SinGAN mainly comprises the steps: carrying out the multi-scale down-sampling of a single training image and condition drilling data, and obtaining a multi-scale data pyramid; and training the constructed attention-enhanced conditional single image generative adversarial network of multiple stages, and performing conditional geological simulation on new conditional drilling data by using the trained attention-enhanced conditional single image generative adversarial network to obtain a high-resolution three-dimensional geological structure model. By implementing the three-dimensional geological structure modeling method based on the attention enhancement SinGAN provided by the invention, the precision and efficiency of multi-scale and conditional constraint geological simulation can be improved.

Description

Technical Field

[0001] This invention relates to the field of geological modeling technology, and more specifically, to a three-dimensional geological structure modeling method based on attention-enhanced SinGAN. Background Technology

[0002] Accurate three-dimensional geological structure models are a crucial foundation for groundwater flow simulation, pollutant transport prediction, and mineral resource assessment. However, the underground medium is highly heterogeneous and uncertain, and how to construct an accurate and geologically sound three-dimensional model under the constraint of sparse borehole data has always been a core challenge in this field.

[0003] Traditional geological modeling methods mainly fall into two categories: deterministic interpolation and stochastic geostatistics. Deterministic methods (such as Kriging) rely on assumptions of spatial continuity and stationarity, making it difficult to characterize complex spatial patterns of geological structures. Stochastic geostatistical methods, especially multi-point geostatistics (MPS), improve the simulation capability of complex geological patterns to some extent by introducing training images as prior knowledge. However, MPS methods (such as the DS algorithm) heavily depend on the quality and representativeness of the training images, and often face problems of high computational cost and overfitting when simulating non-stationary structures and handling large-scale patterns.

[0004] In recent years, deep learning-based generative models, such as Generative Adversarial Networks (GANs), have provided a new paradigm for geological modeling. These methods can learn complex spatial distributions in training images to generate high-resolution, diverse geological representations. However, existing GAN-based methods typically require a large number of training images, which is often difficult to meet in practical applications due to the high cost of acquiring representative training data. Furthermore, most existing methods perform feature learning and conditional control at a single scale, making it difficult to effectively integrate multi-scale geological information from global to local perspectives. This results in insufficient response to sparse borehole conditions and low computational efficiency. Existing single-image GAN models have demonstrated the potential for training and generation using only a single image, providing a solution to the data scarcity problem. However, their original architecture lacks an effective conditional control mechanism, making them difficult to directly apply to geological modeling tasks that require strict borehole constraints. Subsequent improved models have attempted to introduce conditional information, but often simply embed conditional data into the network or impose constraints only through loss functions, failing to achieve deep fusion of conditional information and image features across multiple scales, thus limiting the accuracy and efficiency of modeling. In summary, existing methods still have the following limitations: (1) Model training relies heavily on a large number of representative training images, and the cost of obtaining high-quality training images in real-world scenarios is high; (2) The condition control mechanism is simple and the multi-scale information fusion capability is insufficient, which leads to the unstable performance of the generated results in complex structural regions. (3) The network structure is complex, and gradient instability and pattern collapse are prone to occur during the training process, making it difficult to balance lightweight and high accuracy. (4) The limited ability to perceive and reconstruct key geological features restricts the practicality of the model in geological interpretation.

[0005] Therefore, there is an urgent need in this field for a technical solution that can simultaneously solve the following problems: achieve efficient and high-precision modeling of three-dimensional geological structures under the condition of scarce training images; fully integrate sparse but high-precision borehole condition data; and have multi-scale feature learning capabilities to capture geological structures from macro to micro. Summary of the Invention

[0006] The purpose of this invention is to provide a three-dimensional geological structure modeling method based on attention enhancement SinGAN, which can improve the accuracy and efficiency of multi-scale, conditionally constrained geological simulation.

[0007] The three-dimensional geological structure modeling method based on attention-enhanced SinGAN provided by this invention includes the following steps: S1: Perform multi-scale downsampling on a single training image and conditional borehole data to obtain a multi-scale data pyramid; S2: Based on the attention mechanism and generative adversarial network, a multi-stage attention-enhanced conditional single-image generative adversarial network is constructed; S3: The attention-enhanced conditional single-image generative adversarial network is trained using the multi-scale data pyramid to obtain a trained attention-enhanced conditional single-image generative adversarial network. S4: Using the trained attention-enhanced conditional single-image generative adversarial network, conditional geological simulation is performed on the new conditional borehole data to obtain a high-resolution three-dimensional geological structure model.

[0008] This invention also provides a three-dimensional geological structure modeling system based on attention-enhanced SinGAN, the system comprising the following modules: The data preprocessing module is used to perform multi-scale downsampling on single training images and conditional borehole data to obtain a multi-scale data pyramid. The network building module is used to construct a multi-stage attention-enhanced conditional single-image generative adversarial network based on attention mechanisms and generative adversarial networks. The model training module is used to train the attention-enhanced conditional single image generative adversarial network using the multi-scale data pyramid to obtain the trained attention-enhanced conditional single image generative adversarial network. The geological simulation reasoning module is used to perform conditional geological simulation on new conditional borehole data using the trained attention-enhanced conditional single-image generative adversarial network to obtain a high-resolution three-dimensional geological structure model.

[0009] The three-dimensional geological structure modeling method and system based on attention-enhanced SinGAN provided by this invention has the following beneficial effects: This invention addresses the limitations of traditional geostatistical methods, such as high computational cost, insufficient connectivity of simulation results, and strong dependence on training data, low efficiency of multi-scale feature fusion, and ineffectiveness of conditional constraints. Based on a parallel multi-stage generative adversarial network architecture and a hybrid attention mechanism, this invention proposes an attention-enhanced conditional single natural image generative adversarial network (GAN). The network (AE-CSinGAN) employs a parallel multi-stage generative adversarial network architecture. It uses a single training image and multi-scale borehole data as constraints to construct a multi-scale data pyramid. A parameter-sharing strategy enables cross-stage feature transfer, significantly reducing the dependence on the number of training images and achieving efficient and stable 3D modeling under single-image conditions. A hybrid attention mechanism is introduced into the generator, using adaptive weighting of channel and spatial dimensions to fuse channel attention and spatial attention. This enhances the perception and reconstruction of key geological features, improves the model's ability to model conditional data and prior knowledge, and increases the accuracy of depicting complex structures (such as non-stationary geological structures). The network is trained using a multi-scale data pyramid constructed under the constraints of a single training image and limited corresponding borehole data. Conditional data expansion and joint loss optimization are employed, utilizing the optimization of joint adversarial loss and conditional loss. The goal is to optimize weight parameters, strengthen the constraint effect of sparse borehole data, and ensure the consistency of attributes and geological rationality of the generated results at the conditional locations. The training process adopts gradient penalty and multi-stage initialization strategies to effectively alleviate the problems of pattern collapse and gradient instability, and improve the model convergence speed and generalization performance. After training, only one forward inference is required to directly use borehole data to achieve high-precision reconstruction of 3D underground structures that conform to geological patterns and connectivity characteristics, which greatly improves modeling efficiency. Experimental results show that the present invention can maintain high reconstruction accuracy at borehole locations under various geological scenarios, while ensuring geological continuity and structural authenticity. It outperforms traditional multi-point geostatistical methods in terms of structural similarity and root mean square error, and has higher computational efficiency. Using the present invention for multi-scale, conditionally constrained geological simulation, it can be widely applied in engineering applications such as hydrogeological simulation, resource exploration, and geological hazard assessment. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1This is a flowchart of the three-dimensional geological structure modeling method based on attention-enhanced SinGAN provided by the present invention; Figure 2 This is the overall architecture diagram of AE-CSinGAN provided by the present invention; Figure 3 This is a schematic diagram of the network structure of the generator provided by the present invention; Figure 4 This is a schematic diagram of the structure of the hybrid attention module provided by the present invention; Figure 5 This is a schematic diagram of the network structure of the discriminator provided by the present invention; Figure 6 This is a schematic diagram (XOY plane) of conditional data preprocessing provided by the present invention. Figure 7 These are all the datasets provided by this invention for experimental verification; Figure 8 This is a schematic diagram showing the conditional simulation results, statistical analysis, and error analysis of all datasets provided by this invention on the method of this invention and the traditional DS algorithm. Detailed Implementation

[0011] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0012] Figure 1 A schematic diagram of the attention-enhanced SinGAN-based 3D geological structure modeling method of this embodiment is shown. In this embodiment, the attention-enhanced SinGAN-based 3D geological structure modeling method includes the following steps: S1: Perform multi-scale downsampling on a single training image and conditional borehole data to obtain a multi-scale data pyramid; In one exemplary embodiment, the conditional borehole data includes borehole attribute data and borehole location data; the multi-scale data pyramid includes training images, borehole attribute data, and borehole location data.

[0013] S2: Based on the attention mechanism and generative adversarial network, a multi-stage attention-enhanced conditional single-image generative adversarial network is constructed; In one exemplary embodiment, the attention-enhanced conditional single-image generative adversarial network includes a generator and a discriminator. The generator is used to fuse multi-scale conditional data with training image features and focus on key geological features, and the discriminator is used to distinguish the generated result from the actual training image at the appropriate scale. As an exemplary embodiment, the output of the attention-enhanced conditional single image generative adversarial network is mapped to a probability value by an adaptive average pooling layer and a fully connected layer, and then normalized by a Sigmoid function to determine the authenticity of the input image; In one exemplary embodiment, the generator includes a first convolutional unit, a second convolutional unit, a bottleneck unit, a first deconvolutional unit, a second deconvolutional unit, and an output unit; The first and second convolutional units each include a convolutional layer, a batch normalization layer, a LeakyReLU activation function, and a hybrid attention module connected in sequence. The first and second deconvolution units each include a deconvolution layer, a batch normalization layer, a LeakyReLU activation function, and a hybrid attention module connected in sequence. The bottleneck unit includes a convolutional layer and a hybrid attention module; The output unit includes a deconvolutional layer and a Tanh activation function; In one exemplary embodiment, the generator of the first-stage attention-enhanced conditional single image generative adversarial network further includes a shared convolutional layer, which is used to extract features from Gaussian noise, borehole attribute data, and borehole location data. The extracted features are concatenated and used as input to the first-stage generator. In one exemplary embodiment, the input to the generator includes an upsampled version of the generator's output from the previous stage, Gaussian noise, and conditional borehole data at the current scale. In one exemplary embodiment, the hybrid attention module is used for dynamic weight allocation of key geological features, as shown in the formula: , in, This represents the output feature map of the hybrid attention module; This represents the input feature map of the hybrid attention module; This represents the channel attention mapping of the channel attention submodule; This represents the spatial attention mapping of the spatial attention submodule; Represents element-wise multiplication; In one exemplary embodiment, the hybrid attention module includes a channel attention submodule and a spatial attention submodule, which act on the channel dimension and spatial dimension of the feature, respectively, to achieve adaptive enhancement of cross-scale features; the channel attention submodule generates channel weight vectors through global average pooling and max pooling; the spatial attention submodule generates a spatial weight map through average pooling and max pooling in the channel dimension. In one exemplary embodiment, the discriminator includes multiple cascaded convolutional blocks, each convolutional block including a convolutional layer and a LeakyReLU activation function; S3: The attention-enhanced conditional single-image generative adversarial network is trained using the multi-scale data pyramid to obtain a trained attention-enhanced conditional single-image generative adversarial network. In one exemplary embodiment, the training includes a multi-stage training process, and the total loss function of the training is: , in, This is the total loss function; and They represent the first Adversarial losses and conditional losses in phased networks; and They represent the first The network parameters of the generator and discriminator corresponding to each stage; It is a hyperparameter used to weigh the importance of adversarial loss and conditional loss; In one exemplary embodiment, the training includes a multi-stage training process, and the loss functions corresponding to the discriminator and generator in each stage are as follows: , , , , in, For the first The loss function corresponding to the discriminator in each stage; Indicates from the latent distribution Mid-sampled latent variables Calculate the discriminator's expected output for the generated image; Indicates the first Stage discriminator; No. Stage generator; For the first Simulation implementation of the phase; Indicates the distribution of real data Mid-sampled training images Calculate the expected output of the discriminator for the real image; Indicates the first Training images for each stage; It is the coefficient of the penalty term; Indicates from interpolation distribution Samples from the middle Calculate the expectation of the gradient penalty term; Indicates the discriminator output For interpolated samples gradient calculation, This indicates that the discriminator evaluates the interpolated samples. The output; This represents mixed data consisting of real data and simulated implementations; It is an L2 norm; For the first The loss function corresponding to the generator of the stage; This represents the output of the generator; Indicates the first Gaussian noise at different stages; Indicates the first Drilling properties at each stage; Indicates the first Drilling location at each stage; Indicates the first The upsampling result output by the stage generator; Represents a random number between 0 and 1; , , These represent the true probability distributions of the training image, the latent space, and the mixed data, respectively. In one exemplary embodiment, during the conditional constraint process, the generator is trained and updated, and the corresponding conditional loss is: , in, For the first The corresponding conditional loss of the generator in each stage; Represents element-wise multiplication; In one exemplary embodiment, the conditional borehole data in the multi-scale data pyramid is preprocessed before being input into the attention-enhanced conditional single-image generative adversarial network. The preprocessing includes setting an expansion radius centered on each borehole point. A spherical neighborhood is defined, and the attribute values ​​of all voxels within this spherical neighborhood are set to be the same as those of the borehole points to enhance the constraint effectiveness and visualization effect of sparse conditional data in three-dimensional space. S4: Use the trained attention-enhanced conditional single image generative adversarial network to perform conditional geological simulation on the new conditional borehole data to obtain a high-resolution three-dimensional geological structure model. As an exemplary embodiment, in step S4, conditional geological simulation and multi-scale generation are performed: for new conditional borehole data, a multi-scale conditional pyramid is constructed; starting from the coarsest scale, the data is generated and upsampled step by step by the generator, and the upsampling results of the previous stage are input into the current generator along with the conditional data and Gaussian noise at the current scale, finally obtaining a high-resolution three-dimensional geological structure model. In one exemplary embodiment, the method further includes: changing the Gaussian noise, repeating step S4, and generating multiple equally probable geological realizations for uncertainty analysis.

[0014] This embodiment provides a three-dimensional geological structure modeling system based on attention-enhanced SinGAN, the system including the following modules: The data preprocessing module is used to perform multi-scale downsampling on single training images and conditional borehole data to obtain a multi-scale data pyramid. The network building module is used to construct a multi-stage attention-enhanced conditional single-image generative adversarial network based on attention mechanisms and generative adversarial networks. The model training module is used to train the attention-enhanced conditional single image generative adversarial network using the multi-scale data pyramid to obtain the trained attention-enhanced conditional single image generative adversarial network. The geological simulation reasoning module is used to perform conditional geological simulation on new conditional borehole data using the trained attention-enhanced conditional single-image generative adversarial network to obtain a high-resolution three-dimensional geological structure model.

[0015] In some embodiments, the above-described attention-enhanced SinGAN-based three-dimensional geological structure modeling method can also be implemented in the following ways.

[0016] In this embodiment, the three-dimensional geological structure modeling method based on attention-enhanced SinGAN includes the following steps: Step 1: Data Preparation and Multi-Scale Pyramid Construction: Acquire a single training image representing prior geological knowledge and sparse conditional borehole data; perform multi-scale downsampling on the training image and conditional borehole data to construct a multi-scale data pyramid containing the training image, borehole attribute data, and borehole location data. ,in This represents the total number of stages.

[0017] Step 2: Network Architecture Initialization: Building a network architecture that includes... The Attention-Enhanced Conditional Single Natural Image Generative Adversarial Network (AE-CSinGAN) consists of several stages, each containing a generator. and a discriminator Among them, generator A U-Net-based encoder-decoder structure is adopted, and a hybrid attention module is embedded to fuse multi-source inputs and focus on key geological features; the discriminator... Used to distinguish the generated result from the real training image at the corresponding scale.

[0018] The generator The structure includes: (1) Multi-source feature fusion part: For the first stage The generator extracts conditional drilling attributes through convolutional layers. Drilling location and Gaussian noise The features are then stitched together along the channel dimension; for subsequent stages The generator additionally incorporates upsampling of the results generated in the previous stage. ; (2) Feature extraction and reconstruction: A symmetrical U-Net structure is adopted, which includes a convolution module, a bottleneck module and a deconvolution module; wherein, the convolution module and the deconvolution module are both integrated with a hybrid attention module, which realizes dynamic weight allocation of key geological features through the cascade operation of channel attention and spatial attention.

[0019] The specific operation of the hybrid attention module is as follows: Input feature map sequentially through channel attention mapping Spatial attention mapping Its output feature map The calculation formula is: , in The element-wise multiplication is represented; the channel attention generates channel weight vectors through global average pooling and max pooling; the spatial attention generates a spatial weight graph through average pooling and max pooling along the channel dimension.

[0020] Step 3: Multi-stage adversarial training and parameter inheritance: from the coarsest scale To the finest scale The generator is trained in stages. With discriminator The training process uses a weighted sum of adversarial loss and conditional loss as the total loss function. ,in This is used to force the generated results to be consistent with the known attribute data at the borehole location; and after each stage of training is completed, its network parameters serve as the starting point for the initialization of network parameters in the next stage.

[0021] In the During the training phase, the total loss function can be expressed as:

[0022] in, and They represent the first Adversarial loss and conditional loss in phased networks They represent the first The parameters in the generator and discriminator networks corresponding to each stage; These are hyperparameters used to weigh the importance of adversarial loss and conditional loss. Experiments have shown that... The value of is negatively correlated with r.

[0023] Specifically, during adversarial training, the discriminator and generator are trained and updated sequentially; the corresponding loss functions are respectively, , , , , in, Represents a random number between 0 and 1. They represent the first The parameters in the generator and discriminator networks corresponding to each stage; Indicates the first Training images for each stage For the first The simulation implementation of the stage, Indicates the first Gaussian noise in the stage, Indicates the first Stage condition well location, Indicates the first Stage-specific well attributes; Indicates the first The upsampling result output by the stage generator; It is the coefficient of the penalty term. It is gradient calculation. It is element-wise multiplication. It is a mixture of real data and simulated data; , , These represent the true probability distributions of the training images, the latent space, and the mixed data, respectively.

[0024] During the constraint process, the generator is trained and updated. The corresponding loss function is: , in, It is a hyperparameter used to weigh the importance of adversarial loss versus conditional loss.

[0025] Step 4: Conditional Geological Simulation and Multi-Realization Generation: For new conditional borehole data, construct its multi-scale conditional pyramid; starting from the coarsest scale, generate and upsample step by step through the generator, and input the upsampling results of the previous stage, the conditional data of the current scale, and Gaussian noise into the current generator to finally obtain a high-resolution three-dimensional geological structure model; by changing the Gaussian noise, repeat this process to generate multiple equally probable geological realizations for uncertainty analysis.

[0026] It should be noted that conditional borehole data needs to be preprocessed before being input into the network. Specifically, an expansion radius is set with each borehole point as the center. The attribute values ​​of all voxels within the spherical neighborhood are set to be the same as those of the borehole point to enhance the constraint effectiveness and visualization effect of sparse conditional data in three-dimensional space.

[0027] This embodiment provides a three-dimensional geological structure modeling system, including: (1) Data preprocessing module, used to construct a multi-scale pyramid of training images and conditional borehole data; (2) Network construction module, used to initialize the generator and discriminator of the attention-enhanced conditional SinGAN; (3) Model training module, used to execute the multi-stage adversarial training and parameter inheritance process; (4) Geological simulation reasoning module, used to load new conditional borehole data and generate multiple three-dimensional geological structures.

[0028] In some embodiments, the above-described attention-enhanced SinGAN-based three-dimensional geological structure modeling method can also be implemented in the following ways.

[0029] This embodiment provides an attention-enhanced conditional single image generative adversarial network (AE-CSinGAN) for 3D underground structure modeling, which mainly consists of a multi-stage generator, a discriminator, and a joint loss optimization mechanism.

[0030] 1. The overall process of AE-CSinGAN; like Figure 2 As shown, in this embodiment, the AE-CSinGAN framework adopts a parallel multi-stage generative adversarial network structure, including... Training phases ( Each stage includes multiple generators. and a discriminator ( The generator is responsible for fusing multi-scale conditional data with training image features, while the discriminator is used to distinguish the generated results from the real training images.

[0031] During the training phase, the network input includes a single training image, borehole attribute data, and location data, with each stage's input constructed using a multi-scale pyramid. The generator progressively integrates conditional information and training image features, outputting a coarse-to-fine 3D geological representation that conforms to geological priors and conditional constraints, resulting in a 3D subsurface structure model. The discriminator distinguishes the generated result from the real image, optimizing the generator through adversarial learning.

[0032] During training, a cross-stage parameter sharing strategy is adopted: the generator and discriminator parameters of each stage are used as the initialization for the next stage to accelerate convergence and enhance training stability. Conditional data is embedded in each stage in a multi-scale pyramid format to ensure progressive feature fusion from coarse to fine.

[0033] Once training is complete, given new borehole data, a constrained 3D model can be generated through a single forward inference.

[0034] 2. Generator; Specifically, such as Figure 3 As shown, the first Phase by It consists of several cascaded generators, used to comprehensively acquire and fuse feature information from training images and conditional data across multiple scales.

[0035] To achieve efficient feature representation and reconstruction, each generator adopts a hierarchical U-Net symmetric encoder-decoder network structure, which includes two processes: encoding (downsampling) and decoding (upsampling). The encoding process uses multiple convolutional units to progressively downsample the input data to extract multi-scale structural features from coarse to fine. The decoding process uses deconvolutional units to gradually restore the spatial resolution, reconstructing a geological structure model consistent with the input scale.

[0036] The output of the generator at the previous scale is used as the prior input of the generator at the next scale, thereby realizing the progressive transfer and fusion of information between scales, enabling the network to learn and reconstruct a high-fidelity geological model at the original scale in a progressively refined manner.

[0037] Each generator It consists of two convolutional modules, one bottleneck module, two deconvolutional modules, and one output module. Both the convolutional and deconvolutional modules contain convolutional / deconvolutional layers, batch normalization layers, LeakyReLU activation functions, and hybrid attention layers. The bottleneck module extracts deep semantic features through convolutional layers and hybrid attention layers.

[0038] The generator input differs at different stages: In stage 1, the input includes Gaussian noise. Drilling properties With position Features are extracted through shared convolutional layers and then concatenated, which are then mapped by a generator to form the initial coarse-scale simulation result. .

[0039] In the stage( The input includes upsampling of the output from the previous stage. Gaussian noise Based on the current scale and conditional data, a refined result is generated after feature fusion. .

[0040] The generator internally enhances feature propagation through skip connections and residual connections, alleviates gradient vanishing, and improves detail recovery capabilities.

[0041] 3. Hybrid attention mechanism; like Figure 4 As shown, a hybrid attention module is introduced into the generator, which integrates channel attention and spatial attention mechanisms.

[0042] The hybrid attention module consists of a channel attention submodule and a spatial attention submodule, which act on the channel dimension and spatial dimension of the feature, respectively, to achieve adaptive enhancement of features across scales.

[0043] The channel attention submodule generates channel weights through global average pooling and fully connected layers, adaptively enhancing key feature channels. Specifically, the input features first undergo global pooling to extract channel-level statistics, and then generate channel weight coefficients through linear mapping and nonlinear activation functions. These coefficients are then weighted with the original features channel by channel, thereby highlighting important channel responses related to geological structures, interface morphology, etc., and achieving adaptive adjustment of the contribution of different channels.

[0044] The spatial attention submodule generates a spatial weight map through channel-dimensional pooling and convolution operations, focusing on important geological structural regions such as faults and bedding. Specifically, based on the saliency distribution of features in the spatial dimension, the spatial attention submodule constructs a spatial attention weight matrix to apply higher weights to geological interfaces, bedding boundaries, and key areas with structural abrupt changes, thereby enhancing the model's sensitivity to spatial details and local structural changes.

[0045] By combining channel attention and spatial attention, the input features are weighted by both channel and spatial factors. This effectively enhances the model's ability to capture key information during multi-scale feature learning and reconstruction, resulting in higher accuracy and stability in terms of structural boundaries, spatial continuity, and detail representation. This significantly improves the model's ability to discriminate and reconstruct geological features.

[0046] 4. Discriminator; like Figure 5 As shown, the discriminator employs a PatchGAN-based convolutional neural network structure, consisting of multiple cascaded convolutional blocks. Each convolutional block includes a convolutional layer and a LeakyReLU activation function, achieving feature downsampling through convolution.

[0047] The network output is mapped to probability values ​​through adaptive average pooling and a fully connected layer, and then normalized using the sigmoid function to determine the authenticity of the input image. The discriminator guides the generator optimization during adversarial training, improving the geological realism and structural consistency of the generated results.

[0048] In this embodiment, the discriminator consists of 5 convolutional blocks, with the number of feature channels increasing layer by layer from 64 to 512. After the input image is downsampled by convolution, the discrimination probability is output through adaptive parameters.

[0049] 5. Condition preprocessing; To enhance the constraint effect of sparse borehole data, this embodiment adopts a conditional data expansion strategy: centering on the borehole point and setting a radius... ( Typically, all voxels within a neighborhood of 2-4 are assigned the same attribute, thus expanding the spatial influence range of conditional data (e.g., ...). Figure 6 (As shown). This preprocessing improves the visibility of conditional data and the model's responsiveness, while avoiding hyperparameter sensitivity issues.

[0050] 6. Loss function and training; The loss function includes adversarial loss (with gradient penalty) and conditional loss, and hyperparameters. Take 10, Set the initial learning rate to 0.1. During training, use the Adam optimizer with the following initial learning rate: .

[0051] After research and repeated experiments, the optimization objective of AE-CSinGAN consists of adversarial loss and conditional loss. Therefore, in the first... During the training phase, the total loss function can be expressed as:

[0052] in, and They represent the first Adversarial loss and conditional loss in phased networks They represent the first The parameters in the generator and discriminator networks corresponding to each stage. These are hyperparameters used to weigh the importance of adversarial loss and conditional loss. Experiments have shown that... The value and It is negatively correlated.

[0053] The training employs an alternating update strategy: first, the generator is fixed while the discriminator is updated, then the discriminator is fixed while the generator is updated. Through multi-stage iterative optimization, the model gradually converges, generating high-precision, geologically consistent 3D structures.

[0054] Specifically, during adversarial training, the discriminator and generator are trained and updated sequentially. The corresponding loss functions are, respectively, , , , , in, Represents a random number between 0 and 1. They represent the first The parameters in the generator and discriminator networks corresponding to each stage. Indicates the first Training images for each stage For the first The simulation implementation of the stage, Indicates the first Gaussian noise in the stage, Indicates the first Stage condition well location, Indicates the first The conditional well attributes of the stage. Indicates the first The upsampling result output by the stage generator, It is the coefficient of the penalty term. It is gradient calculation. It is element-wise multiplication. It is a mixture of real data and simulated data. These represent the true probability distributions of the training images, the latent space, and the mixed data, respectively.

[0055] During the constraint process, the generator is trained and updated. The corresponding loss function is,

[0056] in, It is a hyperparameter used to weigh the importance of adversarial loss versus conditional loss.

[0057] 7. Dataset and Experimental Validation To verify the effectiveness of the AE-CSinGAN method, five experimental cases were designed and conducted. Case 1 and Case 2 are two-dimensional stationary river facies models (such as...). Figure 7(a) shown in the middle) and the two-dimensional non-stationary river facies model (as shown in the middle) Figure 7 As shown in (b), each model has a resolution of 250×250 pixels and includes two geological facies: yellow represents mudstone facies and purple represents sandstone facies. In the experiment, 15 borehole points were randomly selected from the original training images as conditional data for modeling analysis. Case 3 is a multi-valued non-stationary ice wedge model (e.g. Figure 7 As shown in (c), the size is 250×250 pixels, containing three geological facies: yellow for sandstone facies, blue for mudstone facies, and purple for siltstone-mudstone facies. Case 4 is a three-dimensional stable fold lithofacies model (e.g., Figure 7 As shown in (d), the size is 100×100×100 pixels, containing two geological facies: blue-green represents lithofacies, and orange-yellow represents the background. Case 5 is a three-dimensional non-stationary deltaic sedimentary deposit model (such as...). Figure 7 As shown in (e), the dataset is from an open-source dataset. Due to computational limitations, a 100×100×75 pixel sub-region was randomly cropped from the model as a training sample, where red represents river channel facies and blue represents natural levee facies. Case 4 and Case 5 each randomly selected 15 borehole points from the original image as conditional data. Meanwhile, the DS algorithm, a typical multi-point geostatistical simulation method based on a single training image, was selected as the benchmark for comparison. It should be emphasized that AE-CSinGAN is a conditional simulation method proposed based on unconditional SinGAN, and like the DS algorithm, it is based on a single training image for modeling. Therefore, choosing the DS algorithm as the comparison object is reasonable and comparable.

[0058] Regarding the quality of the generated data, training images of different types and dimensions from Case 1 to Case 5 were selected, and simulations were conducted under the same 15 conditions of borehole data constraints. For each simulation, one of the generated implementations was randomly selected for demonstration, and variation function analysis and error analysis were performed. For detailed simulation results and related comparative analyses, please refer to [link to relevant documentation]. Figure 8 .

[0059] As shown in Figure 8, both the AE-CSinGAN method and the DS algorithm can meet the conditional constraints with high accuracy. However, in terms of the overall structural features and spatial performance of the generative model, AE-CSinGAN significantly outperforms the DS algorithm in capturing global structural morphology, maintaining spatial connectivity, and reproducing non-stationary geological structures. Specifically, the DS algorithm tends to produce a large number of isolated pixels resembling noise when faced with a small template window and complex training image structures. This is because the DS algorithm matches and copies based on local templates. When there are multiple local regions with large structural differences but similar matching degrees in the training image, the algorithm may randomly select samples from these regions, resulting in problems such as local fragmentation and spatial discontinuity in the generated results. In contrast, the AE-CSinGAN method, by introducing a parallel multi-stage network structure and combining channel attention and spatial attention mechanisms, enables the model to automatically focus on key structural features in the training image, significantly reducing the interference of noise and unimportant information on the simulation results. In terms of spatial modeling, AE-CSinGAN adopts a top-down multi-level structure learning strategy, extracting coarse-grained and fine-grained features step by step, enabling the model to effectively capture the multi-scale structural features of training images and ensuring that the simulation results have higher quality in terms of spatial connectivity, overall consistency and boundary integrity.

[0060] Table 1 systematically compares the training time, inference time, and average generation time of the AE-CSinGAN method and the DS algorithm when generating different numbers of simulation results in 2D Case 3 and 3D Case 5. The results show that in 2D underground structure modeling tasks, the DS algorithm has a certain time advantage when generating a small number of simulations; however, as the scale of the modeling task increases, AE-CSinGAN outperforms the DS algorithm in terms of overall efficiency, generation quality, and stability. Furthermore, in 3D complex underground structure modeling scenarios with significant non-stationarity, AE-CSinGAN demonstrates significant advantages in generation quality, spatial continuity, and average generation speed, fully reflecting its applicability and robustness in high-dimensional modeling tasks.

[0061] Table 1: Time Consumption Comparison with the Comparison Method in Cases 3 and 5

[0062] In this embodiment, a hybrid attention mechanism is integrated into the network structure. Although this mechanism increases memory usage to some extent, it significantly improves model performance in terms of feature representation enhancement, cross-scale information interaction, and boundary structure sensitivity, thereby effectively accelerating the convergence speed of the model and improving the spatial consistency and structural reliability of the generated results.

[0063] It is worth noting that all experiments could be successfully completed in a GPU environment with 24GB of video memory. The complete training time for both the two-dimensional and three-dimensional cases was controlled within 90 minutes, further demonstrating the superiority of this invention in terms of resource utilization and training efficiency.

[0064] In summary, the method of this invention exhibits excellent performance in terms of training efficiency, computational resource utilization, quality of generated results, and accuracy of conditional constraints, and has broad application prospects and high engineering promotion value.

[0065] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A three-dimensional geological structure modeling method based on attention-enhanced SinGAN, characterized in that, Includes the following steps: S1: Perform multi-scale downsampling on a single training image and conditional borehole data to obtain a multi-scale data pyramid; S2: Based on the attention mechanism and generative adversarial network, a multi-stage attention-enhanced conditional single-image generative adversarial network is constructed; S3: The attention-enhanced conditional single-image generative adversarial network is trained using the multi-scale data pyramid to obtain a trained attention-enhanced conditional single-image generative adversarial network. S4: Using the trained attention-enhanced conditional single-image generative adversarial network, conditional geological simulation is performed on the new conditional borehole data to obtain a high-resolution three-dimensional geological structure model.

2. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 1, characterized in that, The attention-enhanced conditional single-image generative adversarial network includes a generator and a discriminator. The generator is used to fuse multi-scale conditional data with training image features and focus on key geological features. The discriminator is used to distinguish the generated result from the real training image at the appropriate scale.

3. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 2, characterized in that, The generator includes a first convolutional unit, a second convolutional unit, a bottleneck unit, a first deconvolutional unit, a second deconvolutional unit, and an output unit. Each of the first and second convolutional units includes a convolutional layer, a batch normalization layer, a LeakyReLU activation function, and a hybrid attention module connected in sequence. Each of the first and second deconvolutional units includes a deconvolutional layer, a batch normalization layer, a LeakyReLU activation function, and a hybrid attention module connected in sequence. The bottleneck unit includes a convolutional layer and a hybrid attention module. The output unit includes a deconvolutional layer and a Tanh activation function.

4. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 3, characterized in that, The generator of the first-stage attention-enhanced conditional single image generative adversarial network also includes a shared convolutional layer, which is used to extract features from Gaussian noise, borehole attribute data, and borehole location data. The extracted features are concatenated and used as the input to the generator of the first stage.

5. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 3, characterized in that, The generator's input includes an upsampled version of the generator's output from the previous stage, Gaussian noise, and conditional borehole data at the current scale.

6. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 3, characterized in that, The hybrid attention module is used for dynamic weight allocation of key geological features, as shown in the formula: ， in, This represents the output feature map of the hybrid attention module; This represents the input feature map of the hybrid attention module; This represents the channel attention mapping of the channel attention submodule; This represents the spatial attention mapping of the spatial attention submodule; This represents element-wise multiplication.

7. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 2, characterized in that, The discriminator comprises multiple cascaded convolutional blocks, each including a convolutional layer and a LeakyReLU activation function.

8. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 1, characterized in that, The training process includes multiple stages, and the total loss function of the training is: ， in, This is the total loss function; and They represent the first Adversarial losses and conditional losses in phased networks; and They represent the first The network parameters of the generator and discriminator corresponding to each stage; It is a hyperparameter used to weigh the importance of adversarial loss versus conditional loss.

9. The three-dimensional geological structure modeling method based on attention-enhanced SinGAN according to claim 1, characterized in that, The conditional borehole data in the multi-scale data pyramid is preprocessed before being input into the attention-enhanced conditional single-image generative adversarial network. The preprocessing includes setting an expansion radius centered on each borehole point. A spherical neighborhood is defined, and the attribute values ​​of all voxels within this spherical neighborhood are set to be the same as those of the borehole points to enhance the constraint effectiveness and visualization effect of sparse conditional data in three-dimensional space.

10. A three-dimensional geological structure modeling system based on attention-enhanced SinGAN, characterized in that, The system includes the following modules: The data preprocessing module is used to perform multi-scale downsampling on single training images and conditional borehole data to obtain a multi-scale data pyramid. The network building module is used to construct a multi-stage attention-enhanced conditional single-image generative adversarial network based on attention mechanisms and generative adversarial networks. The model training module is used to train the attention-enhanced conditional single image generative adversarial network using the multi-scale data pyramid to obtain the trained attention-enhanced conditional single image generative adversarial network. The geological simulation reasoning module is used to perform conditional geological simulation on new conditional borehole data using the trained attention-enhanced conditional single-image generative adversarial network to obtain a high-resolution three-dimensional geological structure model.

Images