Fine representation method for complex geologic body based on dual-cycle generative adversarial network

Abstract

The application discloses a complex geologic body fine characterization method based on a double-loop generative adversarial network, and realizes fine-grained characterization of attribute information through a conditional variational autoencoder generative adversarial network, and introduces a conditional latent regression generative adversarial network, forces and fully utilizes the hidden encoding information, and embeds the hidden encoding information into a generation process. In order to avoid many-to-one mapping and generation mode collapse and other problems, the double-loop generative adversarial network total framework is formed through coupling, the mapping parameters are randomly extracted in a low-dimensional vector space, so that the output distribution mode is effectively modeled and described, and the uncertainty in the modeling process is fully quantified through the double-loop mapping relationship established in the input mode, the hidden space and the output mode. Finally, in order to avoid the problem of attribute information characterization confusion, a joint loss function is defined in the training process to optimize different network parameters, so as to further improve the network simulation performance.

Description

Technical Field

[0001] This invention relates to the field of geological modeling, and in particular to a method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network. Background Technology

[0002] The detailed construction of three-dimensional geological models can clearly depict and describe the geometric morphology, structural trend, and stratigraphic sequence characteristics of geological units, thereby assisting geological experts in making further risk assessment decisions and qualitative and quantitative evaluations.

[0003] Different 3D geological modeling techniques have been widely applied in various fields, playing a crucial indicative role in describing reservoir spatial property laws, numerical simulation of groundwater flow, 3D spatial orebody geometric morphology characterization, and resource reserve estimation. As exploration deepens, underground spatial structures become more complex and diverse, with different geological structural units exhibiting significant heterogeneity and anisotropy. Limited by geological observation data and expert knowledge, traditional modeling results are not only inaccurate and prone to large errors, but also fail to fully quantify the uncertainties of the modeling process. Therefore, stochastic modeling methods have gradually been introduced into the field of 3D modeling. These methods can establish the probability density distribution function or covariance function of regionalized variables, enabling stochastic simulation and real-time reproduction of the structural laws or attribute characteristics of underground space in the study area. By obtaining various simulation results, they assist practitioners in further model optimization and uncertainty assessment.

[0004] Among these, some stochastic modeling methods based on numerical simulation, such as target-based simulation or multi-point geostatistical stochastic simulation, have been widely applied in various scenarios. However, most numerical simulation methods are based on the assumption of stationarity, resulting in poor simulation performance for large-scale, high-precision, and highly non-stationary geological bodies. Therefore, stochastic modeling methods for geological bodies based on deep generative models have gradually attracted widespread attention. These methods can overcome the shortcomings of numerical simulation methods by extracting and fitting the latent distribution patterns of the training dataset through the superposition of multi-layer neural network units, generating simulation results similar to the training dataset, thereby enabling efficient and automatic reconstruction of geological models.

[0005] Among numerous deep generative models, Generative Adversarial Networks (GANs) have gained widespread application in porous media reconstruction, reservoir history fitting, and hydrological numerical simulation due to their adversarial and game-theoretic learning characteristics and powerful fitting and generation capabilities. GANs primarily consist of a generator (G) and a discriminator (D). G learns high-order statistical features of the training dataset to generate similar samples, while D distinguishes the probability of whether the input data is true. Through continuous iterative training, the two reach a "Nash equilibrium," at which point G can randomly generate samples similar to the training dataset, thus enabling rapid and automatic reconstruction of geological models.

[0006] For GANs' stochastic modeling methods for geological bodies, there are two main categories: unconditional simulation and conditional simulation. Compared to unconditional simulation, conditional simulation automatically constructs geological models through given soft and hard data constraints. This not only reduces the uncertainty of the simulation process but also ensures that the simulation results have more geological regularity and background significance.

[0007] However, conditional data is mostly stored in the form of pixels, voxels, or two-dimensional planes. There are certain differences between the data dimension and its type and the corresponding simulation results, making it difficult to establish a mapping relationship between the two. This results in simulation results that are difficult to conform to the given conditional data constraints. Although a deep generative network can be pre-trained to produce results similar to the training dataset, and Markov chain Monte Carlo or structural loss functions can be introduced to search for the optimal vector space to ultimately realize the conditional simulation process, this post-processing method is strictly limited by the distribution pattern of the conditional points. Once new data is introduced, the entire process needs to be updated, resulting in inflexible selection of conditional points and low computational performance. Meanwhile, based on the pix2pix network architecture, multiple training pairs combining conditional data and prior models are established, using the input conditional data as input and outputting the corresponding simulation results. However, its drawback is that given a specific input, it can only generate a deterministic output. Furthermore, the embedding of random noise cannot interfere with the random generation process, and G cannot identify the corresponding perturbation characteristics. As training progresses, this noise is gradually filtered out, at which point the network easily degenerates into a denoising model, causing the generation mode to collapse and failing to fully quantify the uncertainties in the geological modeling process.

[0008] Furthermore, for complex and diverse continuous attribute spaces, the attribute field of each regional variable exhibits a smooth transition within its finite element neighborhood, and adjacent regional variables maintain certain correlations. This places higher demands on the conditional simulation process. Due to the sparse distribution and weak constraint of conditional data, the corresponding output results contain a large amount of local pixel noise and artifacts, leading to chaotic geological attribute field simulation and poor network simulation performance. Summary of the Invention

[0009] To address the problems of inflexible condition point selection, local pixel noise, insufficient diversity representation, and easy collapse of generation patterns in the existing random simulation and automatic construction of geological bodies, this invention introduces a conditional variational autoencoder GAN (cVAE-GAN) to achieve fine-grained representation of attribute information.

[0010] Specifically, this invention provides a method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network.

[0011] The method includes the following steps:

[0012] S1: Establish a prior geological model knowledge base for the study area; for each benchmark model B in the knowledge base, establish corresponding conditional data A to form training pairs (A, B), and all training pairs form a dataset; divide the dataset into a training set and a test set;

[0013] S2: Construct the BicycleGAN network; the BicycleGAN network includes: cLR-GAN network and cVAE-GAN network;

[0014] The cLR-GAN network includes: a U-Net generator G, a first multi-scale discriminator D1, and an encoding function E;

[0015] The cVAE-GAN network includes: a U-Net generator G, a second multi-scale discriminator D2, and an encoding function E;

[0016] S3: Initialize and set the hyperparameters of the BicycleGAN network;

[0017] S4: Train the BicycleGAN network using the training set, optimize the hyperparameters, and obtain the trained network;

[0018] S5: Test the trained network using the test set to obtain the final network, and use the final network to complete the fine characterization of complex geological models.

[0019] The beneficial effects provided by this invention are:

[0020] (1) In this invention, the latent space is reparameterized to adjust and improve its internal statistical features. By randomly extracting mapping parameters in the low-dimensional space, it can effectively characterize and diversely express the output distribution features, thereby realizing the multimodal image-to-image translation problem and preventing the many-to-one mapping problem between the latent code and the output during training. Multiple objectives are combined to encourage bijective consistency between latent variables and output results, thus strengthening their correlation. By perturbing the latent variables and embedding them into the generation process, the diversity of the simulation results is ensured, while also maintaining their similarity to the prior model.

[0021] (2) In this invention, a Conditional Variational Autoencoder Generative Adversarial Network (cVAE-GAN) and a Conditional Latent Regression Generative Adversarial Network (cLR-GAN) are coupled to form the BicycleGAN framework. cVAE-GAN introduces a variational strategy, which encodes the baseline image into a latent vector and reparameterizes it, ensuring that the generator can vaguely "peek" at the desired output, thereby achieving fine-grained representation of attribute information. cLR-GAN fully utilizes the latent encoded information and embeds it into the generation process, while using the encoding function to restore the consistency between the latent distribution and the real-time test distribution. Overall, the input-output flow of both can be summarized as follows: and This establishes a dual-loop strategy among the input mode, the latent space, and the output mode.

[0022] (3) Due to the different data dimensions between conditional data and simulation results, it is difficult to establish a graph-to-graph translation task, and the mapping relationship cannot be accurately expressed. Therefore, the U-Net architecture is used to transform the dimensions of the conditional data, thereby establishing a mapping relationship between the conditional data and the simulation results. The encoding function based on the residual network can effectively solve the ambiguity of the output pattern and avoid the collapse of the generation pattern by randomly sampling its latent space. The discriminator based on the PatchGAN architecture focuses on the characterization and expression of local details by judging the correlation between plates by outputting a multi-scale scalar field, thereby better assisting the reconstruction process of the generator. Under the condition of coupling of the three networks, the network parameters are optimized by using a joint loss function, so that the simulation results not only meet the given conditional constraints, but also maintain a spatial distribution pattern similar to the prior model, thereby realizing the automatic reconstruction and fine representation of complex geological models. Attached Figure Description

[0023] Figure 1This is a schematic diagram of the method flow of the present invention;

[0024] Figure 2 This refers to the overall structure of the BicycleGAN network;

[0025] Figure 3 This is the reconstruction result of a continuous attribute ice wedge hydrogeological model in a two-dimensional case;

[0026] Figure 4 It is a quantitative evaluation result of the ice wedge model reconstruction process;

[0027] Figure 5 It is the reconstruction result of a three-dimensional continuous property lithofacies-fold structure model;

[0028] Figure 6 It is a quantitative evaluation result of the lithofacies-fold model reconstruction process. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0030] Before the formal description, the overall structure of the present invention is summarized as follows.

[0031] This invention first proposes to introduce a conditional variational autoencoder GAN (cVAE-GAN) to achieve fine-grained representation of attribute information.

[0032] Specifically, the real image B is input into the encoding function E, and the latent space distribution Q(z|B) is output.

[0033] Q(z|B) is reparameterized, randomly sampled from it, concatenated with the conditional data A, and then fed into G to obtain the simulation results. Its overall process can be recorded as follows:

[0034] Furthermore, to address the problem of generative pattern collapse, building upon previous research, we propose introducing a Conditional Latent Regression GAN (cLR-GAN). This network forces G to fully utilize and embed latent information, while simultaneously using E to attempt to restore the consistency between the latent distribution and the real-time test distribution. Specifically, we first randomly sample noise from a Gaussian distribution and feed it along with A into G to obtain...

[0035] Then The data is fed into E, and finally a latent vector is obtained. Overall, it can be recorded as

[0036] Finally, cVAE-GAN and cLR-GAN are coupled, some network parameters are shared by weights, and bijective consistency is established among the input pattern, output pattern and latent space. The overall framework is named BicycleGAN.

[0037] Simultaneously, a joint loss function needs to be defined to optimize the network parameters of BicycleGAN, so that the simulation results not only meet the given conditional data constraints, but also maintain the difference from the prior model, thereby realizing the rapid and automatic reconstruction of complex geological models under conditional data constraints.

[0038] Please refer to Figure 1 and Figure 2 , Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is the overall structure of the BicycleGAN network.

[0039] Specifically, it mainly includes two sub-frameworks: Conditional Variational Autoencoder Generative Adversarial Network (cVAE-GAN) and Conditional Latent Regression Generative Adversarial Network (cLR-GAN). The generator adopts the U-Net architecture, including downsampling and upsampling processes, exhibiting a mirror-symmetric structure overall. It can also pass feature maps through skip connections to avoid information loss. The discriminator adopts the PatchGAN architecture, performing multi-scale discrimination by outputting multiple scalar fields. The encoding function based on residual connections can compress the image into a latent space distribution and participate in the random generation process.

[0040] This invention provides a method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network, comprising the following steps:

[0041] S1: Establish a prior geological model knowledge base for the study area; for each benchmark model B in the knowledge base, establish corresponding conditional data A to form training pairs (A, B), and all training pairs form a dataset; divide the dataset into a training set and a test set;

[0042] Combining geological data and multi-source heterogeneous geological data from the study area, a data augmentation strategy was employed to characterize the distribution patterns of underground space, establishing a corresponding geological prior model knowledge base to support the training and validation of subsequent process network models. The knowledge base mainly comprises multiple pairs, each containing a set of conditional data and a corresponding prior model.

[0043] It should be noted that the knowledge base represents one possible distribution pattern of the corresponding underground space. Simultaneously, for each baseline model B within the knowledge base, corresponding conditional data A is established. The type of A includes borehole data, cross-section data, or probabilistic volume data, etc.

[0044] S2: Construct the BicycleGAN network; the BicycleGAN network includes: cLR-GAN network and cVAE-GAN network;

[0045] Specifically, the Bicyc1eGAN network structure includes: a U-Net generator G, two multi-scale discriminators D1 and D2, and an encoding function E; where G and E share weights in cVAE-GAN and cLR-GAN, while D1 and D2 perform discriminative analysis on different input results in the two sub-networks respectively. In subsequent processes, the four networks will be called in a certain order to optimize the parameters respectively.

[0046] Correspondingly, the cLR-GAN network includes: a U-Net generator G, a first multi-scale discriminator D1, and an encoding function E;

[0047] Correspondingly, the cVAE-GAN network includes: a U-Net generator G, a second multi-scale discriminator D2, and an encoding function E;

[0048] S3: Initialize and set the hyperparameters of the BicycleGAN network;

[0049] Define the numerical values ​​of some hyperparameters, mainly including: total training epochs, batch size, and random noise dimension z. dim KL loss term weighting coefficient λ KL λ, the weighting coefficient of the reconstruction loss term rec The weighting coefficient λ of the latent variable loss term latent The weighting coefficient λ of the conditional loss term con With learning rate lr;

[0050] S4: Train the BicycleGAN network using the training set, optimize the hyperparameters, and obtain the trained network;

[0051] First, begin training the discriminators D1 and D2 in cVAE-GAN and cLR-GAN by performing steps (4-1) to (4-4):

[0052] (4-1) Inputting B into E yields a latent distribution Q(z|B). Q(z|B) is then reparameterized; the reparameterization formula can be summarized as follows:

[0053]

[0054] Among them, g θ As latent variables, ε and x (i) ... Let p(*) represent the expected value of the corresponding variable. After random sampling within the space of the reparameterized Q(z|B), this sample is concatenated with A along the channel direction and used as the input to G to obtain the simulation results.

[0055] (4-2) B and The data are fed into D1 respectively, and two sets of group labels, label_1 and label_0, are defined, which are considered as real labels and fake labels respectively. During the discrimination process, B and... The discrimination result D1(B) and The difference between D1 and label_1 and label_0 should be minimized as much as possible, so that D1 can better distinguish between real samples and generated samples. The discriminator loss is calculated using the mean squared error loss, which can be summarized mathematically as follows:

[0056]

[0057] Where m represents the total number of training samples;

[0058] (4-3) Randomly sample a noise vector, denoted as z, from a Gaussian distribution as a random latent vector with dimension z. dim The concatenated tensor of A and z is input into G to obtain the simulation results. Using the same method as in step (4-2), combine B and Input them into D2 respectively, and calculate the mean square error of the corresponding discrimination result compared with label_1 and label_0. This can be expressed mathematically as:

[0059]

[0060] (4-4) will and The losses of D1 and D2 are added together to calculate the total loss, and algorithms such as gradient zeroing, backpropagation, and gradient update are used to update the network parameters of D1 and D2.

[0061] (5) Fix the parameters of D1 and D2, and start training cVAE-GAN and cLR-GAN for G and E training and parameter optimization. You can perform steps (5-1) to (5-6):

[0062] (5-1) Input B into E to obtain Q(z|B). Using the same method as in step (4-1), Q(z|B) is reparameterized and randomly sampled. The corresponding result is embedded into the generation process along with A to obtain the simulation result.

[0063] (5-2) will Input into D1 and calculate the mean square error between the corresponding discrimination result and label_1. This can be summarized mathematically as follows:

[0064]

[0065] if The smaller the value, the higher the similarity between the reconstruction result and the prior model;

[0066] (5-3) Randomly sample a noise z from a Gaussian distribution, with dimension z. dim The simulation results are obtained by concatenating z and A and embedding them into the generation process. Then Input the data into D2 and calculate the mean square error between the corresponding discrimination result and label_1. This can be summarized mathematically as follows:

[0067]

[0068] (5-4) An additional conditional loss term is defined, which is expressed in the following form:

[0069] loss con =||M⊙(G(A, z)-B)||1

[0070] Here, M represents a binary mask tensor, which marks conditional points as 1 and non-conditional points as 0. It can quickly filter out the differences between the attribute information corresponding to the conditional point positions between the simulation results and the prior model, thereby ensuring the accuracy of the reconstruction of local information and making the simulation results conform to the given constraints.

[0071] (5-5) Reset the KL divergence of Q(z|B) to satisfy The corresponding calculation result is KL-div. Then, calculate B and... Between The loss can be expressed by the formula:

[0072]

[0073] in, represents the expected value, and p(*) represents the corresponding distribution characteristic.

[0074] (5-6) will losscon, KL-div and loss rec The values ​​of G and E are added together to calculate the overall loss value, and the network parameters corresponding to G and E are optimized through operations such as gradient zeroing, backpropagation, and gradient update.

[0075] (6) Fix the parameters of E and train G separately in cVAE-GAN and cLR-GAN. Specifically, randomly sample a noise z from a Gaussian distribution, and denote the corresponding dimension as z. dim Then, A and z are concatenated and input into G, and the corresponding output is input into E again to calculate the relationship between the latent distribution and z. Loss, in its form, can be written as:

[0076]

[0077] Then, gradient zeroing, backpropagation, and gradient update are performed separately on G.

[0078] (7) Repeat steps (4) to (6) to train BicycleGAN multiple times. When the loss function converges to a certain interval, stop the training process, save the parameters of different network models, and use the test dataset to verify the simulation performance.

[0079] (8) For the testing process, firstly, a generator needs to be initialized and the saved parameter file needs to be loaded. Then, a pair is randomly selected from the test set, and the conditional data is selected. At the same time, a noise vector is randomly sampled from the Gaussian distribution, and it is concatenated with the conditional data along the channel direction and input into G. Qualitative evaluation is carried out by observing the similarity between the simulation results and the prior model, as well as the accuracy of the reconstruction of the phase attribute information of the conditional point location. In addition, it is still necessary to quantitatively evaluate the simulation performance of the network in terms of spatial variability, phase proportion statistical characteristics, and spatial structure similarity. If the simulation performance is good, it can be used for subsequent random simulation and automatic reconstruction of geological body models; otherwise, it is necessary to return to step 2, reset the range of hyperparameter values, and start a new round of training process for BicycleGAN.

[0080] S5: Test the trained network using the test set to obtain the final network, and use the final network to complete the fine characterization of complex geological models.

[0081] Please refer to Figures 3-6 .

[0082] Figure 3 It is the reconstruction result of a continuous attribute ice wedge hydrogeological model in a two-dimensional case.

[0083] Figure 4 This is a quantitative evaluation result of the ice wedge model reconstruction process. Among them, Figure 4 (a) is the spatial variability analysis between different simulation results and the reference model; Figure 4 Figures (b) and (4)c represent the statistical characteristics of the proportion between multiple sets of simulation results and the reference model, presented in the form of histograms and box plots, respectively. Figure 4 (d) and Figure 4 (e) The MS-SWD distance between multiple sets of simulation results and the reference model was calculated and visualized in a two-dimensional rectangular coordinate system in the form of MDS, including scatter plots and contour plots.

[0084] Figure 5 It is the reconstruction result of a three-dimensional continuous property lithofacies-fold structure model.

[0085] Figure 6 This is a quantitative evaluation result of the lithofacies-fold model reconstruction process. Among them, Figure 6 (a) is the spatial variability analysis between different simulation results and the reference model; Figure 6 (b) and Figure 6 (c) Represents the statistical characteristics of the proportion between multiple sets of simulation results and the reference model, presented in the form of histograms and box plots, respectively; Figure 6 (d) and Figure 6 (e) The MS-SWD distance between multiple sets of simulation results and the reference model was calculated and visualized in a two-dimensional rectangular coordinate system in the form of MDS, including scatter plots and contour plots.

[0086] The beneficial effects of this invention are:

[0087] (1) In this invention, the latent space is reparameterized to adjust and improve its internal statistical features. By randomly extracting mapping parameters in the low-dimensional space, it can effectively characterize and diversely express the output distribution features, thereby realizing the multimodal image-to-image translation problem and preventing the many-to-one mapping problem between the latent code and the output during training. Multiple objectives are combined to encourage bijective consistency between latent variables and output results, thus strengthening their correlation. By perturbing the latent variables and embedding them into the generation process, the diversity of the simulation results is ensured, while also maintaining their similarity to the prior model.

[0088] (2) In this invention, a Conditional Variational Autoencoder Generative Adversarial Network (cVAE-GAN) and a Conditional Latent Regression Generative Adversarial Network (cLR-GAN) are coupled to form the BicycleGAN framework. cVAE-GAN introduces a variational strategy, which encodes the baseline image into a latent vector and reparameterizes it, ensuring that the generator can vaguely "peek" at the desired output, thereby achieving fine-grained representation of attribute information. cLR-GAN fully utilizes the latent encoded information and embeds it into the generation process, while using the encoding function to restore the consistency between the latent distribution and the real-time test distribution. Overall, the input-output flow of both can be summarized as follows: and This establishes a dual-loop strategy among the input mode, the latent space, and the output mode.

[0089] (3) Due to the different data dimensions between conditional data and simulation results, it is difficult to establish a graph-to-graph translation task, and the mapping relationship cannot be accurately expressed. Therefore, the U-Net architecture is used to transform the dimensions of the conditional data, thereby establishing a mapping relationship between the conditional data and the simulation results. The encoding function based on the residual network can effectively solve the ambiguity of the output pattern and avoid the collapse of the generation pattern by randomly sampling its latent space. The discriminator based on the PatchGAN architecture focuses on the characterization and expression of local details by judging the correlation between plates by outputting a multi-scale scalar field, thereby better assisting the reconstruction process of the generator. Under the condition of coupling of the three networks, the network parameters are optimized by using a joint loss function, so that the simulation results not only meet the given conditional constraints, but also maintain a spatial distribution pattern similar to the prior model, thereby realizing the automatic reconstruction and fine representation of complex geological models.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network, characterized in that: The method includes the following steps: S1: Establish a priori geological model knowledge base for the study area; For each baseline model B in the knowledge base, establish corresponding conditional data A to form a training pair (A, B), and all training pairs form a dataset; divide the dataset into a training set and a test set; S2: Construct the BicycleGAN network; The BicycleGAN network includes: cLR-GAN network and cVAE-GAN network; The cLR-GAN network includes: a U-Net generator G, a first multi-scale discriminator D1, and an encoding function E; The cVAE-GAN network includes: a U-Net generator G, a second multi-scale discriminator D2, and an encoding function E; S3: Initialize and set the hyperparameters of the BicycleGAN network; S4: Train the BicycleGAN network using the training set, optimize the hyperparameters, and obtain the trained network; S5: Test the trained network using the test set to obtain the final network, and use the final network to complete the fine characterization of complex geological models; The types of conditional data A include: borehole data, cross-section data, and probability volume data; The training process for the U-Net generator G and the encoding function E is as follows: S51, For the first multi-scale discriminator Second multi-scale discriminator Finally, the parameters obtained from training are fixed. S52. Input the baseline model B into the encoding function E to obtain a hidden distribution. ;right Perform reparameterization, and after reparameterization After randomly sampling within the space, it is compared with the corresponding conditional data. The network is spliced ​​along the direction of the channel to form a U-Net generator. The input is used to obtain the first simulation result. ; S53, The first simulation result Feed into the first multi-scale discriminator In the discrimination process, a loss function is used. The first simulation result The discrimination result and The differences between them should be as small as possible; among them For group tags; S54. Randomly sample a noise from a Gaussian distribution, denoted as . As a random latent vector, its dimension is ;Conditional data and random latent vectors The concatenated hybrid tensor is input into the U-Net generator G to obtain the second simulation result. ; S55, The second simulation results Feed into the second multi-scale discriminator In the discrimination process, a loss function is used. This makes the second simulation result The discrimination result and The differences between them should be as small as possible; S56. Define an additional conditional loss term, which is expressed in the following form: in, This represents a binary mask tensor, where conditional points are marked as 1 and unconditional points are marked as 0. S57, Reset The KL divergence, such that it satisfies The corresponding calculation result is ; calculate again and Between The loss can be expressed as a formula: in, Represents mathematical expectation, This represents the corresponding distribution characteristics; S58, will , , , and Add them together and calculate. and The overall loss value is calculated, and the corresponding U-Net generator G and encoding function are optimized through operations such as gradient zeroing, backpropagation, and gradient update. Network parameters.

2. The method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network as described in claim 1, characterized in that: The knowledge base represents a distribution pattern of underground space in the study area.

3. The method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network as described in claim 1, characterized in that: The hyperparameters include: total training phase Batch training set size Random noise dimension KL loss term weighting coefficient Reconstruction loss term weighting coefficient Latent variable loss term weighting coefficient and the weighting coefficient of the conditional loss term With learning rate .

4. The method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network as described in claim 3, characterized in that: In step S4, the training process of the BicycleGAN network includes: Training of the first multi-scale discriminator D1 and the second multi-scale discriminator D2; Training of the U-Net generator G and encoding function E; Secondary training of the U-Net generator G.

5. The method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network as described in claim 4, characterized in that: The specific training process for the first multi-scale discriminator D1 and the second multi-scale discriminator D2 is as follows: S41. Input the baseline model B into the encoding function E to obtain a hidden distribution. ;right Perform reparameterization, and after reparameterization After randomly sampling within the space, it is compared with the corresponding conditional data. The network is spliced ​​along the direction of the channel to form a U-Net generator. The input is used to obtain the first simulation result. ; S42, Using the benchmark model And the first simulation results They are fed into the first multi-scale discriminator respectively. In the middle, two groups of group tags are defined simultaneously. and These are treated as real labels and fake labels, respectively, and the mean squared error loss function is used in the discrimination process. Make the benchmark model Comparison with simulation results The discrimination result and respectively with and The differences between them should be as small as possible; S43. Randomly sample a noise from a Gaussian distribution, denoted as . As a random latent vector, its dimension is ;Conditional data and random latent vectors The concatenated hybrid tensor is input into the U-Net generator G to obtain the second simulation result. ; S44, Baseline Model Second simulation results They are fed into the second multi-scale discriminator respectively. In step S42, the mean squared error loss function is used. Make the benchmark model Compared with the second simulation results The discrimination result and respectively with and The differences between them should be as small as possible; S45, will and Add them together to calculate the first multi-scale discriminator. Second multi-scale discriminator The overall loss is calculated, and gradient zeroing, backpropagation, and gradient update methods are used to update it. and Network parameters.

6. The method for fine characterization of complex geological bodies based on a dual-loop generative adversarial network as described in claim 1, characterized in that: The secondary training process of the U-Net generator G is as follows: S61. Fix the trained parameters of the encoding function E, and separately test the U-Net generator in the cVAE-GAN and cLR-GAN networks. Conduct training; S62. Randomly sample a noise from a Gaussian distribution. The corresponding dimension size is denoted as ;Will and Input after splicing The corresponding output result is then input back into the system. In the process of calculating the hidden distribution and Between Loss, in its form, is written as: S63. Perform gradient zeroing, backpropagation, and gradient update operations on the U-Net generator G separately to complete the secondary training of the U-Net generator G.

Images