Reservoir model multi-scale fine characterization method based on concurrent generative adversarial network

Abstract

The application provides a reservoir model multi-scale fine characterization method based on a concurrent generative adversarial network. The original training image is characterized in a multi-scale manner through a pyramid structure, and consistent spatial distribution patterns are maintained between the characterization results of different scales. In the case of fixed receptive field, corresponding global information and local information can be extracted in small-scale images and large-scale images respectively. Meanwhile, a concurrent training strategy is applied, and parameter inheritance is used between adjacent stages instead of random initialization, so that the parameters of the network model can be fully and effectively trained. Finally, the Wasserstein distance is used to measure the difference between the generated distribution and the real distribution, a gradient penalty strategy is used to limit the gradient of the loss function within a certain range, the stability during the training process is ensured, and the reconstruction loss is used to repair the detail information of the image. Through the coupling of the two different loss functions, the overall simulation performance of the network is improved.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional geological modeling and geographic information systems, and in particular to a multi-scale fine characterization method for reservoir models based on concurrently generated adversarial networks. Background Technology

[0002] The detailed construction and visualization of reservoir models have always been a key research focus in the field of geoinformation science. They play a crucial role in characterizing subsurface spatial structures, making resource exploration target decisions, and describing regional geological backgrounds, thereby assisting geological experts in making quantitative assessments of subsurface resource migration, formation, and distribution. With deepening exploration, the construction of models for reservoir units with complex spatial structures becomes extremely difficult due to their significant spatial heterogeneity, non-homogeneity, and complex connectivity, especially given limited observational data and expert knowledge. This results in uncertainties in the representation of prior geological knowledge. Numerical simulation methods, based on stochastic function theory and Monte Carlo sampling with equal probability modeling, have been widely applied in reservoir modeling. While this method can quantify the uncertainty of reservoir models, its computational efficiency is low for large-scale, significantly non-stationary, and spatially complex reservoir units due to its stationarity assumptions and CPU-intensive computational characteristics. Although strategies such as local angle rotation and coordinate affinity ratio adjustment can effectively alleviate the difficulty of reservoir model reconstruction under non-stationary conditions, the algorithm requires setting a large number of parameters related to angle and affinity ratio changes in cases with complex and diverse channel geometries, making practical operation very complex and unable to guarantee the accuracy of simulation results. Since reservoir modeling can be viewed as a generative task, modeling methods based on Generative Adversarial Networks (GANs) can effectively overcome the limitations of numerical simulation methods. The development of its basic theory and various improved neural network models have provided technical support for the three-dimensional stochastic simulation and fine characterization of complex geological structures. Existing research shows that GANs and their variants have achieved good applications in porous media reconstruction, hydrogeological simulation, reservoir history fitting, and parameterization.

[0003] The basic idea of ​​GANs is to establish a generator and a discriminator, and achieve random image reconstruction through their adversarial learning characteristics. Specifically, the generator randomly samples noise from a Gaussian distribution and fits it to the high-order statistical distribution features of the training dataset, thus generating a fake sample to "fool" the discriminator. The discriminator performs a binary classification task, outputting a scalar to determine the probability that the corresponding input data comes from the generated sample (labeled False) or the real sample (labeled True). Through multiple iterations of training, the generator and discriminator reach a "Nash equilibrium," where the generator can produce samples similar to the training dataset, while the discriminator cannot distinguish the degree of authenticity of the input data. At this point, the network parameters are saved, and multiple simulation results can be obtained by inputting specific data only once, thus completing the subsequent generation objective.

[0004] GAN-based reservoir modeling methods generally fall into two categories: unconditional simulation and conditional simulation. Compared to the former, the latter involves stochastic simulation under the constraint of observational data, and its simulation results often have more geological significance. However, GANs also have some drawbacks. For example, when the real distribution differs significantly from the generated distribution, there is a saturation range in the JS divergence between the two, making it impossible for the discriminator to reasonably measure the difference, resulting in the gradient vanishing problem. Although the -log scheme can effectively alleviate the problem of insignificant gradients in the early stages of training, the asymmetric KL divergence can lead to inconsistent penalty strength in the discriminator, causing instability in the GAN training process and failing to guarantee the diversity of generated results. However, the Wasserstein distance, due to its smoothness, can reasonably measure the difference between two distributions, thus providing reasonable guidance for the GAN training process. Therefore, it is necessary to establish a reasonable metric to evaluate the difference between the generated and real distributions, thereby ensuring the simulation performance and computational efficiency of GANs.

[0005] In the field of reservoir modeling, prior geological knowledge is limited, while GAN training requires a large training dataset to complete the corresponding generation tasks. Although training datasets can be created through model segmentation, the segmentation results can easily disrupt the correlation between global and local features in the training images. Furthermore, a fixed receptive field may prevent the GAN reconstruction process from simultaneously considering both global and local features of the training images, leading to poor simulation performance. However, multi-scale representation methods based on pyramid structures can overcome the limitation of GAN datasets. The principle lies in downsampling the original images into different sizes while maintaining a similar spatial distribution pattern across all images. This allows the GAN network to extract corresponding feature information from training images at different scales, thus achieving multi-stage, multi-scale random generation tasks.

[0006] Furthermore, in multi-stage generation tasks based on a pyramid structure, multiple generators and discriminators need to be established in the training queue, and image scale conversion needs to be achieved between adjacent generators through upsampling. However, training each stage's generator through random initialization leads to a lack of close correlation between network parameters between adjacent stages. Training each stage's generator separately with a fixed learning rate fails to fully optimize all network model parameters. Moreover, for the generated sequence, under the chain rule, the simulation results of the previous stage affect the reconstruction performance of the subsequent stage. Since the training images corresponding to adjacent stages have similar spatial distribution patterns, it can be inferred that the parameters of adjacent generators also maintain a certain correlation. Therefore, parameter inheritance can be used instead of random initialization to ensure that the training task of the next stage is based on the parameter optimization and adjustment of the previous stage, guaranteeing the realism and reliability of the simulation results.

[0007] In summary, traditional GANs are prone to gradient vanishing and mode collapse during training, fundamentally due to the asymmetry of KL divergence. Therefore, using Wasserstein distance, which has smoothing properties, as a loss metric is crucial for ensuring network training stability. Furthermore, due to the limited prior knowledge representation of reservoir models, support sets are lacking for GAN training. Ensuring sufficient training of GANs under limited dataset conditions is a pressing issue. Additionally, training each stage of the network individually through random initialization may result in ineffective parameter training. Therefore, a strategy is needed to train multiple networks simultaneously, using parameter inheritance instead of random initialization to improve training efficiency. Given the problems and challenges of GANs in complex geological 3D reconstruction analyzed above, it is necessary to propose a generative adversarial network model based on a single training image and a concurrent training strategy to achieve automatic reservoir model reconstruction. Summary of the Invention

[0008] To overcome the limitations of GANs in reservoir modeling, this invention provides a multi-scale fine-grained representation method for reservoir models based on concurrent generative adversarial networks. This method introduces a pyramid structure to represent the original training images at multiple scales, ensuring that training images at different scales maintain similar spatial distribution patterns. With a fixed receptive field, global and local information are captured for small-scale and large-scale training images, respectively. The small-scale training images can be considered as soft data constraints, and subsequent reconstruction is based on detailed information restoration from the small-scale reconstruction. To avoid insufficient training of network parameters due to random initialization, the parameter values ​​of the previous stage are passed to the next adjacent stage after training. Furthermore, for the generators in the training queue, a maximum concurrency is set and different learning rates are assigned to ensure that multiple generators participate in training at each stage, guaranteeing the optimal range of network parameter values. In addition, for the optimization process of each stage, Wasserstein distance and gradient penalty terms are introduced for difference measurement, ensuring that the gradient satisfies the 1-Lipschitz constraint, guaranteeing the training stability of each stage and maintaining the correlation between generated samples and the original training images. Therefore, this method can achieve multi-scale stochastic reconstruction even when the prior reservoir model has limited representation, thus providing new technical means and theoretical support for the stochastic simulation and fine characterization of complex reservoir models.

[0009] The technical solution adopted in this invention is: to provide a multi-scale fine characterization method for reservoir models based on concurrent generative adversarial networks, comprising the following steps:

[0010] S1: Establish a prior reservoir model of the study area as the original training image and extract the corresponding attribute information;

[0011] S2: Define the range of values ​​for the hyperparameters and obtain the dimensions of the original training images;

[0012] S3: Input the original training images into the pyramid model, obtain the multi-scale representation results, and store the multi-scale representation results into the training image queue;

[0013] S4: Create a generator network, randomly initialize its parameters only in the first training phase, and add the initialized generator to the generator training queue;

[0014] S5: Traverse the training image queue, starting with small-scale training images and extracting them sequentially and training them in stages until the traversal process ends; as the training level increases, a concurrent training mode is adopted, and an adversarial learning process is established through the discriminator to optimize the network parameters of multiple generators until convergence.

[0015] S6: When the generators of all stages have completed training, the multi-scale reconstruction results are output based on the network parameters saved in different stages, by inputting the given data.

[0016] Furthermore, it also includes:

[0017] S7: Evaluate the reconstruction performance (spatial variability, connectivity, structural similarity, phase attribute ratio, etc.) of the multi-scale reconstruction results. If the performance threshold condition is met, save all network parameters and realize automatic multi-scale reconstruction of the reservoir model by inputting specific data; otherwise, return to step S2, readjust the parameters and network structure, and retrain the network model.

[0018] Furthermore, in step S1, the prior reservoir model includes a two-dimensional model and a three-dimensional model.

[0019] Furthermore, for the two-dimensional model, the step of inputting the original training image into the pyramid model to obtain multi-scale representation results includes:

[0020] Obtain the dimensions of the training images; the corresponding dimensions of the 2D model are y and y respectively. n = (L, W), where L and W represent the length and width;

[0021] Let n be the number of training stages. Then, for each individual training stage, the size y of the corresponding training image is... i satisfy:

[0022] y i =y n ×m ((n-1) / log n)*log(n-i)+1,i=1,…,n-1.

[0023] For a two-dimensional model, m satisfies:

[0024] m=(min(L1,W1) / min(L*δ,W*δ)) 1 / (n-1) ,

[0025] δ=min(max(L s W s ) / max(L,W),1).

[0026] Where L1 and W1 represent the length and width of the training image in the first stage, respectively, i.e., the minimum size; L s With W s represents the size of the original training image, i.e., the maximum size; min(·) and max(·) represent taking the minimum and maximum values ​​of the values ​​inside the parentheses, respectively; m and δ represent scaling factors.

[0027] Furthermore, for the three-dimensional model, the step of inputting the original training image into the pyramid model to obtain multi-scale representation results includes:

[0028] Obtain the dimensions of the training images; the corresponding size of the 3D model is y. n = (L, W, H), where L and W represent the length and width, and H represents the height of the training image;

[0029] Let n be the number of training stages. Then, for each individual training stage, the size y of the corresponding training image is... i satisfy:

[0030] y i =y n ×m ((n-1) / log n)*log(n-i)+1 ,i=1,…,n-1.

[0031] For a 3D model, m can be defined as:

[0032] m=(min(L1,W1,H1) / min(L*δ,W*δ,H*δ)) 1 / (n-1) ,

[0033] δ=min(max(L s W s H s ) / max(L,W,H),1).

[0034] Where L1, W1, and H1 represent the length, width, and height of the training image in the first stage, respectively, i.e., the minimum size; L s W s With H sThese represent the length, width, and height of the original training image, i.e., the maximum size; min(·) and max(·) represent taking the minimum and maximum values ​​of the values ​​inside the brackets, respectively; m and δ represent scaling factors.

[0035] Furthermore, in step S2, the range of hyperparameter values ​​is defined, including: the total number of training stages (train_stages) and the maximum number of concurrent connections (con). num The parameters include the number of training epochs in a single stage, noise weights, learning rate, gradient penalty term weight coefficient α, and reconstruction loss weight coefficient β.

[0036] Furthermore, in step S5, the training process for each individual stage generator network includes the following steps:

[0037] S51: Initialize a discriminator. If this stage is the first stage, perform random parameter initialization; otherwise, load the discriminator network parameters from the previous stage and perform the discrimination process based on them.

[0038] S52: Define a noise sequence. No noise is added during the first training phase. Starting from the second training phase, random samples are taken from a Gaussian distribution and added to the noise sequence, which then participates in the random reconstruction process of the corresponding phase.

[0039] S53: Define two optimizers to participate in the parameter optimization process of the network in each training stage. Starting from the second stage, the network parameters trained in the previous stage are directly passed to the next adjacent stage. If the number of generators participating in the generator training queue is less than con... num Then, from back to front, the learning rate decreases step by step while training is performed simultaneously; the method of gradually decreasing the learning rate can be written as:

[0040] θ i =θ n *λ n-i i = 1, 2, ..., n-1

[0041] Where, θ n Let L be the learning rate for the nth stage, and λ (0 < λ < 1) be the corresponding scaling factor. If the training queue length L... seq Greater than con num Then pop the (L)th element from the queue. seq -con num Train one generator with fixed parameters and train subsequent generators in the same way.

[0042] S54: The generator and discriminator at each stage calculate the corresponding loss function and use the same optimization method;

[0043] S55: Perform backpropagation and gradient update operations on the results of the loss function calculation to update the network parameters of the generator and discriminator in the current training phase;

[0044] S56: After the training of this stage is completed, input the given data to obtain the simulation results of this stage, upsample the simulation results, and merge them with noise of a certain weight as the input of the generator for the next stage.

[0045] Furthermore, in step S54, the loss function is calculated in the same way for each stage, and the calculation formula is as follows:

[0046]

[0047] in, Indicating resistance to loss, G represents the reconstruction loss. i With D i Represent the network parameters of the generator and discriminator in the i-th stage, respectively, where α is the weight factor. It can be defined in the following form:

[0048]

[0049]

[0050]

[0051] Where, x i This represents the training image input in the i-th stage. This represents the simulation results for the corresponding stage. This represents the result obtained from stage i-1 after upsampling, where D(*) is the discriminator's discrimination result and G(*) is the generator's generation result. Represents the gradient penalty term. This represents the samples selected from real samples and generated samples according to a certain ratio. P represents the gradient magnitude of the mixed samples. g P r and These represent the distribution characteristics among generated data, real data, and sampled data, respectively. This represents the gradient calculation symbol, β represents the correlation coefficient of the gradient penalty term, and E * This represents the expected value of the corresponding data distribution;

[0052] for Using the L2 norm as a measure, it can be defined mathematically as:

[0053]

[0054] Furthermore, a joint loss function is established to optimize and adjust the network parameters in the individual training phases. The Wasserstein distance is used to evaluate the difference between the generated distribution and the true distribution. A gradient penalty strategy is adopted to ensure that the gradient of the loss function satisfies the 1-Lipschitz constraint.

[0055] The beneficial effects of this invention based on its technical solution are as follows:

[0056] (1) In this invention, a pyramid structure is used to downsample the prior reservoir model, which yields corresponding multi-scale representation results. These training images at different scales maintain a consistent spatial distribution pattern. Feature extraction begins with small-scale training images. With a fixed receptive field, more global information can be extracted from small-scale training images, while more detailed information can be captured from large-scale training images. This multi-scale representation approach supports the overall training process of the GAN, overcoming the problem of GAN's ineffective training under limited data conditions.

[0057] (2) This invention employs a concurrent training strategy to ensure effective training of the network model parameters. In the multi-scale representation results, the training images of two adjacent stages are correlated, and the two stages are concatenated through upsampling. Based on this, it can be inferred that the network model parameters of two adjacent stages are also correlated. Therefore, the network model parameters of the previous stage can be directly passed to the next adjacent stage through parameter inheritance. Furthermore, by setting the maximum number of concurrent operations and adjusting the dynamic learning rate, different numbers and stages of generators can be trained simultaneously, ensuring that the network parameters are sufficiently trained.

[0058] (3) In this invention, a joint loss function is established for each individual training stage to optimize the network parameters. The joint loss function mainly includes prior loss and reconstruction loss. For the prior loss, Wasserstein distance is used to measure the difference between the generated distribution and the true distribution, and a gradient penalty strategy is introduced to make the gradient of the loss function satisfy the 1-Lipschitz constraint condition, making the training process more stable and avoiding gradient explosion or mode collapse. For the reconstruction loss, L2 norm is used for calculation, which significantly improves the restoration of detailed information as the scale increases. Attached Figure Description

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0060] Figure 1 This is the overall technical approach of the present invention.

[0061] Figure 2 This is the result of multi-scale characterization of the convolutional pyramid structure in this invention.

[0062] Figure 3 This is the basic structure of the concurrent generative adversarial network in this invention, comprising two parts: a generator and a discriminator. The generator is a sequence of generated data composed of multiple subnetworks, and the discriminator evaluates the multi-scale reconstruction results.

[0063] Figure 4 These are various stochastic simulation results of non-stationary sector deltas in two-dimensional cases in Embodiment 1 of the present invention.

[0064] Figure 5 This is the quantitative evaluation result of the corresponding two-dimensional reconstruction in Embodiment 1 of the present invention. Figure 5 (a) shows the distribution trend of spatial variability curves between different simulation results and the reference model; Figure 5 (b) shows the distribution trend of connectivity curves between different simulation results and the reference model; Figure 5 (c) is the spatial structure similarity measurement result, visualized in MDS form in a two-dimensional Cartesian coordinate system.

[0065] Figure 6 These are various stochastic simulation results of the non-stationary multi-attribute Poyang Lake Delta in three dimensions in Embodiment 2 of the present invention.

[0066] Figure 7 This is the quantitative evaluation result of the corresponding three-dimensional reconstruction in Embodiment 2 of the present invention. Figure 7 (a) shows the distribution trend of spatial variability curves between different simulation results and the reference model; Figure 7 (b) is the result of spatial structure similarity measurement, visualized in two-dimensional Cartesian coordinate system in MDS form. Figure 7 (c) and Figure 7 (d) are schematic diagrams showing the distribution of connectivity curves between the two sedimentary facies of the estuary bar and the tributary channel between different simulation results and the reference model. Detailed Implementation

[0067] 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.

[0068] To overcome the shortcomings and defects of numerical simulation methods in the random reconstruction of non-stationary reservoir models, this invention provides a multi-scale fine representation method for reservoir models based on concurrent generative adversarial networks. A pyramid structure is used to represent the original training images at multiple scales, maintaining a consistent spatial distribution pattern across different scales. With a fixed receptive field, corresponding global and local information can be extracted from both small-scale and large-scale images. Simultaneously, a concurrent training strategy is applied, with parameter inheritance replacing random initialization between adjacent stages, ensuring sufficient and effective training of the network model parameters. Finally, Wasserstein distance is used to measure the difference between the generated and real distributions, and a gradient penalty strategy is employed to limit the gradient of the loss function within a certain range, ensuring stability during training. Reconstruction loss is used to restore image details. Through the coupling of two different loss functions, the overall simulation performance of the network is improved.

[0069] Reference Figure 1 and Figure 3 A multi-scale fine-representation method for reservoir models based on concurrently generated adversarial networks includes the following steps:

[0070] S1: Establish a prior reservoir model of the study area as the original training image and extract the corresponding attribute information;

[0071] S2: Define the range of values ​​for the hyperparameters and obtain the dimensions of the original training images;

[0072] S3: Input the original training images into the pyramid model, obtain the multi-scale representation results, and store the multi-scale representation results into the training image queue;

[0073] S4: Create a generator network, randomly initialize its parameters only in the first training phase, and add the initialized generator to the generator training queue;

[0074] S5: Traverse the training image queue, starting with small-scale training images and extracting them sequentially and training them in stages until the traversal process ends; as the training level increases, a concurrent training mode is adopted, and an adversarial learning process is established through the discriminator to optimize the network parameters of multiple generators until convergence.

[0075] S6: When the generators of all stages have completed training, the multi-scale reconstruction results are output based on the network parameters saved in different stages, by inputting the given data.

[0076] S7: Evaluate the reconstruction performance (spatial variability, connectivity, structural similarity, phase attribute ratio, etc.) of the multi-scale reconstruction results. If the performance threshold condition is met, save all network parameters and realize automatic multi-scale reconstruction of the reservoir model by inputting specific data; otherwise, return to step S2, readjust the parameters and network structure, and retrain the network model.

[0077] The specific implementation process of step S1 is as follows:

[0078] By combining existing geological data and observation data, a priori reservoir model for the corresponding study area is established. At the same time, the corresponding attribute information is extracted, and different types of attribute information are mapped to a regular two-dimensional or three-dimensional grid through relation mapping.

[0079] Step S2 defines the range of hyperparameter values, specifically including: total training stages (train_stages) and maximum concurrency (con). num The values ​​of these hyperparameters, such as the number of training iterations (epcchs) per stage, noise weights, learning rate, gradient penalty term weight coefficient α, and reconstruction loss weight coefficient β, are crucial to the reconstruction performance of the network model.

[0080] In step S3, the original training images are input into the pyramid structure, such as... Figure 2 As shown, the multi-scale representation results are stored in the training image queue. For this step, the multi-scale representation results can be processed using the following steps S31 to S34:

[0081] S31: Obtain the dimensions of the training image, with the corresponding sizes y in two-dimensional and three-dimensional cases respectively. n = (L, W) and y n = (L, W, H), where L and W represent the length and width, and H represents the height of the training image.

[0082] S32: Assuming train_stages = n, then for each individual training stage, the size of the corresponding training image can be defined as:

[0083] y i =y n ×m ((n-1) / log n)*log(n-i)+1 ,i=1,…,n-1.

[0084] S33: For the two-dimensional case, m can be defined as:

[0085] m=(min(L1,W1,H1) / min(L*δ,W*δ)) 1 / (n-1) ,

[0086] where δ=min(max(L) s Ws ) / max(L,W),1).

[0087] S34: For the three-dimensional case, m can be defined as:

[0088] m=(min(L1,W1,H1) / min(L*δ,W*δ,H*δ)) 1 / (n-1) ,

[0089] where δ=min(max(L) s W s H s ) / max(L,W,H),1).

[0090] Where L1, W1, and H1 represent the length, width, and height of the training image in the first stage, respectively, i.e., the minimum size; L s W s With H s These represent the length, width, and height of the original training image, i.e., the maximum size; min(·) and max(·) represent taking the minimum and maximum values ​​of the values ​​inside the brackets, respectively; m and δ represent scaling factors.

[0091] In step S5, for the training process of each individual stage of the network model, steps S51 to S56 can be executed:

[0092] S51: Initialize a discriminator. If this stage is the first stage, perform random parameter initialization; otherwise, load the discriminator network parameters from the previous stage and perform the discrimination process based on them.

[0093] S52: Define a noise sequence. No noise is added during the first training phase. Starting from the second training phase, random samples are taken from a Gaussian distribution and added to the noise sequence, which then participates in the random reconstruction process of the corresponding phase.

[0094] S53: Define two optimizers to participate in the parameter optimization process of the network in each training stage. Starting from the second stage, the network parameters trained in the previous stage are directly passed to the next adjacent stage. If the number of generators participating in the generator training queue is less than con... num Then, from back to front, the learning rate decreases step by step while training is performed simultaneously; the method of gradually decreasing the learning rate can be written as:

[0095] θ i =θ n *λ n-i i = 1, 2, ..., n-1

[0096] Where, θ nLet L be the learning rate for the nth stage, and λ (0 < λ < 1) be the corresponding scaling factor. If the training queue length L... seq Greater than con num Then pop the (L)th element from the queue. seq -con num Train one generator with fixed parameters and train subsequent generators in the same way.

[0097] S54: The generator and discriminator at each stage calculate the corresponding loss function and use the same optimization method;

[0098] S55: Perform backpropagation and gradient update operations on the results of the loss function calculation to update the network parameters of the generator and discriminator in the current training phase;

[0099] S56: After the training of this stage is completed, input the given data to obtain the simulation results of this stage, upsample the simulation results, and merge them with noise of a certain weight as the input of the generator for the next stage.

[0100] S57: Repeat steps S51 to S56 until the traversal of the training image queue is complete.

[0101] Once all stages of the generator have completed training, given input data, it can output simulation results at different scales, evaluating the reconstruction results in terms of spatial variability, connectivity, structural similarity, and phase attribute ratios. If the spatial variability and connectivity curves of the simulation results are close to those of the reference model, with fluctuations not exceeding 1, then all network model parameters are saved. By inputting specific data, multi-scale automatic reconstruction of the reservoir model has been achieved. If the reconstruction performance is poor, otherwise, return to step S2, readjust the parameters and network structure, and retrain the network model.

[0102] Example 1:

[0103] For the automatic reconstruction process in the two-dimensional case, in Example 1, a non-stationary fan-shaped delta is selected as the two-dimensional prior reservoir model. First, the original model is input into the pyramid structure. The size of the first-stage training image is defined as 26×26 (unit: pixels), the total number of training stages is defined as train_stages=7, the initial learning rate θ1=0.0002, the learning rate scaling factor λ=0.1, and the maximum concurrency is con… num=3, gradient penalty term weight coefficient β = 0.1, reconstruction loss term weight coefficient α = 10, single-stage training iterations are 600. The network parameters are optimized using the Adam optimizer with default parameter settings. With the above parameter settings, starting with the smallest training image, the generator is loaded into the training queue in a separate training stage, and a concurrent training strategy is adopted for parameter optimization. As the number of training stages increases, if the training queue length L... seq Greater than co num Then pop the (L)th element from the training queue. seq -con num ( ) generators are used, with their parameters kept fixed, and subsequent generators are trained on them. When training is complete, the network model parameters are saved. Different random simulation results are obtained by inputting random noise, such as... Figure 4 As shown. To further quantify the evaluation, 20 simulation results were randomly generated, and their variation function curves, connectivity function curves, and multidimensional scaling analysis results compared to the reference model are shown below. Figure 5 As shown, Figure 5 (a) shows the distribution trend of spatial variability curves between different simulation results and the reference model; Figure 5 (b) shows the distribution trend of connectivity curves between different simulation results and the reference model; Figure 5 (c) shows the spatial structure similarity measurement results, visualized in MDS form within a two-dimensional Cartesian coordinate system. Figure 5 It can be seen that the different simulation results maintain a high degree of similarity to the reference model in terms of spatial variability, phase attribute connectivity, and spatial distribution pattern.

[0104] Example 2:

[0105] For the automatic reconstruction process in three dimensions, in Example 2, the Poyang Lake fan-shaped delta is selected as the three-dimensional prior reservoir model. First, the original model is input into the pyramid structure. The size of the first-stage training image is defined as 15×15×12 (unit: voxel), the total number of training stages is defined as train_stages = 5, the initial learning rate θ1 = 0.0001, the learning rate scaling factor λ = 0.1, and the maximum concurrency is con... num =3, gradient penalty term weight coefficient β = 0.1, reconstruction loss term weight coefficient α = 10, single-stage training iterations are 400. The network parameters are optimized using the Adam optimizer with default parameter settings. With the above parameter settings, starting with the smallest training image, the generator is loaded into the training queue in a separate training stage, and a concurrent training strategy is adopted for parameter optimization. As the number of training stages increases, if the training queue length L... seq Greater than con num Then pop the (L)th element from the training queue.seq -con num ( ) generators are used, with their parameters kept fixed, and subsequent generators are trained on them. When training is complete, the network model parameters are saved. Different random simulation results are obtained by inputting random noise, such as... Figure 6 As shown. To further quantify the evaluation, 20 simulation results were randomly generated, and their variation function curves, connectivity function curves, and multidimensional scaling analysis results compared to the reference model are shown below. Figure 7 As shown, where, Figure 7 (a) shows the distribution trend of spatial variability curves between different simulation results and the reference model; Figure 7 (b) is the result of spatial structure similarity measurement, visualized in two-dimensional Cartesian coordinate system in MDS form. Figure 7 (c) and Figure 7 (d) These are schematic diagrams showing the connectivity curves of the two sedimentary facies—estuary bar and distributary channel—between different simulation results and the reference model. Figure 7 It can be seen that the different simulation results maintain a high degree of similarity to the reference model in terms of spatial variability, phase attribute connectivity, and spatial distribution pattern.

[0106] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0107] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0108] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-scale fine-representation method for reservoir models based on concurrent generative adversarial networks, characterized in that, Includes the following steps: S1: Establish a prior reservoir model of the study area as the original training image and extract the corresponding attribute information; S2: Define the range of values ​​for the hyperparameters and obtain the dimensions of the original training images; S3: Input the original training images into the pyramid model, obtain the multi-scale representation results, and store the multi-scale representation results into the training image queue; S4: Create a generator network, randomly initialize its parameters only in the first training phase, and add the initialized generator to the generator training queue; S5: Traverse the training image queue, starting with small-scale training images, extract them sequentially and perform stage-by-stage training until the traversal process is complete. As the training progresses, a concurrent training mode is adopted, and an adversarial learning process is established through a discriminator to optimize the network parameters of multiple generators until convergence. S6: When the generators of all stages have completed training, the multi-scale reconstruction results are output based on the network parameters saved in different stages, by inputting the given data. In step S5, the training process for each individual stage generator network includes the following steps: S51: Initialize a discriminator. If this stage is the first stage, perform random parameter initialization; otherwise, load the discriminator network parameters from the previous stage and perform the discrimination process based on them. S52: Define a noise sequence. No noise is added during the first training phase. Starting from the second training phase, random samples are taken from a Gaussian distribution and added to the noise sequence, which then participates in the random reconstruction process of the corresponding phase. S53: Define two optimizers to participate in the parameter optimization process of the network in each training stage. Starting from the second stage, the network parameters trained in the previous stage are directly passed to the next adjacent stage. If the number of generators participating in the generator training queue is less than con... num Then, from back to front, the learning rate decreases step by step while training is performed simultaneously; the method of gradually decreasing the learning rate can be written as: i i =θ n *l n-i ,i=1,2,…,n-1 Where, θ n Let L be the learning rate for the nth stage, and 0 < λ < 1 be the corresponding scaling factor. If the training queue length L... seq Greater than con num Then pop the Lth element from the queue. seq -con num One generator is used, and its parameters are kept fixed. Subsequent generators are trained in the same way. S54: The generator and discriminator at each stage calculate the corresponding loss function and use the same optimization method; S55: Perform backpropagation and gradient update operations on the results of the loss function calculation to update the network parameters of the generator and discriminator in the current training phase; S56: After the training of this stage is completed, input the given data to obtain the simulation results of this stage, upsample the simulation results, and merge them with noise of a certain weight as the input of the generator for the next stage.

2. The multi-scale fine characterization method for reservoir models according to claim 1, characterized in that, Also includes: S7: Evaluate the reconstruction performance of the multi-scale reconstruction results. If the performance threshold condition is met, save all network parameters and realize automatic multi-scale reconstruction of the reservoir model by inputting specific data. Otherwise, return to step S2, readjust the parameters and network structure, and retrain the network model.

3. The multi-scale fine characterization method for reservoir models according to claim 1, characterized in that, In step S1, the prior reservoir model includes a two-dimensional model and a three-dimensional model.

4. The multi-scale fine characterization method for reservoir models according to claim 3, characterized in that, For the two-dimensional model, the step of inputting the original training image into the pyramid model to obtain multi-scale representation results includes: Obtain the dimensions of the training images; the corresponding dimensions of the 2D model are y and y respectively. n = (L, W), where L and W represent the length and width; Let n be the number of training stages. Then, for each individual training stage, the size y of the corresponding training image is... i satisfy: and i / and n ×m ((n-1) / logn)*log(n-i)+1 ,i=1,…,n-1, For a two-dimensional model, m satisfies: m=(min(L1,W1) / min(L*δ,W*δ)) 1 / (n-1) , δ=min(max(L s ,IN s ) / max(L,W),1), Where L1 and W1 represent the length and width of the training image in the first stage, respectively, i.e., the minimum size; L s With W s represents the size of the original training image, i.e., the maximum size; min(·) and max(·) represent taking the minimum and maximum values ​​of the values ​​inside the parentheses, respectively; m and δ represent scaling factors.

5. The multi-scale fine characterization method for reservoir models according to claim 3, characterized in that, For the three-dimensional model, the step of inputting the original training image into the pyramid model to obtain multi-scale representation results includes: Obtain the dimensions of the training images; the corresponding size of the 3D model is y. n = (L, w, H), where L and W represent the length and width, and H represents the height of the training image; Let n be the number of training stages. Then, for each individual training stage, the size y of the corresponding training image is... i satisfy: and i / and n ×m ((n-1) / logn)*log(n-i)+1 ,i=1,…,n-1, For a 3D model, m satisfies: m=(min(L1,W1,H1) / min(L*δ,W*δ,H*δ)) 1 / (n-1) , δ=min(max(L s ,IN s ,H s ) / max(L,W,H),1), Where L1, W1, and H1 represent the length, width, and height of the training image in the first stage, respectively, i.e., the minimum size; L s W s With H s These represent the length, width, and height of the original training image, i.e., the maximum size; min(·) and max(·) represent taking the minimum and maximum values ​​of the values ​​inside the brackets, respectively; m and δ represent scaling factors.

6. The multi-scale fine characterization method for reservoir models according to claim 1, characterized in that, In step S2, the range of hyperparameter values ​​is defined, including: total training stages (train_stages) and maximum concurrency (con). num The number of training epochs in a single stage, noise weight, learning rate, gradient penalty term weight coefficient α, and reconstruction loss weight coefficient β are all considered.

7. The multi-scale fine characterization method for reservoir models according to claim 6, characterized in that, In step S54, the loss function is calculated in the same way for each stage, and the calculation formula is as follows: in, Indicating resistance to loss, G represents the reconstruction loss. i With D i Represent the network parameters of the generator and discriminator in the i-th stage, respectively, where α is the weight factor. Defined in the following form: Where, x i This represents the training image input in the i-th stage. This represents the simulation results for the corresponding stage. This represents the result obtained from stage i-1 after upsampling, where D(*) is the discriminator's discrimination result and G(*) is the generator's generation result. Represents the gradient penalty term. This represents the samples selected from real samples and generated samples according to a certain ratio. P represents the gradient magnitude of the mixed samples. g P r and These represent the distribution characteristics among generated data, real data, and sampled data, respectively. This represents the gradient calculation symbol, β represents the correlation coefficient of the gradient penalty term, and E * This represents the expected value of the corresponding data distribution; for The L2 norm is used for measurement, and it is defined mathematically as follows: .

8. The multi-scale fine characterization method for reservoir models according to claim 1, characterized in that: By establishing a joint loss function to optimize and adjust the network parameters in the individual training phases, the Wasserstein distance is used to evaluate the difference between the generated distribution and the true distribution, and a gradient penalty strategy is adopted to ensure that the gradient of the loss function satisfies the 1-Lipschitz constraint.

Images