Crack expansion agent model construction method based on generative neural network
By constructing a crack expansion agent model (FPGAN) of a generative neural network, the simulation problem of crack expansion in heterogeneous reservoirs is solved, efficient and accurate crack expansion prediction is achieved, and computing efficiency and model applicability are improved.
Patent Information
- Application Number
- CN202510180590.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-19
- Publication Date
- 2025-07-04
AI Technical Summary
The prior art is difficult to accurately simulate fracture expansion in heterogeneous reservoirs. Traditional numerical simulation methods have high calculation costs and low efficiency, and cannot meet the crack prediction needs under complex geological conditions.
A fracture expansion agent model (FPGAN) based on a generative neural network is built. By obtaining input feature parameters, a reservoir physical model is constructed and numerical simulation is performed, the fracture expansion agent model is trained, including generators and discriminators, and the fracture expansion image is output.
Efficient simulation and accurate prediction of fracture expansion in heterogeneous reservoirs are realized, which significantly improves prediction efficiency and reduces calculation time cost. It is suitable for crack expansion optimization design and influencing factor analysis under complex heterogeneous conditions.
Smart Images
Figure CN120257870A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of oil and gas production enhancement, and particularly relates to a method for constructing a fracture propagation surrogate model based on a generative neural network. Background Art
[0002] As a core means for developing low-permeability oil and gas reservoirs, hydraulic fracturing technology can significantly improve the oil and gas recovery rate. Fracture propagation simulation, as a key link in hydraulic fracturing, can predict the geometric shape and distribution of the propagated fractures. By accurately simulating the fracture propagation process, an ideal fracturing effect can be achieved while reducing the generation of ineffective fractures, thereby maximizing the advantages of hydraulic fracturing.
[0003] Under the comprehensive influence of factors such as sedimentation, diagenesis, and later tectonics in actual oil and gas reservoirs, they mostly exhibit heterogeneous characteristics, that is, there are uneven changes in the spatial distribution of oil and gas reservoirs and various internal attribute parameters (such as porosity, permeability parameters, etc.). Heterogeneous reservoirs have significant mineral diversity and bedding distribution differences. The complex interaction between heterogeneous reservoirs and engineering factors will affect the propagation path and shape of fractures, making the prediction of fracture propagation more difficult.
[0004] Firstly, there are significant differences in the mechanical properties (such as elastic modulus and permeability) of rocks in heterogeneous reservoirs, and this difference will affect the opening and propagation modes of fractures. Secondly, the pressure transmission within the formation is usually uneven, and this unevenness will not only cause fractures to initiate simultaneously at multiple points but also form non-linear and unpredictable fracture propagation paths. In addition, the change of in-situ stress also has an important impact on fracture propagation. These factors make it a major challenge to accurately simulate fracture propagation in heterogeneous reservoirs, and the models based on the traditional homogeneous reservoir assumption can no longer meet the accurate simulation requirements of unconventional reservoirs. However, more complex physical models and experimental means are limited by experimental scales and conditions and are difficult to fully simulate real reservoir conditions.
[0005] To address the complexity and challenges of fracture propagation prediction in heterogeneous reservoirs, researchers have gradually met the needs of unconventional oil and gas resource development by developing numerical simulation methods. The finite discrete element method (FDEM) further improves the ability to handle complex geometric shapes and fracture interactions by simulating the responses of reservoir rock masses under different pressure and stress conditions, and can capture the dynamic change process of fractures with higher precision, especially suitable for the study of fracture propagation in heterogeneous reservoirs. However, the numerical simulation has a high computational cost and large demand for computing resources, resulting in difficulty in meeting the efficiency requirements in multiple calculations or real-time simulations. The complex calculation process not only needs to solve problems such as fluid-rock coupling and fracture propagation path prediction, but also must cope with the influence of natural fractures, bedding structures, etc. in the reservoir on the simulation results, which limits the application of numerical simulation in heterogeneous reservoirs. Summary of the Invention
[0006] The object of the present invention is to provide a method for constructing a crack propagation surrogate model based on a generative neural network.
[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0008] In a first aspect, the present invention provides a method for constructing a crack propagation surrogate model based on a generative neural network, which is characterized by including the following steps:
[0009] Step S1: Obtain input characteristic parameters according to the geological characteristics of the target block, construct a reservoir physical model according to the input characteristic parameters, the input characteristic parameters affect the crack propagation pattern in the reservoir, and perform numerical simulation on the physical model to obtain a crack propagation image;
[0010] Step S2: Construct a sample data set according to the input characteristic parameters and the crack propagation image, and use the sample data set to train the crack propagation surrogate model (FPGAN) until the model parameters converge. The crack propagation surrogate model (FPGAN) includes a generator and a discriminator;
[0011] Step S3: Substitute the input characteristic parameters and use the crack propagation surrogate model (FPGAN) to output a crack propagation image.
[0012] In the above embodiment, a sample data set of crack propagation images of the reservoir model is obtained through numerical simulation, and the crack propagation surrogate model (FPGAN) is trained using the sample data set. The crack propagation surrogate model (FPGAN) can achieve rapid prediction of reservoir crack propagation. When the above embodiment is applied to heterogeneous reservoirs, the prediction efficiency can be greatly improved compared with the numerical simulation method in the prior art.
[0013] Further, the reservoir physical model simulates the heterogeneity of formation rocks.
[0014] In the above embodiment, by considering the heterogeneity of formation rocks, the influence of the distribution of heterogeneous rocks on crack propagation under real conditions can be simulated.
[0015] Further, the input characteristic parameters include the mineral composition of formation rocks and the rock mechanical property parameters of each mineral.
[0016] In the above embodiment, the rock mechanical property parameters are determined through rock mechanics experiments and relevant literature, ensuring the accuracy and operability of the heterogeneous reservoir physical model.
[0017] Further, the rock mechanical property parameters include elastic modulus, Poisson's ratio, tensile strength, cohesion, and internal friction angle.
[0018] Further, the heterogeneous reservoir physical model is the first heterogeneous reservoir physical model, and the first heterogeneous reservoir physical model considers the mineral interface.
[0019] In the above embodiment, by considering the cohesion and internal friction angle of the mineral interface, the influence of the mineral boundary on the fracture propagation path can be reflected.
[0020] Further, the input characteristic parameters further include mineral interface parameters, and the mineral interface parameters include the cohesion and internal friction angle of the mineral interface.
[0021] Further, the heterogeneous reservoir physical model is the second heterogeneous reservoir physical model, and the second heterogeneous reservoir physical model ignores the mineral interface.
[0022] In the above embodiment, setting the second heterogeneous reservoir physical model can facilitate the comparison with the case considering the mineral interface. In addition, the second heterogeneous reservoir physical model can also be used to verify the first heterogeneous reservoir physical model.
[0023] Further, the step S1 further includes:
[0024] Construct a homogeneous reservoir physical model according to the input characteristic parameters, and perform numerical simulation to obtain the fracture propagation image of the homogeneous reservoir.
[0025] Further, the MultiFracS software is used for numerical simulation.
[0026] Further, the sample data set is the first data set, the first data set is the basic data set, and the first data set does not include stress field and in-fracture fluid pressure field data.
[0027] The above embodiment can reveal the basic characteristics of fracture propagation in heterogeneous reservoirs and provide initial data support for the training of deep learning models.
[0028] Further, the sample data set is the second data set, the second data set is the complex data set, and the second data set includes stress field and in-fracture fluid pressure field data.
[0029] The above embodiment comprehensively reflects the interaction between fracture propagation and multi-physical fields, and provides rich data support for the multi-dimensional modeling and prediction of deep learning models.
[0030] Further, the generator combines the input characteristic parameters with random noise.
[0031] Further, the step S2 further includes:
[0032] Normalize the crack propagation image in step S1 to the RGB three-channel image format with the first resolution, and use the scaling operation to scale the image with the first resolution to the second resolution.
[0033] In the above embodiment, by normalizing the crack propagation image to the RGB three-channel image format with the first resolution, it is possible to ensure that the data has unified dimensional and channel attributes; using the scaling operation to scale the image with the first resolution to the second resolution can significantly reduce the hardware burden while still retaining the key detail features of the crack propagation image, ensuring that the resolution of the data is sufficient to support the training and prediction of the proxy model.
[0034] Further, the training process in step S2 includes:
[0035] Initialization stage, set the hyperparameters of the network;
[0036] Training stage, alternately optimize the discriminator and the generator. During the optimization process of the discriminator, by sampling real samples and fake samples generated by the generator, calculate the loss function and the gradient penalty term to ensure that the discriminator satisfies 1-Lipschitz continuity; during the optimization process of the generator, generate fake samples by inputting the noise vector and the conditional label, calculate the generator loss function, optimize the generation ability and update the parameters.
[0037] Further, adopt the conditional Wasserstein gradient penalty (CWGAN-GP) generative adversarial network structure to perform adversarial training on the generator and discriminator of the model.
[0038] In the above embodiment, by utilizing the advantages of conditional input and generative adversarial network, while maintaining the high resolution and detail quality of the generated image, the training stability and fast convergence of the model are ensured.
[0039] Further, adopt the Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) as evaluation metrics to evaluate the output crack propagation image.
[0040] In a second aspect, the present invention provides a heterogeneous reservoir crack propagation prediction device based on a generative neural network, including the following steps:
[0041] Crack propagation image sample acquisition module, used to obtain input feature parameters according to the geological characteristics of the target block, construct a physical model of the heterogeneous reservoir according to the input feature parameters, where the input feature parameters affect the crack propagation morphology in the heterogeneous reservoir, and perform numerical simulation on the physical model to obtain crack propagation image samples;
[0042] A crack propagation surrogate model training module, configured to construct a sample data set according to input feature parameters and crack propagation images, and train a crack propagation surrogate model (FPGAN) using the sample data set until the model parameters converge. The crack propagation surrogate model (FPGAN) includes a generator and a discriminator;
[0043] A crack propagation image prediction module, configured to substitute the input feature parameters and output a crack propagation image using the crack propagation surrogate model (FPGAN).
[0044] In a third aspect, the present invention further provides an electronic device, which includes:
[0045] One or more processors;
[0046] A storage device, configured to store one or more programs;
[0047] When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a heterogeneous reservoir crack propagation surrogate model according to any embodiment of the present invention.
[0048] In a fourth aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the method for constructing a heterogeneous reservoir crack propagation surrogate model according to any embodiment of the present invention is implemented.
[0049] In summary, the technical solution provided by the embodiments of the present invention realizes efficient simulation and accurate prediction of the crack propagation behavior of heterogeneous reservoirs by constructing a crack propagation generative neural network surrogate model (FPGAN). Compared with traditional numerical simulation methods, it has significant advantages in computational time efficiency and provides an efficient technical means for engineering applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative efforts.
[0051] Figure 1 It is a schematic flowchart of the method of the present invention;
[0052] Figure 2 It is a schematic diagram of a multi-mineral and interface physical model in the construction method of the embodiment of the present invention;
[0053] Figure 3 It is a schematic diagram of a partial sample data set in the embodiment of the present invention, where (a) is a simple data set and (b) is a complex data set;
[0054] Figure 4 This is the network structure diagram of the FPGAN in the embodiments of the present invention, where (a) is the structure diagram of the generator and (b) is the structure diagram of the discriminator;
[0055] Figure 5 This is the schematic diagram of the target task training process in the construction method of the embodiments of the present invention;
[0056] Figure 6 This is the schematic diagram of the prediction result in the construction method of the embodiments of the present invention. Detailed implementation manners
[0057] The present invention will be described in detail below with reference to the accompanying drawings.
[0058] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0059] In the prior art, numerical simulation methods are mostly used to simulate the fracture propagation morphology. This method not only requires technicians to have extremely high professional modeling capabilities, but also has cumbersome operation steps and slow calculation speeds. Compared with traditional numerical simulation methods, deep learning surrogate models can efficiently capture the non-linear relationships contained in the data, process multi-dimensional data, integrate multi-source information (such as geological parameters and engineering parameters), and achieve rapid prediction of complex fracture networks, and can significantly improve the simulation efficiency of fracture propagation while maintaining accuracy. Among many deep learning methods, generative neural networks have shown unique advantages in the generation of high-fidelity images of fracture propagation morphology and the adaptability to complex conditions.
[0060] Based on this, the present invention provides a construction method for a heterogeneous reservoir fracture propagation surrogate model (FPGAN, fracture propagation generative adversarial network) based on a generative neural network, so as to significantly improve the efficiency of fracture propagation morphology prediction while maintaining the accuracy of numerical simulation.
[0061] As Figure 1 shown, the construction method for a heterogeneous reservoir fracture propagation surrogate model based on a generative neural network provided by the present invention includes the following steps:
[0062] Step S1: Obtain input feature parameters according to the geological characteristics of the target block, construct a physical model of the heterogeneous reservoir according to the input feature parameters, where the input feature parameters affect the fracture propagation morphology in the heterogeneous reservoir, and perform numerical simulation on the physical model to obtain fracture propagation images;
[0063] Based on the actual geological conditions of a shale gas reservoir in Sichuan, the present invention divides the reservoir into multiple randomly distributed mineral regions to simulate the heterogeneous characteristics of the rock. As Figure 2 shown, the mineral components are divided into four main types: quartz, calcite, clay, and organic matter. The rock mechanical properties (such as parameters like elastic modulus, Poisson's ratio, etc.) of each mineral component can be determined through rock mechanics experiments and other means.
[0064] To compare the influence of mineral interfaces on the crack propagation path, as shown in Table 1, the present invention sets two types of physical models. One clearly divides the mineral interface and can simulate the mechanical response of the crack at the interface (as Figure 2 (b) shown); the other assumes that the mineral distribution is continuous and uniform, that is, a homogeneous model, without considering the influence of mineral interfaces on crack propagation. By setting the homogeneous model, it is possible to compare with the heterogeneous model, thereby considering the influence of mineral interfaces on crack propagation. In addition, the homogeneous model can also be used to verify the heterogeneous model.
[0065] Table 1 Basic parameters for fracturing simulation
[0066]
[0067]
[0068] It should be noted that the simulation parameters in Table 1 include elastic modulus, Poisson's ratio, tensile strength, cohesion, and internal friction angle, etc., which are respectively used to describe the mechanical properties of mineral components and interfaces. In the "model considering mineral interfaces", the cohesion and internal friction angle of the interface reflect the influence of the mineral boundary on the crack propagation path; while in the "model not considering mineral interfaces", it is assumed that the reservoir is continuously and uniformly distributed, only simulating the internal properties of the mineral.
[0069] To generate multiple sets of parameter combination data, the present invention sets different step sizes within the parameter range of Table 1. For example, the Poisson's ratio ranges from 0.1 to 0.5 with a step size of 0.1; the internal friction angle ranges from 10° to 50° with a step size of 10°, thereby generating diverse hydraulic fracture propagation data to provide physical input conditions for the construction of subsequent basic data sets and complex data sets.
[0070] When constructing the basic data set, by setting multiple combinations of mineral mechanical parameters, a total of 60 models are established, the simulation time is 10000 s, with a time step of 200 s, and the crack geometry is recorded once every step, generating a total of 2600 crack propagation images. As Figure 3 (a) shown, these images respectively show the crack propagation paths and morphologies under the two conditions of not considering mineral interfaces and considering mineral interfaces.
[0071] It should be noted that the stress field and the fluid pressure in the crack are not introduced when constructing the basic data set here, aiming to reveal the basic characteristics of crack propagation in heterogeneous reservoirs and provide a sample data set for the training of deep learning models.
[0072] To further study the dynamic coupling effect between crack propagation, the stress field, and the fluid pressure in the crack, the present invention constructs a complex data set. The complex data set introduces the stress field and the fluid pressure in the crack and analyzes the dynamic changes of the two physical fields during the crack propagation process. For example, crack propagation may cause local stress concentration or release, changing the distribution of the principal stress direction; at the same time, it affects the fluid migration path, resulting in an increase or decrease in the pressure in the crack. These dynamic changes are mapped onto the crack propagation image through the RGB three channels, representing the evolution processes of the stress field and the fluid pressure in the crack, respectively. As Figure 3 (b) shows, the complex data set comprehensively reflects the interaction between crack propagation and multiple physical fields, providing rich data support for the multi-dimensional modeling and prediction of deep learning models.
[0073] Step S2: Construct a sample data set according to the input characteristic parameters and the crack propagation image, and use the sample data set to train the crack propagation proxy model (FPGAN) until the model parameters converge. The crack propagation proxy model (FPGAN) includes a generator and a discriminator.
[0074] The present invention post-processes the crack propagation data generated by numerical simulation, which specifically includes the following operations: First, all the crack propagation image data generated by numerical simulation are standardized into the RGB three-channel image format of 1024×1024 to ensure that the data have unified dimensions and channel attributes, facilitating subsequent model training and processing. Considering the limitation of hardware computing resources and to avoid video memory overload, the present invention further uses a scaling operation (transform) to scale all images to a size of 512×512. This processing significantly reduces the hardware burden while still retaining the key detail features of the crack propagation image, ensuring that the resolution of the data is sufficient to support the training and prediction of the proxy model. For the convenience of subsequent model training and data management, the present invention performs file label processing on the processed image data. Specifically, the key information involved in the crack propagation simulation process (such as whether to consider the mineral interface, rock mechanical parameters, and time step, etc.) is embedded in the file name through an encoding method. For example, the file name contains identifiers to distinguish different simulation conditions, thereby realizing the direct association between the file name and the model training attribute values.
[0075] Through the above post-processing operations, the present invention not only standardizes the format of the numerical simulation data, ensuring the consistency and usability of the data, but also optimizes the data management process through the label encoding method, laying a foundation for the efficient training and accurate prediction of the subsequent proxy model.
[0076] In this invention, a fracture propagation generative neural network proxy model (FPGAN) is constructed. The FPGAN model includes a generator and a discriminator, and its structure is as Figure 4 shown. The structure of the generator is as Figure 4 (a) shown. The input of the generator is a combination of input feature parameters (rock mechanics parameters) and a random noise vector. Through step-by-step processing, an RGB fracture propagation image with a resolution of 512×512 is generated. The rock mechanics condition parameters in Table 1 are mapped into an embedded vector of a fixed dimension, which is concatenated with the random noise vector to form an initial input vector. After transformation by a linear layer, a high-dimensional feature vector is generated as the input for the subsequent transposed convolution process. There are 8 transposed convolution layers in total. Each layer adjusts the size and number of channels of the feature map through specified convolution kernel sizes (4×4 or 3×3), strides (2 or 1), and padding parameters, gradually reducing from 512 channels to 3 channels to achieve the output of a high-resolution RGB image. After each transposed convolution, batch normalization and the ReLU activation function are combined to ensure the stability of feature extraction and the non-linear expression ability. The Tanh activation function is used in the last layer to limit the pixel values of the output image between [-1, 1], making the distribution of the generated image closer to the real data.
[0077] The structure of the discriminator is as Figure 4 (b) shown, which is responsible for judging the authenticity of the input image. The input image is a real image or a fake image generated by the generator, both with a size of 512×512. The discriminator gradually extracts image features through 6 convolutional layers. The size of each convolutional kernel is 4×4, and the stride is 2. As the convolutional layers deepen, the image size gradually shrinks, and the number of channels increases to capture the detailed features of the image. The extracted convolutional features are flattened into a one-dimensional vector and then input into a fully connected layer, and finally a scalar value is output, indicating the probability that the input image is real data. After each convolution, batch normalization and the LeakyReLU activation function are combined to improve the training stability and avoid gradient vanishing.
[0078] The training process of the FPGAN model is divided into an initialization stage and a training stage. In the initialization stage, the hyperparameters of the network are set, including the learning rates and batch sizes of the generator and the discriminator, etc. In the main loop of the training stage, the discriminator and the generator are alternately optimized. During the optimization process of the discriminator, by sampling real samples and fake samples generated by the generator, the loss function and the gradient penalty term are calculated to ensure that the discriminator satisfies 1-Lipschitz continuity, and the parameters are optimized to improve the ability to distinguish between real and fake samples. During the optimization process of the generator, by inputting a noise vector and a conditional label to generate fake samples, the generator loss function is calculated, the generation ability is optimized, and the parameters are updated. The training terminates after all model parameters converge, and finally the goal of efficiently generating high-resolution fracture propagation morphology images is achieved.
[0079] Through the construction and training of the FPGAN model, leveraging the advantages of conditional input and generative adversarial networks, while maintaining the high resolution and detail quality of the generated images, the stability of model training and rapid convergence are ensured, providing effective technical support for the efficient simulation and accurate prediction of crack propagation behavior.
[0080] In the present invention, to achieve the efficient generation and accurate prediction of crack propagation patterns by the FPGAN model, a generative adversarial network structure with conditional Wasserstein gradient penalty (CWGAN-GP) is adopted to conduct adversarial training on the generator and discriminator of the model. The loss functions of the generator G and discriminator D are defined as follows:
[0081] The loss function of the discriminator is based on the Wasserstein distance, and a gradient penalty term is introduced to ensure that the discriminator satisfies the 1-Lipschitz continuity condition. Its loss function expression is:
[0082]
[0083] where x represents the real sample, G(z|y) represents the fake sample generated by the generator, y is the conditional variable (such as rock mechanics parameters), is the interpolation sample between the real sample and the fake sample, and λ gp is the weight coefficient of the gradient penalty term.
[0084] The loss function of the generator is defined as:
[0085] L G =-E[D(G(z|y)|y)]
[0086] This loss function aims to maximize the misjudgment probability of the discriminator for the generated samples, thereby guiding the generator to generate high-resolution crack propagation images closer to the real distribution.
[0087] Through the above loss function design, the discriminator ensures the stability of model training by optimizing the Wasserstein distance and introducing gradient penalty, avoiding the problems of gradient vanishing and mode collapse in traditional GANs; the generator controls the morphological characteristics of crack propagation through conditional variable input, supporting the generation of specific crack propagation images according to the heterogeneous reservoir geological conditions, thereby improving the controllability of the model and the quality of the generated results. The above optimization objectives ensure the efficiency and applicability of the model under complex geological conditions, providing technical support for the simulation of crack propagation behavior.
[0088] The training process of the FPGAN model of the present invention is shown in Table 2 and is divided into an initialization stage and a training stage. In the initialization stage, the hyperparameters of the network are set, including the learning rates of the generator and discriminator, the batch size, and other parameters, to prepare for the training of the model. In the training stage, the alternating optimization of the discriminator and generator is achieved through the main loop. First, in each loop, the discriminator is updated, including sampling real samples from the dataset and generating fake samples through the generator, calculating the loss function of the discriminator and the gradient penalty term to ensure that the discriminator satisfies the 1-Lipschitz continuity condition. By updating the parameters of the discriminator, its ability to distinguish between real and fake samples is improved. Subsequently, the generator is updated. By inputting the noise vector and conditional labels to generate fake samples, the loss function of the generator is calculated and its parameters are optimized to gradually improve the ability of the generator to generate high-resolution crack propagation images.
[0089] The training terminates after all model parameters converge. At this time, the FPGAN model can efficiently generate high-resolution crack propagation morphology images that meet the input conditions, meeting the accuracy and efficiency requirements of the present invention in simulating the crack propagation behavior of heterogeneous reservoirs. Through the above training process, the present invention has achieved significant improvements in the resolution, detail quality, and training stability of generating crack propagation images, providing an efficient technical means for predicting crack propagation behavior under complex reservoir conditions.
[0090] The training process of the PGAN algorithm in Table 2
[0091]
[0092]
[0093] Figure 5 It is a schematic diagram of the training process of the target task in the embodiment of the present invention. The present invention uses the basic dataset and the complex dataset to train and test the FPGAN model. The dataset is divided into a training set and a test set according to a ratio of 8:2. The training set is used for parameter learning of the model, and the test set is used to evaluate the prediction effect and generalization ability of the model. The design of the training process focuses on balancing the learning rates of the generator and discriminator to ensure the stability and efficiency of the model. The training uses a batch size of 32, the learning rate of the generator is 0.0002, and the learning rate of the discriminator is 0.00006. The above hyperparameter settings can not only make full use of hardware resources but also effectively balance the learning progress of the generator and discriminator, avoiding the risk of video memory overload or mode collapse. The entire training process is carried out for 500 iterations, and the generator and discriminator are alternately updated in each iteration to gradually optimize the model performance.
[0094] In each iteration, first, the real image and the generated image are input into the discriminator. The discriminator calculates the error and updates its parameters by comparing the differences between the generated image and the real image. Subsequently, the generator adjusts its generation strategy according to the feedback of the discriminator to maximize the misjudgment probability of the discriminator for the generated image. During the training process, the generator gradually learns to generate realistic fracture propagation images under heterogeneous reservoir conditions, and the discrimination ability of the discriminator also continuously improves. To monitor the convergence of the model, the loss values of both the generator and the discriminator are recorded after each iteration to ensure that the training process can be continuously optimized.
[0095] Step S3: Substitute the input characteristic parameters and use the fracture propagation proxy model (FPGAN) to output the fracture propagation image.
[0096] After the fracture propagation proxy model is trained, by substituting the set input characteristic parameters, the rapid acquisition of the fracture propagation image can be realized.
[0097] To evaluate the quality and accuracy of the FPGAN model constructed in the present invention when generating fracture propagation images, the present invention introduces two commonly used evaluation metrics, namely the Fréchet Inception Distance (FID) and the Kernel Inception Distance (KID).
[0098] FID is a metric based on the Frechet distance, which is used to measure the similarity between the distribution of the generated images and the distribution of the real images. The features of the images are extracted through a pre-trained Inception network, and the real images and the generated images are embedded into the feature space, assuming that these features follow a multivariate normal distribution in this space. Then, the mean and covariance matrix (μ r , Σ r ) of the real image distribution, and the mean and covariance matrix (μ g , Σ g ) of the generated image distribution are calculated. Finally, the difference between the real image and the generated image distributions is measured by the following formula:
[0099]
[0100] where Tr represents the trace of the matrix. The lower the FID score, the higher the overall similarity between the generated image and the real image distribution. The present invention uses the FID score to evaluate the global feature consistency between the FPGAN model and the real data when generating fracture propagation images.
[0101] Meanwhile, to further evaluate the detailed features of the generated images, the present invention introduces the Kernel Inception Distance (KID) metric. KID is based on the Maximum Mean Discrepancy (MMD). By using the kernel method, the generated images and the real images are mapped into the feature space to calculate their distances. Specifically, after extracting features through a pre-trained Inception network, the squared maximum mean discrepancy between the generated images and the real images is calculated using a polynomial kernel function, and its definition is as follows:
[0102]
[0103] where x i and y i are the feature vectors of the real images and the generated images respectively, m and n are the numbers of the generated images and the real images respectively, and K is the kernel function. Compared with FID, KID does not assume that the feature distribution is a normal distribution, so it is more sensitive to the complex details in the crack propagation images and can capture the differences in local features between the generated images and the real images.
[0104] By combining the evaluation results of FID and KID, the present invention verifies the generation quality of the FPGAN model. When the model performs well in both FID and KID evaluations, it can more reliably judge the similarity between the generated images and the real images in terms of the overall distribution and detailed features, thus proving the effectiveness and accuracy of the present invention in the simulation of crack propagation in heterogeneous reservoirs.
[0105] The present invention verifies the computational efficiency of the FPGAN model and the generation quality of the crack propagation morphology images, and uses the Fréchet Inception Distance (FID) and the Kernel Inception Distance (KID) as evaluation metrics to comprehensively analyze the generation ability, generalization performance, and computational efficiency of the model under heterogeneous conditions. By comparing the generated images with the real numerical simulation images, the similarity of the model in terms of global features and local details is evaluated; by the FID and KID values under different training iteration numbers, the performance change trend of the model is analyzed.
[0106] When generating images with a resolution of 512×512, the reference standards for FID and KID are shown in Table 3. An FID value lower than 50 and a KID value lower than 0.05 indicate excellent quality of the generated images. The model is trained and tested using a complex dataset, and the FID and KID values of the training set and the test set at different iteration numbers are calculated respectively. The results are shown in Table 4.
[0107] Table 3: Standards for FID and KID values for 512×512 resolution images
[0108]
[0109] Table 4: FID and KID values of the training set and test set at different iteration times
[0110]
[0111] At 300 epochs, the FID values of the training set and test set are 28.7080 and 34.9398 respectively, and the KID values are 0.0529 and 0.0590 respectively. At this time, the generated images can already capture the main features of the crack propagation morphology, but the FID and KID values of the test set are relatively high, indicating that there is still room for optimization in the generalization ability and detailed features of the generator. As the training progresses, at 500 epochs, the generation performance of the model is significantly improved, and the FID and KID values of both the training set and test set decrease significantly, indicating that the overall distribution of the generated images is closer to the real images. In particular, the decrease in the KID value reflects that the model's ability to generate local detailed features has been further enhanced.
[0112] At 800 epochs, the KID value of the test set further drops to 0.0439, indicating that the model's ability to generate details of unseen data has been significantly improved. At this time, both the global contour and local details of the generated images are highly consistent with the real crack propagation morphology images, and the generalization performance reaches the best state. However, when the training reaches 1000 epochs, the FID and KID values increase on both the training set and test set, possibly due to model overfitting. When the generator is further optimized on the training set, the generalization performance on the test set slightly decreases, especially showing a certain degradation in detail generation.
[0113] Based on the above results, it can be seen that the FPGAN model performs most ideally in the training range of 500 to 800 epochs. At this time, the FID and KID values of both the training set and test set are relatively low, the quality of the generated images is extremely high, and it can well capture the macroscopic features and local details of crack propagation, possessing excellent generation ability and generalization performance, providing strong support for the simulation and prediction of crack propagation behavior.
[0114] To verify the prediction performance of the FPGAN model constructed in the present invention in the generation of crack propagation morphology images, a qualitative analysis was carried out on the model generation results under the basic dataset and complex dataset. By comparing the numerical simulation images with the images generated by the surrogate model, the generation accuracy and consistency of the model in terms of crack propagation path, morphology, and multi-physical field characteristics were comprehensively evaluated.
[0115] In the basic dataset, the comparison between the numerical simulation images and the images generated by the surrogate model under different time steps and mechanical conditions is as follows Figure 6As shown in (a)-(b). When the mineral interface is not considered (such as Figure 6 (a)), the crack propagation paths, lengths, and morphologies generated by the surrogate model are highly consistent with the numerical simulation results, indicating that the model can accurately capture the overall development trend of cracks. When the influence of the mineral interface is considered (such as Figure 6 (b)), the surrogate model can also reproduce the constraint or promotion effect of the mineral interface on crack propagation. Whether it is the distribution of branched cracks or the orientation and extension of the main crack, the generated results of the model are in good agreement with the numerical simulation results. This shows that the surrogate model has extremely high generation accuracy on the basic dataset and excellent learning ability for crack propagation characteristics.
[0116] In the complex dataset, under the condition of the same mechanical property combination, the comparison between the numerical simulation images at different time steps and the crack propagation images generated by the surrogate model is as Figure 6 (c)-(d) shown. When the mineral interface is not considered (such as Figure 6 (c)), the surrogate model not only highly restores the orientation of the main crack, the propagation length, and the branched morphology of the crack, but also accurately reproduces the RGB value distribution of the stress field and the fluid pressure in the crack. This shows that the model has good capture ability for the dynamic characteristics of multi-physical fields. When the influence of the mineral interface is considered (such as Figure 6 (d)), the surrogate model shows high generation quality. Under the condition of small cohesive force and low internal friction angle of the mineral interface (such as Figure 6 (c)), the crack branches significantly at the interface and extends significantly in different directions, presenting a complex crack morphology. These phenomena are all highly consistent with the numerical simulation results.
[0117] Generally speaking, the analysis results of the complex dataset further prove the generation ability of the FPGAN model in global features and local details. The model can not only reproduce the geometric morphology and propagation trend of cracks, but also accurately capture the spatial distribution of the stress field and the fluid pressure in the crack under multi-physical field conditions, reflecting strong generalization ability and physical consistency. These characteristics make the surrogate model of the present invention show excellent applicability and reliability in the crack propagation simulation task.
[0118] In the specific implementation process of the present invention, to verify the efficiency advantage of the CWGAN model in crack propagation simulation, the time cost of generating crack propagation morphology images by the traditional numerical simulation method and the surrogate model of the present invention was compared and analyzed under the same conditions.
[0119] Using the traditional numerical simulation method, the crack propagation is simulated by the FDEM software MultiFracS. When simulating the crack propagation morphology images under a specific set of parameters containing 50 frames, the numerical simulation method takes about 10 minutes on average to complete. While the surrogate model described in the present invention is based on the FPGAN neural network and can generate the same number of images under the same conditions in only 40 seconds to 60 seconds, and the time to generate a single image is less than 1 second.
[0120] The traditional numerical simulation method requires manual modification of physical parameters (such as elastic modulus, Poisson's ratio, tensile strength, etc.) and re-running the simulation, which not only takes a long time but also requires manual intervention, resulting in low efficiency. In contrast, after the FPGAN surrogate model described in the present invention is trained, users can directly input any set of physical parameters, and the model can automatically generate the crack propagation morphology images under the corresponding parameter conditions. There is no need to repeatedly modify the parameters manually, significantly reducing the time overhead in the data preparation process.
[0121] Through the surrogate model of the present invention, it is possible to achieve a several-fold increase in efficiency while maintaining the quality of high-resolution image generation. Especially under complex heterogeneous conditions, the surrogate model can quickly generate crack propagation morphology images that conform to physical laws and has the ability to generate in real time and simulate large-scale crack propagation. This efficient generation ability makes the present invention show great application potential in the optimization design of crack propagation and the analysis of influencing factors under heterogeneous conditions.
[0122] The efficiency comparison results between the FPGAN model of the present invention and the traditional numerical simulation method are shown in Table 5:
[0123] Table 5: Efficiency comparison results between the FPGAN model and the traditional numerical simulation method
[0124]
[0125] Through the above experimental analysis, it can be seen that the present invention significantly reduces the computational cost of crack propagation simulation, provides technical support for the high-efficiency and automation of crack propagation simulation, and is particularly suitable for the optimization design of crack propagation and the analysis of influencing factors under complex heterogeneous conditions.
[0126] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for constructing a crack propagation surrogate model based on a generative neural network, characterized in that Comprising the following steps: Step S1: Obtain input characteristic parameters according to the geological characteristics of the target block, construct a reservoir physical model based on the input characteristic parameters, where the input characteristic parameters affect the fracture propagation pattern in the reservoir, and perform numerical simulation on the physical model to obtain a fracture propagation image; Step S2: Construct a sample data set according to the input characteristic parameters and the fracture propagation image, and train the fracture propagation proxy model (FPGAN) using the sample data set until the model parameters converge. The fracture propagation proxy model (FPGAN) includes a generator and a discriminator; Step S3: Substitute the input characteristic parameters and use the fracture propagation proxy model (FPGAN) to output a fracture propagation image.
2. The method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1, wherein the reservoir physical model simulates the heterogeneity of formation rocks; Or, the input characteristic parameters include the mineral composition of formation rocks and the rock mechanical property parameters of each mineral.
3. The method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1, wherein the sample data set is a first data set, the first data set is a basic data set, and the first data set does not include stress field and in - fracture fluid pressure field data; And / or, the sample data set is a second data set, the second data set is a complex data set, and the second data set includes stress field and in - fracture fluid pressure field data.
4. The method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1, wherein the generator combines the input characteristic parameters with random noise.
5. The method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1, wherein the following is further included in step S2: Normalize the fracture propagation image in step S1 into an RGB three - channel image format with a first resolution, and use a scaling operation to scale the image with the first resolution to a second resolution.
6. The training process in step S2 of the method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1 includes: Initialization stage, set the hyperparameters of the network; Training stage, alternately optimize the discriminator and the generator. During the optimization process of the discriminator, by sampling real samples and fake samples generated by the generator, calculate the loss function and the gradient penalty term to ensure that the discriminator satisfies 1 - Lipschitz continuity; During the optimization process of the generator, generate fake samples by inputting a noise vector and a conditional label, calculate the generator loss function, optimize the generation ability and update the parameters.
7. The method for constructing a heterogeneous reservoir fracture propagation proxy model according to claim 1, adopts a conditional Wasserstein gradient penalty (CWGAN - GP) generative adversarial network structure to perform adversarial training on the generator and discriminator of the model.
8. A fracture propagation pattern prediction device based on a generative neural network, comprising the following steps: A crack propagation image sample acquisition module, configured to obtain input feature parameters according to the geological characteristics of a target block, construct a heterogeneous reservoir physical model according to the input feature parameters, where the input feature parameters affect the crack propagation pattern in the heterogeneous reservoir, and perform numerical simulation on the physical model to obtain crack propagation image samples; A crack propagation surrogate model training module, configured to construct a sample data set according to input feature parameters and crack propagation images, and train a crack propagation surrogate model (FPGAN) using the sample data set until the model parameters converge, where the crack propagation surrogate model (FPGAN) includes a generator and a discriminator; A crack propagation image prediction module, configured to substitute input feature parameters and output a crack propagation image using the crack propagation surrogate model (FPGAN).
9. An electronic device, the electronic device comprising: One or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a reservoir crack propagation surrogate model according to any one of claims 1-7.
10. A computer-readable storage medium, having stored thereon a computer program, which when executed by a processor implements the method for constructing a reservoir crack propagation surrogate model according to any one of claims 1-7.