Adversarial generative fracture network expansion simulation system and method thereof

Abstract

The invention discloses a generative adversarial fracture network expansion simulation system and a method thereof. The system constructs a digital twinborn model of crack propagation based on a generative adversarial network. The generator adopts a 3D convolutional neural network structure and inputs geomechanical parameter tensors to generate a three-dimensional fracture network form; a discriminator is introduced into a physical constraint module, a fracture mechanical equation is embedded into a discrimination process to serve as a regularization item, and it is ensured that a generated crack meets mechanical rationality. In the training process, a crack fractal dimension evaluation algorithm is developed, the Hausdorff fractal dimension of the generated crack network is calculated, and the Hausdorff fractal dimension is compared with the real crack network fractal dimension to serve as constraint loss. In the application of shale oil horizontal well fracturing design, the matching degree of the fracture network generated by the system and the microseism monitoring result reaches about 89%, and the distribution optimization efficiency of the proppant in the complex fracture network is improved by about 3 times.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas field exploitation and rock mechanics simulation, and in particular to a counter-generative fracture network expansion simulation system and method thereof. Background Art

[0002] Hydraulic fracturing technology is widely used in unconventional oil and gas production. By pumping high-pressure fluid into the formation, a network of fractures is created in the rock formation, thereby increasing oil and gas seepage channels and production. However, the morphology and expansion behavior of the fracture network are controlled by multiple factors, including the ground stress field, lithology, and construction parameters, and are highly nonlinear and uncertain. Traditional fracture simulation methods are mainly based on continuum mechanics or discrete fracture network (DFN) modeling. These methods require pre-assumptions about the geometry or statistical distribution of the fractures, making it difficult to accurately characterize the spatial distribution of complex networks. In addition, the simulation results must be verified and corrected through means such as microseismic monitoring, which is a time-consuming and labor-intensive process with limited accuracy.

[0003] In recent years, fractal geometry has been introduced to describe and quantitatively analyze the complexity of rock fracture networks. Studies have shown that fracture systems in rock masses often have fractal structural characteristics, and their spatial filling degree and heterogeneity can be quantitatively characterized by fractal dimension. For example, existing methods use digital image processing technology to calculate the fractal dimension of rock fractures to assist in the assessment of rock mass structural stability. Fractal dimension has been widely used to describe and characterize rock fracture networks and is an important analytical tool in engineering problems such as coal mining and oil and gas seepage. Based on fractal theory, some researchers have combined microseismic monitoring data with production data to calibrate and characterize the fracture network. Methods such as fractal interpolation have also been used to generate fitted fracture geometries with multi-scale variations, thereby more realistically reproducing the rough bifurcation characteristics of natural fractures. However, existing technologies mainly use fractal geometry as an auxiliary analysis method, and there is no solution to directly incorporate fractal indicators into the fracture generation model.

[0004] With the development of artificial intelligence, data-driven methods have begun to be applied in underground engineering. One study proposed a generative model combining variational autoencoders and generative adversarial networks (GANs) to perform stochastic inversion of fracture network parameter distributions, thereby leveraging prior constraints to improve the rationality of fracture distribution predictions. These methods can, to a certain extent, capture the statistical distribution of complex fracture field parameters and achieve a close approximation of the actual fracture network. However, these methods primarily focus on parameter inversion and do not directly simulate the geometric propagation of fractures. Furthermore, they lack mechanisms to incorporate the physical laws of fracture mechanics into the model. Summary of the Invention

[0005] Technical purpose: In response to the problems of insufficient accuracy and difficulty in integrating measured data in existing fracturing fracture simulations, the present invention discloses an adversarial generative fracture network expansion simulation system and method. It uses a generative adversarial network to integrate the physical laws of rock mechanics and fractal geometric characteristics to generate a digital twin model that is highly matched with the real fracture network, so as to overcome the shortcomings of existing technologies in terms of accuracy and versatility in fracture network simulation.

[0006] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] A countermeasure generative crack network expansion simulation system, comprising a generator and a discriminator;

[0008] The generator is a three-dimensional convolutional neural network model that receives the geomechanical parameter tensor of the formation as input and generates the corresponding three-dimensional fracture network morphology output;

[0009] The discriminator is a discriminative neural network that receives the generated fracture network morphology and real fracture network data and outputs a discriminant signal for the authenticity of the generated fracture network. The discriminator integrates a physical constraint module that embeds the fracture mechanics equation as a regularization constraint into the discriminant process to evaluate the degree of conformity of the generated fracture network with the fracture mechanics criteria in the discriminant output.

[0010] Preferably, the geomechanical parameter tensor includes multiple parameters such as rock stratum elastic modulus, Poisson's ratio, and ground stress magnitude and direction that characterize the mechanical properties of the rock. The generator extracts and reconstructs the geomechanical parameter tensor through multi-layer three-dimensional convolution and upsampling to output the three-dimensional distribution tensor of the fracture network.

[0011] Preferably, the physical constraint module evaluates the generated crack network based on the fracture mechanics criterion. When it is detected that the tip stress intensity factor of the generated crack is lower than the predetermined fracture threshold and produces unreasonable expansion, the discriminator outputs a false discrimination result for the generated sample and adds a corresponding penalty term to the loss function; when the generated crack satisfies the constraints of the fracture mechanics equation, the weight of the penalty term is zero.

[0012] Preferably, it also includes a fractal dimension evaluation module for calculating the fractal dimension of the generated fracture network morphology and comparing it with the fractal dimension of the real fracture network; the training process of the generator is combined with the comparison result for feedback adjustment so that the absolute value of the difference between the fractal dimension of the generated fracture network and the fractal dimension of the real fracture network is less than 0.15.

[0013] Preferably, the generator and discriminator are trained by an adversarial generation algorithm, the loss function of the discriminator includes the physical constraint loss provided by the physical constraint module, and the loss function of the generator includes the fractal dimension loss provided by the fractal dimension evaluation module, so as to jointly constrain the physical rationality and geometric fractal characteristics of the crack network morphology generated by the generator.

[0014] A method for simulating adversarial generative crack network expansion is applied to the adversarial generative crack network expansion system described above, and specifically comprises the following steps:

[0015] S1. Obtain multiple geomechanical parameters of the target reservoir and construct a parameter tensor as input to the generator;

[0016] S2. Input the parameter tensor into the generator to generate the corresponding three-dimensional morphology of the fracture network, and input the real fracture network data into the discriminator to compare and discriminate with the generated three-dimensional morphology of the fracture network;

[0017] S3. During the discrimination process, the deviation degree of the generated crack network from the fracture mechanics equation is calculated, and penalties are imposed on the parts that do not meet the fracture mechanics constraints, so that the discrimination results of the discriminator also reflect physical rationality;

[0018] S4, evaluating the fractal dimension of the generated fracture network, comparing it with the fractal dimension of the real fracture network, and calculating the fractal dimension error;

[0019] S5. adjusting the model parameters of the generator based on the discrimination result and the fractal dimension error, including: when the discrimination result shows that the generated cracks are not real or the fractal dimension error exceeds a threshold, updating the generator parameters to reduce the discrimination loss and the fractal loss;

[0020] S6, repeating steps S1 to S5 until the generated fracture network is judged as real by the discriminator and the absolute value of the difference between its fractal dimension and the fractal dimension of the real fracture network is lower than a preset threshold, thus completing the generation model training;

[0021] S7. Apply the trained generator to actual fracturing simulations, input new geomechanical parameter tensors, generate predicted fracture network extension morphologies, and calibrate and verify the predicted morphologies based on microseismic data monitored in the field to guide fracturing design optimization.

[0022] Preferably, the fracture mechanics equation in step S3 includes a fracture toughness criterion or an energy release rate criterion of the rock material, and the discriminator generates the stress intensity factor K of each crack tip relative to the critical toughness K of the rock. Ic The physical rationality of crack expansion is determined by the relationship between K gen < K Ic When an expansion occurs, a penalty signal is output.

[0023] Preferably, the fractal dimension in step S4 is calculated using a three-dimensional box counting method, which includes meshing the fracture network morphology at different spatial scales and counting the number of cells occupied by fractures, thereby estimating the Hausdorff fractal dimension.

[0024] Preferably, the update of the generator parameters in step S5 adopts the back-propagation algorithm to minimize the comprehensive loss function, wherein the comprehensive loss function includes an adversarial loss term, a physical constraint loss term and a fractal dimension loss term, wherein the weights of each term are adjusted in stages according to the training process, so as to first reduce the fractal dimension error to the target range, and then highlight the reduction of the physical constraint deviation.

[0025] Preferably, in step S7, the trained and calibrated generator is used to predict the fracture network expansion results under different fracturing parameter scenarios, and the predicted results are matched and analyzed with the spatial distribution of events monitored by microseismic monitoring; the fracturing construction parameters are adjusted or the proppant delivery scheme is optimized according to the matching degree, so that the predicted fracture distribution and the actual monitoring results are consistent with the expected target, thereby improving the fracturing production increase effect.

[0026] Beneficial effects: The anti-generative crack network expansion simulation system and method provided by the present invention have the following beneficial effects:

[0027] 1. The present invention embeds fracture mechanics constraints in the generative model, so that the generated crack network strictly follows physical criteria such as energy release rate balance, avoiding the physically unreasonable crack morphology that may occur in pure data-driven models and improving the credibility of the simulation results.

[0028] 2. By developing a fractal dimension evaluation algorithm and integrating it into adversarial training, this invention ensures that the geometric complexity of the generated fracture network is consistent with that of the real network. Adversarial training converges the Hausdorff fractal dimension error of the generated fractures to within 0.15, significantly improving the ability to depict fracture network details (such as branching and roughness), and ensuring that the spatial distribution of the simulated fractures closely matches the fractal characteristics of real rock fractures.

[0029] 3. This system is a digital twin model that dynamically updates data from monitoring during the fracturing process (microseismic events, downhole pressure, etc.). Simulation results match monitoring data by approximately 89%, significantly higher than traditional models. This system can promptly reflect the impact of construction parameter adjustments on crack propagation, guiding on-site decision-making.

[0030] 4. The realistic fracture network generated by this invention allows for optimized proppant placement and distribution within fractures. Due to the more accurate fracture network prediction, the proportion of proppant entering effective fractures is significantly increased. Field applications have shown that, under the same sand loading conditions, the proppant coverage volume optimized by this invention is approximately three times that of conventional designs, significantly improving fracturing efficiency and oil and gas production capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments or the prior art.

[0032] Figure 1 A diagram of a system architecture in one embodiment of the present invention;

[0033] Figure 2 Schematic diagram of the process of the present invention;

[0034] Figure 3 This is a schematic diagram of the matching effect between the model prediction and microseismic monitoring of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be described more clearly and completely below by way of a preferred embodiment in conjunction with the accompanying drawings, but the present invention is not limited to the scope of the embodiment.

[0036] The present invention provides a counter-generative crack network expansion simulation system, comprising a generator and a discriminator;

[0037] The generator is a three-dimensional convolutional neural network model that receives the geomechanical parameter tensor of the formation as input and generates the corresponding three-dimensional fracture network morphology output.

[0038] The generator is a deep generative model based on a 3D convolutional neural network (3D CNN). It takes as input a geomechanical parameter tensor representing formation and rock properties and outputs a simulated 3D fracture network morphology. The geomechanical parameter tensor contains at least Young's modulus, Poisson's ratio, and other parameters that characterize rock rigidity and deformation properties.

[0039] The discriminator is a discriminative neural network that receives the generated fracture network morphology and real fracture network data and outputs a discriminant signal for the authenticity of the generated fracture network. The discriminator integrates a physical constraint module that embeds the fracture mechanics equation as a regularization constraint into the discriminant process to evaluate the degree of conformity of the generated fracture network with the fracture mechanics criteria in the discriminant output.

[0040] Specifically, the physical constraint module evaluates whether the generated fracture network meets the mechanical criteria for rock crack propagation, such as whether the tip stress intensity factor reaches the rock fracture toughness threshold. If deviations from physical laws occur, the discriminator's loss function imposes additional penalties, thereby guiding the generator to produce fracture morphologies that are more consistent with mechanical principles.

[0041] In one embodiment, a generative adversarial training strategy is used to iteratively optimize the generator and discriminator. During the training process, a fractal dimension evaluation algorithm is introduced to measure the morphology of the crack network. The algorithm is based on the three-dimensional box counting method, calculates the Hausdorff fractal dimension of the generated crack network, and compares it with the target fractal dimension of the real crack network to obtain the fractal dimension error. During the training process, this error is used as a component of the generator loss, and the fractal dimension error is continuously reduced through adversarial training. Preferably, when the absolute value of the difference between the fractal dimension of the generated crack network and the fractal dimension of the real network is less than 0.15, the model is judged to have converged.

[0042] Once trained, the generator can be used in conjunction with field monitoring data for real-time simulation and optimization of shale oil horizontal well fracturing designs. By inputting different geological parameter scenarios, the generator can rapidly generate corresponding fracture network propagation predictions. The generated fracture network is then calibrated and matched with microseismic monitoring results, enabling dynamic updates of fracture propagation within the digital twin environment.

[0043] like Figure 2 As shown, the present invention also provides a simulation method for countering generative crack network expansion, which is applied to a countering generative crack network expansion system as described above, and specifically includes the following steps:

[0044] S1. Obtain multiple geomechanical parameters of the target reservoir and construct a parameter tensor as input to the generator;

[0045] S2. Input the parameter tensor into the generator to generate the corresponding three-dimensional morphology of the fracture network, and input the real fracture network data into the discriminator to compare and discriminate with the generated three-dimensional morphology of the fracture network;

[0046] S3. During the discrimination process, the deviation degree of the generated crack network from the fracture mechanics equation is calculated, and penalties are imposed on the parts that do not meet the fracture mechanics constraints, so that the discrimination results of the discriminator also reflect physical rationality;

[0047] S4, evaluating the fractal dimension of the generated fracture network, comparing it with the fractal dimension of the real fracture network, and calculating the fractal dimension error;

[0048] The discriminator receives the real crack network Y and the generated crack network G(X), extracts features through convolution and calculates the discriminant loss L adv , while calculating the physical constraint loss Lphys and fractal constraint loss L frac The loss of physical constraints depends on whether the generated crack violates the fracture mechanics equations, such as the stress intensity factor K at the tip of the generated crack. gen Exceeds rock fracture toughness K Ic The fractal constraint loss is calculated based on the fractal dimension D of the generated fracture network. gen Compared with the fractal dimension D of the real crack network real The difference calculation (such as squared error ).

[0049] S5. adjusting the model parameters of the generator based on the discrimination result and the fractal dimension error, including: when the discrimination result shows that the generated cracks are not real or the fractal dimension error exceeds a threshold, updating the generator parameters to reduce the discrimination loss and the fractal loss;

[0050] The weighted combination of the losses output by the above discriminator forms a total loss function:

[0051]

[0052] in, It is an adversarial loss, which is used to improve the indistinguishability between generated samples and real samples; It is the loss of physical constraints, reflecting the deviation of the generated crack from the mechanical criterion; is the fractal dimension loss, e.g. , which measures the gap between the fractal dimension of the generated crack network and the target fractal dimension. and is a weight coefficient used to balance the contribution of physical constraints and fractal constraints to the overall loss. and It can be adjusted according to training needs and gradually reduced after the fractal constraint converges. To enhance physical consistency.

[0053] Based on the total loss, the model parameters of the generator and discriminator are updated separately through back propagation so that When When convergence and fractal error constraints are satisfied (e.g. ), and the training ends.

[0054] S6, repeating steps S1 to S5 until the generated fracture network is judged as real by the discriminator and the absolute value of the difference between its fractal dimension and the fractal dimension of the real fracture network is lower than a preset threshold, thus completing the generation model training;

[0055] S7. Apply the trained generator to actual fracturing simulations, input new geomechanical parameter tensors, generate predicted fracture network extension morphologies, and calibrate and verify the predicted morphologies based on microseismic data monitored in the field to guide fracturing design optimization.

[0056] The trained generator is then used in actual fracturing simulations. Given a new well's geomechanical parameter tensor input, X', the generator outputs a predicted three-dimensional fracture network model, G(X'). This model can be compared with real-time fracture propagation data from microseismic monitoring to verify the match. If necessary, monitoring data can be fed back to adjust the model input or refine the discriminator, thereby calibrating the digital twin model in real time and ensuring that the simulation results are more consistent with field reality.

[0057] A calibrated digital twin model of the fracture network is used to assist in optimizing fracturing operations. The fracture distribution generated by the present system significantly improves the agreement between fracture geometry (length, height, and complexity) and microseismic cloud maps compared to traditional simulation methods. Based on actual application testing, the match between the simulated fracture volume and the microseismic event cloud volume has increased to approximately 89%, significantly outperforming the results of conventional single-physics models. Leveraging more accurate fracture network predictions, the present invention can guide the optimization of sand proppant placement strategies, achieving efficient matching of proppant distribution with fracture diversion channels, increasing fracturing proppant coverage efficiency by approximately three times and further improving post-fracturing oil and gas production.

[0058] Example 1

[0059] like Figure 1 As shown, the adversarial generative fracture network expansion simulation system provided in this embodiment includes three parts: a generator, a discriminator, and a training feedback module. The generator is a deep generative model based on a 3D convolutional neural network, and its input is a four-dimensional tensor of size L×W×H×C, where L, W, and H correspond to the spatial dimensions of the reservoir simulation grid, and C is the number of input features. In this embodiment, C=12, including the elastic modulus E (unit: GPa), Poisson's ratio ν (dimensionless), triaxial stress components of the reservoir lithology , , (Unit: MPa), natural crack density (unit: bars / m³) and other parameters representing reservoir mechanical and structural properties. The generator consists of several layers of 3D convolution and upsampling, fusing parameter tensor information layer by layer to ultimately output a 3D fracture distribution tensor of the same size as the reservoir model. The value of each voxel in the 3D fracture distribution tensor represents the probability of a fracture at that spatial location or the size of the fracture aperture, thus describing the simulated fracture network morphology.

[0060] The discriminator part consists of two parts: a convolutional discriminant network and a physical constraint module. The convolutional discriminant network structure is similar to that of the generator (the number of convolutional layers may be different), and is used to extract multi-scale features of the input crack network. The discriminator accepts both the real crack network Y and the generated crack network G(X) as input, and outputs the classification judgment result D(X) (true or false) by comparing the feature distribution of the two. During the judgment process, a physical constraint module is introduced to ensure that the judgment result takes into account the rationality of crack mechanics. Specifically, the physical constraint module calculates whether each crack network sample received by the discriminator meets the physical criteria for fracture extension. For example, in this embodiment, the linear elastic fracture mechanics criterion is adopted: for each main crack generated, its tip stress intensity factor K is estimated. gen and the fracture toughness K of rock materials Ic Compare. If K gen ≥K Ic If the crack is still growing, it is considered physically consistent; if K gen Significantly smaller than K Ic However, a large crack appears, indicating that the generated result does not conform to the physical law, and the discriminator will give a "false" judgment. In this way, the physical constraint module is equivalent to adding a physical penalty function L to the loss function of the discriminator. phys , whose form depends on the specific physical criteria. For example, we can set:

[0061]

[0062] where N c is the total number of crack tips in the generated crack network, K gen,i is the stress intensity factor at the i-th crack tip (unit: MPa·m 0.5 ), K Ic is the critical value of fracture toughness of rock matrix material (unit: MPa·m 0.5 ), Ensure that the penalty is only generated when the generated crack does not meet the extension criterion. Through this physical loss definition, when the generated crack meets the mechanical conditions, L phys = 0, and when violated, it will grow with the square of the gap. The total loss of the discriminator L D The weighted sum of adversarial loss and physical loss is obtained:

[0063]

[0064] in is the classic GAN discriminator loss, is the weight (in this embodiment, =1).

[0065] The training feedback module is responsible for coordinating the parameter updates of the generator and discriminator to achieve adversarial training of the model. An iterative training strategy is adopted: in each training batch, the generator is first fixed, and the discriminator parameters are updated using real and generated samples to distinguish between real and fake samples as accurately as possible; then the discriminator is fixed and the generator parameters are updated based on the loss feedback given by the discriminator. In order to incorporate the fractal dimension, a geometric metric, into the training process, the training feedback module introduces a fractal dimension evaluation algorithm to calculate the fractal dimension difference between the current generated crack network and the real crack network, and use it as a constraint signal to apply to the generator update. The specific implementation is as follows:

[0066] At each generator parameter update step, a fractal dimension evaluation algorithm is first applied to the currently generated fracture network G(X). This algorithm performs a three-dimensional box counting analysis on the fracture morphology in G(X): voxel grids of different scales ϵ are selected, and the number of fracture-filling boxes N(ϵ) at each scale is counted. The Hausdorff fractal dimension is then approximated based on fractal theory:

[0067]

[0068] in 、 is the box size at two consecutive scales ( < ), 、 is the number of boxes covering the cracks at the corresponding scale. Multi-scale linear regression can be used to further improve The accuracy of calculation. This embodiment uses three different scales (for example Box counting regression was performed with 2 times, 1 times, and 0.5 times the original grid size, respectively, to reduce random errors.

[0069] Similarly, the same algorithm is applied to the real fracture network Y to obtain the fractal dimension D of the real fracture network. real Then calculate the difference between the two △D = D gen -D real . Define fractal loss as:

[0070]

[0071] The loss in It reaches the minimum value 0 when The larger the value (positive or negative), the greater the loss. The training feedback module adds this fractal loss to the objective function of the generator, thus forming the total loss L of the generator. G :

[0072]

[0073] in is the adversarial loss defined by the generator in GAN to deceive the discriminator (such as using -lnD(G(X)) or Wasserstein distance metric), is the fractal constraint loss weight hyperparameter. In the early stage of training, appropriately increase It can accelerate the generator’s learning of fractal features; in the later stage of training, when D gen Gradually approaching D real When To avoid excessive pursuit of fractal consistency at the expense of other features. ∇θ is calculated by back propagation algorithm. G L G , update the parameters θ of the generator G At this time, the generator not only passes L adv Optimize the overall realism, also through L frac Directly correct the fractal properties of its own output so that the generated crack network gradually approaches the real data in terms of fractal dimension.

[0074] The above training process is repeated until the convergence condition is met. In addition to the conventional adversarial loss stability, the convergence must also meet the requirement that the fractal dimension error is within the allowable range. In this embodiment, the absolute error threshold of the fractal dimension is set to 0.15, that is, when If the performance does not decrease significantly after several iterations, the model is considered to have learned the fractal characteristics of the fracture network and training can be terminated. The resulting generator model is a digital twin model of the fracture network that integrates physical laws and fractal constraints.

[0075] Example 2

[0076] In this example, the model trained in Example 1 is applied to the optimization of the fracturing scheme of a horizontal well in a shale oil reservoir. The well is buried at a depth of about 3,000 meters, and the target layer is dense shale. First, the geomechanical parameters of the well section are obtained through core analysis, well logging, and stress testing: including the Young's modulus E = 25 GPa, Poisson's ratio ν = 0.25, and the maximum horizontal principal stress. =72MPa, minimum horizontal principal stress =68MPa, vertical stress =80MPa, pore pressure P p =45 MPa, along with 12 other parameters, including natural fracture density and toughness, form the geomechanical parameter tensor input X'. X' is fed into the trained generator to generate the predicted three-dimensional fracture network morphology G(X'). The model predicts that each cluster fracturing segment in this well will produce a complex multi-fracture network structure, with the main fracture extending approximately 120 meters along the direction of maximum stress, and branching fractures exhibiting high tortuosity and a fractal branching structure (predicted fractal dimension of approximately 1.85). This non-planar, complex fracture network is consistent with known fracturing behavior in the region.

[0077] In order to verify the accuracy of the model, microseismic monitoring was carried out on site. The microseismic event cloud map during the fracturing operation was projected to the same scale and compared with the fracture geometry predicted by the model. Figure 3 As shown in the figure, the model predicted the direction and length of the main fractures, which were highly consistent with the distribution of microseismic events. The spatial extent of the microseismic event cloud and the simulated fracture network volume overlapped by 89%, significantly higher than the approximately 60% overlap between the traditional KGD model prediction and actual measurement. This indicates that the proposed model effectively captures the distribution trend of the real fracture network. At the same time, the fracture aperture distribution output by the model was combined with proppant migration simulation to optimize the proppant injection and displacement. Compared with the original design, the optimized proppant dosage per cluster was reduced by 10%, while the proppant coverage efficiency within the fracture increased by approximately 3 times, and the proppant placement at the distal end was more uniform. The production test results after the completion of the fracturing operation also verified the effectiveness of the model optimization - the test oil production increased by approximately 25% compared with the adjacent wells, proving that the formation of a more complex fracture network effectively expanded the stimulated volume (SRV).

[0078] As demonstrated in the aforementioned examples, the adversarial generative fracture network expansion simulation system described in this invention can reliably simulate fracture network morphology under diverse reservoir conditions. By introducing physical and fractal constraints, it achieves high-precision approximation of actual fracturing effects. In practical applications, engineers can use this digital twin system to predict fracture distribution and productivity impacts under varying fracturing parameters, assisting in developing optimized fracturing plans and improving the development efficiency of unconventional oil and gas wells. The system can also be expanded to applications in fields requiring fracture expansion simulation, such as coalbed methane and hydraulic geopressure control.

[0079] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A countermeasure generative crack network expansion simulation system, characterized in that: Including generator and discriminator; The generator is a three-dimensional convolutional neural network model that receives the geomechanical parameter tensor of the formation as input and generates the corresponding three-dimensional fracture network morphology output; The discriminator is a discriminative neural network that receives the generated fracture network morphology and real fracture network data and outputs a discriminant signal for the authenticity of the generated fracture network. The discriminator integrates a physical constraint module that embeds the fracture mechanics equation as a regularization constraint into the discriminant process to evaluate the degree of conformity of the generated fracture network with the fracture mechanics criteria in the discriminant output.

2. The anti-generative crack network expansion simulation system according to claim 1, characterized in that: The geomechanical parameter tensor includes multiple parameters that characterize the mechanical properties of rock, such as the elastic modulus of the rock formation, Poisson's ratio, and the magnitude and direction of ground stress. The generator extracts and reconstructs the geomechanical parameter tensor through multi-layer three-dimensional convolution and upsampling to output the three-dimensional distribution tensor of the fracture network.

3. The anti-generative crack network expansion simulation system according to claim 1, characterized in that: The physical constraint module evaluates the generated crack network based on the fracture mechanics criteria. When it is detected that the tip stress intensity factor of the generated crack is lower than the predetermined fracture threshold and produces unreasonable expansion, the discriminator outputs a false discrimination result for the generated sample and adds a corresponding penalty term to the loss function. When the generated crack satisfies the constraints of the fracture mechanics equation, the weight of the penalty term is zero.

4. The anti-generative crack network expansion simulation system according to claim 1, characterized in that: It also includes a fractal dimension evaluation module for calculating the fractal dimension of the generated fracture network morphology and comparing it with the fractal dimension of the real fracture network; the training process of the generator is combined with the comparison result for feedback adjustment so that the absolute value of the difference between the fractal dimension of the generated fracture network and the fractal dimension of the real fracture network is less than 0.

15.

5. The anti-generative crack network expansion simulation system according to claim 4, characterized in that: The generator and discriminator are trained through an adversarial generation algorithm. The loss function of the discriminator includes the physical constraint loss provided by the physical constraint module, and the loss function of the generator includes the fractal dimension loss provided by the fractal dimension evaluation module, so as to jointly constrain the physical rationality and geometric fractal characteristics of the crack network morphology generated by the generator.

6. A method for simulating the expansion of a counter-generative crack network, characterized in that: A system for resisting generative crack network expansion as claimed in any one of claims 1 to 5, comprising the following steps: S1. Obtain multiple geomechanical parameters of the target reservoir and construct a parameter tensor as input to the generator; S2. Input the parameter tensor into the generator to generate the corresponding three-dimensional morphology of the fracture network, and input the real fracture network data into the discriminator to compare and discriminate with the generated three-dimensional morphology of the fracture network; S3. During the discrimination process, the deviation degree of the generated crack network from the fracture mechanics equation is calculated, and penalties are imposed on the parts that do not meet the fracture mechanics constraints, so that the discrimination results of the discriminator also reflect physical rationality; S4, evaluating the fractal dimension of the generated fracture network, comparing it with the fractal dimension of the real fracture network, and calculating the fractal dimension error; S5. adjusting the model parameters of the generator based on the discrimination result and the fractal dimension error, including: when the discrimination result shows that the generated cracks are not real or the fractal dimension error exceeds a threshold, updating the generator parameters to reduce the discrimination loss and the fractal loss; S6, repeating steps S1 to S5 until the generated fracture network is judged as real by the discriminator and the absolute value of the difference between its fractal dimension and the fractal dimension of the real fracture network is lower than a preset threshold, thus completing the generation model training; S7. Apply the trained generator to actual fracturing simulations, input new geomechanical parameter tensors, generate predicted fracture network extension morphologies, and calibrate and verify the predicted morphologies based on microseismic data monitored in the field to guide fracturing design optimization.

7. The method for simulating the expansion of a counter-generative crack network according to claim 6, characterized in that: The fracture mechanics equation in step S3 includes the fracture toughness criterion or energy release rate criterion of the rock material. The discriminator generates the stress intensity factor K of each crack tip according to the stress intensity factor K of each crack tip. gen Relative to the critical toughness K of rock Ic The physical rationality of crack expansion is determined by the relationship between K gen < K Ic When an expansion occurs, a penalty signal is output.

8. The method for simulating the expansion of a counter-generative crack network according to claim 6, characterized in that: The fractal dimension in step S4 is calculated using a three-dimensional box counting method, which includes meshing the fracture network morphology at different spatial scales and counting the number of cells occupied by fractures, thereby estimating the Hausdorff fractal dimension.

9. The method for simulating the expansion of a counter-generative crack network according to claim 6, characterized in that: The update of the generator parameters in step S5 adopts the back-propagation algorithm to minimize the comprehensive loss function, which includes the adversarial loss term, the physical constraint loss term and the fractal dimension loss term. The weights of each term are adjusted in stages according to the training process to first reduce the fractal dimension error to the target range and then highlight the reduction of the physical constraint deviation.

10. The method for simulating the expansion of a counter-generative crack network according to claim 6, characterized in that: In step S7, the trained and calibrated generator is used to predict the fracture network expansion results under different fracturing parameter scenarios, and the predicted results are matched and analyzed with the spatial distribution of events monitored by microseismic monitoring. According to the matching degree, the fracturing construction parameters are adjusted or the proppant placement plan is optimized so that the predicted fracture distribution and the actual monitoring results are consistent with the expected target, thereby improving the fracturing production increase effect.

Images