Fracture network optimization characterization and uncertainty control method and system

Abstract

The invention discloses a fracture network optimization characterization and uncertainty control method and system, and the method comprises the steps: S1, constructing an SA-UNet substitution model and a DAOCRN deep learning inversion model based on a self-attention mechanism according to the multi-source monitoring data of a high-level radioactive waste disposal site and a typical fracture network sample; s2, according to the SA-UNet substitution model, initial population generation and simulation calculation, mismatch degree evaluation and optimal set screening are carried out through genetic algorithm inversion, and a fracture probability result is obtained; and S3, according to the fracture probability result and a DAOCRN deep learning inversion model, through deep learning inversion optimization representation, obtaining a final fracture network representation result with the minimum mismatch degree. By adopting the technical scheme of the invention, accurate characterization and uncertainty reduction of the fracture network in a high-level radioactive waste deep geological disposal scene are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of high-level waste disposal, and particularly relates to a fissure network optimization characterization and uncertainty control method and system. BACKGROUND

[0002] In the field of high-level waste deep geological disposal, the complexity of geological structure and the wide distribution of fissured rock mass constitute the core technical challenge. The seepage dynamics simulation of groundwater in the fissure network and the accurate characterization of the spatial form of the fissure network directly determine the reliability of the nuclide transport path and diffusion risk assessment, which is the key technical support to ensure the long-term safety of the disposal project. In addition, the concealment of the underground geological structure leads to significant sparsity of the observation data, and the characterization of the fissure network based on such data easily introduces a large amount of uncertainty, which in turn causes systematic deviation in the numerical simulation of nuclide transport. How to efficiently characterize the fissure network and reduce uncertainty relying on limited observation data has become a core problem restricting the safety evaluation and engineering decision of high-level waste deep geological disposal.

[0003] To improve the reliability of fissure network characterization and support the safety analysis of high-level waste deep geological disposal, data assimilation methods are widely used to integrate multi-source observation data, and stochastic inversion methods are used as mainstream technologies to derive representative algorithms such as Markov Chain Monte Carlo and genetic algorithm under the support of the Bayesian probability inversion framework. However, stochastic inversion methods are easily affected by the ill-posedness of the inverse problem, and may produce multiple local optimal solutions. Moreover, the high-dimensional parameter space involved in the characterization of the fissure network leads to high computational cost, which seriously limits its application efficiency in the high-precision simulation scenario of high-level waste deep geological disposal. At the same time, deep learning inversion methods have rapidly risen due to their excellent data processing and high-dimensional information extraction capabilities, and have performed outstandingly in nonlinear inversion tasks. However, this method is strongly dependent on massive reliable data sets, which forms a prominent contradiction with the high difficulty and high cost of obtaining observation data in high-level waste disposal sites, greatly hindering its engineering promotion process.

[0004] The application of the Bayesian probability framework related to high-level waste deep geological disposal and uncertainty analysis requires repeated iteration to update and fit the observation data and high-fidelity numerical model results, consuming a large amount of computing resources. To solve this problem, surrogate models have emerged, which can approximate the input-output relationship of high-fidelity models at a lower computational cost, significantly improving the inversion efficiency. Currently, traditional types such as Gaussian process and support vector machine have been developed. However, traditional surrogate models generally face the "dimension disaster", with computational cost increasing exponentially with input dimension, making it difficult to adapt to the demand of high-dimensional fissure network characterization in the high-level waste disposal scenario. In recent years, deep learning surrogate models for image regression have made key breakthroughs in handling high-dimensional and nonlinear problems, effectively improving the computational efficiency. However, in the specific scenario of high-level waste deep geological disposal, the adaptability to complex geological conditions and the stability of the simulation results still need to be further optimized and improved. SUMMARY

[0005] To solve the problems in the prior art, the application provides a fissure network optimization characterization and uncertainty control method and system.

[0006] To achieve the above-mentioned object, the application provides the following scheme: A fissure network optimization characterization and uncertainty control method, comprising: Step S1, constructing a SA-UNet substitute model based on a self-attention mechanism and a DAOCRN deep learning inversion model according to multi-source observation data of a high-level waste disposal site and a typical fissure network sample; Step S2, generating an initial population and performing simulation calculation, evaluating a mismatch degree and screening an optimal set through genetic algorithm inversion according to the SA-UNet substitute model, to obtain a fissure probability result; Step S3, optimizing characterization through deep learning inversion according to the fissure probability result and the DAOCRN deep learning inversion model, to obtain a final fissure network characterization result with the minimum mismatch degree.

[0007] Preferably, in step S1, the multi-source observation data comprises direct observation data such as drilling logs, observation data of radionuclide concentration, and regional geological prior information.

[0008] Preferably, in step S2, the initial fissure network population is generated with the observation data and the geological prior information as constraints; for each generated fissure network in the population, a pre-trained SA-UNet substitute model is used to quickly simulate the corresponding physical process; a comprehensive mismatch degree function is used to calculate the deviation of each fissure network simulation result from the observation data, and an optimal fissure network set with a lower mismatch degree is screened out; the optimal fissure network set is iteratively output to obtain a final fissure probability result through genetic algorithm.

[0009] Preferably, in step S3, the fissure probability result output by the genetic algorithm and the site observation data are input into the pre-trained DAOCRN model to extract spatial feature information of the fissure network; when the model training reaches a preset iteration round or the loss function value tends to be stable, the optimization process of the DAOCRN model is terminated, and a final fissure network characterization result with the minimum mismatch degree is output.

[0010] The application further provides a fissure network optimization characterization and uncertainty control system, comprising: A first processing module is configured to construct a SA-UNet substitute model based on a self-attention mechanism and a DAOCRN deep learning inversion model according to multi-source observation data of a high-level waste disposal site and a typical fissure network sample; The second processing module is configured to generate an initial population and perform simulation calculation, evaluate a mismatch degree, and screen an optimal set according to the SA-UNet substitute model through genetic algorithm inversion, so as to obtain a fracture probability result. The third processing module is configured to perform deep learning inversion optimization representation according to the fracture probability result and the DAOCRN deep learning inversion model through deep learning inversion, so as to obtain a final fracture network representation result with the minimum mismatch degree.

[0011] Preferably, the multi-source observation data includes direct observation data such as drilling logs, observation data of radionuclide concentration, and regional geological prior information.

[0012] Preferably, the second processing module is configured to generate an initial fracture network population under the constraint of observation data and geological prior information; for each generated fracture network in the population, a pre-trained SA-UNet substitute model is used to quickly simulate the corresponding physical process; a comprehensive mismatch degree function is used to calculate the deviation of each fracture network simulation result from the observation data, and an optimal fracture network set with a lower mismatch degree is screened out; and the optimal fracture network set is iteratively output to obtain a final fracture probability result through genetic algorithm.

[0013] Preferably, the third processing module is configured to input the fracture probability result output by the genetic algorithm and the site observation data into the pre-trained DAOCRN model to extract spatial feature information of the fracture network; when the model training reaches a preset iteration round or the loss function value tends to be stable, the optimization process of the DAOCRN model is terminated, and a final fracture network representation result with the minimum mismatch degree is output.

[0014] Compared with the prior art, the present application has the following advantages: The present application combines the global optimization capability of genetic algorithm with the pattern recognition advantage of deep learning, and uses a robust substitute model containing a self-attention mechanism to improve the calculation efficiency; first, the genetic algorithm inversion model is combined with observation data and prior information to generate fracture network samples as a training data set, and then the deep learning inversion model is used to further reduce the uncertainty of the genetic algorithm generation result, so as to realize high-precision representation of the fracture network. The present application can significantly reduce the mean square error loss in the 2D real fracture network scene, effectively reduce the representation uncertainty, and provide key technical support for radionuclide transport simulation, safety evaluation and engineering decision-making of high-level waste deep geological disposal. The present application is suitable for precise representation and uncertainty reduction of fracture networks in the scene of high-level waste deep geological disposal, especially suitable for scenes of groundwater seepage and radionuclide transport simulation, long-term safety geological evaluation and engineering design of disposal engineering, and can be applied to other hydrogeological engineering fields involving fractured rock mass. BRIEF DESCRIPTION OF DRAWINGS

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the method for optimizing and characterizing the crack network and controlling uncertainty according to an embodiment of the present invention; Figure 2 A schematic diagram illustrating the characterization of a multi-stage fracture network based on real outcrops; Figure 3 This is a diagram illustrating the uncertainty analysis at each stage. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 like Figure 1 As shown, this invention provides a method for optimizing the characterization and uncertainty control of fracture networks. By constructing a comprehensive framework of "data preparation and model pre-training - genetic algorithm inversion to generate fracture networks - deep learning inversion to optimize characterization," it integrates the technical advantages of stochastic methods and deep learning to achieve accurate characterization and uncertainty reduction of fracture networks in deep geological disposal scenarios for high-level radioactive waste. The method includes: Step S1: Data Preparation and Model Pre-training Stage Data Collection and Preprocessing: Multi-source observation data from the high-level radioactive waste disposal site were collected, including direct observation data such as borehole logs, indirect observation data such as radionuclide concentrations, and prior regional geological information (such as fracture statistics and permeability parameters). The raw data underwent cleaning, outlier removal, and standardization to clarify key attributes such as spatial coordinates and parameter magnitudes, forming a standardized dataset to provide data support for subsequent model training and inversion calculations.

[0020] Dual deep learning model pre-training: A SA-UNet surrogate model based on self-attention mechanism is constructed to process the pre-processed geological data as observations. By learning the mapping relationship between the fracture network and the simulation results of seepage and nuclide transport, the model's ability to capture multi-scale geological features is improved, and the model structure is optimized to meet the low computational cost requirement. The pre-training of the surrogate model provides computational support for the efficient iteration of the subsequent genetic algorithm. The DAOCRN deep learning inversion model is constructed, and based on the field observation data and typical fracture network samples, the model parameters are initialized. Through preliminary training, the model has basic fracture network feature extraction and pattern recognition ability, laying the foundation for subsequent precise optimization.

[0021] Further, the constructed SA-UNet model aims to accurately predict the spatial distribution of concentration and pressure. The model introduces a self-attention (SA) module based on the original UNet architecture. The core consists of SA modules, encoder layers, decoder layers, and skip connection layers. The encoder layer is responsible for extracting spatial features from the fracture network data, and the decoder layer reconstructs the extracted features to complete the simulation prediction. Through the SA module, the model's ability to capture multi-scale spatial features is enhanced, and the local and global features are processed together, multi-scale analysis is performed, and the information retention mechanism of skip connection is used. The model can efficiently adapt to the complex spatial pattern and heterogeneous characteristics of the fracture network.

[0022] Further, the constructed DAOCRN model integrates convolution and self-attention (SA) mechanisms to cooperatively extract local and global patterns from input data, enhancing multi-scale feature representation. Relying on the SA mechanism, the model can efficiently assimilate observation data (including borehole data, observed well concentration, and pressure data), optimize the fracture probability map (FPM) generated in stage 2, and more accurately approximate the real fracture network. The core of the model consists of a deep convolutional encoder, an SA module, and a deep convolutional decoder: the deep convolutional encoder is responsible for converting the FPM and observation data into latent feature vectors; the SA module processes the vector by assigning global feature weights, improving model accuracy and stability; the decoder reconstructs the fracture network that approximates the true value based on the latent feature vector. The deep residual OctRRDRB is constructed as the basis, which integrates Gaussian noise and adopts a multi-level residual structure, effectively alleviating overfitting and improving training stability; at the same time, residual connection promotes gradient flow during the training process, and dense connection enables feature reuse, significantly enhancing the model's ability to learn complex spatial patterns of fracture networks.

[0023] Step S2, genetic algorithm inversion to generate fracture network Initial population generation and simulation: Based on the pre-processed observation data and geological prior information, the initial fracture network population is generated. For each generated fracture network in the population, the pre-trained SA-UNet surrogate model is used to quickly simulate its corresponding physical processes (such as groundwater seepage and radionuclide migration), avoiding the high computational cost of high-fidelity models.

[0024] In the first generation, the genetic algorithm model determines the fracture location at the observation well based on the borehole data and the characterized fracture network, combining the number of fractures and their length, orientation, and range. This random fracture network set constitutes the initial fracture network population, which also serves as the training data set for the deep learning (DL) inversion model.

[0025] Mismatch degree evaluation and optimal set screening: The comprehensive mismatch degree function is used to calculate the deviation between the simulation results of each fracture network and the observation data, and the optimal fracture network set with lower mismatch degree is selected. Based on this set, the fracture probability map is calculated, which reflects the characterization results of the genetic algorithm and serves as the input data for the subsequent deep learning inversion, while also providing a basis for uncertainty analysis.

[0026] Mismatch degree evaluation is quantified by mean square error loss, which compares the simulation data obtained by the randomly generated fracture network with the aid of the surrogate model with the observation data to guide iterative updates. Accordingly, the objective function optimized by the genetic algorithm is transformed into: ; ; Iterative update to termination condition: Based on the selection, crossover, and mutation update operations of the genetic algorithm, the optimal fracture network set of the previous generation is used to generate a new generation of fracture network population. The simulation calculation, mismatch degree evaluation, and fracture probability map generation steps are repeated until the preset iteration number is reached, the genetic algorithm iteration is terminated, and the final fracture probability result is output.

[0027] In subsequent iterative updates, 30% and 50% of the new fracture networks are taken from the optimal solution and randomly generated results of the previous generation to improve model robustness; the remaining 20% is generated through the natural selection, crossover, and mutation process (see equations 3-5). This process randomly selects two individuals from the optimal candidate set of the previous iteration, and their fracture characteristics are fused through crossover to generate a new network, thus simulating the survival of the fittest principle in biological evolution.

[0028] ; ; ; where representing the random sampling from the optimal candidate set, set is the selected crack proportion from the parent crack network ( ), with a range of 1 / 3-2 / 3, set corresponds to the randomly selected 20% cracks in the child crack network ( ). In addition, the iteration number is set to 10, and the population size is 64.

[0029] Step S3, deep learning inversion optimization representation stage Feature extraction and parameter updating: input the crack probability results output by the genetic algorithm and the site observation data into the pre-trained DAOCRN model, use the multi-scale feature capture ability and long-distance dependence relationship recognition ability of the model to extract the spatial feature information of the crack network. Through the loss function analysis of the deviation between the model output results and the real observation, combined with the gradient descent mechanism to iteratively update the model parameters, the representation accuracy of the crack network is continuously optimized.

[0030] ; wherein, is the total number of input data, represents the binary label of the i-th pixel, represents the predicted probability of the existence of cracks at the i-th pixel. The training parameters of the DAOCRN model are set as follows: the batch size is 32, the gradient descent optimizer uses the Adam algorithm, and the initial learning rate is set to 0.001.

[0031] Termination of iteration and result output: when the model training reaches the preset iteration number, or the loss function value tends to be stable, the optimization process of the DAOCRN model is terminated, and the final crack network representation result with the smallest mismatch degree is output (as shown in Figure 2 ), which effectively reduces the uncertainty caused by randomness of the genetic algorithm.

[0032] Uncertainty analysis and verification: as shown in Figure 3 , based on the crack network set generated by the genetic algorithm and the final result optimized by DAOCRN, the uncertainty reduction degree is quantified using indicators such as mean square error (MSE) loss, and the effectiveness and stability of the sequence framework in the 2D crack network scenario are verified to ensure that the representation results meet the engineering application requirements of high-level waste deep geological disposal.

[0033] Through the synergistic operation of the above three stages, the application realizes the "genetic algorithm global exploration-depth learning precise optimization" fracture network characterization mode, which not only plays the global optimization advantage of the genetic algorithm, but also makes up for the uncertainty defects of the random method with the pattern recognition ability of the deep learning, and provides precise fracture network basic data for the safety evaluation of high-level waste deep geological disposal.

[0034] The application has the following technical effects: 1. High characterization accuracy and low uncertainty: the "genetic algorithm-depth learning" sequence framework is innovatively constructed, the basic sample is generated through global exploration of the genetic algorithm first, and then the DAOCRN model is precisely optimized, the mean square error loss is significantly reduced in the 2D scene, the characterization uncertainty is greatly reduced, and the method is better than the single random or deep learning inversion method.

[0035] 2. High computational efficiency and high-dimensional scene adaptation: the SA-UNet replacement model with self-attention mechanism is introduced, which breaks through the "dimension disaster" of the traditional replacement model, supports the efficient iteration of the genetic algorithm with low computational cost, and at the same time, the sequence process reduces redundant calculation and adapts to the high-dimensional fracture parameter scene of high-level waste disposal.

[0036] 3. Strong data adaptability and breakthrough in application bottleneck: the training sample is generated by combining the genetic algorithm with limited observation data and prior information, which solves the problem of deep learning dependence on massive measured data, fits the actual situation of difficult data acquisition in disposal site, and promotes the popularization of technology engineering.

[0037] 4. High degree of technology fusion and strong universality: the advantages of genetic algorithm, deep learning and replacement model are complementary through serialization design, forming an organic system of "global exploration-precise optimization-efficient calculation", which accurately matches the core needs of high-level waste deep geological disposal for characterization accuracy, efficiency and reliability. In addition, this method is not only suitable for high-level waste deep geological disposal engineering, but also can be extended to other underground space engineering, such as carbon dioxide geological sequestration, slope stability evaluation and other multidisciplinary applications.

[0038] Embodiment 2 The application also provides a fracture network optimization characterization and uncertainty control system, comprising: A first processing module is used to construct a SA-UNet replacement model based on a self-attention mechanism and a DAOCRN deep learning inversion model according to multi-source observation data of a high-level waste disposal site and typical fracture network samples; A second processing module is used to generate an initial population and perform simulation calculation, mismatch degree evaluation and optimal set screening through genetic algorithm inversion according to the SA-UNet replacement model, to obtain a fracture probability result; The third processing module is configured to obtain a final fracture network representation result with the minimum mismatch degree by deep learning inversion optimization according to the fracture probability result and the DAOCRN deep learning inversion model.

[0039] As an embodiment of the present application, the multi-source observation data includes direct observation data such as drilling logs, observation data of radionuclide concentration, and regional geological prior information.

[0040] As an embodiment of the present application, the second processing module is configured to generate an initial fracture network population by taking the observation data and the geological prior information as constraints, simulate the corresponding physical process of each generated fracture network in the population by using the pre-trained SA-UNet substitute model, calculate the deviation of each fracture network simulation result from the observation data by using a comprehensive mismatch degree function, and screen out an optimal fracture network set with a lower mismatch degree, and output the final fracture probability result by iterating the optimal fracture network set through a genetic algorithm.

[0041] As an embodiment of the present application, the third processing module is configured to input the fracture probability result output by the genetic algorithm and the site observation data into the pre-trained DAOCRN model to extract the spatial feature information of the fracture network, terminate the optimization process of the DAOCRN model when the model training reaches a preset iteration round or the loss function value tends to be stable, and output the final fracture network representation result with the minimum mismatch degree.

[0042] The above-described embodiments are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for optimizing the characterization and uncertainty control of fracture networks, characterized in that, include: Step S1: Based on multi-source observation data and typical fracture network samples from high-level radioactive waste disposal sites, construct an SA-UNet substitution model and a DAOCRN deep learning inversion model based on a self-attention mechanism. Step S2: Based on the SA-UNet substitution model, the initial population is generated and simulated using genetic algorithm inversion, mismatch degree is evaluated and optimal set is selected to obtain the crack probability result; Step S3: Based on the crack probability results and the DAOCRN deep learning inversion model, optimize the representation through deep learning inversion to obtain the final crack network representation result with the minimum mismatch.

2. The method for optimizing the characterization and uncertainty control of fracture networks as described in claim 1, characterized in that, In step S1, the multi-source observation data includes: direct observation data such as borehole logs, observation data on radionuclide concentrations, and regional geological prior information.

3. The method for optimizing the characterization and uncertainty control of fracture networks as described in claim 2, characterized in that, In step S2, an initial fracture network population is generated based on observation data and prior geological information. For each fracture network generated in the population, the corresponding physical process is quickly simulated using a pre-trained SA-UNet replacement model. The deviation between the simulation results and the observed data of each fracture network is calculated using a comprehensive mismatch function, and the optimal fracture network set with low mismatch is selected. The optimal fracture network set is then iteratively output using a genetic algorithm to produce the final fracture probability result.

4. The method for optimizing the characterization and uncertainty control of fracture networks as described in claim 3, characterized in that, In step S3, the crack probability results output by the genetic algorithm and the site observation data are input into the pre-trained DAOCRN model to extract the spatial feature information of the crack network. When the model training reaches the preset number of iterations, or when the loss function value tends to stabilize, the optimization process of the DAOCRN model is terminated, and the final fracture network representation result with the minimum mismatch is output.

5. A crack network optimization characterization and uncertainty control system, characterized in that, include: The first processing module is used to construct an SA-UNet substitution model based on a self-attention mechanism and a DAOCRN deep learning inversion model based on multi-source observation data and typical fracture network samples from high-level radioactive waste disposal sites. The second processing module is used to generate the initial population and perform simulation calculations, mismatch assessment and optimal set selection based on the SA-UNet substitution model through genetic algorithm inversion, and obtain the crack probability results. The third processing module is used to optimize the representation through deep learning inversion based on the crack probability results and the DAOCRN deep learning inversion model, and obtain the final crack network representation result with the minimum mismatch.

6. The crack network optimization characterization and uncertainty control system as described in claim 5, characterized in that, Multi-source observation data includes: direct observation data such as borehole logs, observation data on radionuclide concentrations, and regional geological prior information.

7. The crack network optimization characterization and uncertainty control system as described in claim 6, characterized in that, The second processing module is used to generate an initial fracture network population based on observation data and prior geological information. For each fracture network generated in the population, the corresponding physical process is quickly simulated using a pre-trained SA-UNet replacement model. The deviation between the simulation results and the observed data of each fracture network is calculated using a comprehensive mismatch function, and the optimal fracture network set with low mismatch is selected. The optimal fracture network set is then iteratively output using a genetic algorithm to produce the final fracture probability result.

8. The crack network optimization characterization and uncertainty control system as described in claim 7, characterized in that, The third processing module is used to input the crack probability results output by the genetic algorithm and the site observation data into the pre-trained DAOCRN model to extract the spatial feature information of the crack network. When the model training reaches the preset number of iterations, or when the loss function value tends to stabilize, the optimization process of the DAOCRN model is terminated, and the final fracture network representation result with the minimum mismatch is output.

Images