A method for inverting dominant flow channels based on generative adversarial networks and tracer experiments

Abstract

This invention discloses a method for inverting dominant flow channels based on generative adversarial networks (GANs) and tracer experiments, belonging to the fields of hydrogeology and deep learning. The method involves determining the regional permeability coefficient values ​​and boundary conditions of the target site, obtaining observational data based on field tracer experiments, generating a prior permeability coefficient field set using a GAN, and performing forward simulation based on the prior set and a numerical model. Using the observational data as constraints, the latent variable set of the generator is iteratively updated, and the permeability coefficient field is continuously optimized through the updated latent variables. Finally, a posterior set reflecting the spatial distribution of dominant flow channels and its optimal estimate are obtained. This invention overcomes the limitation of traditional Gaussian random field priors in characterizing non-Gaussian channel structures. By using latent variable inversion, it improves computational efficiency and inversion accuracy, achieving accurate and reliable inversion of the spatial distribution of dominant flow channels in aquifers. This method has significant application value for the fine characterization and remediation of groundwater-contaminated sites.

Description

Technical Field

[0001] This invention belongs to the fields of hydrogeology and deep learning, and in particular, it is a method for retrieving dominant flow channels based on generative adversarial networks and tracer experiments. Background Technology

[0002] Groundwater pollution is characterized by its high degree of concealment, long remediation cycle, and high treatment cost, making it a significant environmental problem hindering regional sustainable development. In alluvial aquifers, influenced by paleochannel deposition or alluvial processes, the permeability field often exhibits strong spatial heterogeneity, with high-permeability areas interconnected to form dominant flow channels. The permeability difference between these dominant flow channels and the surrounding low-permeability matrix can reach several orders of magnitude, leading to rapid migration of pollutants along these channels and seriously threatening downstream water sources and the safety of drinking water for residents. Therefore, accurately identifying the spatial distribution of dominant flow channels is crucial for developing scientific pollution prevention and remediation plans.

[0003] Currently, commonly used methods for identifying dominant flow channels include geophysical exploration, pumping tests, and tracer tests. Geophysical exploration utilizes the differences in resistivity, wave velocity, and other physical properties of different lithological media to detect the heterogeneous structure of aquifers. This method can quickly obtain preliminary information on the stratigraphic structure and is suitable for large-scale surveys. However, its resolution is limited, making it difficult to accurately identify the connectivity and branching characteristics of dominant flow channels. Pumping tests analyze the spatiotemporal changes in drawdown during pumping to invert the hydraulic parameter distribution of the aquifer. In application, this method can only obtain average parameters within the influence range of the pumping well and cannot characterize the spatial distribution details of dominant flow channels. Tracer tests involve releasing a conservative tracer into the injection well, observing the tracer production curve in the monitoring well, and inverting the permeability field using groundwater flow and solute transport models. This method can directly detect the actual flow path of groundwater and is widely used for identifying dominant flow channels. The accuracy of tracer tests largely depends on the accurate characterization of the spatial distribution of dominant flow channels.

[0004] However, traditional methods typically employ Gaussian random fields to describe the spatial variability of the permeability coefficient, assuming that the parameters follow a log-normal distribution. This assumption fails to characterize the channel-like connectivity structure and non-Gaussian distribution characteristics of the dominant flow channels, leading to discrepancies between the inversion results and actual conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for inverting dominant flow channels based on generative adversarial networks and tracer experiments, enabling accurate inversion of dominant flow channels in heterogeneous groundwater systems.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for inverting dominant flow channels based on generative adversarial networks and tracer experiments, comprising:

[0008] S1. Determine the regional permeability coefficient and boundary conditions of the target site, and obtain observation data based on field tracer tests. The observation data includes tracer concentration production curves and hydraulic head data of each monitoring well.

[0009] S2. Based on a pre-trained spectral normalized Wasserstein generative adversarial network, a set of latent variables is sampled from the standard normal distribution and input into the generator of the generative adversarial network to generate a set of prior permeability coefficient fields for characterizing the channel-like connected structure of the dominant flow channel.

[0010] S3. Construct a numerical model of groundwater flow and solute transport. Combine the site boundary conditions and input the a priori permeability coefficient field set into the numerical model for forward modeling to obtain the corresponding simulated tracer concentration and simulated water head.

[0011] S4. An iterative local update set smoother algorithm based on multiple data assimilation is adopted to iteratively update the latent variables with the observed data as constraints. In each iteration, the updated latent variables are input into the generator of the generative adversarial network to obtain the updated set of permeability coefficient fields and perform forward simulation again until the convergence condition is met to obtain the posterior permeability coefficient field set.

[0012] S5. Perform statistical analysis on the posterior permeability coefficient field set to determine the spatial distribution of the dominant flow channels.

[0013] Furthermore, the process of determining the regional permeability coefficient value and boundary conditions of the target site includes:

[0014] Based on site geological drilling, stratigraphic profile and groundwater monitoring data, the aquifer thickness, groundwater level and hydraulic gradient were determined.

[0015] The boundary conditions are determined based on the surrounding water bodies and groundwater flow direction. The boundary conditions include constant head boundaries and impermeable boundaries.

[0016] Based on pumping test data, the permeability coefficient values ​​of the dominant flow channel region and the matrix region were determined respectively.

[0017] Furthermore, the process of obtaining observational data based on the field tracer experiment includes:

[0018] Along the direction of groundwater flow, injection wells are deployed upstream of the study area or near the pollution source, and multiple monitoring wells are deployed downstream.

[0019] The dosage of tracer is calculated based on the minimum detection limit of the tracer equipment, the well spacing, the aquifer thickness, and the porosity.

[0020] The tracer solution is injected into the injection well at a constant flow rate, and water samples are collected in the monitoring well at set time intervals to obtain the tracer concentration production curve and record the head data simultaneously.

[0021] Furthermore, S2 includes:

[0022] S21. Obtain training images containing dominant flow channel features, and randomly crop multiple two-dimensional aquifer plane samples from the training images to form a training dataset.

[0023] S22. Train the spectral normalized Wasserstein generative adversarial network using the training dataset, so that the generator of the generative adversarial network learns the mapping relationship from latent variables to the permeation coefficient field image.

[0024] S23. After the network training is completed, multiple latent variables are sampled from the standard normal distribution and input into the trained generator to obtain multiple permeability coefficient field images;

[0025] S24. The generated permeability field image is binarized, and the channel region and matrix region are respectively assigned the region permeability coefficient values ​​determined in step S1 to form the prior permeability field set.

[0026] Furthermore, the numerical model for groundwater flow and solute transport includes the groundwater flow equation, Darcy's law equation, and the convection-dispersion equation for describing solute transport; and different hydrodynamic dispersion parameters are set for the dominant flow channel region and the matrix region respectively.

[0027] Furthermore, in S5, the mean field of the posterior permeability coefficient field set is used as the optimal estimate of the spatial distribution of the dominant flow channel.

[0028] Furthermore, the monitoring wells deployed in the field tracer experiment in S1 are divided into assimilation wells and verification wells. In S3, only the simulated tracer concentration and simulated head at the location of the assimilation well are obtained. The iterative update process in S4 uses the observation data of the assimilation well as constraints.

[0029] Furthermore, it also includes S6:

[0030] The mean field of the posterior permeability coefficient field set is input into the numerical model for forward modeling to obtain the simulated tracer concentration and simulated water head at the location of the verification well in the field tracer test.

[0031] The simulated tracer concentration, simulated water head, and observation data from the verification well were compared to calculate the coefficient of determination and root mean square error, which were used as indicators of the accuracy of the spatial distribution of the dominant flow channels obtained by S5.

[0032] Secondly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for inverting the dominant flow channel based on generative adversarial networks and tracer experiments.

[0033] Thirdly, the present invention provides a computer electronic device, including a memory and a processor;

[0034] The memory is used to store computer programs;

[0035] The processor is configured to implement the aforementioned method for inverting the dominant flow channel based on generative adversarial networks and tracer experiments when executing the computer program.

[0036] The beneficial effects of this invention are:

[0037] This invention combines Spectral Normalized Wasserstein Generative Adversarial Network (SN-WGAN) with the Iterative Local Update Set Smoothing (ILUES-MDA) data assimilation algorithm. First, SN-WGAN is used to learn and generate a prior permeability coefficient field with a realistic channel-like connectivity structure, overcoming the limitation of traditional Gaussian random fields in characterizing non-Gaussian geological structures. Then, the optimization object is transformed from massive permeability coefficient field parameters into low-dimensional latent variables, and the ILUES-MDA algorithm is used for efficient inversion, significantly reducing computational complexity. Using field tracer test data as constraints, this invention ultimately obtains a convergent posterior permeability coefficient field set, thereby achieving accurate prediction of the distribution structure of dominant flow channels. This method effectively solves the bottleneck faced by traditional inversion techniques in characterizing dominant flow channels in aquifers. The overall process is clear and efficient, with clear engineering application value, and can provide a reliable scientific basis for the detailed characterization and remediation of groundwater contaminated sites. Attached Figure Description

[0038] Figure 1 This is a flowchart of a method for inverting the dominant flow path based on generative adversarial networks and tracer experiments;

[0039] Figure 2 This is a diagram showing the boundary conditions and monitoring well settings for the numerical model;

[0040] Figure 3 This is a schematic diagram of the original training images for SN-WGAN;

[0041] Figure 4 This is a diagram showing the results of the dominant flow channel inversion;

[0042] Figure 5 These are tracer production curves and head fitting plots;

[0043] Figure 6This is a comparison chart of simulated and observed values ​​from the verification well;

[0044] Figure 7 This is a schematic diagram of a computer electronic device. Detailed Implementation

[0045] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0046] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0047] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0048] like Figure 1 As shown, this invention proposes a method for inverting dominant flow channels based on generative adversarial networks and tracer experiments. This method integrates the inversion method of spectral normalized Wasserstein generative adversarial networks and iterative data assimilation algorithms. By using field tracer experiment data to iteratively update the latent variables of the generative network, the distribution structure of dominant flow channels in the aquifer can be accurately inverted.

[0049] The main steps include:

[0050] Step 1: Site Survey and Parameter Determination

[0051] Hydrogeological survey: Based on site geological drilling, stratigraphic profile and groundwater monitoring data, determine aquifer thickness, groundwater level and hydraulic gradient, etc.

[0052] Taking a chemical plant site in City B of Province A as an example, the site area is approximately 1.33 km². 2 The site previously experienced a serious groundwater pollution incident, with preliminary investigations showing that cadmium and lead concentrations in the groundwater exceeded the Class IV water quality standard by 5-30 times. Based on the site's hydrogeological survey results, the aquifer thickness is approximately 15 m, the groundwater level is 4-5 m deep, and fluctuations are relatively small. The underlying strata mainly consist of gravel and coarse sand, belonging to river alluvial deposits.

[0053] Based on the distribution of water bodies around the site and long-term groundwater monitoring data, the boundary condition types were determined as follows: the western and eastern boundaries are constant head boundaries (Dirichlet type), with the western boundary being a uniform constant head and the eastern boundary divided into three constant head segments to reflect the observed hydraulic gradient differences; the northern and southern boundaries are impermeable boundaries, such as... Figure 2 As shown.

[0054] Permeability coefficient determination: Water level drawdown data were obtained through single-well pumping tests. Combined with lithological characteristics revealed by formation drilling, the Theis plumb line method was used to calculate the permeability coefficient. Based on lithological differences, the aquifer was divided into a dominant flow channel zone (gravel, coarse sand) and a matrix zone (fine sand, silt), with a permeability coefficient of 2 × 10⁻⁶ for each zone. -2 m / s and 4.9×10 -3 The difference between the two is about an order of magnitude, which is consistent with the typical characteristics of the development of a dominant flow channel.

[0055] Step 2: Acquisition of tracer test data

[0056] Deployment of injection wells and monitoring wells: Along the groundwater flow direction, injection wells are deployed upstream of the study area or near the pollution source, and multiple monitoring wells are deployed downstream; the dosage of tracer is calculated based on the tracer equipment's minimum detection limit, well spacing, aquifer thickness, and porosity; the tracer solution is injected into the injection wells at a constant flow rate, and water samples are collected from the monitoring wells at set time intervals to obtain the tracer concentration production curve, and the head data is recorded simultaneously as monitoring well observation data.

[0057] In this embodiment, 3-fluorobenzoic acid was selected as the tracer based on the site soil conditions and pollutant characteristics. This tracer has advantages such as high sensitivity, good stability, and ease of detection. Its molecular structure contains fluorine atoms, allowing for high-sensitivity detection by high-performance liquid chromatography (HPLC). Furthermore, it is not easily degraded in the environment, resulting in minimal background interference. A background value survey of the site was conducted before the experiment, and the results showed that the concentration of 3-fluorobenzoic acid was close to the detection limit (approximately 0 ng / mL). In this embodiment, one injection well and 30 monitoring wells were installed within the site. The injection well was located near the pollution source (e.g., Figure 2 The brown discrete points in the image are at a depth of 15 m, and the monitoring wells are at a depth of 10-20 m with a diameter of 100 mm.

[0058] Calculate the tracer injection volume using the Brigham-Smith formula:

[0059]

[0060] In the formula: This refers to the amount of effective tracer substance added. To ensure the coefficient; The limit of detection (ng / ml) for tracer devices; The average well spacing between injection wells and monitoring wells; Effective aquifer thickness; Porosity.

[0061] The tracer dosage was determined to be 1 kg based on calculations. 1052 g of 95% pure 3-fluorobenzoic acid was dissolved in 100 L of deionized water to prepare a standard solution, which was then uniformly injected into the aquifer at a constant flow rate of 1.4 L / min over 1.2 hours. Groundwater samples were collected from each monitoring well at 24, 48, 72, 96, 120, 144, 168, and 192 hours after injection, resulting in a total of 248 samples. All samples were immediately stored at 0–4℃ and sent to the laboratory within 48 hours for quantitative analysis of the tracer concentration using high-performance liquid chromatography (HPLC). Hydraulic head data could be directly measured based on field conditions.

[0062] In one specific embodiment of the present invention, the monitoring wells deployed in the field tracer experiment are divided into assimilation wells and verification wells, and the observation data of 70% of the monitoring wells (22 wells) are used as assimilation wells (e.g., Figure 2 The green discrete points in the data are used for data assimilation, and 30% of the monitoring wells (9 in total) are used as validation wells (e.g., Figure 2 The red discrete points in the graph are used for independent verification.

[0063] Step 3: Train the SN-WGAN network and generate a prior set.

[0064] The SN-WGAN network is trained using training images representing the channel-like connected structure of the dominant flow channels. During training, the generator learns to incorporate low-dimensional latent variables. (A random noise vector following a standard normal distribution) is mapped to a permeation coefficient field image K. The discriminator learns to distinguish between real training images and generated images. The discriminator uses Wasserstein distance as a loss function and applies spectral normalization constraints to the weight matrix W of each layer of the discriminator.

[0065] The game process between the generator and the discriminator can be expressed by the following objective function:

[0066]

[0067] In the formula: It is random noise; Indicates a generator; Indicates the discriminator; The probability distribution representing the real sample; Represents the probability distribution of the generated samples; E represents the expectation. and They represent Judging the real data and The probability that the generated sample is real. For the objective function... , With the goal of maximizing it, The goal is to minimize it.

[0068] The training dataset for SN-WGAN is constructed by randomly cropping 10,000 two-dimensional aquifer plane samples from the original training images (TI). The original training images are shown below. Figure 3 As shown, the image presents a complex, non-periodic, channel-like connected network. White lines represent dominant flow channels, and the black background represents low-permeability matrix, effectively characterizing the randomness, connectivity, and path tortuosity of dominant flow channels in natural aquifers. Here, the original training images do not depend on the specific site data for this inversion. The original training images selected in this embodiment are from classic literature in the field of statistics, possessing high geological representativeness and serving as recognized standard reference images for characterizing the spatial structure of dominant flow channels in heterogeneous aquifers.

[0069] After training, samples are taken from the standard normal distribution N(0,1). A latent variable is input into the trained generator to obtain... A permeability field image with the same size as the water layer plane sample is generated, and the resulting image is binarized.

[0070] The binarization process is to convert the continuous grayscale image output by the generator into a binary image containing only two distinct regions: channels and matrix. The white-represented channel regions are assigned a dominant flow channel permeability coefficient value (2×10⁻⁶). -2 m / s), assigning the matrix permeability coefficient value (4.9 × 10 m / s) to the non-channel areas represented by black. -3 (m / s), forming a priori permeability coefficient field set, while retaining the corresponding latent variable set for subsequent data assimilation.

[0071] Step 4: Construct a numerical model and perform forward simulation.

[0072] The governing equations of the numerical model for groundwater flow and solute transport include Darcy's law equation for describing groundwater flow and convection-dispersion equation for describing solute transport; and different hydrodynamic dispersion parameters are set for the dominant flow channel region and the matrix region respectively.

[0073] Specifically, this invention simulates tracer transport in a two-dimensional heterogeneous aquifer under steady-state flow conditions. Assuming solute transport is primarily controlled by convection and dispersion, and neglecting chemical reactions, the governing equations for the numerical model of groundwater flow and solute transport are:

[0074]

[0075]

[0076]

[0077] The first formula above is the equation for groundwater flow, the second formula is Darcy's law, and the third formula is the equation for solute transport. Where: This refers to the solute concentration. For time; The hydrodynamic dispersion tensor is derived from the pore flow velocity. Longitudinal diffusion and lateral dispersion Sure; For source and sink volumetric flow rates; Source and sink concentrations; Permeability coefficient; For the water head, This is the gradient operator.

[0078] Considering the scale effect of dispersion and soil texture characteristics, different dispersion values ​​were set for the dominant flow channel and the matrix: the longitudinal and lateral dispersions of the dominant flow channel were 100 m and 10 m, respectively, while the longitudinal and lateral dispersions of the matrix were 1 m and 0.1 m, respectively. A two-dimensional aquifer model covering the study area was constructed using a 9 m × 9 m spatial discrete grid, forming a 128 × 128 grid structure. The numerical simulation lasted for 8 days (192 hours) with a time step of 1 day (24 hours), containing a total of 8 time nodes. The MODFLOW and MT3DMS solvers were used for numerical solutions.

[0079] Each permeability coefficient field in the prior set is input into the numerical model for forward simulation to obtain the corresponding head field and tracer concentration field at each time step, and the simulated concentration and head value at the assimilation well location are extracted.

[0080] Step 5: ILUES-MDA Data Assimilation Algorithm Inversion

[0081] The ILUES-MDA algorithm is used to iteratively update the latent variables with the observation data of the assimilated well as the constraints. In each iteration, the updated latent variables are input into the generator of the generative adversarial network to obtain the updated permeability coefficient field set and perform forward simulation again until the convergence condition is met to obtain the posterior permeability coefficient field set.

[0082] The ILUES-MDA algorithm uses the perturbation observation error covariance matrix to perform multiple data assimilations on the latent variables. After each iteration updates the latent variables, they are input into the SN-WGAN generator to obtain a new permeability coefficient field. The update formula is:

[0083]

[0084]

[0085] In the formula: For the first In the nth iteration A vector of latent variables for each set member; This is the covariance matrix between the simulated and predicted values; The autocovariance matrix of the predicted values; Let be the expansion factor for the k-th iteration; The observation error covariance matrix; The observations after adding perturbations; This represents the simulated value obtained by inputting latent variables into the SN-WGAN generator to obtain the permeability coefficient field, and then calculating it using a numerical model. The size of the set; This refers to the number of iterations. In this embodiment, the set size... =5000, number of iterations =8.

[0086] Step Six: Obtain and validate the posterior result.

[0087] After iterative convergence, the mean field of all members in the posterior set is calculated as the optimal estimate of the dominant flow channel distribution, and the results are as follows: Figure 4 As shown, the dominant flow channels obtained by inversion exhibit a typical branched connectivity structure, demonstrating significant spatial heterogeneity.

[0088] Simultaneously, the mean field of the posterior permeability coefficient field set is input into the numerical model for forward simulation to obtain the simulated tracer concentration and simulated hydraulic head at the locations of the verification wells in the field tracer experiment. The simulated tracer concentration and simulated hydraulic head are compared with the observation data of the verification wells, and the comparison results are as follows: Figure 5 As shown, the fitting curves of most verification wells are in good agreement with the observation data in terms of peak concentration and arrival time, indicating that the present invention can effectively capture the complex migration behavior of tracers in non-Gaussian dominant flow systems.

[0089] Using the coefficient of determination (R) 2 The accuracy of the inversion is quantitatively assessed using the root mean square error (RMSE).

[0090]

[0091]

[0092] in, Indicates the first Observations from one verification well; This represents the average of the observed values; Indicates the numerical model for the first Simulated values ​​from one verification well; This represents the average value predicted by the numerical model. This indicates the total number of verification wells.

[0093] Evaluation results as follows Figure 6 As shown, the R-values ​​of the tracer concentration fitting for each verification well are... 2 All exceed 0.7, R0 verified by head. 2 The value reached 0.98. Most regression lines were close to the ideal y = x reference line, indicating good agreement between the simulated and observed values. These results demonstrate that the dominant flow channels obtained by this invention can accurately characterize the hydraulic behavior and tracer transport features of the study area.

[0094] It should also be noted that the dominant flow path inversion method based on generative adversarial networks and tracer experiments in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a dominant flow path inversion system based on generative adversarial networks and tracer experiments, corresponding to the dominant flow path inversion method based on generative adversarial networks and tracer experiments provided in the above embodiments, comprising:

[0095] The data acquisition module is used to determine the regional permeability coefficient and boundary conditions of the target site, and to obtain observation data based on field tracer tests. The observation data includes tracer concentration production curves and hydraulic head data of each monitoring well.

[0096] The prior set generation module is communicatively connected to the data acquisition module. It includes a pre-trained spectral normalized Wasserstein generative adversarial network, which samples a set of latent variables from a standard normal distribution and inputs them into the generator of the generative adversarial network to generate a prior permeability coefficient field set for characterizing the channel-like connectivity structure of the dominant flow channel.

[0097] The numerical simulation module is communicatively connected to the data acquisition module and the prior set generation module. It is used to construct a numerical model of groundwater flow and solute transport. Combined with the site boundary conditions, the prior permeability coefficient field set is input into the numerical model for forward simulation to obtain the corresponding simulated tracer concentration and simulated hydraulic head.

[0098] The data assimilation and inversion module is communicatively connected to the data acquisition module, the prior set generation module, and the numerical simulation module. It is used to iteratively update the latent variables using the observed data as constraints by employing an iterative local update set smoother algorithm. In each iteration, the updated latent variables are input into the generator of the generative adversarial network to obtain the updated permeability coefficient field set and perform forward simulation again until the convergence condition is met to obtain the posterior permeability coefficient field set.

[0099] The results analysis module is communicatively connected to the data assimilation and inversion module and is used to perform statistical analysis on the posterior permeability coefficient field set to determine the spatial distribution of the dominant flow channels.

[0100] It is understood that the dominant flow path inversion method based on generative adversarial networks and tracer experiments in the above embodiments can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the dominant flow path inversion method based on generative adversarial networks and tracer experiments provided in the above embodiments. This product includes a computer program / instruction that, when executed by a processor, can implement the dominant flow path inversion method based on generative adversarial networks and tracer experiments as described in the above embodiments.

[0101] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the dominant flow channel inversion method based on generative adversarial networks and tracer experiments provided in the above embodiments, such as... Figure 7 As shown, it includes a memory and a processor;

[0102] The memory is used to store computer programs;

[0103] The processor is configured to implement the dominant flow channel inversion method based on generative adversarial networks and tracer experiments in the above embodiments when executing the computer program.

[0104] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0105] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the dominant flow channel inversion method based on generative adversarial networks and tracer experiments provided in the above embodiments. The storage medium stores a computer program that, when executed by a processor, can realize the dominant flow channel inversion method based on generative adversarial networks and tracer experiments in the above embodiments.

[0106] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0107] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0108] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0109] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. An advantage flow channel inversion method based on a generative adversarial network and a tracer experiment, characterized in that, include: S1. Determine the regional permeability coefficient and boundary conditions of the target site, and obtain observation data based on field tracer tests. The observation data includes tracer concentration production curves and hydraulic head data of each monitoring well. S2. Based on a pre-trained spectral normalized Wasserstein generative adversarial network, a set of latent variables is sampled from the standard normal distribution and input into the generator of the generative adversarial network to generate a set of prior permeability coefficient fields for characterizing the channel-like connected structure of the dominant flow channel. S3. Construct a numerical model of groundwater flow and solute transport. Combine the site boundary conditions and input the a priori permeability coefficient field set into the numerical model for forward modeling to obtain the corresponding simulated tracer concentration and simulated water head. S4. An iterative local update set smoother algorithm based on multiple data assimilation is adopted to iteratively update the latent variables with the observed data as constraints. In each iteration, the updated latent variables are input into the generator of the generative adversarial network to obtain the updated set of permeability coefficient fields and perform forward simulation again until the convergence condition is met to obtain the posterior permeability coefficient field set. S5. Perform statistical analysis on the posterior permeability coefficient field set to determine the spatial distribution of the dominant flow channels.

2. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, The process of determining the regional permeability coefficient and boundary conditions of the target site includes: Based on site geological drilling, stratigraphic profile and groundwater monitoring data, the aquifer thickness, groundwater level and hydraulic gradient were determined. The boundary conditions are determined based on the surrounding water bodies and groundwater flow direction. The boundary conditions include constant head boundaries and impermeable boundaries. Based on pumping test data, the permeability coefficient values ​​of the dominant flow channel region and the matrix region were determined respectively.

3. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, The process of obtaining observational data based on field tracer experiments includes: Injection wells were installed upstream of the study area or near the pollution source, and multiple monitoring wells were installed downstream, following the direction of groundwater flow. The dosage of tracer is calculated based on the minimum detection limit of the tracer equipment, the well spacing, the aquifer thickness, and the porosity. The tracer solution is injected into the injection well at a constant flow rate, and water samples are collected in the monitoring well at set time intervals to obtain the tracer concentration production curve and record the head data simultaneously.

4. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, S2 include: S21. Obtain training images containing dominant flow channel features, and randomly crop multiple two-dimensional aquifer plane samples from the training images to form a training dataset. S22. Train the spectral normalized Wasserstein generative adversarial network using the training dataset, so that the generator of the generative adversarial network learns the mapping relationship from latent variables to the permeation coefficient field image. S23. After the network training is completed, multiple latent variables are sampled from the standard normal distribution and input into the trained generator to obtain multiple permeability coefficient field images; S24. The generated permeability field image is binarized, and the channel region and matrix region are respectively assigned the region permeability coefficient values ​​determined in step S1 to form the prior permeability field set.

5. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, The numerical model for groundwater flow and solute transport includes the groundwater flow equation, Darcy's law equation, and the convection-dispersion equation for describing solute transport; and different hydrodynamic dispersion parameters are set for the dominant flow channel region and the matrix region respectively.

6. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, In S5, the mean field of the posterior permeability coefficient field set is used as the optimal estimate of the spatial distribution of the dominant flow channel.

7. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 1, characterized in that, The monitoring wells deployed in the field tracer test in S1 are divided into assimilation wells and verification wells. In S3, only the simulated tracer concentration and simulated water head at the location of the assimilation well are obtained. The iterative update process in S4 uses the observation data of the assimilation well as constraints.

8. The dominant flow path inversion method based on generative adversarial networks and tracer experiments according to claim 7, characterized in that, Also includes S6: The mean field of the posterior permeability coefficient field set is input into the numerical model for forward modeling to obtain the simulated tracer concentration and simulated water head at the location of the verification well in the field tracer test. The simulated tracer concentration, simulated water head, and observation data from the verification well were compared to calculate the coefficient of determination and root mean square error, which were used as indicators of the accuracy of the spatial distribution of the dominant flow channels obtained by S5.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the dominant flow channel inversion method based on generative adversarial networks and tracer experiments as described in any one of claims 1 to 8.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the dominant flow channel inversion method based on generative adversarial networks and tracer experiments as described in any one of claims 1 to 8.

Images