Hydrogeological parameter intelligent identification and inversion system

Abstract

The present application relates to the technical field of hydrogeology, and specifically relates to a hydrogeological parameter intelligent identification and inversion system, comprising: a parameter field generation module for generating processing based on prior geological information and initial observation data of a target area to obtain multiple mutually competitive candidate hydrogeological parameter fields; a scheme generation module for counterfactual deduction of the multiple candidate hydrogeological parameter fields to generate a counterfactual observation scheme; a parameter field update module for, after executing the counterfactual observation scheme to obtain execution observation data on the target area, performing confidence evaluation and updating of each candidate hydrogeological parameter field based on the execution observation data and each candidate hydrogeological parameter field to obtain multiple updated candidate hydrogeological parameter fields; and a target screening module for counterfactual deduction of each updated candidate hydrogeological parameter field to obtain a target hydrogeological parameter field, thereby solving the problem of low practical application value of hydrogeological parameters caused by sparse observation data.

Description

Technical Field

[0001] This invention relates to the field of hydrogeology, specifically to an intelligent identification and inversion system for hydrogeological parameters. Background Technology

[0002] Accurate identification and inversion of hydrogeological parameters are the core foundation for groundwater resource assessment, pollutant transport prediction, and geological safety assessment of major engineering projects. The reliability of the results directly affects the effectiveness of decision-making in water resource management, environmental protection, and engineering construction.

[0003] Existing methods for identifying and inverting hydrogeological parameters have evolved from traditional manual trial-and-error methods to automated optimization inversion methods represented by data assimilation and machine learning. These methods establish numerical models between parameter fields and observation data, and use optimization algorithms to automatically adjust parameters so that the model output matches the observation data, thereby inferring the spatial distribution of underground media properties.

[0004] However, these existing methods are essentially a passive fitting paradigm. Their common prominent problem is that under complex geological conditions, when the observational data available for inversion is sparse, the subsurface structure is highly uncertain. There are often multiple parameter fields with different geological structures that can fit the existing observational data to the same extent. This makes it impossible for the inversion process to escape the dilemma of multiple solutions. The results heavily depend on the initial guess and have low credibility in geological interpretation, which restricts the practical application value of hydrogeological parameters. Summary of the Invention

[0005] To address the technical problem of low practical application value of hydrogeological parameters due to sparse observation data, this application provides an intelligent identification and inversion system for hydrogeological parameters.

[0006] The intelligent identification and inversion system for hydrogeological parameters provided in this application adopts the following technical solution: A system for intelligent identification and inversion of hydrogeological parameters includes: The parameter field generation module is used to generate multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area. The scheme generation module is used to perform counterfactual inference on multiple candidate hydrogeological parameter fields and generate counterfactual observation schemes; The parameter field update module is used to perform confidence assessment and update each candidate hydrogeological parameter field based on the executed observation data and each candidate hydrogeological parameter field after obtaining the executed observation data and the candidate hydrogeological parameter fields, so as to obtain multiple updated candidate hydrogeological parameter fields. The target screening module is used to perform counterfactual inference on each updated candidate hydrogeological parameter field to obtain the target hydrogeological parameter field.

[0007] Furthermore, the steps for generating and processing multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area include: The prior geological information is analyzed and processed to extract the prototype field of geological structure that differs from the preset geological structural features; Based on the geological structure prototype field and initial observation data, collaborative data assimilation processing is performed to generate multiple initial parameter fields; The initial parameter fields are subjected to adversarial differentiation enhancement to optimize their geological structure characteristics, resulting in multiple candidate hydrogeological parameter fields.

[0008] Furthermore, the steps for generating multiple initial parameter fields through collaborative data assimilation processing based on the geological structure prototype field and initial observation data include: Based on the geological structure prototype field, multiple parametric field models with differences in preset geological structure characteristics were constructed; Numerical simulations were performed on the field models of each parameter based on the initial observation data to obtain the physical field prediction data of each field model. The model data matching degree between the initial observation data and the predicted data of each physical field is calculated, and the difference between each parameter field model and the geological structure prototype field is calculated to obtain the model structure fidelity. Based on model data matching degree and model structure fidelity, a multi-objective collaborative optimization function is constructed, and then the parameter field models are updated through the multi-objective collaborative optimization function to obtain multiple initial parameter fields.

[0009] Furthermore, the steps for generating a counterfactual observation scheme by performing counterfactual inferences on multiple candidate hydrogeological parameter fields include: Numerical simulations were conducted based on multiple candidate hydrogeological parameter fields to obtain the physical field prediction responses of each candidate hydrogeological parameter field under simulation combinations of multiple preset observation locations and preset time series. Based on the pairwise physical field prediction responses under the same simulation combination, the differences in the predicted responses between each candidate hydrogeological parameter field are calculated to obtain the distinguishability index. Global optimization is performed based on the distinguishability index to obtain the target observation location and target time series, and then a counterfactual observation scheme is generated based on the target observation location and target time series.

[0010] Furthermore, the steps for obtaining executed observation data on the target area by implementing a counterfactual observation scheme include: Based on the target observation location and target time series in the counterfactual observation scheme, a sensor network is deployed in the target area, and measured physical field data of the target area are collected through the sensor network; and, based on the counterfactual observation scheme, a high-fidelity numerical simulation is performed to calculate the simulated physical field data of the target area under the current hydrogeological conditions. Spatiotemporal registration and uncertainty quantification are performed on measured and simulated physical field data to generate fused physical field data. Extract fusion feature vectors from the fused physical field data and combine the fused feature vectors to form the execution observation data output.

[0011] Furthermore, based on the observed data and each candidate hydrogeological parameter field, the steps for performing confidence assessment and updating each candidate hydrogeological parameter field to obtain multiple updated candidate hydrogeological parameter fields include: From the physical field prediction response of each candidate hydrogeological parameter field, extract the prediction feature vector corresponding to the fused feature vector in the observation data; Based on the fused feature vector and each predicted feature vector, the matching index of each candidate hydrogeological parameter field is calculated. The confidence scores of each candidate hydrogeological parameter field in the current iteration are obtained by fusion calculation based on each matching index and the cumulative confidence in historical iterations. After eliminating candidate hydrogeological parameter fields corresponding to confidence scores lower than the preset confidence scores, Gaussian random perturbation is applied to the confidence scores corresponding to the remaining candidate hydrogeological parameter fields. Then, the preset geological structure features among the remaining candidate hydrogeological parameter fields are cross-recombined to obtain the updated candidate hydrogeological parameter fields.

[0012] Furthermore, the steps for obtaining the target hydrogeological parameter field by performing counterfactual deduction on each updated candidate hydrogeological parameter field include: Based on each updated candidate hydrogeological parameter field and cumulative confidence, several first hydrogeological parameter fields that satisfy the preset convergence criteria are obtained from multiple updated candidate hydrogeological parameter fields. Based on the cumulative confidence of each first hydrogeological parameter field, a second hydrogeological parameter field is obtained from several first hydrogeological parameter fields. The robustness of the second hydrogeological parameter field was verified in multiple scenarios, and the predicted stability index of the second hydrogeological parameter field under various preset hydrogeological boundary conditions was obtained. Based on the predicted stability index, the parameter values ​​in the second hydrogeological parameter field are spatially smoothed and adapted to each preset hydrogeological boundary condition to obtain the target hydrogeological parameter field.

[0013] Beneficial effects achieved: This application provides an intelligent identification and inversion system for hydrogeological parameters, comprising: a parameter field generation module, used to generate multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area; a scheme generation module, used to perform counterfactual inference on the multiple candidate hydrogeological parameter fields to generate counterfactual observation schemes; a parameter field update module, used to perform confidence assessment and update of each candidate hydrogeological parameter field based on the executed observation data and each candidate hydrogeological parameter field after obtaining the executed observation data on the target area through the counterfactual observation scheme, to obtain multiple updated candidate hydrogeological parameter fields; and a target screening module, used to perform counterfactual inference on each updated candidate hydrogeological parameter field to obtain the target hydrogeological parameter field.

[0014] In this application, a parameter field generation module constructs multiple competing candidate hydrogeological parameter fields, covering the highly uncertain solution space present in the case of sparse observation data, rather than relying on a single guess. Then, a scheme generation module performs counterfactual inference to generate counterfactual observation schemes specifically designed to distinguish the candidate hydrogeological parameter fields. This allows the system to proactively plan and acquire the new data with the highest discriminative value, thereby injecting key information to address the problem of different parameters having the same effect. After executing the counterfactual observation scheme and obtaining the executed observation data, the parameter field update module uses the executed observation data to evaluate and update the confidence level of each candidate hydrogeological parameter field, thereby quantifying the results. This approach effectively reduces candidate hydrogeological parameter fields that are inconsistent with the observed data, achieving efficient pruning and convergence of the solution space. The final target selection module iteratively optimizes the above steps to ensure that the output target hydrogeological parameter field is the most reliable solution after multiple rounds of key observation verification. Therefore, this application can transform passive data fitting into an observation design and hypothesis testing loop driven by reducing cognitive uncertainty. This allows the geological truth to be gradually locked in through a systematic active learning process, even under conditions of sparse observation data, thereby significantly improving the geological credibility and practical application value of the target hydrogeological parameter field obtained by inversion identification. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the modules of the intelligent identification and inversion system for hydrogeological parameters in this application; Figure 2 A schematic diagram illustrating the steps involved in generating multiple candidate hydrogeological parameter fields for this application; Figure 3 A flowchart illustrating the steps involved in generating a counterfactual observation scheme for this application; Figure 4This is a schematic diagram of the steps involved in this application to assess and update the confidence level of each candidate hydrogeological parameter field based on the counterfactual observation scheme, and to obtain multiple updated candidate hydrogeological parameter fields. Figure 5 This is a schematic diagram illustrating the steps involved in obtaining the target hydrogeological parameter field from multiple updated candidate hydrogeological parameter fields in this application.

[0016] Explanation of icon numbers: 10. Parameter field generation module; 20. Scheme generation module; 30. Parameter field update module; 40. Target filtering module. Detailed Implementation

[0017] The following combination Figures 1 to 5 This application will be described in further detail.

[0018] This application discloses an intelligent identification and inversion system for hydrogeological parameters.

[0019] Please refer to Figure 1 The intelligent identification and inversion system for hydrogeological parameters proposed in this embodiment includes: The parameter field generation module is used to generate multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area. The scheme generation module is used to perform counterfactual inference on multiple candidate hydrogeological parameter fields to generate counterfactual observation schemes. The parameter field update module is used to perform confidence assessment and update each candidate hydrogeological parameter field based on the executed observation data and each candidate hydrogeological parameter field after executing the counterfactual observation scheme to obtain the executed observation data on the target area, thereby obtaining multiple updated candidate hydrogeological parameter fields. The target selection module is used to perform counterfactual inference on each updated candidate hydrogeological parameter field to obtain the target hydrogeological parameter field.

[0020] The purpose of this embodiment is to change the inherent pattern of passively relying on limited observation data for fitting in traditional hydrogeological parameter identification and inversion. By constructing a systematic cognitive framework of active exploration, verification, and evolution, it aims to solve the problems of high ambiguity and low reliability of identification and inversion results under conditions of sparse observation data.

[0021] The parameter field generation module addresses and formalizes the uncertainty of initial cognition. Instead of providing a single guess, it systematically generates multiple competing candidate hydrogeological parameter fields, all possible under initial observation data and prior geological information. This provides clear candidate hydrogeological parameter field objects for subsequent identification. The scheme generation module undertakes the core function of proactive cognition. Its purpose is not to passively receive information, but to actively question the candidate hydrogeological parameter fields generated by the previous module. Through counterfactual deduction, it designs counterfactual observation schemes that maximize the exposure of differences between candidate hydrogeological parameter fields, thus transforming the blind or fixed data collection in traditional inversion into a targeted search for optimal information acquisition. The parameter field update module aims to introduce an adjudication and learning mechanism based on the executed observation data obtained from the counterfactual observation scheme. Using the counterfactual observations obtained by the scheme generation module, it assesses and updates the confidence of each candidate hydrogeological parameter field, thereby filtering, strengthening, and eliminating previous cognition based on the executed observation data, achieving a cognitive evolution from "many possibilities" to "more likely." Ultimately, the goal of the target selection module is to converge from the updated candidate hydrogeological parameter field and lock in the most thoroughly tested and robust solution as the target hydrogeological parameter field output after going through multiple rounds of the above-mentioned "hypothesis-test-update" game cycle.

[0022] The combination of the above modules constructs a complete, logically embedded intelligent identification and inversion closed loop: the parameter field generation module proposes multiple candidate hydrogeological parameter fields, the scheme generation module generates counterfactual observation schemes, the parameter field update module evaluates the candidate hydrogeological parameter fields based on the counterfactual observation schemes and corresponding executed observation data, and the target selection module confirms the target hydrogeological parameters after counterfactual deduction. This combination is necessary because a single module cannot independently solve the complex problem of obtaining reliable insights from sparse observation data. A collaborative workflow of "exploration (parameter field generation module), verification (scheme generation module), evaluation (parameter field update module), and convergence (target selection module)" is required to transform passive fitting into active learning. This allows the system to actively guide data acquisition (counterfactual observation schemes) to obtain the most discriminative information even with extremely limited initial observation data, and iteratively eliminate erroneous assumptions and approximate the true geological structure. This significantly reduces the ambiguity of the identification and inversion results, enhancing their reliability and decision support value in subsequent practical applications such as water resource assessment and pollution prediction.

[0023] In one feasible implementation, refer to Figure 2 As shown, the specific construction method of the parameter field generation module includes steps S11 to S13: Step S11: Analyze and process the prior geological information to extract the prototype field of the geological structure that differs from the preset geological structural features.

[0024] First, the prior geological information, including geological maps of the target area, borehole lithological records, geophysical interpretation data, sedimentary facies model diagrams, and other diverse and heterogeneous information, is analyzed to extract the descriptions, ranges, and uncertainties of tectonic and sedimentary characteristics (such as fault dip and extension range, main sand body distribution direction, unconformity spatial morphology, etc.). Then, based on the preset geological structural characteristics (such as "whether the fault cuts through the main aquifer" and "whether the connectivity pattern of the main channel sand body is network-like or zonal"), the qualitative descriptions and quantitative ranges with uncertainties are transformed into multiple spatially specific and mathematically expressible initial parameter fields.

[0025] For example, when the feature of "main channel extension direction" has two equally probable prior geological information directions, namely northeast and northwest, the analytical processing will not give a fuzzy mean field, but will generate a geological structure prototype field that clearly represents the spatial distribution of the northeast-oriented channel and another geological structure prototype field that clearly represents the spatial distribution of the northwest-oriented channel. In this way, qualitative, fuzzy, and even contradictory prior geological information is systematically transformed into multiple clear geological structure prototype fields that can be directly used by computers for forward modeling and optimization calculations. Thus, in the initial stage of the inversion process, the geological uncertainty space is completely defined and encapsulated in a structured form, providing a clear, specific and geologically meaningful set of candidate objects for the subsequent game and identification process based on competitive assumptions. This ensures the geological rationality and completeness of the exploration of diversity in the final inversion results from the source.

[0026] Step S12: Based on the geological structure prototype field and initial observation data, perform collaborative data assimilation processing to generate multiple initial parameter fields.

[0027] The purpose of collaborative data assimilation in this step is to fuse and constrain the multiple geological structure prototype fields extracted in the previous step with the initial observation data (such as measured hydrogeological data such as borehole water level, spring flow or tracer concentration) already acquired in the target area, so as to generate multiple initial parameter fields (i.e., spatially distributed hydrogeological parameter fields, such as permeability coefficient field or water storage coefficient field) that are both faithful to the prior geological information and the preset geological structure characteristics and can fit the initial observation data.

[0028] By employing the mathematical framework of collaborative data assimilation, and simultaneously applying spatial structural constraints provided by the geological structure prototype field and process response constraints provided by the initial observation data, extreme solutions that merely fit the data but have unreasonable geological structures or merely conform to the structure but cannot explain the data are avoided. This ensures that each generated initial parameter field is a reasonable candidate that is geologically and hydrologically consistent with the existing evidence. These generated initial parameter fields not only cover the key uncertainties in prior geological knowledge, but also pass the preliminary test of the initial observation data. This provides a data foundation for the subsequent scheme generation module to design counterfactual observation schemes and for the parameter field update module to conduct confidence assessments. Thus, from the initial stage, it lays the foundation for the possibility of the entire identification and inversion process converging in a geologically reliable and data-driven direction.

[0029] Furthermore, step S12 may specifically include steps S121 to S124: Step S121: Construct multiple parameter field models that differ in preset geological structure characteristics based on the geological structure prototype field.

[0030] Each extracted geological structure prototype field is used as a structural constraint framework. Under the defined geometric shape and spatial distribution rules of each geological structure prototype field, a parameter field model of multiple hydrogeological parameters (such as permeability coefficient or water storage coefficient) is generated.

[0031] For example, for a prototype geological structure representing a northeast-trending river channel, multiple concrete implementations of permeability coefficient fields can be generated based on the uncertainties in its channel axis, width, and boundaries. These implementations all maintain the structural characteristics of a northeast-trending river channel, but the spatial distribution of internal parameters exhibits random fluctuations, thus forming a series of parameter field models with different specific parameter configurations under the same preset structural characteristics. Similarly, for another prototype geological structure representing a northwest-trending river channel, another set of parameter field models with distinct structural characteristics can be generated. This transforms multiple differentiated prototype geological structures resolved from prior geological information into multiple concrete, computable, and structurally distinct parameter field models. These models maintain fundamental differences in preset geological structural characteristics (such as different river channel directions), thus providing a diverse and geologically meaningful model foundation for subsequent candidates. This ensures that the identification and inversion process can cover the geological uncertainty space from the source and lays the foundation for subsequent optimization and identification based on competitive assumptions.

[0032] Step S122: Numerical simulation is performed on each parameter field model based on the initial observation data to obtain the physical field prediction data of each parameter field model.

[0033] Using initial observation data from multiple initial observation points, including the spatial location of the initial observation points, the initial observation time, and initial conditions (the spatial distribution of system state variables throughout the simulation area at the start of the hydrogeological numerical simulation) and boundary conditions (such as initial head distribution, river water level boundaries, or pumping well flow rates), as input conditions to drive and control the numerical simulation of groundwater flow and solute transport, this data is used as the attribute input for each parameter field model constructed in the previous step. Under the simulation scenario defined by identical initial observation data, based on the geological structure of the target area and simulation requirements, the continuous simulation scenario is discretized into a spatial grid system composed of grid cells and a time series composed of time steps. Then, the control partial differential equations (such as the control equations composed of Darcy's law and the mass conservation equation describing groundwater flow) are used. ,in, The head (state variable) to be determined. and They are respectively and Permeability coefficient in the direction, For source and sink items, The water storage rate is discretized and approximated at each grid cell and each time step. It is a time variable.

[0034] For example, when approximating the spatial and temporal derivatives in the equations using finite difference forms, the simulation scene is first divided into rectangular meshes, and the mesh nodes are set. The head at time layer n is For the reciprocal of a space, for example, for the term In grid nodes To approximate the value, a central difference scheme is used to discretize it into... , , After discretizing the equations, substitute all the discretized equations into the control partial differential equations, and establish the discretized equations as shown above for each grid node in the simulation scenario, while also incorporating the known source and sink terms. Substituting this into the equation, we obtain an unknown head for all grid nodes at the current time level n. The system of nonlinear algebraic equations formed by this.

[0035] Next, the initial conditions and boundary conditions provided in the initial observation data are used as the boundary conditions for the nonlinear algebraic equation system. This includes assigning the initial conditions (i.e., the spatial distribution of state variables in the entire region at the start of the simulation) to all grid nodes of the initial time layer, and applying the boundary conditions (such as known head boundaries or flux boundaries) to the boundary grid nodes at the corresponding locations. Specifically, the initial conditions, i.e., the spatial distribution of head in the entire simulation scenario at the start of the simulation (corresponding to time layer 0), are applied. The values ​​are directly assigned to all network node variables at the corresponding time level in the nonlinear algebraic equation, and the boundary conditions are applied to the boundary grid nodes at the corresponding locations. For example, for a known head boundary, the boundary conditions at the boundary grid nodes are directly applied. Assuming known constant values ​​or time function values, for known flow boundaries, these are transformed into difference forms through boundary terms and substituted into the discrete equations. Then, this system of nonlinear algebraic equations containing known initial values ​​and boundary conditions is solved, typically using iterative methods such as the Newton-Raphson method. This method linearizes the system and solves a modified system of linear algebraic equations in each iteration. Once the iterations converge, the numerical solution of the head for each internal grid node at each discrete time step n can be calculated. .

[0036] Finally, for each initial observation point recorded in the initial observation data (which typically does not precisely coincide with the coordinates of discrete grid nodes), the grid cell in which the initial observation point falls is first determined. For a regular rectangular grid, by comparing the spatial position of the initial observation point with the coordinates of all grid nodes, the grid cell in which it falls can be quickly located, i.e., finding the rectangular region whose east-west coordinates lie between two adjacent grid lines and whose north-south coordinates lie between two other adjacent grid lines; the four corner nodes of this grid cell constitute the reference points for interpolation; then, the numerical solution of the head is extracted... After determining the known head values ​​at the four corner grid nodes at the current discrete time step, two linear interpolations are performed. Specifically, in the east-west direction, based on the horizontal position ratio of the initial observation point relative to the left and right grid lines, the head values ​​at the two upper corner nodes and the two lower corner nodes are linearly interpolated once to obtain two intermediate estimates. Subsequently, in the north-south direction, based on the vertical position ratio of the initial observation point relative to the upper and lower grid lines, these two intermediate estimates are linearly interpolated a second time. The calculated result is the estimated head value of the initial observation point at its spatial location and at the current discrete time step.

[0037] Subsequently, alignment is performed in the time dimension. Since the initial observation time is a continuous time point, it may not be exactly equal to a discrete time step. Therefore, it is necessary to perform another linear interpolation in the time dimension based on the estimated water head values ​​at two discrete time steps adjacent to the initial observation time. This allows for the final calculation of the water head value at the initial observation point at its initial observation time and corresponding initial observation spatial location. Specifically, let the initial observation time be... It falls exactly within the discrete time step. and Between, among them, for Previous adjacent discrete time steps, for The subsequent adjacent discrete time steps, and it is known that at... The estimated head value at the initial observation point at time (after spatial interpolation) is ,exist The estimated head value at the initial observation point at time (after spatial interpolation) is Then, calculate the initial observation time. The time position scaling factor relative to these two discrete time steps Through linear interpolation formula in the time dimension The calculations were performed, and the results were obtained. That is, the initial observation time corresponding to the initial observation point. And the final head value at its corresponding initial observation spatial location.

[0038] The above interpolation calculation process is repeated for all initial observation points recorded in the initial observation data. The resulting set of water head values ​​with clear spatiotemporal labels, which correspond one-to-one with the observation records, constitutes the physical field prediction data of the corresponding parameter field model. This provides a set of physical field prediction data that can be directly quantified and compared under the same observation conditions for each parameter field model established based on different preset geological structure characteristics assumptions. In this way, the parameter field model is transformed into prediction time series or spatial distribution data that are completely consistent with the dimensions of the measured data. This provides a comparison benchmark for objectively evaluating the model data matching degree between each parameter field model and the real observation data in subsequent steps.

[0039] Step S123: Calculate the model data matching degree between the initial observation data and the predicted data of each physical field, and calculate the difference between each parameter field model and the geological structure prototype field to obtain the model structure fidelity.

[0040] The physical field prediction data of each parameter field model is quantitatively compared point by point with the actual measured values ​​recorded in the initial observation data. The root mean square error of the difference between all corresponding physical field prediction data and actual measured values ​​is calculated to quantify the overall deviation between the data. The overall deviation obtained by this calculation is the model data matching degree of the corresponding parameter field model, and its value directly characterizes the accuracy of the parameter field model in reproducing existing observation phenomena.

[0041] Simultaneously, key spatial structural features (such as the spatial distribution direction, geometric morphology, or connectivity pattern of major high-permeability zones) are extracted for each parameter field model and its corresponding geological structural prototype field. A predefined structural similarity metric function quantifies the degree of consistency between the two in spatial structural patterns. The calculated similarity metric value represents the model structure fidelity of the parameter field model. Specifically, the formula for the predefined structural similarity metric function can be expressed as the root mean square error based on the difference in variogram functions, i.e. ,in, Represents a structural similarity metric. It is a predefined number of lag distance tiers. It is the k-th lag distance vector. It is the parametric field model in the lag distance The experimental variation function value at that location, The prototype field of the geological structure is at the same lag distance The experimental variogram values ​​at each location. Experimental variogram values ​​are calculated for the gridded parameter data of the parameter field model and the geological structure prototype field, respectively; that is, for each preset lag distance vector... traverse all pairs of ... For each grid point pair, the square of the difference in parameter values ​​between the pairs is calculated. Then, the square of the differences between all such pairs is averaged to obtain a series of experimental variogram values ​​for the two fields. and These variogram sequences characterize the spatial autocorrelation structure of their respective parameter field models. Then, the two variogram sequences are aligned point-by-point on the same lag distance sequence, and the values ​​for each pair are calculated according to the formula described above. and The squared difference is summed over all lag distance levels and then divided by the total number of levels. Obtain the mean, then take the square root of that mean to get the final result. The value represents the model structure fidelity of the parametric field model. The smaller the model structure fidelity, the closer the parametric field model is to the spatial variation pattern of the geological structure prototype field, and the higher its structure fidelity.

[0042] For each parameter field model, two complementary and key quantitative evaluation indicators are generated: model data matching degree measures its goodness of fit with the dynamic observation process, and model structure fidelity measures its consistency with static geological knowledge. This provides optimization objectives for subsequent steps to construct a multi-objective collaborative optimization function, enabling the system to drive the evolution of the parameter field model in the iterative update process in two directions: "improving the ability to interpret observation data" and "maintaining the rationality of geological structure". This ensures that the generated initial parameter field is a high-quality solution under the joint constraints of data and knowledge.

[0043] Step S124: Based on the model data matching degree and model structure fidelity, after constructing a multi-objective collaborative optimization function, the model of each parameter field is updated through the multi-objective collaborative optimization function to obtain multiple initial parameter fields.

[0044] Model data matching degree for each parameter field model Model structure fidelity and other parameters are used as two optimization objectives, which are aggregated into a scalarized multi-objective collaborative optimization function through a weighted summation function. The specific formula is as follows: ,in, Representative parameter field model, and These are pre-defined positive weighting coefficients used to balance the relative importance of the two objectives: model data fit and model structure fidelity. Next, the parameter field model is adjusted using an optimization algorithm. The spatial parameter distribution is used to generate new parameter field models. Taking the differential evolution algorithm as an example, a population containing multiple parameter field models is maintained. In each iteration, for each parameter field model in the population... Randomly select three other different parameter field models from the population , and By generating mutation vectors ,in, It is a scaling factor, and then the mutated vector is mutated through a crossover operation. With the current parameter field model The experimental vector is generated by mixing the variant vectors or the parameter values ​​of the current model with a certain crossover probability. Then calculate the test vector. Corresponding function value and with the current model function value Comparison, if Then use experimental vectors in the next generation. replace Otherwise, keep Through this process of mutation, crossover, and selection, the permeation coefficient values ​​of the parameter field model at the grid nodes are gradually adjusted to generate a new parameter field model. This process is repeated until the convergence criterion is met (such as the function value change being less than a threshold or reaching the maximum number of iterations). The new parameter field model generated at this point is the initial parameter field.

[0045] By simultaneously driving the models of each parameter field under an automated optimization framework to evolve in the direction of improving the fitting accuracy of observation data and maintaining the rationality of geological structure, a set of multiple initial parameter fields that achieve the best trade-off between data matching and structural fidelity are finally generated, providing a high-quality and geologically reasonable initial solution set for subsequent hydrogeological inversion or simulation prediction.

[0046] Step S13: Perform adversarial differentiation enhancement on each initial parameter field to optimize the geological structure characteristics of each initial parameter field and obtain multiple candidate hydrogeological parameter fields.

[0047] An optimization process is defined with the goal of maximizing the distinguishability of different initial parameter fields in terms of preset geological structural features.

[0048] First, the eigenvectors of each initial parameter field are defined and extracted. Specifically, the eigenvectors are defined as core attributes that characterize the spatial variation patterns of the initial parameter field, including the main modes of spatial distribution, dominant permeation directions, and connected body scale. For the main modes of spatial distribution, principal component analysis is used to process the gridded data of the initial parameter field, calculate its covariance matrix, and perform eigenvalue decomposition. The eigenvectors corresponding to the first few largest eigenvalues ​​are extracted as the main spatial modes, and the coefficients or corresponding variance contribution rates of these eigenvectors are used as the quantitative indicators of the main modes. For the dominant permeation direction, the spatial gradient of the initial parameter field at each grid node is calculated to construct a structural tensor field, and the structural tensor field is subjected to regional averaging or eigenvalue decomposition. The eigenvector direction angle corresponding to the largest eigenvalue is used as the quantitative indicator of the dominant permeation direction. For the connected body scale, high-permeability or low-permeability connected regions are identified based on the spatial distribution of the initial parameter field. Then, the equivalent diameter or area distribution statistics (such as average area, maximum Feret diameter) of these connected regions are calculated as the quantitative indicators of the connected body scale. In this way, each initial parameter field is transformed into a set of numerical eigenvectors.

[0049] Then, an adversarial loss function is constructed. The core term of this function is to calculate the difference between any two different initial parameter fields on the aforementioned feature vectors, and to average the difference between all pairs of initial parameter fields. Specifically, for all pairs of initial parameter fields (i.e., any two different initial parameter fields), the difference between their feature vectors is calculated. Finally, the difference between all pairs is aggregated, and the average is calculated to reflect the overall difference. This aggregation is achieved using the adversarial loss function, the specific formula of which is expressed as follows: ,in, It is the number of initial parameter fields. It is the first The feature vector formed by feature extraction from the initial parameter field. It is an eigenvector and The degree of difference between them, for example, using Euclidean distance Or based on the difference degree of cosine similarity This forms the core term of the adversarial loss function.

[0050] Next, under the constraint that the matching degree of the original model data and the fidelity of the model structure of each initial parameter field are not significantly reduced, the spatial parameter distribution of each initial parameter field is adjusted (i.e., the parameter values ​​on the grid nodes are fine-tuned) through an iterative optimization algorithm (e.g., introducing an adversarial regularization term in the gradient descent framework). The direction of adjustment is to increase the value of the adversarial loss function, that is, to actively expand the structural feature differences between different initial parameter fields. Specifically: First, for each initial parameter field... Define its original objective function value The objective function The value represents the collaborative state of data matching and structure fidelity achieved in step S124; next, a total loss function including an adversarial regularization term is constructed. ,in, Based on the set of all initial parameter fields { The calculated adversarial loss function value, It is a positive regularization weight coefficient; the negative sign indicates that we need to maximize the adversarial loss (i.e., enhance the difference), but as a whole loss function, we still adopt the minimization framework, so it can be defined as follows: To achieve an equivalent result, in minimizing At that time, the optimizer will try to reduce But increase Then, iterative optimization is performed within the gradient descent framework, and the total loss function is calculated in each iteration. Regarding the gradient of the permeability coefficient parameter value at each grid node, the gradient consists of two parts, one part being... The gradient, whose direction tends to maintain or improve the data matching and structure fidelity of a single field, is another part of... Its direction tends to adjust To increase overall combat losses (i.e., highlighting the characteristic differences between different fields); parameter updates follow ,in, It is the learning rate. Through this update rule, the spatial parameter distribution of each initial parameter field is synchronously fine-tuned, and the direction of adjustment is... Under the constraint that the growth does not exceed the preset tolerance threshold (i.e., does not significantly reduce the original performance), the adversarial loss function value L_adv is increased as much as possible, thereby achieving the purpose of actively expanding the structural feature differences between different initial parameter fields.

[0051] Finally, when the optimization process converges, a set of multiple candidate hydrogeological parameter fields with higher distinguishability in key geological structural features is obtained after adversarial differentiation enhancement. This actively strengthens the diversity within the initial parameter field set generated in step S12, ensuring that the generated candidate hydrogeological parameter fields are not only different in random details, but also have significant divergences in geological structural patterns (such as different channel configurations or fault networks) that can be revealed by subsequent differences in physical response. This provides input for the counterfactual inference in the next stage, and improves the overall ability and efficiency of the inversion process to identify and approximate the real geological structure.

[0052] In one feasible implementation, refer to Figure 3 As shown, the specific implementation of the scheme generation module includes steps S21 to S23: Step S21: Numerical simulation is performed based on multiple candidate hydrogeological parameter fields to obtain the physical field prediction response of each candidate hydrogeological parameter field under simulation combinations of multiple preset observation locations and preset time series.

[0053] First, the multiple candidate hydrogeological parameter fields generated in the above steps are used as the input model for numerical simulation. Each candidate hydrogeological parameter field defines the hydrogeological attribute value of each grid node on the spatial grid of the study area.

[0054] Next, a unified numerical simulation control condition is set, including identical initial conditions (such as initial head distribution), boundary conditions (such as constant head boundary or flow boundary), and source-sink terms (such as the flow rate of pumping wells). A specific groundwater flow and solute transport simulator (such as MODFLOW based on the finite difference method or FEFLOW based on the finite element method) is selected as the forward modeling engine. Then, a preset observation location containing multiple possibilities (such as a series of potential borehole coordinates or surface monitoring point locations) and a preset time series (such as a series of discrete observation times, from day 1, day 10 to day 100 after the start of the simulation) are defined. For each candidate hydrogeological parameter field, the numerical simulation is run under the exact same numerical simulation control condition to calculate the physical field state (such as groundwater head field or solute concentration field) that evolves over time.

[0055] Subsequently, for each combination of a preset observation location and a preset time series (i.e., a specific "where and when to observe" scheme), the physical field state values ​​(such as water head or concentration values) at each time point in the corresponding preset time series at the corresponding preset observation location are extracted from the simulation output. This numerical sequence arranged in chronological order is the physical field prediction response of the corresponding candidate hydrogeological parameter field under the simulation combination.

[0056] This process traverses all candidate parameter fields and all preset observation locations and preset time series combinations, ultimately obtaining a three-dimensional response data volume. Its dimensions correspond to different candidate hydrogeological parameter fields, different preset observation locations, and different preset time series. This provides a predictive experimental library for subsequent evaluation and counterfactual observation schemes, fully quantifying and parallelizing the performance of different candidate hydrogeological parameter fields under various possible future observation scenarios. This provides a complete data foundation for proactively designing the optimal observation scheme that can most effectively distinguish these competing hypotheses.

[0057] It should be noted that the physical field prediction response specifically refers to the numerical value of a physical quantity (such as water level, pressure, or concentration) that can be observed at a specific location and time, obtained through numerical simulation; the preset observation location refers to the set of spatial coordinates of the actual measurement points set in advance; and the preset time series refers to the set of time points set in advance for observation.

[0058] Step S22: Based on the pairwise physical field prediction responses under the same simulation combination, calculate the difference in prediction responses between each candidate hydrogeological parameter field to obtain the distinguishability index.

[0059] For each simulation combination, the physical field prediction responses of all candidate hydrogeological parameter fields under the simulation combination are extracted. This process iterates through all possible candidate field pairs (i.e., any two different candidate hydrogeological parameter fields). For each candidate field pair, the difference between their physical field prediction responses under the same simulation combination is calculated. The difference calculation typically uses a predefined difference metric function, such as calculating the Euclidean distance or root mean square error between two candidate hydrogeological parameter fields. A specific formula could be: ,in, and They represent the first The candidate field predicted response and the first Each candidate field predicts the response in time. The predicted response values ​​are summed and iterated through all preset time points in the simulation combination. Then, the difference values ​​calculated for all candidate field pairs are aggregated. The aggregation method can be to calculate the average of all difference values, thus obtaining the distinguishability index for the simulation combination. For example, the distinguishability index D is defined as the average difference of all field pairs. ,in, This represents the total number of candidate fields. The potential utility of each simulation combination is transformed into a single, quantifiable numerical indicator, providing a direct decision-making basis for the global optimization selection of the optimal simulation combination in subsequent steps. This enables the system to automatically identify simulation combinations that best highlight the differences between different candidate hydrogeological parameter fields in terms of spatial location and temporal arrangement, thereby guiding data acquisition to maximize information acquisition efficiency.

[0060] It should be noted that the distinguishability index is a quantitative indicator that characterizes the overall difference between the physical field prediction responses of different candidate hydrogeological parameter sites under the simulation combination. The larger the distinguishability index, the more effectively the simulation combination can distinguish different candidate hydrogeological parameter sites.

[0061] Step S23: Perform global optimization based on the distinguishability index to obtain the target observation location and target time series, and then generate a counterfactual observation scheme based on the target observation location and target time series.

[0062] The specific method of global optimization is as follows: all preset observation locations and all preset time series are regarded as a discrete decision space, where each simulation combination is a candidate point in the decision space, and its corresponding distinguishability index value is the utility value. By traversing and comparing the utility values ​​of all candidate points, the simulation combination with the largest distinguishability index value is selected. The preset observation location corresponding to this simulation combination is the target observation location, and the preset time series corresponding to it is the target time series.

[0063] Then, the target observation location and target time series are used as core instructions and encapsulated into a counterfactual observation plan containing operational details. This counterfactual observation plan at least clearly specifies what kind of sensors need to be deployed or what kind of test (such as a pumping test) needs to be conducted at the target observation location, the specific time specified in the target time series for observation, and the specific physical quantities (such as water head and concentration) that need to be recorded. In this way, the distinguishability index calculated in the previous steps is transformed into an immediately executable data acquisition action plan, namely the counterfactual observation plan, which provides reliable data input for the subsequent parameter field update module.

[0064] In one feasible implementation, refer to Figure 4 As shown, the specific implementation of the parameter field update module includes steps S31 to S38: Step S31: Based on the target observation location and target time series in the counterfactual observation scheme, deploy a sensor network in the target area and collect measured physical field data in the target area through the sensor network.

[0065] The target observation locations specified in the counterfactual observation scheme are transformed into on-site deployment schemes. Corresponding hydrogeological sensors (such as pressure sensors for measuring head, conductivity sensors, or samplers for measuring concentration) are installed or activated at the corresponding coordinate points in the target area. The type, range, and accuracy of the sensors must match the type of physical field (such as head field or concentration field) required by the counterfactual observation scheme. All sensors are connected to a unified network via wired or wireless means. This network is responsible for power supply, command control, and data transmission.

[0066] According to the target time series clearly defined in the counterfactual observation scheme, the control unit of the sensor network synchronously sends a trigger command to all sensors at each specified observation time. Each sensor then performs a measurement and converts the measured signal (such as voltage and frequency) into a value with physical units (such as water head value, in meters; or concentration value, in milligrams per liter) through its internal or network-integrated calibration curve. This value, along with its corresponding sensor identifier and timestamp, is packaged and transmitted to the central data receiving and storage unit through the data network to form measured physical field data.

[0067] This allows the counterfactual observation scheme obtained through computational optimization in the preceding steps to be executed in the physical world, thereby obtaining a set of highly targeted and spatiotemporally accurate measured physical field data. This measured physical field data serves as the data source for hypothesis testing and confidence assessment in subsequent steps, ensuring that the cognitive update of the entire identification and inversion system is based on real and targeted physical observations, rather than relying solely on simulations or speculations.

[0068] Step S32: Based on the counterfactual observation scheme, perform a fidelity numerical simulation to calculate the simulated physical field data of the target area under the current hydrogeological conditions.

[0069] First, a hydrogeological conceptual model is established that is independent of the candidate hydrogeological parameter field and represents the most reliable current understanding of the target area. This hydrogeological conceptual model is built based on all available prior geological information, historical observation data and engineering experience, and defines the geometric structure, boundary conditions, initial conditions and all material properties and source and sink terms of the target area except for the parameters to be inverted (such as permeability coefficient).

[0070] Then, based on the established hydrogeological conceptual model, a high-resolution, high-precision numerical simulation model is constructed. A validated numerical simulator capable of accurately solving the relevant physical process control equations (such as a three-dimensional unsteady flow finite element simulator) is selected. The target observation location and target time series specified in the counterfactual observation scheme are used as the output control conditions of the numerical simulator. That is, the simulation settings explicitly specify that the physical field state variables (such as head or concentration) of each moment of the target time series need to be output at the grid nodes or cells of the target observation location. Then, a fidelity numerical simulation is run under the set initial and boundary conditions to calculate the spatiotemporal evolution of the physical field in the target area throughout the simulation period.

[0071] Finally, simulated physical field data at the target observation location and at each moment of the target time series are extracted from the simulation output data. This enables the generation of a highly reliable simulated physical field data based on the best current geological knowledge under the same counterfactual observation scheme. This simulated physical field data does not depend on any candidate parameter fields to be evaluated, thus providing a data benchmark for subsequent steps to integrate with measured physical field data. This allows the system to evaluate the degree of support or refutation of each candidate hydrogeological parameter field by comparing the differences between measured data and simulated physical field data, thereby enhancing the reliability of the entire evaluation process.

[0072] Step S33: Perform spatiotemporal registration and uncertainty quantification on the measured physical field data and the simulated physical field data to generate fused physical field data.

[0073] When performing spatiotemporal registration, since the measured physical field data comes from discrete sensors, the timestamps recorded by the data may not be perfectly aligned with the time output of the simulated physical field data due to clock deviations and transmission delays. Therefore, it is necessary to interpolate or resample the time series of the measured physical field data to ensure that its time points are accurately aligned with the target time series output by the simulated physical field data. At the same time, it is necessary to ensure that the spatial coordinates (sensor positions) of the measured physical field data are matched with the grid nodes used for extracting the simulated physical field data in space through coordinate transformation or projection, thereby eliminating the system position deviation introduced by coordinate system or grid discretization.

[0074] Next, when performing uncertainty quantification, the process includes analyzing the error sources of the measured physical field data (such as sensor accuracy error, reading drift, and environmental interference), and assigning a quantification index (such as standard deviation, used to quantify the measurement or prediction error range of the data point) to each data point (i.e., the spatial coordinates corresponding to the measured physical field data and the grid nodes corresponding to the simulated physical field data). At the same time, the error sources of the simulated physical field data are analyzed (such as conceptual model simplification error, numerical discretization error, and boundary condition uncertainty), and corresponding model prediction uncertainty indices (such as numerical discretization error, used to reflect the source of model prediction error) are assigned to them.

[0075] Then, based on the spatiotemporally registered measured and simulated physical field data, and their respective indices (i.e., uncertainty quantification indices, such as standard deviation), and model prediction uncertainty indicators, such as numerical discretization error. The method employs a variance-based weighted average fusion: First, the square of the standard deviation is used as the basis for calculating the fusion weights. The weights of the measured physical field data are calculated as follows: The weights of the simulated physical field data are calculated as follows: The principle of this weight allocation is to assign higher feasibility to data points with lower uncertainty. Then, the fusion estimate is calculated, i.e., fusion estimate = This generates a statistically optimal fusion estimate, and calculates the variance corresponding to this fusion estimate, which integrates the uncertainties of observation and model. This variance comprises the weights used in the weighted average fusion, the uncertainty quantification index, and the model prediction uncertainty index. Following the variance propagation law of linear combinations, and assuming that the uncertainty quantification index and the model prediction uncertainty index are independent, the formula for calculating the variance of the fusion estimate is as follows: The variance calculation formula for this fusion estimate is derived from the definition of weights in the optimal weighted average and is equivalent to... Calculated This is the square root of the variance of the fusion estimate. This is the standard deviation of the fusion estimate. It directly measures the uncertainty of the fused estimate, and its value includes the effects of both observation error and model prediction error. Moreover, due to the complementarity of information, its value is usually smaller than that of any single independent estimate. or Based on this, the confidence interval of the fused estimate can be constructed as follows: ,in, It is a constant selected based on the required confidence level (e.g., for an approximately normal distribution). =1.96 corresponds to approximately 95% confidence level. Finally, for each spatiotemporal point (i.e., target observation location - target observation time), a triplet record will be generated, namely the fusion estimate, the standard deviation of the fusion estimate, and the confidence interval of the fusion estimate. This dataset, consisting of triplet records corresponding to all candidate hydrogeological parameter fields, is the fusion physical field data. It lays the data foundation for extracting fusion feature vectors and performing hypothesis testing in subsequent steps, effectively improving the robustness and scientificity of the entire parameter field update module in evaluation and decision-making.

[0076] Step S34: Extract the fusion feature vector from the fused physical field data, and combine the fusion feature vector to form the execution observation data output.

[0077] For each spatiotemporal point in the fused physical field data, a set of quantitative indicators (including fused estimates, their standard deviations, and possible confidence intervals) is designed and calculated to characterize the core features of the physical response at the corresponding spatiotemporal point, based on the needs of subsequent hypothesis testing. These core physical response features typically include, but are not limited to, the statistical characteristics of the target time series (calculated from a set of fused estimates at different observation times for the same target observation location. For example, for location A, all fused estimates at observation times T1, T2, ..., Tn of the target time series are extracted to form a numerical sequence. Then, statistical analysis is performed on this numerical sequence to calculate its mean and variance, and the slope of its changing trend is fitted by linear regression), the response amplitude at a specific time, and the time to reach the peak (extracted from the fused estimate sequence; the response amplitude can be directly taken from the specific time series). The calculation includes the fusion estimate at a fixed time (such as the stimulus end time set in the simulation; the time to reach the peak is determined by finding the observation time corresponding to the maximum value of the fusion estimate in the target sequence), the standardized residual of the simulation reference value (directly from the simulated physical field data output in step S32. During the calculation, the simulated physical field data at the same spatiotemporal point (i.e., the same location and the same time) as the fusion physical field data are first extracted from the simulated physical field data as the simulation reference value. Then, the fusion estimate at that spatiotemporal point is subtracted from the simulation reference value to obtain the initial residual. After that, the original residual is divided by the standard deviation of the fusion estimate at that spatiotemporal point to obtain the standardized residual), as well as its own indicators, and for multiple target observation positions deployed in space, it is also necessary to calculate the spatial relationship characteristics between the target observation positions, such as the correlation coefficient or phase difference of the time series between different target observation positions.

[0078] The quantitative indicators obtained above are arranged into an ordered numerical array according to a preset fixed order. This numerical array is a fusion feature vector extracted from the fused physical field data. It is a dimensionality reduction and information summary of the spatiotemporal data, encapsulating key information used to distinguish different candidate hydrogeological parameter fields. Then, the fusion feature vectors are combined as the output of the execution observation data. The specific method is as follows: in the current iteration, the fusion feature vectors extracted from all spatiotemporal points obtained according to the counterfactual observation scheme are connected and combined into a longer, global execution observation data according to the rule of spatial location priority and temporal order. This execution observation data can reduce the computational burden of a large number of repetitive feature comparisons in subsequent steps, while ensuring that the data sent to the hypothesis testing stage is redundancy-free and high-fidelity data. Thus, while maintaining the integrity of information, the processing efficiency and numerical stability of the entire parameter field update module are significantly improved.

[0079] Step S35: Extract the predicted feature vector corresponding to the fused feature vector in the observation data from the physical field prediction response of each candidate hydrogeological parameter field.

[0080] Following the same set of feature extraction rules, calculation order, and spatiotemporal framework used when defining the fusion feature vector in step S34, this template is applied to the predicted response value of each candidate hydrogeological parameter field.

[0081] For example, if the first element of the fused feature vector is "the response amplitude at location A at time T", then the predicted response value at location A at time T is extracted from the predicted response value of the first candidate hydrogeological parameter field as the first element of the predicted feature vector. If the second element is "the time series mean at location B", then the predicted response mean of the corresponding candidate hydrogeological parameter field at location B over the entire target time series is calculated as its second element. This process is repeated sequentially to generate a predicted feature vector for each candidate hydrogeological parameter field that corresponds to its physical field predicted response and is completely isomorphic to the fused feature vector in the executed observation data in terms of dimension and feature meaning. This provides data input for the next step of directly calculating the quantified matching index to update the candidate hydrogeological parameter field.

[0082] Step S36: Based on the fused feature vector and each predicted feature vector, calculate the matching index of each candidate hydrogeological parameter field.

[0083] For each candidate hydrogeological parameter field, the corresponding fused feature vector and predicted feature vector are compared component by component. Based on the uncertainty of each feature component, a weighted distance metric function is used to calculate the overall difference between the fused feature vector and the predicted feature vector.

[0084] For example, the formula for calculating the Mahalanobis distance as the matching metric is as follows: First, based on the standard deviation of each feature component in the fused feature vector, a diagonal covariance matrix Σ is constructed, where the elements on the diagonal are the squares of the standard deviations of each feature component; then, for the ... For each candidate hydrogeological parameter field, calculate its predicted eigenvector. With fused feature vectors The difference vector between Next, calculate the square of the Mahalanobis distance. , The matching index for the currently calculated candidate hydrogeological parameter field. The smaller the value, the higher the consistency between the predicted response to the candidate hydrogeological parameter field and the fused observation after taking uncertainty into account.

[0085] This transforms the qualitative question of whether the fused feature vector and the predicted feature vector are consistent into a quantitative value, thus providing an objective and operable parameter for subsequent steps.

[0086] Step S37: Based on the matching indicators and the cumulative confidence scores in historical iterations, perform a fusion calculation to obtain the confidence scores of each candidate hydrogeological parameter field in the current iteration round.

[0087] It should be noted that the cumulative confidence score in the historical iterations refers to the confidence score accumulated for each candidate hydrogeological parameter field after several rounds of counterfactual observations and updates before the current iteration round, such as the confidence score output at the end of the previous iteration round.

[0088] First, the cumulative confidence score from previous iterations is used as the confidence score for the current round. Then, the matching index calculated in the previous step is converted into a likelihood function value. The transformation method involves mapping the matching index to a probability density function based on a chi-square distribution, representing a numerical value indicating the probability of observing the currently executed observation data under the condition that the candidate hydrogeological parameter field is true. Specifically, this can be achieved using... Likelihood function value The likelihood decreases monotonically as the matching index increases, meaning the worse the match, the lower the likelihood.

[0089] Next, the confidence score is calculated according to Bayes' theorem. The denominator is the normalized constant of all candidate hydrogeological parameter fields, ensuring that the sum of all posterior confidence scores is 1. This is the full ergodic index in the normalized denominator, used to perform ergodic summation on all candidate hydrogeological parameter fields.

[0090] Finally, the calculated confidence score will be... As the confidence score output of the corresponding candidate hydrogeological parameter field in the current iteration round, the confidence score can gradually evolve with the iteration process. It can not only quickly respond to the strong identification results of new observations, but also resist excessive fluctuations caused by the random errors of single observations. This allows for a more robust and accurate characterization of the credibility of each candidate field relative to the real underground structure, providing a scientific and continuous decision-making basis for the elimination and evolution based on the confidence score in the subsequent step S38.

[0091] Step S38: After eliminating the candidate hydrogeological parameter fields corresponding to confidence scores lower than the preset confidence scores, Gaussian random perturbation is applied to the confidence scores corresponding to the remaining candidate hydrogeological parameter fields, and the preset geological structure features among the remaining candidate hydrogeological parameter fields are cross-recombined to obtain the updated candidate hydrogeological parameter fields.

[0092] First, based on the confidence scores of each candidate hydrogeological parameter field output in the previous step in the current iteration, a preset confidence score threshold is set. For example, the preset confidence score threshold can be set as the median or a certain percentile of the confidence scores of all candidate hydrogeological parameter fields. All candidate hydrogeological parameter fields with confidence scores lower than the preset confidence score threshold are removed from the current set of all candidate hydrogeological parameter fields, because these candidate hydrogeological parameter fields have been determined to have low consistency with reality, and continuing to retain them would dilute the quality and increase the computational burden.

[0093] Then, for the remaining candidate hydrogeological parameter fields, a Gaussian random perturbation is applied to their respective confidence scores. That is, a Gaussian random noise with a mean of zero and a preset small standard deviation is added to the confidence scores of the remaining candidate hydrogeological parameter fields to obtain the perturbed confidence scores. The purpose of this step is to introduce appropriate randomness in the subsequent reorganization operation to prevent the candidate hydrogeological parameter fields from losing diversity due to over-selection, thereby avoiding premature convergence to a local optimum.

[0094] Next, based on the perturbed confidence scores and the differences in preset geological structural features (such as river channel direction, fault distribution, and lens morphology) among the remaining candidate hydrogeological parameter fields, a cross-recombination operation is performed. That is, a weighted random pairing is performed based on the perturbed confidence scores (candidate hydrogeological parameter fields with higher scores have a greater probability of being selected for recombination). For each pair of candidate hydrogeological parameter fields, based on the spatial distribution of their preset geological structural features, structural features of a portion of the region (such as the permeability coefficient distribution of a certain sub-region) are extracted from a candidate hydrogeological parameter field. Structural features of the remaining areas are extracted from another candidate hydrogeological parameter field. These two fields are then combined and localized smoothing is applied to eliminate boundary abrupt changes, generating a new candidate hydrogeological parameter field. This pairing and recombination process is repeated until multiple new candidate hydrogeological parameter fields are generated, restoring the total number of updated candidate hydrogeological parameter fields to the number before elimination. Finally, these newly generated candidate hydrogeological parameter fields are combined with the candidate hydrogeological parameter fields that were not eliminated and did not participate in recombination to form an updated candidate hydrogeological parameter field. By eliminating low-confidence candidate hydrogeological parameter fields to achieve a process of natural selection, maintaining necessary exploration randomness through Gaussian random perturbation to avoid getting trapped in local optima, and achieving the inheritance and innovation of structural patterns through cross-recombination based on geological structural features, the synergistic effect of these three factors allows the candidate hydrogeological parameter fields to evolve towards a more reliable, diverse, and competitive direction in each iteration, thereby continuously improving the ability and efficiency of the inversion system to approximate the real hydrogeological parameter field.

[0095] In one feasible implementation, refer to Figure 5 As shown, the specific implementation of the target filtering module includes steps S41 to S44: Step S41: Based on each updated candidate hydrogeological parameter field and the cumulative confidence level, obtain several first hydrogeological parameter fields that satisfy the preset convergence criteria from multiple updated candidate hydrogeological parameter fields.

[0096] The updated candidate hydrogeological parameter fields are judged to have reached convergence based on a preset convergence criterion. In this embodiment, the preset convergence criterion is as follows: find the maximum cumulative confidence score among all updated candidate hydrogeological parameter fields, and check whether the maximum cumulative confidence score exceeds a first preset threshold. At the same time, check whether the sum of the cumulative confidence scores of the top few (e.g., the top three) exceeds a second preset threshold. If both conditions are met, the updated candidate hydrogeological parameter fields are considered to have converged. The first preset threshold is used to determine whether a dominant solution exists. The upper threshold for cumulative confidence requires that the maximum cumulative confidence score must be greater than or equal to the first preset threshold (e.g., 0.85) to be considered the maximum cumulative confidence score. The second preset threshold is the lower limit threshold for judging whether the probability quality has been contracted to a few candidates. It requires that the sum of the confidence scores of the top few updated candidate hydrogeological parameter fields sorted by cumulative confidence score must be greater than or equal to the second preset threshold (e.g., 0.95) to prevent false convergence caused by a single candidate with a high cumulative confidence score but the remaining updated candidate hydrogeological parameter fields are still scattered.

[0097] Once the preset convergence criterion is met, several updated candidate hydrogeological parameter fields with a cumulative confidence level higher than a certain cutoff threshold (e.g., a cumulative confidence level higher than half of the maximum cumulative confidence level or higher than the cumulative confidence level of the average cumulative confidence level of all updated candidate hydrogeological parameter fields) are extracted from each updated candidate hydrogeological parameter field. These candidate hydrogeological parameter fields are then marked as the first hydrogeological parameter field.

[0098] If the preset convergence criterion is not met, the process returns to continue executing a new round of counterfactual deduction and parameter field update iteration until the preset convergence criterion is met.

[0099] In this step, by using quantitative convergence criteria, candidate hydrogeological parameter fields that have become stable and have significantly leading confidence after multiple rounds of active learning and game evolution are automatically identified. This allows subsequent operations to focus on the candidate hydrogeological parameter field (i.e., the first hydrogeological parameter field) that is most likely to represent the real geological structure, avoiding wasting computational resources on scattered solution sets that have not yet converged. It also lays the foundation for the final output of a highly reliable target hydrogeological parameter field.

[0100] Step S42: Based on the cumulative confidence of each first hydrogeological parameter field, a second hydrogeological parameter field is selected from several first hydrogeological parameter fields.

[0101] Based on the cumulative confidence scores of each first hydrogeological parameter field, the first hydrogeological parameter field with the highest cumulative confidence score can be directly selected as the second hydrogeological parameter field. This is because the first hydrogeological parameter field with the highest cumulative confidence score means that it has the highest credibility in the current convergence and is the most representative single optimal solution. In this way, the multiple first hydrogeological parameter fields with high cumulative confidence scores obtained after multiple rounds of active learning and evolution are further refined into a second hydrogeological parameter field that is the most representative or statistically optimal. This provides input objects for the multi-scenario robustness verification, spatial smoothing and boundary adaptation of subsequent steps, thereby ensuring that the final output target hydrogeological parameter field is based on reliable cognition.

[0102] Step S43: Perform multi-scenario robustness verification on the second hydrogeological parameter field to obtain the predicted stability index of the second hydrogeological parameter field under various preset hydrogeological boundary conditions.

[0103] First, multiple hydrogeological boundary conditions are preset, which cover the reasonable range of changes that the target area may face in actual operation, such as different river boundary water levels, different rainfall infiltration recharge intensities, different pumping well flow rates, and different lateral inflow and outflow fluxes.

[0104] Next, for each hydrogeological boundary condition, the second hydrogeological parameter field is used as the only parameter input. The initial conditions, source and sink terms, and other settings of the numerical simulation (such as mesh generation, time step, and solver parameters) are kept completely unchanged. An independent forward modeling simulation of the physical process is run (for example, using the same numerical solution method as in step S122, such as the finite difference method or finite element method to solve the groundwater flow control equation) to obtain the physical field prediction data under each hydrogeological boundary condition.

[0105] Then, based on the physical field prediction data obtained from the above calculations, the prediction stability index of the second hydrogeological parameter field is calculated. For example, one or more target locations are selected (e.g., the location of typical observation wells within the target area, the location of important ecologically sensitive areas, or the average head of the entire simulation domain). The mean of all physical field prediction data is calculated, and the standard deviation of the dispersion of all physical field prediction data around the mean is calculated. The standard deviation is then divided by the mean to obtain the coefficient of variation of the physical field prediction data under all hydrogeological boundary conditions. The coefficients of variation of all selected target locations are then averaged or the maximum value is taken as the prediction stability index of the second hydrogeological parameter field. This prediction stability... The smaller the index, the more consistent and robust the response of the second hydrogeological parameter field to changes in different hydrogeological boundary conditions. Based on the prediction stability index, the fluctuation and reliability of the prediction results of the second hydrogeological parameter field when facing uncertain boundary conditions that may occur in the future are quantitatively evaluated. This effectively verifies whether the second hydrogeological parameter field has become vulnerable due to overfitting a single iterative condition, ensuring that the final output target hydrogeological parameter field not only performs well under the specific boundary conditions used in the iteration, but also can cope with various possible boundary condition changes in practical engineering applications, maintain stable prediction capabilities, and improve the robustness and credibility of the inversion results in actual decision support.

[0106] Step S44: Based on the predicted stability index, spatial smoothing and adaptation of the parameter values ​​in the second hydrogeological parameter field are performed, and the target hydrogeological parameter field is obtained.

[0107] An adaptive anisotropic diffusion filter is adopted, with the reciprocal of the prediction stability index as the control factor for smoothing intensity. When the prediction stability index of the second hydrogeological parameter field is high (i.e., the prediction is unstable), a stronger smoothing intensity is applied to suppress local parameter abnormal spikes or sawtooth abrupt changes caused by numerical optimization. When the prediction stability index of the second hydrogeological parameter field is low (i.e., the prediction is stable), a weaker smoothing intensity is applied to preserve the original hydrogeological structure characteristics. Thus, while eliminating unreasonable parameter value oscillations in the second hydrogeological parameter field, the hydrogeological structure information is preserved to the maximum extent.

[0108] Next, the parameter values ​​located at the boundary grid nodes or adjacent boundary cells in the second hydrogeological parameter field are compared with the physical compatibility requirements implied by the hydrogeological boundary condition types (such as constant head boundary, constant flow boundary, general head boundary, etc.). For example, for the permeability coefficient near the constant head boundary, it is necessary to ensure that the parameter value does not cause non-physical extreme values ​​in the calculated flux at the boundary. The parameter values ​​of all internal grid nodes not near the boundary in the second hydrogeological parameter field can be completely frozen. Only the parameter values ​​(such as permeability coefficient or water storage coefficient) on grid nodes a certain number of layers away from the boundary of the second hydrogeological parameter field (e.g., one or two grid layers away from the boundary) are set as adjustable variables, and their maximum adjustment range relative to the original parameter value is limited (e.g., not exceeding 10% of the original value) to ensure the small-amplitude nature of the correction. Based on the geological parameter field, a complete forward simulation of the physical process is run (the same forward simulation process as step S122). After calculating the head value (for constant head boundary) or flux value (for constant flux boundary) at all boundary nodes, for constant head boundary nodes, the square of the difference between the forward simulation head value and the preset head value is calculated. For constant flux boundary nodes, the square of the difference between the forward simulation flux value and the preset flux value is calculated. All squares are summed to obtain a scalar error value. Then, with the goal of minimizing the scalar error value, the parameter values ​​near the aforementioned adjustable boundary are iteratively adjusted. In each iteration, the parameter values ​​are updated, the forward simulation is rerun, and the total error is recalculated until the total error is reduced to below the preset tolerance threshold or the maximum number of iterations is reached.

[0109] Finally, after the above optimization and adjustment, the parameter values ​​near the boundary of the second hydrogeological parameter field were corrected to values ​​that can make the forward modeling results physically consistent with all hydrogeological boundary conditions, thereby obtaining the target hydrogeological parameter field that is consistent with each hydrogeological boundary condition.

[0110] While preserving the core structural characteristics of the second hydrogeological parameter field, it eliminates the local discontinuities and boundary inconsistencies that may be introduced by the numerical optimization process, making the second hydrogeological parameter field reliable and practical for direct application in actual water resource assessment, pollution prediction, or engineering geological analysis.

[0111] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A system for intelligent identification and inversion of hydrogeological parameters, characterized in that, include: The parameter field generation module is used to generate multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area. The scheme generation module is used to perform counterfactual inference on multiple candidate hydrogeological parameter fields and generate counterfactual observation schemes; The parameter field update module is used to perform confidence assessment and update each of the candidate hydrogeological parameter fields based on the execution observation data and each of the candidate hydrogeological parameter fields after obtaining the execution observation data on the target area by executing the counterfactual observation scheme, so as to obtain multiple updated candidate hydrogeological parameter fields. The target screening module is used to perform counterfactual deduction on each of the updated candidate hydrogeological parameter fields to obtain the target hydrogeological parameter field. The step of performing counterfactual deduction on multiple candidate hydrogeological parameter fields to generate a counterfactual observation scheme includes: Numerical simulations were performed based on multiple candidate hydrogeological parameter fields to obtain the physical field prediction responses of each candidate hydrogeological parameter field under simulation combinations of multiple preset observation locations and preset time series. Based on the predicted responses of the physical fields in the same simulation combination, the differences in the predicted responses between the candidate hydrogeological parameter fields are calculated to obtain the distinguishability index. After performing global optimization based on the distinguishability index to obtain the target observation location and target time series, the counterfactual observation scheme is generated based on the target observation location and the target time series. The step of performing counterfactual deduction on each of the updated candidate hydrogeological parameter fields to obtain the target hydrogeological parameter field includes: Based on the updated candidate hydrogeological parameter fields and cumulative confidence scores, several first hydrogeological parameter fields that satisfy the preset convergence criteria are obtained from the multiple updated candidate hydrogeological parameter fields. Based on the cumulative confidence level of each of the first hydrogeological parameter fields, a second hydrogeological parameter field is obtained by screening from the plurality of first hydrogeological parameter fields; Multi-scenario robustness verification of the second hydrogeological parameter field was conducted to obtain the predicted stability index of the second hydrogeological parameter field under various preset hydrogeological boundary conditions. Based on the predicted stability index, the parameter values ​​in the second hydrogeological parameter field are spatially smoothed and adapted to each of the preset hydrogeological boundary conditions to obtain the target hydrogeological parameter field.

2. The intelligent identification and inversion system for hydrogeological parameters according to claim 1, characterized in that, The step of generating and processing multiple competing candidate hydrogeological parameter fields based on prior geological information and initial observation data of the target area includes: The prior geological information is analyzed and processed to extract the geological structure prototype field that differs from the preset geological structure features; Based on the geological structure prototype field and the initial observation data, a collaborative data assimilation process is performed to generate multiple initial parameter fields; The initial parameter fields are subjected to adversarial differentiation enhancement to optimize the geological structure characteristics of each initial parameter field, thereby obtaining multiple candidate hydrogeological parameter fields.

3. The intelligent identification and inversion system for hydrogeological parameters according to claim 2, characterized in that, The step of generating multiple initial parameter fields by performing collaborative data assimilation processing based on the geological structure prototype field and the initial observation data includes: Based on the geological structure prototype field, multiple parameter field models that differ in their representation of the preset geological structure features are constructed. Numerical simulations are performed on each of the parameter field models based on the initial observation data to obtain the physical field prediction data of each of the parameter field models; The model data matching degree between the initial observation data and each of the physical field prediction data is calculated, and the difference between each of the parameter field models and the geological structure prototype field is calculated to obtain the model structure fidelity. Based on the model data matching degree and the model structure fidelity, a multi-objective collaborative optimization function is constructed, and then the parameter field models are updated through the multi-objective collaborative optimization function to obtain multiple initial parameter fields.

4. The intelligent identification and inversion system for hydrogeological parameters according to claim 1, characterized in that, The step of obtaining execution observation data on the target area by executing the counterfactual observation scheme includes: Based on the target observation location and target time series in the counterfactual observation scheme, a sensor network is deployed in the target area, and measured physical field data in the target area are collected through the sensor network; and, Based on the aforementioned counterfactual observation scheme, a high-fidelity numerical simulation is performed to calculate the simulated physical field data of the target area under the current hydrogeological conditions; The measured physical field data and the simulated physical field data are subjected to spatiotemporal registration and uncertainty quantization to generate fused physical field data; The fused feature vector is extracted from the fused physical field data, and the fused feature vector is combined to form the execution observation data output.

5. The intelligent identification and inversion system for hydrogeological parameters according to claim 4, characterized in that, The step of evaluating and updating the confidence level of each candidate hydrogeological parameter field based on the observed data and each candidate hydrogeological parameter field to obtain multiple updated candidate hydrogeological parameter fields includes: From the physical field prediction response of each of the candidate hydrogeological parameter fields, extract the prediction feature vector corresponding to the fused feature vector in the execution observation data; Based on the fused feature vector and each of the predicted feature vectors, the matching index of each of the candidate hydrogeological parameter fields is calculated; The confidence scores of each candidate hydrogeological parameter field in the current iteration are obtained by fusion calculation based on the matching indicators and the cumulative confidence scores in historical iterations. After eliminating candidate hydrogeological parameter fields corresponding to confidence scores lower than the preset confidence scores, the confidence scores corresponding to the remaining candidate hydrogeological parameter fields are subjected to Gaussian random perturbation, and the preset geological structure features among the remaining candidate hydrogeological parameter fields are cross-recombined to obtain the updated candidate hydrogeological parameter fields.

Images