Method for predicting redox capacity of aquifer based on lithology and conventional physicochemical parameters

Abstract

This invention relates to a method for predicting aquifer redox capacity based on lithology and conventional physicochemical parameters. Addressing the shortcomings of existing EEC (Earth Expansion Capacity) detection technologies, which require on-site sampling and laboratory testing, resulting in high costs, cumbersome operations, low batch testing efficiency, lack of estimation schemes for incomplete existing exploration data, and low utilization of exploration data, this invention uses the lithology and conventional physicochemical parameters of the medium sample with corresponding measured true values ​​as input, and the measured EEC as the training target. A prediction model is constructed through a machine learning regression algorithm. The model incorporates multi-level prediction modes to adapt to archived historical exploration data with varying degrees of completeness. Predicted redox capacity values ​​can be output simply by inputting existing archived lithology and physicochemical parameters. This invention provides an indirect measurement method that does not rely on new medium sampling or laboratory physicochemical testing, achieving rapid and low-cost redox capacity calculation. It is particularly suitable for the efficient assessment of the spatial distribution of redox capacity in large-scale, multi-depth, and large-sample-volume contaminated sites.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction technology of hydrogeological environment, specifically involving a method for predicting the redox capacity of aquifers based on lithology and conventional physicochemical parameters. Background Technology

[0002] Electron Exchange Capacity (EEC) of an aquifer refers to the total amount of electrons that all redox-active components (such as iron and manganese oxides, organic carbon, and sulfur-containing minerals) in an aquifer can provide and accept during redox reactions. It is a core indicator comprehensively characterizing the electron-donating and accepting capacity and redox buffering potential of an aquifer. Aquifers generally exhibit both structural and chemical heterogeneity. Spatial differences in lithology, pore structure, grain size composition, mineral type, and hydrogeochemical conditions jointly control the spatial distribution and evolution of the aquifer's EEC. Among these, lithology determines the occurrence state of primary minerals and the composition of particle size, serving as a fundamental geological factor regulating the content of redox-active components and electron exchange capacity. The EEC of an aquifer directly affects the migration and transformation processes of various pollutants in groundwater and is an indispensable basic parameter for conducting in-situ groundwater remediation, determining the dosage of remediation agents, and quantitatively assessing the effectiveness of contaminated site remediation.

[0003] Currently, the mature methods for quantitatively obtaining the redox capacity of aquifer media are all direct laboratory measurements, mainly divided into two categories: dielectric electrochemical testing and chemical probe spectrophotometric testing. The first category is dielectric electrochemical testing methods. For example, CN118883690A discloses a method for detecting the redox capacity of heterogeneous aquifer media, which uses a ruthenium-iridium-titanium MMO electrode to build a three-electrode testing system, quantitatively determining the electron donation and acceptance capacity of the medium through dielectric electrochemical reactions. While this approach broadens the range of applicable samples and improves the detection accuracy of single samples, the entire detection process requires high-value specialized equipment such as an electrochemical workstation, dedicated working electrodes, and an oxygen-free glove box. The operation process is complex, the single test cycle is long, and the efficiency of batch sample detection is low. The second category is chemical probe spectrophotometry. For example, CN118583847A discloses a rapid detection method for the redox capacity of heterogeneous aquifer media. This method uses ethyl viologen (EV) and 2,2'-adiazonium bis(3-ethylbenzothiazoline-6-sulfonic acid) diammonium salt (ABTS) as reaction probes, and calculates the redox capacity of the medium based on the change in absorbance of the solution before and after the reaction. This method does not require large-scale electrochemical equipment, but it requires multiple pretreatment steps before detection, including probe activation, reagent purification, and light-protected isothermal reaction. These manual steps are numerous and cannot meet the needs of simultaneous determination of large batches of site samples.

[0004] Both of the aforementioned direct measurement methods share common and unavoidable limitations: on-site drilling and sampling at the target site are necessary before testing, followed by transporting the samples to a laboratory for accompanying chemical or electrochemical experiments. For investigations and assessments of multi-level, large-scale contaminated sites, point-by-point sampling and testing significantly increases manpower, material costs, and time. Furthermore, the aquifer medium exhibits pronounced spatial heterogeneity and significant spatial differentiation of redox capacity; relying solely on discrete point measurements makes it difficult to comprehensively and continuously characterize the spatial distribution of regional redox capacity, thus posing a significant challenge to large-area site assessments.

[0005] Against the backdrop of the ongoing national survey of key industrial enterprises' land use and the Third National Soil Census, various regions across China have conducted site investigations in accordance with industry standards such as the "Technical Guidelines for Soil Pollution Status Investigation of Construction Land" (HJ 25.1-2019) and the "Technical Regulations for the Third National Soil Census," accumulating a massive historical database of aquifer media. During the investigation, lithological information of each stratum was preserved, along with the collection and recording of various conventional physicochemical parameters, covering multiple indicators such as porosity, density, particle size, iron, manganese, carbon, nitrogen, and sulfur mineral composition, pH, redox potential, electrical conductivity, and water content. Although these lithological and physicochemical indicators cannot be directly converted into EEC values, stable and quantifiable statistical correlations exist between them.

[0006] The existing testing system can only conduct laboratory tests on complete and complete media samples, which cannot be adapted to the data incompleteness that is common in exploration archives: for historical exploration points that only retain lithological information, lack all physical and chemical parameters, or only record some physical and chemical indicators, there is currently no mature and feasible redox capacity estimation scheme in the industry.

[0007] The existing industry technology system relies entirely on field sampling combined with laboratory measurements to obtain EEC values. This not only fails to fully exploit the existing archived lithological and physicochemical historical exploration data, but also fails to develop rapid prediction technology solutions that can adapt to exploration data with different levels of completeness. As a result, a large amount of collected and archived site survey data cannot be reused, and the overall utilization rate of exploration data resources is low. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting aquifer redox capacity (EEC) based on lithology and conventional physicochemical parameters. This method utilizes media samples with corresponding measured values ​​to establish a quantitative mapping relationship between lithology, conventional physicochemical parameters, and EEC through machine learning, and constructs a prediction model. It reuses existing site exploration historical data for prediction. This method solves the problems of existing EEC testing methods, which require on-site sampling and laboratory testing, involve high equipment investment, cumbersome procedures, low batch testing efficiency, lack of estimation methods for incomplete exploration data, and difficulty in evaluating large-scale sites. It eliminates the need for additional sampling and testing, quickly and accurately obtaining EEC and improving the utilization rate of existing exploration data.

[0009] To achieve the above-mentioned objectives, the technical solution adopted is as follows:

[0010] This invention discloses a method for predicting the redox capacity of aquifers based on lithology and conventional physicochemical parameters, comprising the following steps: S1, collecting multiple sets of detection data of aquifer media, each set of detection data containing the measured true values ​​of lithology, conventional physicochemical parameters, and redox capacity obtained simultaneously from the same medium; S2, using the lithology and conventional physicochemical parameters in the detection data as input features, and the measured true values ​​of redox capacity in the detection data as training targets, establishing a quantitative mapping relationship that can characterize the physicochemical correlation between lithology, conventional physicochemical parameters, and redox capacity through a machine learning algorithm, and constructing a redox capacity prediction model; the redox capacity prediction model has a built-in multi-level prediction mode adapted to different levels of completeness of historical test data; S3, acquiring historical test data of the aquifer media, and inputting the historical test data into the multi-level prediction mode matched by the redox capacity prediction model according to the completeness of the historical test data, and outputting the redox capacity prediction value.

[0011] Based on the above steps, this invention breaks away from the limitations of traditional on-site sampling and laboratory testing, transforming the acquisition of the redox capacity of aquifer media from complex chemical experiments into the conversion of conventional physicochemical parameters. It provides a brand-new indirect calculation method, eliminating the need for supporting professional testing equipment and chemical reagents, saving sample pretreatment and testing procedures, simplifying the overall evaluation process, reducing testing costs, and effectively improving the efficiency of batch evaluation. It is suitable for rapid evaluation of large-scale contaminated sites of different depths and large sample sizes. At the same time, it fully utilizes existing exploration data to achieve secondary development and utilization of historical exploration archives.

[0012] Further, in step S1, the lithology refers to soil texture type, which includes clay, clay loam, loam, sandy loam, sandy loam, silty clay, silty clay loam, and silty loam.

[0013] Further, in step S1, the conventional physicochemical parameters include structural parameters, texture component parameters, geochemical parameters, and hydrogeochemical parameters; the structural parameters include at least one of permeability coefficient, porosity, bulk density, and average particle size; the texture component parameters include at least one of clay content, silt content, and sand content; the geochemical parameters include at least one of iron content, manganese content, carbon content, nitrogen content, and sulfur content; and the hydrogeochemical parameters include at least one of pH value, redox potential, electrical conductivity, and water content.

[0014] Further, in step S2, the measured true value of the redox capacity corresponding to the training target is obtained by dielectric electrochemical testing; the dielectric electrochemical testing includes dielectric electrochemical oxidation testing and dielectric electrochemical reduction testing; the testing adopts a three-electrode system, constant potential chronoamperometry, and is carried out under anaerobic conditions; the mediators used include 2,2'-azinobis(3-ethylbenzothiazoline-6-sulfonic acid) diammonium salt and ethyl viologen.

[0015] Furthermore, the redox capacity prediction model is any one of random forest, gradient boosting decision tree, extreme gradient boosting, lightweight gradient boosting machine, categorical gradient boosting, and support vector machine, and is trained after optimization.

[0016] Furthermore, the redox capacity prediction model is a random forest; during the model training process in step S2, the random forest is used as the benchmark model, and at least one of the following strategies is used for optimization: grid search, random search, feature selection, outlier removal by interquartile range, and ensemble optimization. Training and evaluation are completed through 5-fold cross-validation.

[0017] By optimizing model parameters and structure, the prediction error can be effectively reduced, the fitting accuracy and generalization ability of the model can be improved, and the EEC prediction needs in complex aquifer environments can be adapted.

[0018] Furthermore, in step S3, before inputting the historical data to be tested into the prediction model, a step of preprocessing the conventional physicochemical parameters in the historical data to be tested is also included; the preprocessing includes a missing value imputation operation, and the missing value imputation adopts the median interpolation method based on the sampling point location and lithology.

[0019] Furthermore, the redox capacity prediction model incorporates three prediction modes adapted to different levels of completeness of historical test data: a first prediction mode, a second prediction mode, and a third prediction mode. The redox capacity prediction model automatically matches the corresponding prediction mode based on the completeness of the conventional physicochemical parameters of the input historical test data: the first prediction mode uses only lithology as an input feature and is suitable for historical test data where all conventional physicochemical parameters are missing; the second prediction mode uses one or more of the following as input features: structural parameters, texture component parameters, geochemical parameters, and hydrogeochemical parameters, and is suitable for historical test data where some conventional physicochemical parameters are missing; the third prediction mode uses all conventional physicochemical parameters as input features and is suitable for complete historical test data where no conventional physicochemical parameters are missing.

[0020] This invention sets up a three-level prediction mode, which can adapt to existing exploration archives with incomplete parameters, fill the gap in the lack of incomplete data estimation schemes in the existing technology, make full use of archived historical exploration data, and improve the efficiency of exploration data use.

[0021] Furthermore, the redox capacity prediction model has a built-in feature importance score module, which is used to quantify the contribution of each conventional physicochemical parameter to the redox capacity prediction. After the model is trained, the feature importance score corresponding to each conventional physicochemical parameter is output through the feature importance score module. In the prediction process of step S3, the conventional physicochemical parameters participating in the prediction are sorted from high to low according to the pre-stored feature importance scores.

[0022] This design facilitates the identification of the main physicochemical factors affecting redox capacity and can accurately quantify the contribution of each physicochemical parameter to the aquifer EEC. Based on achieving efficient prediction, it further reveals the intrinsic mechanism of spatial variation of EEC in heterogeneous aquifers and enhances the guidance value of this method for site engineering applications.

[0023] The beneficial effects of this invention are as follows:

[0024] Compared to existing EEC detection technologies, which heavily rely on on-site sampling and laboratory testing, resulting in high equipment and reagent costs, cumbersome sample pretreatment and testing procedures, low efficiency in batch sample testing, difficulty in fully characterizing the spatial distribution of redox capacity in aquifer media in heterogeneous sites through discrete point measurements, difficulty in reusing existing site survey data, and challenges in implementing large-scale site evaluation, this invention abandons the traditional testing method that relies solely on laboratory testing. Instead, it indirectly predicts the redox capacity of aquifer media through archived lithological and conventional physicochemical data from the site. This eliminates the need for specialized testing equipment and chemical reagents, removes all laboratory sample pretreatment and testing steps, simplifies the overall site evaluation process, reduces comprehensive testing costs, and effectively improves the efficiency of batch evaluation. It can meet the needs of rapid evaluation of large-scale contaminated sites and fully utilize existing survey data, thereby improving the utilization rate of site investigation resources.

[0025] The following detailed description of the aquifer redox capacity prediction method based on lithology and conventional physicochemical parameters, with reference to the embodiments in the accompanying drawings, further discloses the present invention. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the steps of the aquifer redox capacity prediction method based on lithology and conventional physicochemical parameters of the present invention.

[0027] Figure 2 Scatter plot comparing the prediction performance of various machine learning models on the redox capacity (EEC) of aquifer media; where a represents the prediction performance of the Random Forest (RF) model, b represents the prediction performance of the CatBoost model, c represents the prediction performance of the Lightweight Gradient Boosting Machine (LGBM) model, e represents the prediction performance of the Gradient Boosting Decision Tree (GBDT) model, f represents the prediction performance of the Support Vector Machine (SVM) model, and g represents the prediction performance of the Extreme Gradient Boosting (XGBoost) model.

[0028] Figure 3 This is a scatter plot comparing the prediction performance of random forest models with different optimization strategies for the redox capacity (EEC) of aquifer media in this invention; where a is the prediction performance of the baseline model, b is the prediction performance after grid search optimization, c is the prediction performance after random search optimization, e is the prediction performance after feature selection optimization, f is the prediction performance after outlier removal optimization, and g is the prediction performance after ensemble optimization.

[0029] Figure 4 This is a bar chart showing the ranking of feature importance in the random forest model based on outlier removal optimization according to the present invention.

[0030] Figure 5 This is a schematic diagram of the interface of the first prediction mode of the redox capacity prediction model of the present invention.

[0031] Figure 6 This is a schematic diagram of the interface of the structure parameter sub-model in the second prediction mode of the redox capacity prediction model of the present invention.

[0032] Figure 7 This is a schematic diagram of the interface of the geochemical parameter sub-model in the second prediction mode of the redox capacity prediction model of the present invention.

[0033] Figure 8 This is a schematic diagram of the interface of the hydrogeochemical parameter sub-model in the second prediction mode of the redox capacity prediction model of the present invention.

[0034] Figure 9 This is a schematic diagram of the interface of the texture component parameter sub-model in the second prediction mode of the redox capacity prediction model of the present invention.

[0035] Figure 10 This is a schematic diagram of the interface of the third prediction mode of the redox capacity prediction model of the present invention.

[0036] Figure 11 This is a spatial distribution map of EEC stratification obtained by applying the method of the present invention to two actual sedimentary porous aquifer sites. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with embodiments of this invention. Obviously, the described embodiments are one embodiment of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0038] like Figure 1 As shown, this invention discloses a method for predicting the redox capacity of aquifers based on lithology and conventional physicochemical parameters, comprising the following steps: S1, collecting multiple sets of detection data of aquifer media, each set of detection data containing the measured true values ​​of lithology, conventional physicochemical parameters, and redox capacity obtained simultaneously from the same medium; S2, using the lithology and conventional physicochemical parameters in the detection data as input features, and the measured true values ​​of redox capacity in the detection data as training targets, establishing a quantitative mapping relationship that can characterize the physicochemical correlation between lithology, conventional physicochemical parameters, and redox capacity through machine learning algorithms, and constructing a redox capacity prediction model; the redox capacity prediction model has a built-in multi-level prediction mode adapted to different levels of completeness of historical test data; S3, acquiring historical test data of the aquifer media, and inputting the historical test data into the multi-level prediction mode matched by the redox capacity prediction model according to the completeness of the historical test data, and outputting the redox capacity prediction value.

[0039] Traditional methods for assessing aquifer redox capacity require individual indoor electrochemical testing, relying on specialized equipment and chemical reagents. This results in lengthy and costly field sampling and testing cycles, and yields only single-point data, hindering large-scale site assessment. This new method trains a machine learning model using media samples that have undergone simultaneous testing. It can directly predict redox capacity using existing site survey archives, eliminating the need for repeated sampling and indoor measurements. Furthermore, it addresses the common problem of incomplete indicators in historical survey data by incorporating preprocessing techniques and employing a three-tiered prediction architecture to adapt to archives with varying levels of completeness. The model can also quantify the contribution of various physicochemical indicators to the prediction results, thus combining batch prediction with intrinsic mechanism analysis.

[0040] The redox capacity (EEC) of an aquifer essentially depends on the content and occurrence state of active redox components (such as iron and manganese oxides, organic carbon, sulfur-containing minerals, etc.) in the medium, and the distribution and enrichment of these components are controlled by the lithological type, grain size composition, pore structure and hydrogeochemical conditions of the medium.

[0041] Specifically: Lithology (soil texture type) determines the mineral composition and specific surface area of ​​the medium, thus affecting the adsorption and storage capacity of active components; structural parameters (porosity, permeability coefficient, average particle size, etc.) reflect the pore structure and connectivity of the medium, controlling the spatial accessibility and reaction kinetics of redox active components; texture parameters (clay, silt, and sand content) quantitatively describe particle size distribution, and together with lithology, determine the specific surface area and cation exchange capacity of the medium, affecting the enrichment degree of electron transport media; geochemical parameters (iron, manganese, carbon, nitrogen, and sulfur content) directly characterize the abundance of major redox active components in the medium. Iron and manganese oxides are the main carriers of electron acceptors, and organic carbon is the core source of electron donors; hydrogeochemical parameters (pH, redox potential, conductivity, and water content) reflect the chemical microenvironment of the medium, regulating the valence transformation of active components, mineral dissolution-precipitation equilibrium, and the direction of redox reactions.

[0042] The aforementioned lithology and four types of conventional physicochemical parameters jointly characterize the redox activity of the aquifer medium from different dimensions. There are quantifiable synergistic and restrictive coupling relationships among various indicators. Therefore, by using them as input features and EEC measured values ​​as training targets, a quantitative mapping model reflecting the intrinsic physicochemical relationship between the two can be established through machine learning algorithms.

[0043] During step S1, when collecting samples and conducting simultaneous testing, the soil texture types of the sample media were classified into eight categories: clay, clay loam, loam, sandy loam, sandy loam, silty clay, silty clay loam, and silty loam. The conventional physicochemical parameters obtained from the simultaneous testing were divided into four categories: the first category was structural parameters, with selectable indicators including permeability coefficient, porosity, bulk density, and average particle size; the second category was texture component parameters, with selectable indicators including clay content, silt content, and sand content; the third category was geochemical parameters, with selectable indicators including iron content, manganese content, carbon content, nitrogen content, and sulfur content; and the fourth category was hydrogeochemical parameters, with selectable indicators including pH value, redox potential, electrical conductivity, and water content. All samples collected in step S1 underwent simultaneous testing of lithology and all the above-mentioned physicochemical indicators, and the dataset used for model training had no missing indicators.

[0044] The true value of the redox capacity obtained in step S1 was detected by the dielectric electrochemical method. The entire test system was set up with a three-electrode device. In an oxygen-free environment, the dielectric electrochemical oxidation test and dielectric electrochemical reduction test were carried out simultaneously using the constant potential chronoamperometry method. The mediators used in the experiment were 2,2'-adiazonium bis(3-ethylbenzothiazoline-6-sulfonic acid) diammonium salt and ethyl viologen.

[0045] In step S2, when building the redox capacity prediction model, the available machine learning algorithms include random forest, gradient boosting decision tree, extreme gradient boosting, lightweight gradient boosting machine, categorical gradient boosting, and support vector machine. All algorithms are optimized before being used for training. In this embodiment, a total of 214 complete samples from two sites are used as the training dataset. For the six algorithms, the number of decision trees is uniformly set to 100 and the random seed to 42. Five-fold cross-validation is used to compare performance.

[0046] Figure 2 A scatter plot comparing the prediction performance of six machine learning models is presented. In the plot, a corresponds to Random Forest, b to Class Gradient Boosting, c to Lightweight Gradient Boosting Machine, e to Gradient Boosting Decision Tree, f to Support Vector Machine, and g to Limiting Gradient Boosting. The horizontal axis represents the measured redox capacity, the vertical axis represents the model's predicted value, and the diagonal dashed line represents the ideal fitting curve. The scatter plot distribution shows that Random Forest's sample points are closest to the diagonal, resulting in smaller prediction bias. Quantitative evaluation metrics for each model are shown in Table 1.

[0047] Table 1. Performance Comparison of Five-Fold Cross-Validation for Six Machine Learning Models

[0048]

[0049] As shown in Table 1, the comprehensive coefficient of determination (R²) 2Of the three indicators—root mean square error (RMSE), mean absolute error (MAE)—random forest has the best overall prediction accuracy and stability. Therefore, this embodiment selects random forest as the basic prediction model.

[0050] The random forest regression model used in this embodiment has the following key hyperparameter settings: the number of decision trees (n_estimators) is set to 100; the maximum depth of a single decision tree (max_depth) is set to 10, in order to control...

[0051] To control model complexity and prevent overfitting, the minimum number of samples required for internal node re-split (min_samples_split) is set to 5; the minimum number of samples required for leaf nodes (min_samples_leaf) is set to 2; the mean squared error (MSE) is used as the node splitting criterion; the size of the randomly selected feature subset (max_features) during each node split is set to the square root of the total number of input features (i.e., sqrt(n_features)); out-of-bag (OOB) evaluation is enabled. The above hyperparameter values ​​are obtained by traversing preset candidate intervals (number of decision trees 50~500, maximum depth 3~20, minimum number of samples required for internal node re-split 2~10, etc.) using a grid search method, with the optimization objective of minimizing the root mean squared error (RMSE) under five-fold cross-validation.

[0052] After determining random forest as the base model, this embodiment sets up six differentiated optimization schemes: no optimization baseline model, grid search optimization, random search optimization, feature selection optimization, interquartile range optimization, and so on.

[0053] Constant values ​​were eliminated and integrated optimization was performed. All schemes were evaluated using five-fold cross-validation to assess their predictive performance. Figure 3 The table shows the scatter plots of predictions for the random forest model under different optimization strategies, where a represents the baseline model, b represents grid search optimization, c represents random search optimization, e represents feature selection optimization, f represents outlier removal optimization, and g represents ensemble optimization. The diagonal dashed lines represent the ideal fit curves. Quantitative performance data for each optimization strategy are shown in Table 2.

[0054] Table 2 Performance Comparison of Different Optimization Strategies for Random Forest Model

[0055]

[0056] Table 2 shows that the coefficient of determination (R²) of the integrated optimization scheme 2The accuracy is slightly higher, but multiple parameter interventions can easily cause the model to overfit the local features of the sample. The interquartile range method removes outliers only by removing extreme outliers, preserving the natural physical and chemical distribution patterns of the site medium to the greatest extent, and has stronger engineering applicability. Therefore, this embodiment uses the random forest optimized by outlier removal as the final prediction model.

[0057] After model training is completed, the redox capacity prediction model incorporates a feature importance score module. During the full training phase of the model, this module calculates and stores the feature importance score corresponding to each parameter based on all input conventional physicochemical parameters. When performing the prediction operation in step S3, the model reads the pre-stored scores and sorts the physicochemical parameters that actually participated in the calculation in the current historical test data from highest to lowest according to their contribution.

[0058] Figure 4 This is a bar chart showing the feature importance ranking obtained from the outlier removal-optimized random forest in this embodiment. Porosity, average particle size, and electrical conductivity rank as the top three in importance, contributing over 56% to the overall score, making them the core factors controlling redox capacity. Nitrogen content, iron content, and redox potential play a synergistic regulatory role, while permeability coefficient, water content, and clay content are secondary influencing factors. The importance values ​​of various physicochemical indicators only represent the relative explanatory power of the prediction results and do not represent the direct causal control effect of a single indicator on redox capacity.

[0059] Before performing the prediction operation in step S3, the input historical aquifer data needs to be preprocessed. This preprocessing includes missing value imputation, which uses a stratified grouping median interpolation method based on sampling location and lithology. The historical data consists of existing site survey archives, and due to limitations in past sampling conditions, physicochemical indicators are generally missing. In contrast, the modeling samples in step S1 are synchronous and complete detection data, requiring no missing value imputation. The execution logic for stratified grouping median interpolation is as follows: missing items are filled primarily using the median of the same type of indicator within the same sampling location; if no valid samples are available at the same location, lithology is used as the grouping basis, and the median of all samples within that lithology is used to complete the imputation.

[0060] To more clearly illustrate the specific operation of the stratified grouped median interpolation method, an interpolation example is given below. Assume that the lithology of a certain historical exploration point (point number A-12) is clay, and the recorded porosity (0.42) and unit weight (1.85 g / cm³) are... 3The data included indicators such as iron content (0.63%) and pH value (6.8), but lacked data on permeability coefficient (m / d) and carbon content. The interpolation process was as follows: First, it was checked whether there were valid values ​​for permeability coefficient or carbon content recorded in the same sampling point (A-12). If not, the median of permeability coefficient and carbon content (assumed to be 0.15 m / d and 0.82%, respectively) of all clay lithology samples was extracted using "lithology = clay" as the grouping condition, and the missing items were filled with this median. If the number of valid samples in the same lithology group was too small and not statistically representative, the grouping condition was further relaxed to the next higher lithology category (including clay, silty clay, etc.), and the median of the samples in that category was used to complete the filling. This interpolation strategy ensured that the filled values ​​were all derived from the actual test data distribution of the same type of medium, avoiding the systematic bias introduced by blindly filling across lithologies, and preserving the original physiological and chemical distribution characteristics of the site medium to the greatest extent.

[0061] To verify that the filling method does not distort the original physicochemical distribution of the medium, statistics were conducted on the mean values ​​of each index before and after interpolation, the range of mean value changes, and the number of samples with out-of-bounds filling. The statistical results are shown in Table 3.

[0062] Table 3 Comparison of statistical characteristics of missing physicochemical properties of aquifer media before and after imputation.

[0063]

[0064] Note: The permeability coefficient is only used as a grouping criterion for stratified grouping median interpolation and is not involved in redox capacity modeling. Therefore, this table does not include its original data, the mean after interpolation, and the rate of change of the mean. The values ​​of all indicators after interpolation do not exceed the natural range of the medium, and there are no out-of-bounds interpolation samples.

[0065] As shown in Table 3, after interpolation using the median of stratified groupings, no out-of-bounds interpolation samples were found for any of the physicochemical indicators, and the filled values ​​all conformed to the natural range of aquifer media indicators. The permeability coefficient was only used as a statistical reference for grouping. The mean values ​​of porosity, bulk density, Fe, Mn, H, and S all varied by less than 5%, with only N and C showing the highest mean variation of 11.7% due to missing local samples. Overall, the data disturbance was controllable. This filling logic can completely supplement incomplete exploration data, adapt to historical archives with varying degrees of indicator completeness, and preserve the original physical and geochemical distribution characteristics of the site's media to the greatest extent possible, ensuring the accuracy and reliability of the prediction results.

[0066] The redox capacity prediction model incorporates three prediction modes. The model can automatically match the corresponding mode based on the completeness of conventional physicochemical parameters in the input historical test data. The first prediction mode only requires inputting lithology and is suitable for historical test data lacking all conventional physicochemical parameters; the second prediction mode requires inputting one or more types of conventional physicochemical parameters and is suitable for historical test data lacking some indicators; the third prediction mode requires inputting all conventional physicochemical parameters and is suitable for historical test data with complete and unmissing indicators.

[0067] The first prediction mode is a rapid prediction mode, which can complete the preliminary estimation of redox capacity based solely on the lithology of the medium, without requiring any physicochemical index testing. It is suitable for rapid screening scenarios where only the soil classification of the site is known and no indoor testing data is available. To intuitively distinguish the redox capacity baseline levels of different texture media, the statistical data of the mean EEC, quartiles, and fluctuation ranges corresponding to the eight lithologies are shown in Table 4.

[0068] Note: The lithology codes in this table are abbreviations of soil texture in English. Si represents silt, C represents clay, and L represents loam; CL = clay, CLm = clay loam, L = loam, LS = sandy loam, SL = sandy loam, SiC = silty clay, SiCL = silty clay loam, and SiL = silty loam.

[0069] Table 4. EEC Statistics and Fluctuation Characteristics of Aquifer Media with Different Lithologies

[0070]

[0071] As shown in Table 4, the EEC values ​​of clay and silty clay, two types of fine-grained media, are the highest among all soil types, while the EEC values ​​of loam and sandy loam, two types of coarse-grained media, are the lowest overall. The redox buffering capacity of different texture media is significantly different. A preliminary assessment of a large-scale site can be completed based solely on lithological classification, completely eliminating the need for sampling and a full set of electrochemical testing procedures.

[0072] The second prediction mode is a sub-item adaptive prediction mode, which independently constructs four sets of sub-item sub-models according to the four major physicochemical parameter categories. The corresponding sub-model can be flexibly selected according to the indicators already detected on site, eliminating the need to conduct full indicator testing and significantly reducing the workload of indoor experiments. The optimal algorithm and prediction performance of the four types of sub-item sub-models are shown in Table 5.

[0073] Table 5. Performance Comparison of Five-Fold Cross-Validation for the Four Types of Sub-Models

[0074]

[0075] As shown in Table 5, the sub-model based on structural parameters has the best prediction accuracy and can be used only for site physicochemical data, thus reducing the workload of testing.

[0076] The third prediction mode is the precise prediction mode, which inputs all four types of conventional physicochemical parameters and obtains the highest accuracy prediction results through complete data. After the prediction is completed, the model simultaneously outputs the contribution ranking of the physicochemical parameters involved in the prediction, as well as the R², RMSE, and MAE evaluation indicators. The prediction errors of the three modes show a gradient distribution (error values ​​are based on statistics from 214 sedimentary porous aquifer samples in this embodiment): the third precise mode has the lowest average error, while the first mode has a larger error fluctuation range. Engineers can conduct batch redox capacity assessments based on the completeness of indicators in historical exploration archives and the corresponding interfaces matched to the site's exploration stage.

[0077] The samples used to train the model in this embodiment were taken from two sedimentary porous aquifers, covering eight soil textures. The model is suitable for sedimentary porous aquifers. However, direct use of the model for karst, fractured bedrock, high-salt, and high-organic-matter aquifers may lead to prediction bias. When applying the model to other sites, it is necessary to supplement the model with local samples for recalibration.

[0078] This method integrates a complete implementation process, and the entire solution can complete batch prediction of redox capacity using existing survey archives without the need for repeated sampling and complex electrochemical experiments. It is equipped with missing data filling methods and a three-level adaptive prediction architecture, and has a built-in feature importance module to quantify the contribution of each indicator, taking into account the dual value of batch engineering assessment and aquifer heterogeneity mechanism analysis.

[0079] like Figures 5-10 As shown, the redox capacity prediction model of this invention has a built-in three-level adaptive prediction mode. Different modes correspond to independent operation interfaces. Based on the completeness of indicators in the existing historical survey archive data, the practical prediction process of each mode is as follows:

[0080] Figure 5 This is the user interface for the first prediction mode (Level 1 model). Only the lithological information of the medium needs to be entered to complete a rapid estimation. In this example, after selecting SiC-silty clay from the drop-down menu and clicking "Start Prediction," the interface outputs a predicted redox capacity value of 0.4166 mmol e. - / g, the interface is labeled as a first-level model (lithological median method), which is suitable for rapid screening scenarios where historical exploration archives only record soil classification and have no physical and chemical index records.

[0081] Figures 6 to 9 This is the user interface for the four sub-models corresponding to the second prediction mode. Predictions can be made independently based on single-category physicochemical indicators preserved in historical exploration archives.

[0082] Figure 6 The interface for the structural parameter sub-model is as follows: input permeability coefficient 1e-5 cm / s, porosity 35%, bulk density 15 kN / m³, and average particle size 70 μm. After clicking "predict," the output result is 0.2618 mmol e.- / g, the model type is labeled as a level 2 model;

[0083] Figure 7 For the geochemical parameter sub-model interface, the following parameters were input sequentially: Fe content 1.2 mmol / g, Mn content 0.2 mmol / g, N content 1.3 mmol / g, C content 1.5 mmol / g, and S content 0.05 mmol / g. The predicted redox capacity was 0.2551 mmol e⁻ / g.

[0084] Figure 8 For the hydrogeochemical parameter sub-model interface, input parameters include pH 7.12, redox potential Eh = 130 mV, water content 20%, and conductivity 100 μS / m. The output predicted value is 0.1731 mmol e. - / g;

[0085] Figure 9 For the texture component parameter sub-model interface, inputting clay content of 30%, silt content of 40%, and sand content of 30%, the predicted result is 0.2373 mmol e. - / g;

[0086] Figure 10 This is the interface for the third prediction mode (precise prediction mode). All four categories of conventional physicochemical parameters from the historical exploration archives must be completely entered. The parameters to be entered this time are: Fe content 0.5 mmol / g, Mn content 0.1 mmol / g, C content 2.0 mmol / g, N content 1.3 mmol / g, S content 0.05 mmol / g, permeability coefficient 1e-6 cm / s, porosity 20%, and bulk density 15 kN / m³. 3 The parameters are: average particle size 18 μm, redox potential Eh = 110 mV, pH 7.1, moisture content 25%, electrical conductivity 500 μS / m, clay content 30%, silt content 40%, and sand content 30%. After clicking "Start Precise Prediction," the model outputs a high-precision prediction result of 0.5113 mmol e. - / g, the interface indicates the model type as a three-level ensemble prediction model. This mode achieves optimal prediction accuracy based on complete archival data. After prediction is completed, the contribution ranking of the physicochemical parameters involved in the calculation and R are output simultaneously. 2 Multiple model evaluation indicators, including RMSE and MAE.

[0087] By using the three-level adaptive prediction model of this invention to batch process the existing historical survey archives of two sites, continuous spatial distribution results of redox capacity of multi-layer aquifers across the entire region can be generated, such as... Figure 11 As shown. Site 1 has an area of ​​0.45 km². 2The strata are covered at depths of 0–5 m, divided into five layers: 0–1 m, 1–2 m, 2–3 m, 3–4 m, and 4–5 m; the site area is 0.65 km². 2 The strata cover a depth of 0–8 m, divided into six layers: 0–1 m, 1–2 m, 2–3 m, 3–4 m, 4–5 m, and 5–8 m. Based on existing archives, EEC prediction results for each location are obtained in batches. A spatial interpolation algorithm is used to draw a layered isopleth distribution map. The map includes a scale to characterize the EEC values ​​of different regions, with varying gray levels corresponding to redox capacities, thus comprehensively depicting the spatial differentiation characteristics of the redox capacity in the plane and vertical directions. Compared to traditional evaluation methods that rely on point-by-point drilling and sampling or laboratory electrochemical measurements to obtain only discrete single-point data, this approach eliminates the need for additional field sampling and indoor electrochemical experiments. It can output continuous EEC spatial distribution results across the entire area and multiple layers in a single operation, significantly reducing the manpower, material, and time costs of evaluating large-area aquifer sites and achieving efficient batch site assessment.

[0088] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0089] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting the redox capacity of aquifers based on lithology and conventional physicochemical parameters, characterized in that, Includes the following steps: S1. Collect multiple sets of detection data for aquifer media. Each set of detection data includes the measured true values ​​of lithology, conventional physicochemical parameters and redox capacity obtained by synchronous detection of the same medium. S2. Using the lithology and conventional physicochemical parameters in the test data as input features, and the measured true value of the redox capacity in the test data as the training target, a quantitative mapping relationship that can characterize the physicochemical correlation between lithology, conventional physicochemical parameters and redox capacity is established through machine learning algorithms, and a redox capacity prediction model is constructed; the redox capacity prediction model has a built-in multi-level prediction mode that adapts to different levels of completeness of historical test data. S3. Obtain historical test data of the aquifer medium. Based on the completeness of the historical test data, input the historical test data into the multi-level prediction mode matched by the redox capacity prediction model, and output the redox capacity prediction value.

2. The method according to claim 1, characterized in that, In step S1, the lithology refers to soil texture type, which includes clay, clay loam, loam, sandy loam, sandy loam, silty clay, silty clay loam, and silty loam.

3. The method according to claim 1, characterized in that, In step S1, the conventional physicochemical parameters include structural parameters, texture component parameters, geochemical parameters, and hydrogeochemical parameters; the structural parameters include at least one of permeability coefficient, porosity, bulk density, and average particle size; the texture component parameters include at least one of clay content, silt content, and sand content; the geochemical parameters include at least one of iron content, manganese content, carbon content, nitrogen content, and sulfur content; and the hydrogeochemical parameters include at least one of pH value, redox potential, electrical conductivity, and water content.

4. The method according to claim 1, characterized in that, In step S2, the true value of the redox capacity corresponding to the training target is obtained by dielectric electrochemical testing; the dielectric electrochemical testing includes dielectric electrochemical oxidation testing and dielectric electrochemical reduction testing; the testing adopts a three-electrode system, constant potential chronoamperometry, and is carried out under anaerobic conditions; the mediators used include 2,2'-azinobis(3-ethylbenzothiazoline-6-sulfonic acid) diammonium salt and ethyl viologen.

5. The method according to claim 1, characterized in that, The redox capacity prediction model is any one of random forest, gradient boosting decision tree, extreme gradient boosting, lightweight gradient boosting machine, categorical gradient boosting, and support vector machine, and is trained after optimization.

6. The method according to claim 5, characterized in that, The redox capacity prediction model is a random forest. In the model training process of step S2, the random forest is used as the benchmark model, and at least one of the following strategies is used for optimization: grid search, random search, feature selection, outlier removal by interquartile range, and ensemble optimization. The training and evaluation are completed through 5-fold cross-validation.

7. The method according to claim 1, characterized in that, In step S3, before inputting the historical data to be measured into the prediction model, a step of preprocessing the conventional physicochemical parameters in the historical data to be measured is also included; the preprocessing includes missing value imputation operation, and the missing value imputation adopts the median interpolation method based on the sampling point location and lithology.

8. The method according to claim 1, characterized in that, The redox capacity prediction model incorporates three prediction modes adapted to different levels of completeness of historical test data: a first prediction mode, a second prediction mode, and a third prediction mode. The redox capacity prediction model automatically matches the corresponding prediction mode based on the completeness of the conventional physicochemical parameters of the input historical test data. The first prediction model uses only lithology as an input feature and is suitable for historical test data where all conventional physicochemical parameters are missing. The second prediction model uses one or more of the following as input features: structural parameters, texture component parameters, geochemical parameters, and hydrogeochemical parameters. It is suitable for historical data that lacks some conventional physicochemical parameters. The third prediction mode uses all conventional physicochemical parameters as input features and is suitable for complete historical test data with no missing conventional physicochemical parameters.

9. The method according to claim 8, characterized in that, The redox capacity prediction model has a built-in feature importance score module, which is used to quantify the contribution of each conventional physicochemical parameter to the redox capacity prediction. After the model is trained, the feature importance score corresponding to each conventional physicochemical parameter is output through the feature importance score module. In step S3, during the prediction process, the conventional physicochemical parameters participating in this prediction are sorted from high to low according to their contribution based on the pre-stored feature importance scores.

Images