Model for predicting rate constant of non-metallic material for catalyzing persulfate to degrade pollutants based on machine learning and application

Abstract

The invention discloses a model for predicting a rate constant of a non-metallic material for catalyzing persulfate to degrade pollutants based on machine learning and application, and the model is constructed for the degradation rate of the non-metallic material for catalyzing the persulfate to degrade the pollutants by comprehensively adopting a simple and complex integrated learning machine learning model. Different model evaluation indexes are calculated respectively, the model with the optimal performance is selected according to R2 and RMSE values to serve as a final model for pollutant migration and transformation prediction, the model can achieve rapid prediction of the pollutant degradation rate, and the model effect is good.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and more specifically, to a model and application for predicting the rate constant of persulfate degradation of pollutants by non-metallic materials based on machine learning. Background Art

[0002] In recent years, with the development of industrialization, various pollutants are inevitably discharged into the aquatic environment, resulting in pollution of aquatic ecosystems and threatening human survival. The degradation of pollutants is a complex and important process, which requires comprehensive consideration of multiple factors and effective measures to accelerate the degradation process and reduce environmental pollution, which is crucial for environmental protection and pollution control.

[0003] Currently, advanced oxidation pathways (AOPs) based on persulfate (PS) catalyzed by non-metallic carbon-based materials (MFCMs) are a promising method for transforming pollutants in water. However, due to the different structural and surface characteristics of different MFCMs, it is still challenging and controversial to determine their inherent catalytic sites, and their different abilities to affect the migration and transformation of pollutants need further study. In addition, during the PS activation process, MFCMs tend to mediate the singlet oxygen ( 1 O2) oxidation, but few studies have reported selective production 1 The above problems directly affect the application of MF CMs catalyzing PS to treat pollutants in water.

[0004] Machine learning (ML) is an emerging data-driven approach that can process and learn large-scale, complex, and high-dimensional data to develop predictive models. In recent years, ML has been widely used in environmental remediation, but it has not involved the prediction of the efficiency of MFCMs in catalyzing persulfate PS degradation of pollutants. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a model and application based on machine learning for predicting the rate constant of persulfate degradation of pollutants by non-metallic materials.

[0006] The first object of the present invention is to provide a model based on a machine learning model for predicting the rate constants of persulfate degradation of pollutants catalyzed by non-metallic materials.

[0007] The second object of the present invention is to provide an application of any of the above models in predicting the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

[0008] The third object of the present invention is to provide a method for predicting the rate constant of persulfate degradation of pollutants catalyzed by non-metallic materials based on a machine learning model.

[0009] A fourth object of the present invention is to provide a system for predicting the rate constants of persulfate degradation of pollutants catalyzed by non-metallic materials.

[0010] A fifth object of the present invention is to provide a computer device.

[0011] A sixth object of the present invention is to provide a computer-readable storage medium.

[0012] In order to achieve the above object, the present invention is implemented by the following technical solutions:

[0013] The present invention firstly claims a model for predicting the rate constant of pollutant degradation by persulfate catalyzed by non-metallic materials based on a machine learning model, wherein the model uses the data value of relevant features as input values ​​and the rate constant of pollutant degradation as output values, and is obtained by training the machine learning model.

[0014] The related features include the following three types:

[0015] (1) Catalyst properties: pyrolysis temperature, carbon content, oxygen content, hydrogen content, nitrogen content, ratio of nitrogen to carbon, ratio of oxygen to carbon, pyridinic N content, pyrrolic N content, graphitic N content, N oxide content, specific surface area, total pore volume and ratio of Raman D peak intensity to G peak intensity.

[0016] (2) Experimental conditions: catalyst concentration, PMS concentration, PDS concentration, reaction temperature, pH and pollutant concentration,

[0017] (3) Properties of the pollutants: molecular weight, pKa, highest occupied molecular orbital, lowest unoccupied molecular orbital, and molecular energy gap;

[0018] The machine learning model is one or more of a K nearest neighbor model, a decision tree model, a random forest model, a LightGBM model, an XGBoost model, a gradient boosting model, a CatBoost model or an extreme random forest model.

[0019] Preferably, the data values ​​of the relevant features are used as input values ​​for machine learning model training after missing data is filled.

[0020] Preferably, the data values ​​of the features are normalized and then used as input values ​​for machine learning model training.

[0021] Preferably, the specific method of filling the missing data is:

[0022] S1. For features with missing data less than or equal to 10% of the total number of samples, the fill value is the mean of all non-missing data of the feature;

[0023] S2. For features where the amount of missing data is greater than 10% of the total sample size, fill in the missing data based on the results of the Pearson correlation analysis:

[0024] S21. Select simulation features:

[0025] Input features without missing values ​​(including input features that have completed missing value filling using the mean method) are selected as candidate features. At the same time, candidate features that meet the following two points are selected. At least 3 candidate features with the highest correlation (largest absolute value of correlation coefficient) are selected as simulated features, and their data values ​​are simulated variables:

[0026] (a) The candidate features have the same category as the input features whose missing values ​​need to be supplemented;

[0027] (b) The candidate features are highly correlated with the data values ​​of the features that need to supplement the missing values;

[0028] S22. Use the algorithm to establish the mapping relationship between the simulated variables and the missing variables, establish the model, and use the obtained model to fill the missing data values ​​of the input features with missing values:

[0029] For all data samples: the data values ​​of the simulated variables of the samples whose missing variable values ​​are non-missing values ​​are taken as x_train, the data values ​​of the missing variables whose missing variable values ​​are non-missing values ​​are taken as y_train, the data values ​​of the simulated variables of the samples whose missing variable values ​​are missing values ​​are taken as x_test, and the data values ​​of the missing variables whose missing variable values ​​are missing values ​​are taken as y_test.

[0030] By using x_train as the training input and y_train as the training output, the random forest regression algorithm is used to train the model to obtain the trained model. x_test is input into the trained model for prediction, and the predicted result y_test is used to fill the missing values.

[0031] Preferably, Pearson correlation analysis is used, and an absolute value of a correlation coefficient above 0.8 can be considered to be highly correlated, between 0.5 and 0.8 is moderately correlated, between 0.3 and 0.5 is lowly correlated, and less than 0.3 is considered to be uncorrelated.

[0032] Preferably, the data values ​​of the candidate features and the features whose missing values ​​need to be supplemented are analyzed using Pearson correlation, and the absolute value of the correlation coefficient is greater than or equal to 0.3.

[0033] Preferably, in the training, the ratio of the number of samples in the training set to that in the test set is 7:3.

[0034] As a specific implementation scheme, the data values ​​of the features of the 700 samples in the following table are used for training:

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Among them, NAN is missing data. The catalyst properties and pollutant properties of each data sample can be found in the original literature.

[0046] Tp, C, O, H, N, Tr and pH are input features with a low degree of missingness. The missing values ​​are filled using the mean method, that is, the filling value is the mean of all non-missing data of the input feature.

[0047] Pyd-N, Pyr-N, Gph-N, N-oxi, Id / Ig, SSA, TPV and pKa are input features with missing values ​​and input features with a high degree of missingness. The corresponding relationship between the input features with a high degree of missingness and the corresponding simulation variables is shown in the following table.

[0048]

[0049] Preferably, the parameters of the extreme random forest model are: n_estimators=202, max_depth=14, min_samples_leaf=1, min_samples_split=2, max_features='sqrt', random_state=42.

[0050] Preferably, the parameters of the K nearest neighbor model are: algorithm='kd_tree', metric='manhattan', n_neighbors=3, weights='uniform'.

[0051] Preferably, the parameters of the decision tree model are: max_depth = None, min_samples_leaf = 1, min_samples_split = 3, max_features = 'sqrt', random_state = 42

[0052] Preferably, the parameters of the random forest model are: n_estimators=114, max_depth=17, min_samples_leaf=1, min_samples_split=2, max_features=0.1, random_state=42.

[0053] Preferably, the parameters of the LightGBM model are: n_estimators=145, num_leaves=9, max_depth=8, learning_rate=0.02, colsample_bytree=0.7000000000000002, subsample=0.1, reg_alpha=0, reg_lambda=10, random_state=42.

[0054] Preferably, the parameters of the XGBoost model are: n_estimators=194, learning_rate=0.03, max_depth=11, min_child_weight=1, colsample_bytree=0.79999999999999999, subsample=0.89999999999999999, reg_alpha=0.1, reg_lambda=0, gamma=0, random_state=42.

[0055] Preferably, the parameters of the gradient boosting model are: n_estimators=100, max_depth=9, learning_rate=0.02, min_samples_split=2, min_samples_leaf=1,

[0056] max_features=0.30000000000000004, alpha=0.1, random_state=42.

[0057] Preferably, the parameters of the CatBoost model are: learning_rate=0.03, iterations=100, depth=12, l2_leaf_reg=3, bagging_temperature=0, random_strength=1, loss_function='RMSE', random_state=42.

[0058] The present invention also claims the use of any of the above models in predicting the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

[0059] The present invention also claims protection for the following product: a method for predicting the rate constant of pollutant degradation by persulfate catalyzed by non-metallic materials based on a machine learning model, using any of the machine learning-based models to predict the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

[0060] A system for predicting the rate constant of persulfate degradation of pollutants by non-metallic materials, comprising an acquisition unit, a storage unit and a processing unit, wherein the acquisition unit is used to acquire the numerical value of the characteristic of the non-metallic material sample to be predicted,

[0061] The features include the following three types:

[0062] (1) Catalyst properties: pyrolysis temperature, carbon content, oxygen content, hydrogen content, nitrogen content, ratio of nitrogen to carbon, ratio of oxygen to carbon, pyridinic N content, pyrrolic N content, graphitic N content, N oxide content, specific surface area, total pore volume and ratio of Raman D peak intensity to G peak intensity.

[0063] (2) Experimental conditions: catalyst concentration, PMS concentration, PDS concentration, reaction temperature, pH and pollutant concentration,

[0064] (3) Properties of the pollutants: molecular weight, pKa, highest occupied molecular orbital, lowest unoccupied molecular orbital, and molecular energy gap;

[0065] The storage unit stores program instructions that can be executed by the processing unit, and the program instructions input the numerical value of the feature of the non-metallic material sample to be predicted obtained by the obtaining unit into the processing unit;

[0066] The processing unit includes any of the above models and executes the program instructions to obtain the numerical value of the characteristic of the non-metallic material sample to be predicted as a result.

[0067] A computer device comprises a memory and a processor, wherein the memory stores a computer program executable on the processor;

[0068] When the computer program is executed by the processor, the operation of the system is realized.

[0069] A computer-readable storage medium stores a computer program that can be executed by a processor. When the computer program is executed by the processor, the operation of the system is realized.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] The present invention provides a prediction model for the degradation rate of pollutants based on machine learning, which comprehensively adopts simple and complex integrated learning machine learning models to construct a model for the degradation rate of pollutants catalyzed by persulfate of non-metallic materials, and calculates different model evaluation indicators respectively. 2 The model with the best performance was selected based on the RMSE value as the final model for pollutant migration and transformation prediction. This model can quickly predict the degradation rate of pollutants and has a good effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Flowchart for constructing a machine learning-based model for predicting the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

[0073] Figure 2 Pearson correlation analysis for each feature.

[0074] Figure 3 Visual output of each prediction model: A is the LGB model, B is the KNN model, C is the CatBoost model, D is the XGB model, E is the DT model, F is the RF model, G is the GBR model, H is the ET model, and I is the Pearson feature correlation analysis;

[0075] Figure 4 To compare the performance of prediction models for each pollutant;

[0076] Figure 5 Model analysis based on model ET, A is the SHAP feature importance diagram, showing the average SHAP value of each factor, and B is the SHAP summary diagram, showing the importance of the features and their impact on the pollutant migration and transformation process.

[0077] Figure 6The mapping relationship between the top eight individual features based on the individual conditional expectation (ICE) of the data and the pollutant degradation rate constants, as well as the interaction analysis of Pyr-N and Gph-N, revealed the influence of the top eight features on the degradation process. DETAILED DESCRIPTION

[0078] The present invention is further described in detail below in conjunction with the accompanying drawings and specific examples of the specification. The examples are only used to explain the present invention and are not used to limit the scope of the present invention. The test methods used in the following examples are conventional methods unless otherwise specified; the materials and reagents used are reagents and materials that can be obtained from commercial channels unless otherwise specified.

[0079] Library information used in Python:

[0080]

[0081] Example 1 Data acquisition and filling of missing feature values

[0082] 1. Experimental Methods

[0083] 1. Data Collection

[0084] Using keywords to search in Web of Science and Google Scholar, relevant literature reviews were found. The keywords included "non-metals, carbon, persulfate, and singlet oxygen". The literature reviews were collected from 2017 to 2022.

[0085] A total of 51 papers were selected to extract data on carbon material properties, experimental conditions, and pollutant properties, resulting in a total of 700 data samples (Table 1), including 25 input features and 1 output variable (Table 2). Pollutants include: antibiotics, hormones, and dyes.

[0086] Table 1:

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] NAN indicates missing data. The catalyst properties and pollutant properties of each data sample can be found in the original literature.

[0097] Among them, 4-CP: 4-chlorophenol; TC: tetracycline; MB: methylene blue; Phel: phenol; SMX: sulfamethoxazole; OG: orange G; BPA: bisphenol A; RhB: rhodamine B; NOR: norfloxacin; AO7: orange II; TMP: trimethoprim; BPS: bisphenol S; 2,4-DCP: 2,4-dichlorophenol; 2,4-D: 2,4-dichlorophenol; X-RG: reactive yellow X-RG YellowX-RG); MET: metolachlor; MP: methyl paraben; LOS: losartan potassium; E1: Estrone; DS: diclofenac sodium.

[0098] 2. Classify and count the collected data

[0099] The data classification and statistics are shown in Table 2.

[0100] Table 2:

[0101]

[0102] 3. Correlation analysis of input feature values

[0103] Pearson correlation analysis (Pandas 1.4.4) is used to analyze the correlation between input features. The specific formula is:

[0104] The r in the formula is the PCC value, which is used to reflect the strength of the relationship between two variables. and is the variable x i ori The average value of r ranges from -1 to 1, 0 means no correlation, and the larger the absolute value, the stronger the correlation;

[0105] The absolute value of the correlation coefficient above 0.8 can be considered a high correlation, between 0.5 and 0.8 is a moderate correlation, between 0.3 and 0.5 is a low correlation, and less than 0.3 is considered no correlation. Figure 2 Pearson correlation results for each feature.

[0106] 4. Filling missing values ​​of input features

[0107] According to the missing degree of the input feature values ​​and the correlation between the input feature values, the collected data is processed to fill the missing values ​​to form a complete data set. The specific method is as follows:

[0108] 4.1. First, for input features with a low degree of missingness (the amount of missing data is less than or equal to 10% of the total number of samples, such as Tp, C, O, H, N, Tr and pH), the missing values ​​are filled using the mean method, that is, the filling value is the mean of all non-missing data of the input feature;

[0109] 4.2. For input features with a high degree of missingness (the amount of missing data is greater than 10% of the total sample size), missing data is filled according to the results of Pearson correlation analysis:

[0110] 4.2.1. Select simulation features:

[0111] Input features without missing values ​​(including input features that have completed missing value filling using the mean method) are selected as candidate features. At the same time, candidate features that meet the following two points are selected. At least 3 candidate features with the highest correlation (largest absolute value of correlation coefficient) are selected as simulated features, and their data values ​​are simulated variables:

[0112] (a) The candidate features have the same category as the input features whose missing values ​​need to be supplemented;

[0113] (b) The candidate features are highly correlated with the data values ​​of the features that need to supplement the missing values;

[0114] The corresponding relationship between input features with a higher degree of missingness and candidate features is shown in Table 3.

[0115] Table 3

[0116]

[0117] 4.2.2. Use the random forest regression algorithm to establish a mapping relationship between simulated variables and missing variables (the data values ​​of input features that need to supplement missing values) to build a model, and use the model to fill in the missing data values ​​of input features with missing values.

[0118] Among them, the method of establishing the mapping relationship using the random forest regression algorithm is:

[0119] (a) For all data samples: the data values ​​of the simulated variables of the samples whose missing variable values ​​are non-missing values ​​are taken as x_train, the data values ​​of the missing variables of the samples whose missing variable values ​​are non-missing values ​​are taken as y_train, the data values ​​of the simulated variables of the samples whose missing variable values ​​are missing values ​​are taken as x_test, and the data values ​​of the missing variables of the samples whose missing variable values ​​are missing values ​​are taken as y_test.

[0120] Taking the missing value of "Pyd-N" as an example: the four features "Tp, C, N and N / C" are data without missing values, and the Pearson correlation analysis results show that the absolute value of the correlation coefficient between the values ​​of the four features "Tp, C, N and N / C" and the value of "Pyd-N" is the largest (the absolute value of the correlation coefficient is greater than or equal to 0.3); the mapping relationship between "Tp, C, N and N / C" and "Pyd-N" is established through the random forest regression algorithm, and the data of the four features "Tp, C, N and N / C" whose values ​​of "Pyd-N" in the sample are not missing values ​​are used as x_train, the data of the values ​​of "Pyd-N" whose values ​​of "Pyd-N" in the sample are not missing values ​​are used as y_train, the data of the four features "Tp, C, N and N / C" whose values ​​of "Pyd-N" are missing values ​​are used as x_test, and the data of the values ​​of "Pyd-N" whose values ​​of "Pyd-N" are missing values ​​are used as y_test;

[0121] (b) Then, by using x_train as the training input and y_train as the training output, the model is trained using the random forest regression algorithm to obtain the trained model. x_test is input into the trained model for prediction, and the predicted result y_test is used to fill the missing value.

[0122] 5. Standardize the data

[0123] In order to (1) improve the performance of the algorithm, standardization can ensure that different features have similar contributions to the algorithm's weights, thereby improving the performance of the algorithm; (2) improve the convergence speed (i.e., increase the speed of calculation); (3) avoid the influence of outliers, after the missing values ​​are filled, the data should be standardized before model training. Through standardization, the influence of outliers can be weakened and the robustness of the model can be improved.

[0124] The formula for standardization is: Among them, x i represents the value of input feature i, x i * Represents x i The normalized value of , s represents x i The standard deviation of x i The average value of .

[0125] Specifically, Standard (a module in data preprocessing in the scikit-learn 1.0.2 learning library) is used to normalize all data so that the features have the same dimension and range (this processing can eliminate the dimensional differences between features, making the comparison between different features more fair and accurate).

[0126] Standardization makes the data values ​​of each input feature have the same dimension and range to eliminate the dimensional differences between features. A feature data set containing 700 samples was obtained.

[0127] Example 2 Construction of a pollutant degradation rate prediction model

[0128] 1. Experimental Methods

[0129] The process of constructing a model for predicting the degradation rate of pollutants by persulfate catalyzed by non-metallic materials (pollutant degradation rate prediction model) based on machine learning is shown in Figure 1 , the specific steps are as follows:

[0130] The feature data set containing 700 samples obtained in Example 1 is randomly divided into a training set and a test set in a ratio of 7:3.

[0131] Based on different machine learning algorithms, the input features and the corresponding rate constants (k obs ), the model algorithms include K nearest neighbor model (KNN), decision tree model (DT), random forest model (RF), LightGBM model (LGB), XGBoost model (XGB), gradient boosting model (GBR), CatBoost model (CatBoost) or extreme random forest model (ET).

[0132] After training each model using the training set data, we used grid search (GridSearchCV) combined with R with 5-fold cross validation. 2 Specifically, through grid search, each parameter combination will be evaluated using 5-fold cross validation to obtain R 2 , R 2The highest parameter combination is the best parameter combination, which has the highest prediction effect of the model and has good performance and higher accuracy on different data sets.

[0133] The specific method of 5-fold cross validation is as follows: the feature data set containing 700 samples obtained in Example 1 is divided into 5 subsets of equal size. In each model training, 4 subsets are used as training sets, and the remaining 1 subset is used as a validation set. This process is repeated 5 times, and each subset is used as a validation set once.

[0134] All programs were run in Python (version 3.9.13). The optimal parameters of each algorithm are shown in Table 4. Figure 3 (A to H) show the visual output of each prediction model. These images are generated by matplotlib (version 3.5.2) and seaborn (version 0.11.2). Among them, matplotlib library is used to draw basic graphs and charts, while seaborn library is used to provide more beautiful and informative statistical graphics to better show the distribution and relationship of data.

[0135] Table 4 Optimal parameters of the pollutant migration and transformation prediction models established by each algorithm

[0136]

[0137] Then, the sklearn.metrics.r2_score function in the Scikit-learn (1.0.2) library calculates R2, the mean_squared_error function in the Scikit-learn (1.0.2) library calculates MSE, and then the np.sqrt function in the NumPy library calculates the square root of MSE, which is RMSE. 2 The specific calculation formulas for RMSE are:

[0138] and

[0139] in, represents the predicted value, and y represents the actual value. Represents the average value of the target feature. The variable n represents the data point of a specific instance, and N represents the total number of data points. Furthermore, in order to ensure the robustness of the parameter adjustment results, one of the parameters is modified based on the model with the optimal parameters after parameter adjustment, the model is fine-tuned, and the performance of the model is re-evaluated (i.e., R is calculated). 2This process was repeated three times, each time adjusting different parameters and re-evaluating the R of the model on the Training set, Validation set, and Test set. 2 and RMSE, where the test set is the validation set of the 5-fold cross validation above. Finally, the R obtained in the three times is 2 The RMSE and P values ​​were averaged and the standard errors of these results were calculated to assess the stability and performance of the model.

[0140] 2. Experimental Results

[0141] Results Figure 4 Figure and Table 5, Figure 4 Shows the average R of different models on Training, Validation, and Test sets 2 The model after parameter adjustment has good prediction accuracy and generalization ability. Table 5 shows the performance score (R 2 and RMSE).

[0142] from Figure 4 As can be seen from Table 5, for the pollutant migration and transformation prediction model, the R 2 They are the highest among other models, while the RMSE value is the lowest compared with other model algorithms. Therefore, the ET model is used as the final model for the prediction of pollutant migration and transformation.

[0143] Table 5 R of the pollutant migration and transformation prediction model 2 and RMSE

[0144] Model LGB KNN Catboost XGB DT RF GBR ET <![CDATA[Training set R 2 > 0.5428 0.8658 0.8794 0.9970 0.9880 0.9696 0.9789 0.9997 <![CDATA[Test set R 2 > 0.4122 0.5532 0.6265 0.6229 0.6412 0.7108 0.7259 0.7719 <![CDATA[Mean R 2 of validation set]]> 0.4290 0.5723 0.6086 0.6282 0.6103 0.6962 0.7102 0.8001 Training set RMSE 0.3997 0.2166 0.2053 0.0322 0.0647 0.1000 0.0859 0.0105 Test set RMSE 0.3742 0.3262 0.2983 0.2997 0.2924 0.2600 0.2556 0.2331 Mean RMSE of validation set 0.3956 0.3129 0.2945 0.3058 0.3731 0.2659 0.2606 0.2491

[0145] Example 3 Evaluation of the prediction effect of the pollutant degradation rate prediction model

[0146] 1. Experimental Methods

[0147] According to the missing variables and simulated variables when filling data in Example 1, the data values ​​of the simulated variables are used as input values, and the data values ​​of the missing variables are used as output values. 2 According to the training set and test set divided in Example 2, we obtain the training set R 2 and Test (test set) R 2 , the score after 5-fold cross validation is Validation R 2 , to evaluate the filling effect and prediction effect, where Traing R 2 、Test R 2and Validation R 2 It is calculated by calling the sklearn.metrics.r2_score function in the Scikit-learn library.

[0148] 2. Experimental Results

[0149] The results are shown in Table 6. The results show that the filled data show a higher Train R 2 This indicates that the model fit of the mapping relationship between the missing variables and the simulated variables is good. In addition, the higher Test R 2 and Validation R 2 This reflects that the method of filling missing values ​​has achieved significant results in improving data completeness and model accuracy.

[0150] Table 6

[0151]

[0152] Example 4: Contribution of output features to the model for predicting pollutant degradation rate

[0153] 1. Experimental Methods

[0154] For the pollutant degradation rate prediction model established by the ET algorithm in Example 2, two machine learning model interpreters were used: Shapley value method interpretation (SHAP) and partial dependence graph (PDP). In the Shapley value method, the shap learning library (version 0.41.0) was used to construct the arrangement of input features' contribution to the model. In the partial dependence graph analysis, the PDP box learning library (version 0.2.1) was called to construct the mapping relationship between feature input and target output. .

[0155] 2. Experimental Results

[0156] SHAP is used to explain the output of the machine learning model. The importance ranking of the features based on the model ET is listed, and the contribution of the features to the model is listed, such as Figure 5 As shown, the top eight features in terms of model contribution are N-oxi (+0.16)>TPV (+0.12)>Pyr-N (+0.03)>Gph-N (+0.03)>LUMO (+0.02)>N (+0.02)>ΔG (+0.01)>Id / Ig (+0.01).

[0157] Figure 6 The results show how each feature affects the migration and transformation of pollutants. Under certain conditions, when LUMO and ΔG are both above -0.04 and 0.11, N-oxi>5%, and TPV is between 0.45-0.55cm 3 / g range, Pyr-N>9.5%, Gph-N<9.5%, N>2.0%, Id / Ig between 1.0 and 1.1, the degradation effect of pollutants is significantly improved. At the same time, Pyr-N and Gph-N are allotropes of N element. Figure 6 It shows that they have opposite trends in affecting the pollutant transformation process, and Pyr-N and Gph-N compete at the reaction site during the reaction process. By combining SHAP and PDP analysis, the influence of the top eight features on the model output is revealed. It shows that the pollutant degradation rate is mainly affected by N-oxi, TPV, Pyr-N, Gph-N, N and Id / Ig of carbon materials in the properties of carbon materials, and is also dominated by the lowest unoccupied molecular orbital LUMO and ΔG in the properties of pollutants. These findings provide important insights and reliable data support for in-depth exploration of the migration and transformation process of pollutants.

[0158] Example 4 Prediction of the rate constant of pollutant degradation by persulfate catalyzed by non-metallic materials using the pollutant degradation rate prediction model

[0159] 1. Experimental Methods

[0160] Another 122 data samples of non-metallic materials catalyzing persulfate degradation of pollutants were obtained from published literature, and the pollutant degradation rate prediction model established using the ET algorithm in Example 2 was used to calculate the rate constants of non-metallic materials catalyzing persulfate degradation of pollutants.

[0161] And calculate R according to the method of Example 2 2 and RMSE.

[0162] 2. Experimental Results

[0163] The detailed information of the other 122 data samples, as well as the actual and predicted values ​​of the rate constants, are shown in Table 7.

[0164] Table 7:

[0165]

[0166]

[0167] The catalyst properties and pollutant properties of each data sample can be found in the original literature.

[0168] The results show that the pollutant degradation rate prediction model established by the ET algorithm in Example 2 has a data set of R 2 and RMSE are 0.6163 and 0.2152 respectively, while in Example 2, the R 2and RMSE are 0.8001 and 0.2491 respectively.

[0169] Although R is obtained by using the data sample of this embodiment 2 The value obtained using the 700 data samples in Example 1 is slightly lower, but it still shows that the pollutant degradation rate prediction model established by the ET algorithm in Example 2 has good predictive ability; at the same time, the pollutant degradation rate prediction model established by the ET algorithm in Example 2 can explain about 61.63% of the variance of the target variable, further indicating that it has certain consistency and stability, and can maintain relatively consistent predictive ability on different data sets.

[0170] The RMSE obtained by using the data samples of this embodiment is 0.2154, and the prediction accuracy is higher than that of the 700 data samples in Example 1.

[0171] In summary, the pollutant degradation rate prediction model established by the ET algorithm in Example 2 can predict the rate constant of pollutant degradation by persulfate catalyzed by non-metallic materials, and has good prediction ability and moderate generalization ability.

Claims

1. A model based on a machine learning model for predicting the rate constant of persulfate degradation of pollutants catalyzed by non-metallic materials, characterized in that: The model uses the data value of the relevant features as the input value and the rate constant of pollutant degradation as the output value, and is obtained by training the machine learning model. The related features include the following three types: (1) Catalyst properties: pyrolysis temperature, carbon content, oxygen content, hydrogen content, nitrogen content, ratio of nitrogen to carbon, ratio of oxygen to carbon, pyridinic N content, pyrrolic N content, graphitic N content, N oxide content, specific surface area, total pore volume and ratio of Raman D peak intensity to G peak intensity. (2) Experimental conditions: catalyst concentration, PMS concentration, PDS concentration, reaction temperature, pH and pollutant concentration, (3) Properties of the pollutants: molecular weight, pKa, highest occupied molecular orbital, lowest unoccupied molecular orbital, and molecular energy gap; The machine learning model is one or more of a K nearest neighbor model, a decision tree model, a random forest model, a LightGBM model, an XGBoost model, a gradient boosting model, a CatBoost model or an extreme random forest model.

2. The model according to claim 1, characterized in that The data values ​​of the relevant features are filled with missing data and used as input values ​​for machine learning model training.

3. The model according to claim 1, characterized in that The data values ​​of the features are normalized and then used as input values ​​for machine learning model training.

4. The model according to claim 2 or 3, characterized in that The specific method of filling missing data is: S1. For features with missing data less than or equal to 10% of the total number of samples, the fill value is the mean of all non-missing data of the feature; S2. For features where the amount of missing data is greater than 10% of the total sample size, fill in the missing data based on the results of the Pearson correlation analysis: S21. Select simulation features: Input features without missing values, including input features that have completed missing value filling using the mean method, are used as candidate features. At the same time, candidate features that meet the following two points are selected. At least 3 candidate features with the highest correlation are selected as simulated features, and their data values ​​are simulated variables: (a) The candidate features have the same category as the input features whose missing values ​​need to be supplemented; (b) The candidate features are highly correlated with the data values ​​of the features that need to supplement the missing values; S22. Use the algorithm to establish the mapping relationship between the simulated variables and the missing variables, establish the model, and use the obtained model to fill the missing data values ​​of the input features with missing values: For all data samples: the data values ​​of the simulated variables of the samples whose missing variable values ​​are non-missing values ​​are taken as x_train, the data values ​​of the missing variables whose missing variable values ​​are non-missing values ​​are taken as y_train, the data values ​​of the simulated variables of the samples whose missing variable values ​​are missing values ​​are taken as x_test, and the data values ​​of the missing variables whose missing variable values ​​are missing values ​​are taken as y_test. By using x_train as the training input and y_train as the training output, the random forest regression algorithm is used to train the model to obtain the trained model. x_test is input into the trained model for prediction, and the predicted result y_test is used to fill the missing values.

5. The model according to claim 1, characterized in that In the training, the ratio of the number of samples in the training set and the test set is 7:

3.

6. Use of the model described in any one of claims 1 to 5 in predicting the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

7. A method for predicting the rate constant of pollutant degradation by persulfate catalyzed by non-metallic materials based on a machine learning model, characterized in that: The machine learning-based model described in any one of claims 1 to 5 is used to predict the degradation rate of pollutants by persulfate catalyzed by non-metallic materials.

8. A system for predicting the rate constant of persulfate degradation of pollutants by non-metallic materials, comprising an acquisition unit, a storage unit and a processing unit, characterized in that: The acquisition unit is used to acquire the numerical value of the characteristic of the non-metallic material sample to be predicted, The features include the following three types: (1) Catalyst properties: pyrolysis temperature, carbon content, oxygen content, hydrogen content, nitrogen content, ratio of nitrogen to carbon, ratio of oxygen to carbon, pyridinic N content, pyrrolic N content, graphitic N content, N oxide content, specific surface area, total pore volume and ratio of Raman D peak intensity to G peak intensity. (2) Experimental conditions: catalyst concentration, PMS concentration, PDS concentration, reaction temperature, pH and pollutant concentration, (3) Properties of the pollutants: molecular weight, pKa, highest occupied molecular orbital, lowest unoccupied molecular orbital, and molecular energy gap; The storage unit stores program instructions that can be executed by the processing unit, and the program instructions input the numerical value of the feature of the non-metallic material sample to be predicted obtained by the obtaining unit into the processing unit; The processing unit comprises the model according to any one of claims 1 to 5, and executes the program instructions to obtain the numerical value of the characteristic of the non-metallic material sample to be predicted as a result.

9. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program executable on the processor; When the computer program is executed by a processor, the operation of the system of claim 8 is realized.

10. A computer-readable storage medium storing a computer program executable by a processor, wherein when the computer program is executed by the processor, the operation of the system of claim 8 is implemented.

Images