Method for predicting heavy metal adsorption capacity of microplastics

Abstract

The invention discloses a method for predicting the heavy metal adsorption capacity of micro-plastics, which comprises the following steps of: screening heavy metal and micro-plastic adsorption data from literature and professional databases, and arranging the data into a sample data set; performing preprocessing including data cleaning, missing value processing and data standardization on the sample data set; constructing an adsorption capacity prediction model, and performing training and evaluation to determine an optimal prediction model; using random search, Bayesian optimization or grid search to adjust hyper-parameters of the model; and carrying out adsorption capacity prediction on microplastics and heavy metals with unknown adsorption capacities by using the adsorption prediction model. The method can be practically applied to research on adsorption of the heavy metal ions to the micro-plastics, interaction between the heavy metal ions and the micro-plastics can be visually understood, and ecological risks caused by coexistence of the heavy metal ions and the micro-plastics can be evaluated.

Description

Technical Field

[0001] The invention relates to a method for predicting environmental pollution, in particular to a method for predicting the adsorption capacity of microplastics for heavy metals. Background Art

[0002] The widespread use of plastics has resulted in a large amount of plastic waste being discharged into the environment. The continuous fragmentation of plastic waste has led to the widespread presence of microplastics (microplastics) in different areas such as rivers, lakes and soil. In aquatic environments, the concentration of microplastics ranges from a few to tens of thousands per cubic meter. Due to their hydrophobicity and large specific surface area, microplastics can serve as effective carriers of organic pollutants and heavy metal ions, thus affecting the fate of these pollutants and their ecological risks to aquatic systems.

[0003] Heavy metal ions always coexist and interact with microplastics. For example, polystyrene (PS) can effectively adsorb Pb 2+ , Cu 2+ 、Cd 2+ 、Ni 2+ and Zn 2+ As polystyrene ages, its adsorption capacity increases due to surface wear and the formation of oxygen-containing functional groups. Microplastic properties, environmental conditions, and heavy metal ion properties all affect the adsorption of heavy metal ions on microplastic surfaces. Considering that the adsorption behaviors and mechanisms of different microplastics and heavy metal ions under different environmental conditions vary, studying the adsorption process through laboratory methods is time-consuming and laborious. Using the preliminarily determined data set to quickly predict the adsorption of heavy metal ions on microplastics will help to efficiently assess ecological risks.

[0004] At present, the adsorption relationship between heavy metals and microplastics is mainly studied through experimental methods, which are usually time-consuming and costly. In addition, the adsorption behavior of different types of microplastics and heavy metals under different experimental conditions, and the adsorption of different types of heavy metals by the same microplastic are significantly different, resulting in poor universality of the experimental results. Although traditional experimental methods can provide certain data support, it is difficult to quickly and accurately predict the adsorption capacity of microplastics for heavy metals under different conditions.

[0005] In the study of microplastics and heavy metal adsorption, machine learning technology can predict the adsorption behavior of heavy metals by different microplastics under different environmental conditions by training models. This can not only improve research efficiency, but also provide a scientific basis for actual environmental governance. However, existing prediction methods have problems such as low heavy metal prediction accuracy, low efficiency and low universality. Summary of the invention

[0006] The purpose of the present invention is to provide a method for predicting the adsorption capacity of microplastics for heavy metals, so as to solve the problems of low heavy metal prediction accuracy, low efficiency and poor universality in existing prediction methods.

[0007] The object of the present invention is achieved in that:

[0008] A method for predicting the adsorption capacity of microplastics for heavy metals comprises the following steps:

[0009] S1. Obtain sample data sets: Screen heavy metal and microplastic adsorption data from literature databases and professional databases, including heavy metal types, microplastic types, experimental conditions, and corresponding adsorption capacity data and characteristic data affecting adsorption capacity; organize the acquired data into a sample data set for machine learning;

[0010] S2. Data preprocessing: Perform preprocessing on the sample data set, including data cleaning, missing value processing and data standardization, remove abnormal data, and divide the data in the sample data set into a training set and a test set in a ratio of 8:2;

[0011] S3, model construction: use random forest RF, artificial neural network ANN, distributed gradient boosting XGBoost, decision tree DT, gradient boosting decision tree GBDT and multivariate linear regression MLR algorithm to build adsorption capacity prediction model; then use the training set data and the selected machine learning algorithm to train and evaluate the adsorption capacity prediction model to determine the adsorption capacity prediction model;

[0012] S4. Model optimization: Use random search, Bayesian optimization or grid search to adjust the hyperparameters of the model, measure the fit of the adsorption capacity prediction model to the observed data based on the calculated determination coefficient R2; measure the average prediction error of the adsorption capacity prediction model based on the calculated root mean square error RMSE indicator;

[0013] S5. Use the adsorption prediction model to predict the adsorption capacity of microplastics and heavy metals with unknown adsorption capacity.

[0014] The present invention further comprises the following steps:

[0015] S6. Model interpretation: The SHAP method was used to interpret the prediction results of the optimal model. Through the SHAP value, the contribution of each feature to the prediction results can be quantified, so as to analyze the specific influencing factors of microplastics and heavy metal adsorption behavior and evaluate the important characteristics of the model for heavy metal ion adsorption.

[0016] The prediction method of the present invention collects experimental data of microplastics and heavy metals from literature databases and professional databases, selects data obtained from experimental tests, and then screens out important features through feature selection methods, constructs an adsorption capacity prediction model and optimizes the model, uses cross-validation to evaluate the model performance, uses determination coefficient, mean absolute error and root mean square error to measure the model, uses SHAP to interpret the model, and compares experimental results. The prediction method of the present invention has the following advantages:

[0017] 1. High accuracy: Since the data set is based on a large amount of sample data from heavy metal and microplastic adsorption tests screened and collected from literature databases and professional databases, the rich data base covers actual adsorption test results under a variety of different conditions, different types of heavy metals and microplastics. By analyzing, summarizing and learning from many real and diverse samples, it is possible to capture various key factors that affect adsorption behavior and their complex relationships, so that the constructed prediction model can be more in line with the actual situation, thereby ensuring a high prediction accuracy.

[0018] 2. Low cost: Traditional research on the adsorption behavior of heavy metals on microplastics often requires a large number of tedious and costly adsorption experiments, involving the configuration of experimental materials, reagents, reaction solutions, detection equipment, etc., and the cost of manpower, money, materials and time is very high. However, the present invention does not require a large number of actual experiments, data acquisition is relatively easy, and does not involve a large amount of manpower, money and materials consumption.

[0019] 3. Green and environmentally friendly: The present invention does not generate any experimental waste, avoids heavy metal pollution, and is in line with the concept of green and environmental protection.

[0020] The present invention can explain the factors that affect relevant properties and quickly guide experimental synthesis. It has the characteristics of fast calculation speed, high precision, and convenient structural analysis. It can be actually applied to the study of the adsorption of heavy metal ions on microplastics, which helps to intuitively understand the interaction between heavy metal ions and microplastics and evaluate the ecological risks brought about by the coexistence of the two. It overcomes the shortcomings of traditional experimental and theoretical calculation methods, saves time and resources, and is interpretable, making the prediction process more transparent, and can provide guidance for the laboratory synthesis of new microplastics. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a diagram showing the effect of environmental conditions on the adsorption of heavy metals by microplastics.

[0022] Figure 2 Schematic diagram of the top 14 features with the greatest impact on adsorption capacity and their SHAP values. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0024] The method for predicting the adsorption capacity of microplastics for heavy metals of the present invention comprises the following steps:

[0025] S1. Obtain sample data sets: Use "microplastics", "heavy metals" and "interactions" as keywords to retrieve and collect publications from literature databases and professional databases such as Web of Science and Google Scholar. Publications with low relevance or lack of clear data were excluded from the database. A total of 72 articles were identified, and a total of 1,302 data on heavy metal and microplastic adsorption were screened out; 16 parameters were extracted from them, including: particle size of microplastics, microplastic density, glass transition temperature, microplastic type, aging degree, atomic mass, electronegativity, first ionization energy, valence state, initial concentration of heavy metal ions, initial concentration of microplastics, reaction time, solution pH, temperature, salinity and adsorption capacity and other related parameter information; also including heavy metal types, microplastic types, experimental conditions and corresponding adsorption capacity data and characteristic data affecting adsorption capacity.

[0026] The acquired data is organized into a sample data set for machine learning. The sample data set includes experimental sample data, which includes the physical and chemical properties of microplastics, the physical and chemical properties of heavy metals, and experimental conditions.

[0027] S2. Data preprocessing: The sample data set is preprocessed, including data cleaning, missing value processing and data standardization, to ensure the consistency and reliability of the data. Data cleaning is used to remove outliers and noise data; missing value processing can use mean filling or interpolation methods; data standardization can use 0-1 standardization, Z-score standardization or linear scale standardization.

[0028] By using the correlation coefficient to determine the independence between each feature, the Pearson correlation coefficient (PCC) is used to evaluate the linear relationship between the input variable and the target variable. The mathematical expression of the Pearson correlation coefficient is as follows:

[0029]

[0030] Among them, r xy is the PCC value; x i and i is the i observation in the sample; and are the means of the x and y variables respectively; n is the sample size.

[0031] The present invention uses rigorous data cleaning and estimation techniques to solve the problem of missing variables in the original data set. Since the algorithm cannot use non-numeric input values, two non-numeric input variables (microplastic type and aging degree) are encoded as numerical values ​​using a random assignment method. 14 types of MP were randomly assigned numerical values. Specifically, polystyrene (PS), polypropylene (PP), polyethylene terephthalate (PET), nylon (Nylon), polyvinyl chloride (PVC), polyamide (PA), ethylene-vinyl acetate copolymer (EVA), polyvinyl alcohol (PVA), polyethylene (PE), high-density polyethylene (HDPE), low-density polyethylene (LDPE), chlorinated polyethylene (CPE), polylactic acid (PLA) and polymethyl methacrylate (PMMA) are represented from 1 to 14. The aging state values ​​of unaged and aged microplastics are 0 and 1, respectively.

[0032] After removing the abnormal data in the sample data set, the data in the sample data set is divided into a training set and a test set in a ratio of 8:2. That is, the preprocessed sample data set is randomly divided into two parts, 20% (260 data) is used as a test set, and the remaining 80% (1042 data) is used as a training set.

[0033] S3, model construction: random forest RF, artificial neural network ANN, distributed gradient boosting XGBoost, decision tree DT, gradient boosting decision tree GBDT and multivariate linear regression MLR algorithms are used to build adsorption capacity prediction models. Then, the training set data is used to train and evaluate the adsorption capacity prediction model using the selected machine learning algorithm to determine the adsorption capacity prediction model.

[0034] The adsorption capacity prediction models constructed include the following:

[0035] 1. Random Forest RF:

[0036] There are T decision trees, each decision tree t gives a regression prediction value y for sample x t (x), the final regression prediction value y of the random forest RF (x) is the average value of the predictions of each decision tree:

[0037]

[0038] 2. Artificial Neural Network ANN:

[0039]

[0040] Among them, x i (i=1, 2, 3) is the input value, represents the activation value of the i-th neuron in the k-th layer, N krepresents the number of neurons in the kth layer. For the input layer (k = 1), The bias term is represented by x0=1, The weight of the kth layer is Represents, where j represents the jth neuron in the (k+1)th layer and i represents the neuron in the kth layer.

[0041] 3. Decision Tree DT:

[0042]

[0043] in, is the predicted output for input x; M is the total number of leaf nodes; c m It is a leaf node R m The predicted output value of I(x∈R m ) is an indicator function. If x belongs to a leaf node R m , then the function is equal to 1, otherwise it is equal to 0.

[0044] 4. Gradient Boosting Decision Tree GBDT:

[0045]

[0046] in, is the predicted output for input x, M is the number of decision trees, and f m (x) is the prediction function of the mth tree.

[0047] 5. Multiple Linear Regression MLR Algorithm:

[0048] γ=β0+β1x1+β2x2+…+β n x n +ε

[0049] Among them, γ represents the dependent variable, x1,x2,...,x n represents the independent variable, β0,β1,β2,...,β n represents the regression coefficient and ε represents the error term.

[0050] Then, the training set data is used to train and evaluate the adsorption capacity prediction model using the selected machine learning algorithm to determine the adsorption capacity prediction model. Machine learning models such as random forest (RF), XGBoost, and decision tree (DT) are used for training, and the feature data is used as input to output the predicted adsorption capacity. The output predicted adsorption capacity is fitted with the real data in the validation set. The present invention finally determines that the random forest RF machine learning model is the optimal model for adsorption capacity prediction.

[0051] The training process of the model is: using the training set data after feature selection, the selected machine learning algorithm is applied to the data. During the training process, the adsorption capacity prediction model will learn how to predict the adsorption amount of heavy metals on microplastics based on the input features.

[0052] The feature data is used as input and the predicted adsorption capacity is output. This process can be programmed in Python and implemented in PyCharm Community Edition. Model parameter training: Through the optimization algorithm, the parameters of the adsorption capacity prediction model are continuously updated to minimize the loss function.

[0053] S4. Model optimization: Use random search, Bayesian optimization or grid search in Python to adjust the hyperparameters of the model. Among them, grid search is a method of selecting parameters based on the minimum root mean square value on the validation data set to achieve the best performance of RF. In the scikit-learn software package, RF contains RF classifiers and regressors. Bagging and CART decision trees respectively contain hyperparameters of two parts. The parameters used in the present invention include n_estimators (100, 200, 300, 500, 1000), max_depth (3, 5, 8, 15, 20, 25, 30, None), min_samples_leaf (1, 2, 5, 10), criterion ("squared_error", "absolute_error", "poisson") and max_features ("log2", "sqrt"). During the search process, the sample data set is still randomly divided into training set and test set at a ratio of 8:2. According to the calculated coefficient of determination R 2 The fit of the adsorption capacity prediction model to the observed data is measured. The average prediction error of the adsorption capacity prediction model is measured according to the calculated root mean square error RMSE index.

[0054] Coefficient of determination R 2 The calculation method is as follows:

[0055]

[0056] in, is the predicted value of the target variable, y i is the actual value of the target variable, is the mean value of the target variable and n is the total number of data points.

[0057] The root mean square error RMSE is calculated as follows:

[0058]

[0059] in, is the predicted value of the target variable, y i is the actual value of the target variable, is the mean value of the target variable and n is the total number of data points.

[0060] Use random search, Bayesian optimization, or grid search in Python to adjust the model's hyperparameters and optimize the prediction model to improve prediction accuracy. Optimize Random Forest (RF), MLR, XGBoost, DT, GBDT, and ANN. The coefficient of determination R of the test set 2 The closer it is to 1, the smaller the mean absolute error and root mean square error values ​​are, and the more accurate the model is.

[0061] The adsorption capacity prediction model was evaluated by cross-validation. Specifically, the sample data set was randomly divided into 5 groups, 4 groups for training and 1 group for testing. The determination coefficient R 2 , mean absolute error (MAE) and root mean square error (RMSE) were used as evaluation indicators, and the optimal model was selected as the final adsorption capacity prediction model.

[0062] The method for evaluating the adsorption capacity prediction model is: comparing the results of the training set and the test set predicted by the microplastics adsorption heavy metal prediction model with the experimental results, and evaluating the microplastics adsorption heavy metal prediction model; the evaluation index includes the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE.

[0063] Using the coefficient of determination R 2 The model is measured by two evaluation indicators: RMSE and RMSE. The specific results of the training set and test set are shown in Table 1:

[0064] Table 1. Evaluation results of the adsorption capacity prediction model

[0065]

[0066] Table 1 compares the accuracy of six machine learning methods in predicting adsorption capacity. The traditional regression model MLR and machine learning models (ANN, XGBoost, GBDT, DT, and RF) were tested based on the data set. MLR performed poorly on the test set, with a determination coefficient R 2 The RMSE is 0.21 and the DT performance is poor on the dataset. This may be because the traditional linear model is too simple to capture the complexity of the dataset, resulting in insufficient fit. The prediction performance of the machine learning models also varies greatly. DT performs poorly on the dataset, and the determination coefficient R 2 The determination coefficient R of the test set is 0.91. 2is 0.87; many data points fall outside the 95% confidence interval. Although the coefficient of determination R 2 reached 0.98, but the coefficient of determination R 2 The determination coefficient R of the training set is only 0.72, which may be due to the poor generalization ability of GBDT on the dataset. 2 The determination coefficients of the test set reached 0.87 and 0.99 respectively. 2 The prediction accuracy may be limited by the generalization ability, making it difficult for the algorithm to learn and reproduce all possible relationships between input variables and target variables. RF performs significantly better, with the coefficient of determination R of the training set and the test set being 0.86 and 0.90, respectively. 2 The RMSE of the training set and the test set are 0.93 and 0.93 respectively, and the RMSE of the training set and the test set are 0.03 and 0.04 respectively. At the same time, most of the predicted data points are within the 95% confidence interval. At the same time, the RF model has the best prediction performance due to its robustness in dealing with outliers and missing data, which reduces the complexity of data preprocessing and parameter adjustment.

[0067] S5. Use the adsorption prediction model to predict the adsorption capacity of microplastics and heavy metals with unknown adsorption capacity. The present invention verifies the adsorption capacity prediction model with validation set data, and the accuracy of the model prediction reaches 93%, and the total root mean square error value is within 0.03.

[0068] S6. Model interpretation: The prediction results of the optimal model were explained using the SHAP method. Through the SHAP value, the contribution of each feature to the prediction results can be quantified, so as to analyze the specific influencing factors of microplastics and heavy metal adsorption behavior and evaluate the important characteristics of the model for heavy metal ion adsorption. The selected input variables related to the physicochemical properties of microplastics, the chemical properties of heavy metal ions and environmental conditions have a certain influence on adsorption. Figure 1 As shown, environmental conditions (63.0%) had the greatest impact on adsorption, followed by the physicochemical properties of microplastics (27.4%) and the chemical properties of heavy metal ions (9.6%).

[0069] Using SHAP (SHapley Additive explanation) to explain the selected optimal model, we can specifically analyze the impact of changes in each characteristic value on the adsorption capacity, thereby explaining the influence of microplastic structure on heavy metal adsorption behavior. Figure 2 The top 14 features with the greatest impact on adsorption capacity and their SHAP values ​​are listed. The change from blue to red indicates that the feature value increases from small to large. Figure 2 It can be seen that the change of each characteristic value affects the SHAP value, that is, the adsorption capacity, thus explaining the model.

[0070] For each sample, the model generates a prediction and assigns a SHAP value to each feature in the sample. Specifically, the SHAP value is determined by calculating the difference between the model predictions with and without the feature, and then averaging these differences to get the SHAP value of the feature. The calculated SHAP value can be given by the following formula:

[0071]

[0072] Among them, S i is the subset of all features that includes feature i, |s| is the number of elements in set s, v(s) is the prediction made by the model considering feature i, and v(sU{i}) is the prediction made by the model without considering feature i.

[0073] Use the partial dependence method to draw the PDP diagram, and explain each eigenvalue. The partial dependence function is given by:

[0074]

[0075] in, Represents the response variable for the feature subset x s The partial dependence of is the prediction result for the remaining features x c The expected value of the distribution. The calculation is for the probability distribution P(x c ) is the expected value of x. c Represents a subset x s The complement of x, that is, the complement of s The remaining features in dP(x c ) represents the feature x c The probability distribution of .

[0076] The "model.feature_importances" in the scikit-learn library is used to extract the feature importance of the RF model. Its theoretical basis is to reduce the node purity (Gini impurity or entropy reduction). SHAP is an additive interpretation model inspired by the Shapley value, which explains the feature parameters in the RF model by evaluating the contribution of each feature to the predicted output. Partial dependency analysis was performed based on the Shapley value theory to further explore the intricate relationship between the input variables and the target variable. SHAP and Scikit-learn libraries are used to calculate Shapley values ​​and partial dependency analysis. Use SHAP to interpret the optimal model,

[0077] The specific effect of each feature on the adsorption capacity is analyzed to make the prediction process of the model more transparent.

[0078] In the model interpretation work, the performance of traditional models (MLR), machine learning models (ANN, XGBoost, GBDT, DT and RF) in predicting the adsorption of heavy metal ions by microplastics was evaluated. Among them, the RF model provided the best data fitting and prediction performance. The model interpretation shows that the trend of feature importance is as follows: initial concentration of heavy metal ions> type of microplastics> reaction time> solution pH> atomic mass of heavy metal ions> salinity> degree of aging of microplastics> initial concentration of microplastics> particle size of microplastics> first ionization energy of heavy metals> density of microplastics> glass transition temperature> experimental temperature> valence state of heavy metal ions.

[0079] Since solution pH and atomic mass are positively correlated with adsorption, while salinity is negatively correlated with adsorption, electrostatic effects may have a significant effect on the adsorption of heavy metal ions. Complexation between microplastic surface functional groups and heavy metal ions is also involved in adsorption, as aging promotes the adsorption of heavy metal ions.

[0080] The present invention uses laboratory synthesized microplastic structure and heavy metal data for model training and verification, and further illustrates the accuracy and reliability of the model by comparing the experimental values ​​and predicted values.

Claims

1. A method for predicting the adsorption capacity of microplastics for heavy metals, characterized in that: The following steps are involved: S1. Obtain sample data sets: Screen heavy metal and microplastic adsorption data from literature databases and professional databases, including heavy metal types, microplastic types, experimental conditions, and corresponding adsorption capacity data and characteristic data affecting adsorption capacity; organize the acquired data into a sample data set for machine learning; S2. Data preprocessing: Perform preprocessing on the sample data set, including data cleaning, missing value processing and data standardization, remove abnormal data, and divide the data in the sample data set into a training set and a test set in a ratio of 8:2; S3, model construction: random forest RF, artificial neural network ANN, distributed gradient boosting XGBoost, decision tree DT, gradient boosting decision tree GBDT and multivariate linear regression MLR algorithms were used to build adsorption capacity prediction models; Then, the training set data is used to train and evaluate the adsorption capacity prediction model using the selected machine learning algorithm to determine the adsorption capacity prediction model; S4. Model optimization: Use random search, Bayesian optimization, or grid search to adjust the model's hyperparameters based on the calculated coefficient of determination R 2 Measure the fit of the adsorption capacity prediction model to the observed data; measure the average prediction error of the adsorption capacity prediction model based on the calculated root mean square error RMSE index; S5. Use the adsorption prediction model to predict the adsorption capacity of microplastics and heavy metals with unknown adsorption capacity.

2. The prediction method according to claim 1, characterized in that: The following steps are also included: S6. Model interpretation: The SHAP method was used to interpret the prediction results of the optimal model. Through the SHAP value, the contribution of each feature to the prediction results can be quantified, so as to analyze the specific influencing factors of microplastics and heavy metal adsorption behavior and evaluate the important characteristics of the model for heavy metal ion adsorption.

3. The prediction method according to claim 1, characterized in that: The sample data set includes experimental sample data, and the experimental sample data includes the physical and chemical properties of microplastics, the physical and chemical properties of heavy metals and experimental conditions.

4. The prediction method according to claim 1, characterized in that: The adsorption capacity prediction model constructed in step S3 is the preferred random forest RF machine learning model: There are T decision trees, each decision tree t gives a regression prediction value y for sample x t (x), the final regression prediction value y of the random forest RF (x) is the average value of the predictions of each decision tree:

5. The prediction method according to claim 1, characterized in that: The coefficient of determination R in step S4 2 The calculation method is as follows: in, is the predicted value of the target variable, y i is the actual value of the target variable, is the mean value of the target variable and n is the total number of data points.

6. The prediction method according to claim 1, characterized in that: The root mean square error RMSE in step S4 is calculated as follows: in, is the predicted value of the target variable, y i is the actual value of the target variable, is the mean value of the target variable and n is the total number of data points.

7. The prediction method according to claim 1, characterized in that: In step S4, the evaluation of the adsorption capacity prediction model is to use the cross-validation method to evaluate the optimized adsorption capacity prediction model. Specifically, the sample data set is randomly divided into 5 groups, 4 groups are used for training, and 1 group is used for testing. The determination coefficient R 2 , mean absolute error (MAE) and root mean square error (RMSE) were used as evaluation indicators, and the optimal model was selected as the final adsorption capacity prediction model.

8. The prediction method according to claim 1, characterized in that: The method for evaluating the adsorption capacity prediction model in step S4 is: comparing the results of the training set and the test set predicted by the microplastics adsorption heavy metal prediction model with the experimental results to evaluate the microplastics adsorption heavy metal prediction model; the evaluation index includes the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE.

Images