A resin synthesis method, system, medium and product based on machine learning

Abstract

The present invention discloses a resin synthesis method, system, medium and product based on machine learning, which relates to the field of perfluoroalkyl compound pollution control. The method includes obtaining multiple groups of influencing factors of PFAS removal rate, and based on the multiple groups of influencing factors, using a PFAS adsorption model to obtain PFAS removal rates of multiple resins with different carbon fluorine chain lengths, and then screening the PFAS removal rates of each resin with different carbon fluorine chain lengths to obtain the optimal PFAS removal rate, and based on the optimal PFAS removal rate, determining the optimal influencing factors, and finally synthesizing resins based on the optimal influencing factors, and removing PFAS through the synthesized resins. The present invention can provide an effective solution for the efficient removal of PFAS.

Description

Technical Field

[0001] The present invention relates to the field of perfluoroalkyl compound pollution control, and in particular to a resin synthesis method, system, medium and product based on machine learning. Background Art

[0002] Per- and polyfluoroalkyl substances (PFAS) are a group of more than 10,000 synthetic chemicals with a history of more than 50 years. PFAS substances are classified according to their carbon chain length: perfluoroalkyl sulfonic acids with carbon fluorine ≥8 and perfluoroalkyl carboxylic acids with carbon fluorine ≥7 are considered long-chain PFAS, while those with shorter chain lengths are considered short-chain PFAS. Their particularities in durability, dielectric properties, water resistance and heat resistance give them advantages in various industrial and household applications. Due to their extreme persistence and stability, PFAS are widely distributed in the environment and accumulate in the human body, and are commonly detected in trace concentrations in drinking water resources. However, PFAS are associated with negative health effects (e.g., developmental and immune disorders, infertility, thyroid disease, organ toxicity, and cancer). Therefore, PFAS pollution has not only become a growing environmental problem, but also an important global public health issue.

[0003] Adsorption is an efficient and simple technology for removing PFAS from water environments, with the advantages of low cost, environmental protection and simple design. Among them, the removal of PFAS from drinking water by anion exchange has been demonstrated, and disposable, chloride-form, gel-type strong base resins functionalized with quaternary ammonium are usually used. Several studies conducted in the laboratory (pilot or full scale) have shown that anion exchange resins can effectively remove long-chain and short-chain PFAS. The mechanism of PFAS removal using ion exchange resins is mainly due to electrostatic interactions leading to ion exchange between the resin and the PFAS charged head group, as well as hydrophobic interactions. The adsorption capacity of ion exchange resins for PFAS compounds depends on the resin characteristics, PFAS molecular characteristics, water chemistry (such as pH, ionic strength, concentration of natural organic matter, and the presence of other impurities or contaminants) and operating conditions. However, most current studies focus on limited influencing factors, and the complex interactions between resins, PFAS and environmental conditions are often oversimplified. Although ion exchange technology has become an effective method for removing long-chain PFAS, there are still some challenges in removing short-chain PFAS. Moreover, the removal of PFAS by ion exchange resins also shows slow adsorption kinetics, especially in polystyrene beads, which takes 168 hours to reach equilibrium. In addition, ion exchange resins have the ability to bind to other molecules in the water, which can reduce their ability to remove PFAS and require more frequent regeneration processes to meet the required standards for PFAS removal. In full-scale applications, the large amount of waste generated by resin regeneration and its disposal should also be considered.

[0004] Machine Learning (ML) algorithms are able to manage complex, multidimensional data sets with powerful fitting capabilities and have gained increasing attention in the field of new material development. Several ML models have been developed and successfully deployed to predict membrane properties, adsorbent performance, etc. under manufacturing conditions. However, for so many types of organic pollutants, these ML models only use limited data points to predict their removal, an approach that may lead to uncertainty in the prediction results because some types of emerging organic pollutants, such as PFAS, are compounds composed of a large number of analogs or homologs, and their physicochemical properties are different from those of traditional organic pollutants, including proteophilicity, amphiphilicity, and ionicity. In addition, most experimental studies limit the analysis to a relatively small number of perfluoroalkyl carboxylates (PFCAs), perfluoroalkyl sulfonates (PFSAs), and other PFAS included in widely implemented analytical methods. However, traditional machine learning methods still have limitations in the removal of PFAS.

[0005] Therefore, a resin synthesis method for PFAS removal is urgently needed to achieve efficient removal of PFAS. Summary of the invention

[0006] The purpose of the present invention is to provide a resin synthesis method, system, medium and product based on machine learning, which can provide an effective solution for the efficient removal of PFAS.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A resin synthesis method based on machine learning, the method comprising:

[0009] Obtain multiple groups of influencing factors of PFAS removal rate; the influencing factors include resin properties, PFAS properties and adsorption system properties; the PFAS properties include SMILES descriptors of PFAS structural formulas;

[0010] Based on multiple groups of influencing factors, the PFAS adsorption model was used to obtain the PFAS removal rates of multiple resins for different carbon-fluorine chain lengths;

[0011] The PFAS removal rates of the resins with different carbon fluorine chain lengths were screened to obtain the optimal PFAS removal rate;

[0012] Based on the optimal PFAS removal rate, determining optimal influencing factors;

[0013] Resin is synthesized based on the optimal influencing factors, and PFAS is removed through the synthesized resin.

[0014] Optionally, after obtaining multiple groups of influencing factors of PFAS removal rate, the method further includes preprocessing the multiple groups of influencing factors; the preprocessing of the multiple groups of influencing factors specifically includes:

[0015] Performing dummy variable processing on multiple groups of the influencing factors to obtain numerical variables of resin properties, numerical variables of PFAS properties, and numerical variables of adsorption system properties;

[0016] Molecular fingerprint encoding method was used to convert the numerical variables of PFAS properties into molecular fingerprints.

[0017] Optionally, the training process of the PFAS adsorption model specifically includes:

[0018] Constructing a data set and dividing the data set into a training set and a test set; the data set includes experimental influencing factors and corresponding resin removal rates of PFAS with different carbon fluorine chain lengths;

[0019] Constructing initial models based on a plurality of different machine learning algorithms, and training and testing each of the initial models through a training set and a test set to determine the performance of each of the initial models;

[0020] Screening each of the initial models based on the performance of each of the initial models to obtain a plurality of initial optimization models;

[0021] The hyperparameters, molecular fingerprint encoding methods and molecular fingerprint parameters of each of the initial optimization models are optimized and adjusted by the Bayesian algorithm to obtain the Bayesian optimization results corresponding to each of the initial optimization models; the molecular fingerprint encoding methods include a binary molecular fingerprint encoding method and a counting molecular fingerprint encoding method; the parameters of the molecular fingerprint include a molecular fingerprint length and a molecular fingerprint radius; the Bayesian optimization results include an optimal hyperparameter combination, an optimal molecular fingerprint encoding method and optimal molecular fingerprint parameters;

[0022] The initial optimization model is screened according to the Bayesian optimization results to obtain the optimal model; the optimal model is the PFAS adsorption model.

[0023] Optionally, the optimal molecular fingerprint encoding method is a counting molecular fingerprint encoding method.

[0024] Optionally, the removal rates of PFAS with different chain lengths of each of the resins are screened to obtain the optimal PFAS removal rate, specifically including:

[0025] Based on the PFAS removal rate of each resin for different carbon fluorine chain lengths, the Shapley value of each influencing factor was calculated;

[0026] According to the Shapley values ​​of the influencing factors, a single factor SHAP diagram of different carbon fluorine chain lengths on PFAS removal rate was drawn to obtain the carbon fluorine chain length range with the best PFAS removal rate;

[0027] Based on the PFAS removal rates of the resins with different carbon fluorine chain lengths, a single factor SHAP diagram of the resin properties versus the PFAS removal rate was drawn to obtain the resin property characteristics with the best PFAS removal rate;

[0028] The optimal PFAS removal rate is determined based on the carbon fluorochain length range with the optimal PFAS removal rate and the resin property characteristics with the optimal PFAS removal rate.

[0029] To achieve the above objectives, the present invention also provides a computer system, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned resin synthesis method based on machine learning.

[0030] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned resin synthesis method based on machine learning.

[0031] To achieve the above objectives, the present invention also provides a computer program product, including a computer program, which implements the above-mentioned resin synthesis method based on machine learning when executed by a processor.

[0032] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention obtains the optimal hyperparameter combination, the optimal molecular fingerprint encoding method and the combination result of the parameters of the optimal molecular fingerprint by adopting the Bayesian optimization method, selects the best performing machine learning algorithm according to the combination result to construct a PFAS adsorption model, and analyzes the effects of resin properties and PFAS carbon fluorine chain length on the removal rate based on the results of feature importance sorting, and obtains the optimal resin synthesis and preparation method suitable for a wide range of PFAS removal, which can effectively predict the adsorption of PFAS with different carbon fluorine chain lengths by the resin, realizes the rational design and efficient functional preparation of the resin, and provides an effective solution for the efficient removal of PFAS. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A schematic diagram of a process for synthesizing a resin based on machine learning provided in Example 1 of the present invention;

[0035] Figure 2 The average dice similarity heat map of different types of PFAS; a is the heat map obtained based on binary molecular fingerprints; b is the heat map obtained based on counting molecular fingerprints;

[0036] Figure 3 Schematic diagram of the enhancement effect of different molecular fingerprint parameters on the performance of the ML model; where a is a molecular fingerprint radius of 5; b is a molecular fingerprint radius of 4; c is a molecular fingerprint radius of 3; d is a molecular fingerprint radius of 2; e is a molecular fingerprint radius of 1; and f is a molecular fingerprint radius of 0;

[0037] Figure 4 Schematic diagram of the change of the removal effect of resin on PFAS with different carbon fluorine chain lengths; a is a schematic diagram of the effect of PFAS with different carbon fluorine chain lengths on the removal rate; b is a schematic diagram of the effect of the resin polymer matrix and functional groups on the removal rate;

[0038] Figure 5 Schematic diagram of the parameter tuning process corresponding to the selection of the best hyperparameter combination for Bayesian optimization parameter tuning; a is a schematic diagram of the validation set RMSE corresponding to the selection of the best hyperparameter combination for Bayesian optimization parameter tuning when the RMSE is the smallest, and C1 is the point with the smallest RMSE of the validation set; b is a schematic diagram of the Bayesian optimization parameter tuning when the fitting error is the smallest corresponding to the selection of the best hyperparameter combination, and D2 is the point with the smallest fitting error;

[0039] Figure 6 Schematic diagram of the performance of the PFAS adsorption model built based on the XGBoost algorithm;

[0040] Figure 7 Schematic diagram of feature importance analysis of XGBoost model based on resin adsorption of PFAS; a is a schematic diagram of feature importance analysis of XGBoost model for all PFAS removal by resin; b is a schematic diagram of feature importance analysis of XGBoost model based on resin adsorption of short-chain PFAS;

[0041] Figure 8 This is a trend chart showing the effect of PFAS carbon fluorine chain length on removal rate;

[0042] Fig. 9 Schematic diagram of the effect of the polymer matrix and functional groups of the resin on the removal rate of PFAS;

[0043] Fig.10It is a schematic diagram of the predicted results of the interaction between polyacrylic acid resin and carbon fluorine chain on the removal rate; wherein, a is a 3D diagram of the predicted results of the interaction between polyacrylic acid resin and PFAS carbon fluorine chain on the removal rate; b is a 2D diagram of the predicted results of the interaction between polyacrylic acid resin and PFAS carbon fluorine chain on the removal rate;

[0044] Fig.11 It is a schematic diagram of the predicted results of the interaction between polystyrene resin and carbon fluorine chain on the removal rate; wherein, a is a 3D diagram of the predicted results of the interaction between polystyrene resin and PFAS carbon fluorine chain on the removal rate; b is a 2D diagram of the predicted results of the interaction between polystyrene resin and PFAS carbon fluorine chain on the removal rate;

[0045] Fig.12 It is a schematic diagram of the predicted results of the interaction between the average diameter of the resin and the carbon fluorine chain on the removal rate; wherein, a is a 3D diagram of the predicted results of the interaction between the average diameter of the resin and the carbon fluorine chain of PFAS on the removal rate; b is a 2D diagram of the predicted results of the interaction between the average diameter of the resin and the carbon fluorine chain of PFAS on the removal rate;

[0046] Fig.13 It is a schematic diagram of the predicted results of the interaction between the resin loaded with complex amine groups and the carbon fluorine chain on the removal rate; wherein, a is a 3D diagram of the predicted results of the interaction between the resin loaded with complex amine groups and the carbon fluorine chain of PFAS on the removal rate; b is a 2D diagram of the predicted results of the interaction between the resin loaded with complex amine groups and the carbon fluorine chain of PFAS on the removal rate;

[0047] Fig.14 This is a diagram of the internal structure of a computer device. DETAILED DESCRIPTION

[0048] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0049] The purpose of the present invention is to provide a resin synthesis method, system, medium and product based on machine learning. First, a data set of literature related to the removal of different types of PFAS by resin is mined and established, the factors affecting the removal rate are preprocessed, the SMILES descriptors of PFAS are encoded using molecular fingerprints, and the correlation between the performance enhancement of the ML model and the parameters of the molecular fingerprints is analyzed. An ML model is constructed based on a variety of conventional ML regression algorithms, several ML algorithms with the best performance are screened out, and the hyperparameter combination of the ML model, the encoding method of the molecular fingerprint and the parameters of the molecular fingerprint are adjusted using Bayesian optimization to determine the ML model with the best performance, and the contribution of the influencing factors to the removal of PFAS is quantified by calculating the SHAP value of each influencing factor to the label. Finally, the interaction between important features is visualized by a partial dependence graph, and important features related to the properties of the resin are screened out based on visual analysis, and the optimal synthesis conditions are obtained for the design of functionalized resins, and the synthesized resin is used to achieve efficient removal of PFAS.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example 1

[0052] like Figure 1 As shown, a resin synthesis method based on machine learning in this embodiment includes the following steps:

[0053] Step S1: Obtain multiple groups of influencing factors of PFAS removal rate; the influencing factors include resin properties, PFAS properties and adsorption system properties; the PFAS properties include SMILES descriptors of PFAS structural formula. Among them, the resin properties include average diameter (MD, μm), water content (MR, %), total exchange capacity (TC, eq / L), functional groups and polymer matrix; PFAS properties include PFAS carbon fluoride chain length and SMILES descriptors of PFAS structural formula; adsorption system properties include PFAS initial concentration (C 0 , ppm), resin dosage (M resin , g / L), contact time (A_time, h), equilibrium time (CT, min), solution pH and temperature (A_tem, ℃), solution matrix and concentration (mg / L) and adsorption oscillation frequency (rpm).

[0054] Step S2: Based on the multiple groups of influencing factors, the PFAS adsorption model is used to obtain the PFAS removal efficiency of multiple resins for different carbon fluorine chain lengths. The calculation of the removal efficiency (Remove efficiency, RF, %) of the resin for PFAS with different carbon fluorine chain lengths is shown in formula (1):

[0055] Remove efficiency=(C 0 -C e ) / C 0 ×100% (1)

[0056] Among them, C 0 represents the initial concentration of PFAS, C e It indicates the equilibrium concentration (ppm) of PFAS at adsorption equilibrium.

[0057] Step S3: Screening the PFAS removal rates of the resins with different carbon fluorine chain lengths to obtain the optimal PFAS removal rate.

[0058] Step S4: Based on the optimal PFAS removal rate, determine the optimal influencing factors.

[0059] Step S5: synthesizing resin based on the optimal influencing factors, and removing PFAS through the synthesized resin.

[0060] Furthermore, after step S1, the method further includes: preprocessing the multiple groups of influencing factors; the preprocessing the multiple groups of influencing factors specifically includes:

[0061] The multiple groups of influencing factors are processed as dummy variables to obtain numerical variables of resin properties, numerical variables of PFAS properties and numerical variables of adsorption system properties.

[0062] The molecular fingerprint encoding method is used to convert the numerical variables of PFAS properties into molecular fingerprints. Specifically, the SMILES descriptors of PFAS structural formulas are converted into molecular fingerprints.

[0063] Specifically, the text variables corresponding to the multiple groups of influencing factors (such as solution matrix, functional groups of resin, polymer matrix, average diameter and water content, etc.) are converted into numerical variables through dummy variable processing, and then the SMILES descriptors of the PFAS structural formula are converted into molecular fingerprints in Python based on the RDKit software package; wherein, the command: "AllChem.GetMorganFingerprintAsBitVect" can be used to obtain the corresponding binary molecular fingerprints (Binary Morgan fingerprints, BMF); the following command: "AllChem.GetHashedMorganFingerprint" can be used to obtain the corresponding count molecular fingerprints (Count-based Morgan fingerprint, CMF).

[0064] Furthermore, the training process of the PFAS adsorption model specifically includes the following steps:

[0065] Step S201: construct a data set, and divide the data set into a training set and a test set; the data set includes experimental influencing factors and the corresponding resin removal rate of PFAS with different carbon fluorine chain lengths. Among them, by mining and establishing a data set of relevant literature based on resin adsorption of perfluorooctanoic acid (PFOA), perfluorooctane sulfonic acid (PFOS), perfluorohexanoic acid (PFHxA), perfluorohexane sulfonic acid (PFHxS), perfluorobutyric acid (PFBA) and perfluorobutane sulfonic acid (PFBS) with different chain lengths of per- and polyfluoroalkyl substances (PFAS), the data set is split into a training set and a test set at a ratio of 8:2.

[0066] Step S202: construct an initial model based on a plurality of different machine learning algorithms, and train and test each of the initial models through a training set and a test set to determine the performance of each of the initial models; wherein, eight algorithms based on random forest (RF), support vector machine (SVM), neural network (MLP), K nearest neighbor method (KNN), ridge regression (Ridge), gradient boosting decision tree (GBDT), category gradient boosting (CatBoost) and extreme gradient boosting (XGBoost) are used to construct the initial model.

[0067] Step S203: Screening the initial models based on their performance to obtain multiple initial optimization models; wherein the model performance is determined by the coefficient of determination (R 2 ), root mean square error (RMSE) and mean absolute error (MAE), and the three best performing ML algorithms (RF, CatBoost and XGB) were selected based on the results of R2, RMSE and MAE of the comprehensive ML model.

[0068] Step S204: optimizing and adjusting the hyperparameters, molecular fingerprint encoding methods and parameters of the molecular fingerprint of each of the initial optimization models through the Bayesian algorithm to obtain the Bayesian optimization results corresponding to each of the initial optimization models; the molecular fingerprint encoding methods include a binary molecular fingerprint encoding method and a counting molecular fingerprint encoding method; the parameters of the molecular fingerprint include a molecular fingerprint length and a molecular fingerprint radius; the Bayesian optimization results include an optimal hyperparameter combination, an optimal molecular fingerprint encoding method and parameters of the optimal molecular fingerprint.

[0069] Specifically, the adjustment process is as follows: first, the SMILES symbols of PFAS are converted into molecular fingerprints, then the objective function is defined, the mean root square error (RMSE) on the validation set is set as the target loss, and the ML regression model is constructed. A five-fold cross-validation is performed on the training set and the hyperparameter space of the model is defined (i.e., a space containing any possible hyperparameter values), as shown in Table 1. Subsequently, Bayesian optimization is performed, and the best hyperparameter combination is obtained by iterating 1000 times to minimize the target loss. The final result of Bayesian optimization is determined by the RMSE (RMSE validation set and RMSE training set) on the validation set and training set under different hyperparameter combinations obtained by five-fold cross-validation, as well as the fitting error (RMSE validation set and RMSE training set). By optimizing each initial optimization model with Bayesian optimization, the hyperparameters of the corresponding ML algorithm are adjusted, and the overfitting problem of the ML model is alleviated.

[0070] Table 1 Model hyperparameter tuning ranges corresponding to different ML algorithms

[0071]

[0072]

[0073] Step S205: Screen the initial optimization model according to the Bayesian optimization result to obtain the optimal model; the optimal model is the PFAS adsorption model. The Bayesian optimization results of the ML models constructed by the RF, CatBoost and XGBoost algorithms are compared, and the ML algorithm with the best performance is selected. The ML model is constructed using the best hyperparameter combination, the optimal molecular fingerprint encoding method and the optimal molecular fingerprint parameters. The quality of the final training and testing results of the model is determined by R 2 , RMSE and MAE jointly determine that the performance of the ML model built based on the XGBoost algorithm is the best.

[0074] Preferably, the optimal molecular fingerprint encoding method is a counting molecular fingerprint encoding method.

[0075] Specifically, the BMF and CMF encoding methods were used for the PFAS molecular structure to obtain the molecular fingerprints of different PFAS structures. Based on the molecular fingerprints of different PFAS structures, the average dice similarity was calculated using formula (2) to obtain the structural differences of different PFAS.

[0076]

[0077] Among them, S A,B represents the average dice similarity, x jA represents the jth feature of molecule A, x jB represents the j-th feature of molecule B, and n represents the number of molecules.

[0078] Further references Figure 2 It can be seen that different molecular fingerprint encoding methods have quite different representations of the same PFAS molecular structure, while the representation results based on BMF are the same for the representation of PFAS molecular structures of the same type and different numbers of groups, and CMF can usually represent information including the number of molecular groups, among which the counting information of atomic groups plays a vital role in certain activities and properties. Therefore, it is speculated that the counting information of atomic groups may help improve the training and prediction performance of ML models. Among them, Figure 2 The horizontal and vertical axes are the names of different PFAS molecular types. Specifically, F-53B is chlorinated polyfluoroalkyl ether sulfonic acid; FTS is fluorotelomer sulfonate; PFUnDA is perfluoroundecanoic acid; PFDoDA is perfluorododecanoic acid; PFNA is perfluorononanoic acid; PFO5DoA is perfluoro(3,5,7,9,11-pentaoxadodecanoic acid); FTOH is fluorotelomer alcohol; FBSA is perfluorobutylsulfonamide; FHxSA is perfluorohexanesulfonamide; FOSA is perfluoroalkylsulfonamide; PFPrS is perfluorobutanesulfonic acid; NVHOS is perfluoroethoxysulfonic acid; PEPA is C 16 H 16 F 2 N 2 O 4 S 2 ;TAmPr-FHxSA is N-trimethylaminopropyl perfluoroconch sulfonamide; 5:1:2FTB is 2-[(3,4,4,5,5,6,6,7,7,8,8,8-difluorooctyl)dimethylamino]acetate; AmPr-FHxSA is N-dimethylaminopropyl perfluorohexane sulfonamide; 6:2FTSA-PrB is 6:2 fluorotelomer sulfonic acid; GenX is perfluoro(2-methyl-3-oxohexanoic acid)ammonium; N-E tFOSAA is N-ethylperfluoro-1-octanesulfonic acid aminoacetic acid; PFOA is perfluorooctanoic acid; PFDA is perfluorodecanoic acid; PFHpA is perfluoroheptanoic acid; PFHxA is perfluorohexanoic acid; PFBA is perfluorobutyric acid; PFPeA is perfluoropentanoic acid; PFHpS is perfluorobutanesulfonic acid; PFHxS is perfluorohexanesulfonic acid; PFOS is perfluorooctanesulfonic acid; PFPeS is fluorotelomer phosphate monoester; PFBS is perfluorobutanesulfonate.

[0079] Furthermore, the effects of different molecular fingerprint encoding methods on the performance of the ML model were analyzed. Table 2 shows the Bayesian optimization results of the ML model corresponding to the PFAS structure represented by BMF and CMF. As shown in Table 2, the CatBoost model constructed using the PFAS molecular fingerprint obtained based on CMF performed best (the fitting error between the training set and the validation set was the smallest: RMSE Val -RMSETra =9.835-5.542=4.293), and the Bayesian optimization results of the CatBoost model show that when the radius of CMF is 2 and the length is 2172, the corresponding ML model performs best. Therefore, in the specific embodiment of the present invention, CMF is finally used to represent the PFAS molecular structure and construct the CatBoost model.

[0080] Table 2 Bayesian optimization results of different ML models corresponding to different molecular fingerprint encoding methods

[0081]

[0082]

[0083] Furthermore, the influence of different molecular fingerprint parameters on the performance of ML models is analyzed. Specifically, Morgan fingerprint encodes the atomic group of a chemical substance into a binary vector with length and radius as its two parameters. The length refers to the length of the vector, while the radius represents the size of the atomic group. The larger the radius, the larger the atomic group it represents. Each bit in the vector represents a unique or single atomic group. The structural information of the molecule is represented by flexibly adjusting the radius and length of the molecular fingerprint. Since the RDKit package can easily generate the Morgan fingerprint of a molecule, the package can also be used to obtain the atomic group in each bit. Based on this, ML models based on different molecular fingerprints were developed on the dataset. After obtaining the RMSE on the ML model test set under molecular fingerprints of different lengths and radii, the measured ML model performance enhancement was obtained based on formula (3), thereby obtaining the performance enhancement achieved by CMF under different lengths and different radii.

[0084] Enhancement(%)=RMSE test (BMF)-RMSE test (CMF) / RMSE test (BMF)×100% (3)

[0085] Further references Figure 3 , Figure 3 When the radius of the molecular fingerprint changes from 0 to 5, the different molecular fingerprint lengths enhance the performance of the ML model. The results show that the combination of molecular fingerprint length and radius will affect the integrity and accuracy of molecular structure information, thereby affecting the performance of the ML model. Therefore, the parameter values ​​of the molecular fingerprint that give the best model performance are further adjusted through Bayesian optimization.

[0086] Furthermore, step S3 specifically includes the following steps:

[0087] Step S31: Calculate the Shapley value of each influencing factor based on the PFAS removal rate of each resin for different fluorocarbon chain lengths.

[0088] Step S32: According to the Shapley value of each of the influencing factors, a single factor SHAP diagram of different carbon fluorine chain lengths versus PFAS removal rate is drawn to obtain the carbon fluorine chain length range with the optimal PFAS removal rate.

[0089] Step S33: Based on the PFAS removal rates of the resins with different carbon fluorine chain lengths, a single factor SHAP diagram of the resin properties versus the PFAS removal rate is drawn to obtain the resin property characteristics with the best PFAS removal rate.

[0090] Step S34: Determine the optimal PFAS removal rate based on the carbon fluorine chain length range with the optimal PFAS removal rate and the resin property characteristics with the optimal PFAS removal rate.

[0091] Specifically, the SHAP method is used to calculate the Shapley value of each influencing factor, and the importance ranking of the influencing factors is calculated. The working principle of the SHAP method is to check the difference in the predicted results before and after removing the corresponding influencing factors. The Shapley value calculation of the influencing factor x (among N influencing factors) of the predicted p given by the established ML model is shown in Formula 4:

[0092]

[0093] in, represents the Shapley value, N represents the set of all influencing factors in the training set; n represents the dimension of N; S represents the subset of all influencing factors with influencing factor x; p(SUx) represents the predicted value of the established ML model considering influencing factor x; p(S) represents the predicted value without considering influencing factor x.

[0094] Based on the trained ML model and the calculated SHAP values ​​of various influencing factors on the label (removal rate), a single-factor SHAP diagram of the PFAS carbon-fluorine chain on the removal rate was drawn to visualize the effect of the PFAS carbon-fluorine chain length on the removal rate.

[0095] Based on the trained ML model, partial dependence plots were drawn to visualize the relationship between the carbon fluorocarbon chain length and resin properties (polymer matrix, functional groups, and average diameter) of PFAS, and to analyze the degree of correlation between characteristic interactions and removal rates. The carbon fluorocarbon chain length ranged from 3 to 12, and the resin properties included polymer matrix (polyacrylic acid and polystyrene), average resin particle size, and functional groups (complex amine groups, polymeric amine groups, tertiary amine groups, quaternary amine groups, and sulfonic acid groups).

[0096] Technical effects of the present invention:

[0097] 1) The hyperparameters of machine learning models built based on multiple algorithms were optimized by Bayesian optimization to select the best performing ML algorithm, and the models were screened based on the Bayesian optimization results including the best hyperparameter combination, the optimal molecular fingerprint encoding method, and the parameters of the optimal molecular fingerprint. The resulting PFAS adsorption model had high prediction accuracy and good prediction effect.

[0098] 2) The present invention can greatly save experimental process and cost. It does not need to conduct experiments to complete the determination of the removal rate of PFAS by the resin. Instead, it only needs to input the influencing factors into the trained PFAS adsorption model to obtain the corresponding removal rate of PFAS by the resin, determine the various properties of the resin, and prepare the functional resin, and then use the prepared resin to remove PFAS. Compared with traditional trial and error experimental exploration, this method can guide the design and optimization of functional resins through more targeted ML model predictions, providing an effective solution for the efficient removal of PFAS.

[0099] 3) The PFAS adsorption model established based on ML in the present invention can effectively predict the adsorption of PFAS with different carbon chain lengths by the resin, realize the rational design and efficient functional preparation of the resin, provide a reference for the establishment of various pollutant adsorption models, and realize the application of ML and statistical methods in analyzing the structural characteristics and application scenarios of persistent organic pollutants.

[0100] Example 2

[0101] The effectiveness of the resin synthesis method based on machine learning in the above-mentioned Example 1 of the present invention is further verified by Example 2.

[0102] First, if Figure 4 As shown, the changing pattern of the resin's removal effect on PFAS with different carbon fluorine chain lengths was preliminarily explored. Figure 4 (a) in the figure shows the average removal rate of the resin for perfluoroalkyl and polyfluoroalkyl substances and various perfluorocarboxylic acids (PFCA) and perfluorosulfonic acids (PFSA) compounds. Among various PFAS compounds, the removal rates of the resin for PFCA and PFSA are comparable. For PFCA compounds, the removal rate seems to decrease with the decrease of the PFAS carbon fluorine chain length, especially between the two fluorine carbon chain orders of magnitude. ( Figure 4 The numbers in brackets after the names of different types of PFAS in the horizontal axis represent the carbon fluorine chain length of PFAS. It can be seen that the carbon fluorine chain lengths of PFAS are significantly different when they are 8, 6, and 4. Further analysis found that similar change trends were observed for PFSA compounds as for PFCA, but the removal rates of PFBS, PFHxS, and PFOS did not change much. Figure 4(b) shows the effect of the polymer matrix, pore structure and functional groups of the resin on the PFAS removal rate. It can be seen that the different properties of the resin have a regular effect on the removal of PFAS, among which the macroporous resin has a significantly higher PFAS removal effect than the gel-type resin. The above results further show that the complex resin-PFAS-adsorption condition interaction has a great influence on the removal and chemical properties of PFAS compounds.

[0103] Further, Figure 5 The following are the results of Bayesian optimization under five-fold cross validation. During the optimization process, 1,000 sets of hyperparameters were selected using the Bayesian optimization algorithm, and then tested to obtain the average RMSE of the 1,000 training sets (RMSE training set) and the average RMSE of the validation set (RMSE validation set). The RMSE training set and RMSE validation set initially decreased, and finally stabilized or even increased. The black dotted line in the figure represents the minimum value of the RMSE validation set, C represents the RMSE value of the validation set, and D represents the difference between the RMSE of the validation set and the RMSE of the training set (fitting error). Figure 5 (b) shows the zoomed-in pattern around the minimum RMSE validation set, with C1 marking the hyperparameter combination that achieves the lowest RMSE validation set, and C2 marking the hyperparameter combination that achieves the fifth lowest RMSE validation set. The degree of the RMSE downward trend is much higher before position C2 than after position C2. Specifically, from C1 to C2, as the RMSE training set increases from 2.402 to 3.880, it only causes the RMSE validation set to rise from 9.567 to 9.625, while the corresponding fitting error (RMSE validation set - RMSE training set) decreases from 7.165 to 5.744. In other words, the ML model fits more noise or irrelevant information from C2 to C1, and C1 should have a more serious overfitting problem than C2. Therefore, in the present invention, we plot the hyperparameter adjustment results to select hyperparameters to control overfitting, rather than directly selecting the hyperparameter combination that achieves the minimum RMSE validation set. As shown Figure 6 As shown in the figure, by comparing the random forest, CatBoost, and XGBoost algorithms, it is found that the XGB model built based on the optimal hyperparameter combination performs best. 2 , RMSE and MAE are 0.906, 9.875 and 6.755 respectively.

[0104] Furthermore, after validating the ML models on the training and validation sets, the SHAP interpretation method was used to interpret the XGBoost regression model to clarify how the influencing factors affect the removal rate. Figure 7 As shown, Figure 7 The SHAP plot of the CMF-based model. The X-axis has Shapley values, where positive values ​​indicate Figure 7 (a) and Figure 7 The specific influencing factors listed in (b) can increase the removal rate, while negative values ​​indicate a reduction in the removal rate. Each influencing factor is a feature. The size of each feature value changes from blue to red, corresponding to the minimum and maximum values. The pattern of each feature is composed of small dots, and each dot represents a sample containing the feature. The most critical factors for predicting removal rate are resin dosage, contact time, initial PFAS concentration, adsorption oscillation frequency, PFAS fluorocarbon chain length, average diameter of the resin, and polyacrylic acid matrix. Figure 7 The results in (a) show that the characteristic values ​​of resin dosage, contact time, carbon fluoride chain length, deionized water matrix and fp_788 (CF) are positively correlated with the predicted values, and higher characteristic values ​​are conducive to the removal of PFAS. In contrast, the initial concentration of PFAS, adsorption oscillation frequency, polyacrylic acid matrix and higher characteristic values ​​of fp_1080 (CH) are usually distributed on the left and negatively correlated with the predicted values.

[0105] PFAS head groups and chain lengths affect PFAS removal through electrostatic interactions. In previous studies, the comparison of removal efficiency was based on the total number of carbons rather than the total number of fluorinated carbons, which limited the understanding of the effects of head groups and fluorinated chain lengths on PFAS removal by anion exchange adsorbents. Therefore, the present invention uses SHAP values ​​to visualize the feature importance ranking of carbon-fluorinated chain lengths in the range of 3-6, referring to Figure 7 In (b), it was found that the importance of the polypropylene matrix and the average diameter of the resin was ranked higher than Figure 7 The feature ranking in (a) is different from that in (a), and there is also a big difference between the molecular fingerprint ranking of different groups of short-chain PFAS and the SHAP value ranking corresponding to the total PFAS removal. Among them, the SHAP values ​​of PFAS groups fp_623 (sulfonic acid group) and fp_587 (carboxyl group) are significantly improved, which means that these two groups have a more targeted effect on the removal of short-chain PFAS by the resin.

[0106] Then, a visual analysis of the importance of single factors was conducted on the effects of PFAS carbon fluorine chains and resin properties on the removal rate. The results are as follows: Figure 8 and Fig. 9 shown. Figure 8 This is the changing trend of the effect of PFAS carbon fluoride chain length on the removal rate. As the carbon fluoride chain length increases from 3 to 12, the SHAP value first increases and then decreases, and reaches the highest in the range of carbon fluoride chain lengths of 8 and 9. This result indicates that the resin has a significantly higher removal effect on long-chain PFAS (such as PFOA, PFOS and PFDA) than short-chain PFAS (such as PFBA and PFBS). Fig. 9A visual analysis was conducted to investigate how the polymer matrix and functional groups of the resin affect the removal of PFAS by the resin. The results showed that resins with complex groups and polypropylene matrix were more conducive to the removal of PFAS.

[0107] In order to further analyze how the interaction between features affects the removal of PFAS by the resin, 3D and planar graphs of the partial dependence diagram were drawn to predict the effect of the interaction between the two on the removal rate based on the carbon fluorine chain length of PFAS and the properties of the resin. Figure 10-13 As shown in the figure, as the length of the carbon fluorine chain increases, the interaction between the characteristics promotes the removal rate. When the carbon fluorine chain length is greater than 8, the removal rate increases significantly. The effects of the polymer matrix and functional groups of the resin on the removal rate are also consistent with the results of the single factor importance analysis obtained above. And as the length of the carbon fluorine chain increases, the resin with an average particle size in the range of 430-697nm has a better removal effect on PFAS.

[0108] The above results indicate that resin dosage, contact time, initial PFAS concentration, adsorption oscillation frequency, carbon fluoride chain length and resin polymer matrix are the main influencing factors that determine the removal of PFAS by the resin, that is, the optimal influencing factors. Therefore, when preparing the resin, adjustments can be made from several aspects such as resin dosage, contact time, initial PFAS concentration, adsorption oscillation frequency, carbon fluoride chain length and resin polymer matrix. The resin can be synthesized based on the optimal influencing factors, and PFAS can be removed by the synthesized resin.

[0109] Example 3

[0110] A computer system comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the resin synthesis method based on machine learning in Example 1.

[0111] Example 4

[0112] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the resin synthesis method based on machine learning in Example 1.

[0113] Example 5

[0114] A computer program product includes a computer program, which implements the resin synthesis method based on machine learning in embodiment 1 when executed by a processor.

[0115] Example 6

[0116] A computer device, which may be a database, may have an internal structure as shown in Fig.14As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Wherein, the processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. Wherein, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the resin synthesis method based on machine learning in Example 1 is implemented.

[0117] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided by the present invention may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited thereto. The processor involved in each embodiment provided by the present invention may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited thereto.

[0118] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0119] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A resin synthesis method based on machine learning, characterized in that: The method comprises: Obtain multiple groups of influencing factors of PFAS removal rate; the influencing factors include resin properties, PFAS properties and adsorption system properties; the PFAS properties include SMILES descriptors of PFAS structural formulas; Based on multiple groups of influencing factors, the PFAS adsorption model was used to obtain the PFAS removal rates of multiple resins for different carbon-fluorine chain lengths; The PFAS removal rates of the resins with different carbon fluorine chain lengths were screened to obtain the optimal PFAS removal rate; Based on the optimal PFAS removal rate, determining optimal influencing factors; Synthesize a resin based on the optimal influencing factors, and remove PFAS through the synthesized resin; The training process of the PFAS adsorption model specifically includes: Constructing a data set and dividing the data set into a training set and a test set; the data set includes experimental influencing factors and corresponding resin removal rates of PFAS with different carbon fluorine chain lengths; Constructing initial models based on a plurality of different machine learning algorithms, and training and testing each of the initial models through a training set and a test set to determine the performance of each of the initial models; Screening each of the initial models based on the performance of each of the initial models to obtain a plurality of initial optimization models; The hyperparameters, molecular fingerprint encoding methods and molecular fingerprint parameters of each of the initial optimization models are optimized and adjusted by the Bayesian algorithm to obtain the Bayesian optimization results corresponding to each of the initial optimization models; the molecular fingerprint encoding methods include a binary molecular fingerprint encoding method and a counting molecular fingerprint encoding method; the parameters of the molecular fingerprint include a molecular fingerprint length and a molecular fingerprint radius; the Bayesian optimization results include an optimal hyperparameter combination, an optimal molecular fingerprint encoding method and optimal molecular fingerprint parameters; The initial optimization model is screened according to the Bayesian optimization results to obtain the optimal model; the optimal model is the PFAS adsorption model.

2. The resin synthesis method based on machine learning according to claim 1, characterized in that: After obtaining multiple groups of influencing factors of PFAS removal rate, it also includes pre-processing the multiple groups of influencing factors; The preprocessing of the multiple groups of influencing factors specifically includes: Performing dummy variable processing on multiple groups of the influencing factors to obtain numerical variables of resin properties, numerical variables of PFAS properties, and numerical variables of adsorption system properties; Molecular fingerprint encoding method was used to convert the numerical variables of PFAS properties into molecular fingerprints.

3. The resin synthesis method based on machine learning according to claim 2, characterized in that: The optimal molecular fingerprint encoding method is a counting molecular fingerprint encoding method.

4. The resin synthesis method based on machine learning according to claim 1, characterized in that: The removal rates of PFAS with different chain lengths of each resin were screened to obtain the optimal PFAS removal rate, specifically including: Based on the PFAS removal rate of each resin for different carbon fluorine chain lengths, the Shapley value of each influencing factor was calculated; According to the Shapley values ​​of the influencing factors, a single factor SHAP diagram of different carbon fluorine chain lengths on PFAS removal rate was drawn to obtain the carbon fluorine chain length range with the best PFAS removal rate; Based on the PFAS removal rates of the resins with different carbon fluorine chain lengths, a single factor SHAP diagram of the resin properties versus the PFAS removal rate was drawn to obtain the resin property characteristics with the best PFAS removal rate; The optimal PFAS removal rate is determined based on the carbon fluorochain length range with the optimal PFAS removal rate and the resin property characteristics with the optimal PFAS removal rate.

5. A computer system comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the resin synthesis method based on machine learning according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the resin synthesis method based on machine learning described in any one of claims 1 to 4 is implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the resin synthesis method based on machine learning described in any one of claims 1 to 4 is implemented.

Images