Atmospheric reaction kinetic parameter intelligent prediction method and system fusing multilevel quantum chemical characterization

Abstract

The invention discloses an atmospheric reaction kinetic parameter intelligent prediction method and system fusing multilevel quantum chemical characterization. The method comprises the following steps: acquiring experimental measurement data, respectively acquiring a descriptor table data set and a plurality of quantum chemistry descriptors according to the experimental measurement data, further respectively acquiring a new feature set serving as model input and a reaction rate constant data set serving as model output, and then acquiring an optimal machine learning model according to the input and the output, and the optimal machine learning model is used for predicting the reaction rate constant of the reaction of the organic compound to be detected and the hydroxyl radical. According to the method, parallel training is carried out by integrating models of multiple decision trees, the complex nonlinear relation between molecular structure parameters and reaction kinetics is effectively analyzed, compared with a traditional regression model, the analysis capacity of high-dimensional data such as interpretable quantum chemistry descriptors is remarkably improved, and the calculation complexity is greatly reduced while the prediction precision is guaranteed.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and environmental chemistry, and specifically relates to an intelligent prediction method and system for atmospheric reaction dynamics parameters that integrates multi-level quantum chemical characterization. Background Technology

[0002] Currently, the gas-phase reaction of volatile organic compounds (VOCs) with oxidants such as hydroxyl radicals (OH) is the main pathway for their removal from the atmosphere, with a bimolecular rate coefficient k. OH The structure-activity relationship (QSAR) determines the lifespan of VOCs and influences the formation of secondary organic aerosols and ozone. Due to instrumental limitations in experimental measurements and the time-consuming and expensive nature of high-level quantum chemical calculations, quantitative structure-activity relationship (QSAR) models have become a key solution balancing accuracy and efficiency. With the development of artificial intelligence, research trends in this field have shifted from early linear algorithms to nonlinear machine learning algorithms. Nonlinear algorithms demonstrate stronger generalization ability and robustness when handling large-scale, diverse datasets, but their inherent "black box" nature leads to a lack of transparency in prediction results. Especially in atmospheric chemistry research requiring clear reaction mechanisms and scientific evidence, traditional characterization systems often focus on macroscopic physicochemical properties or two-dimensional topological structures, lacking in-depth characterization of the microscopic local electronic effects crucial to reaction rates, making it difficult for models to capture the deep-seated reaction patterns of complex molecules. Therefore, there is an urgent need to develop a prediction framework that combines multi-level quantum chemical characterization. By introducing microscopic electronic features with clear physical meaning, this framework can compensate for the deficiencies of existing characterization systems, ensuring high accuracy of nonlinear algorithms while achieving model interpretability, thus overcoming the shortcomings of current technologies.

[0003] Patent CN202410354797.9 discloses a method for predicting the chlorination degradation rate of pollutants based on machine learning. The process involves obtaining the second-order reaction rate constants of compound standards through chlorination degradation experiments, calculating molecular descriptors to construct a basic database, and then training a prediction model. The drawbacks of this method are: 1. It only inputs molecular descriptors into the prediction model, lacking a deep characterization of local electronic effects; 2. Chlorination degradation is a typical electrophilic or oxidation reaction, but the model lacks information on the electronic reactivity of reaction sites; 3. Its rate constants are derived from specific chlorination systems and operating conditions, and model predictions can usually only be reliably extrapolated in similar systems.

[0004] Patent CN202010380633.5 discloses a method for predicting the reaction rate constant of organic matter with singlet oxygen in aqueous solution under different pH conditions. The process involves collecting rate constant data at different pH values, dividing the data into training and test sets, performing statistical regression to establish and validate a QSAR model, inputting the molecular structure to be tested to obtain the predicted rate constant, and simultaneously distinguishing between molecular and ionic states and providing corresponding models. The disadvantages of this method are: 1. It uses a multiple linear regression algorithm, which may limit the model's interpretability and predictive reliability when dealing with complex nonlinear relationships due to the linear assumption; 2. It requires specific commercial software to obtain descriptor data, resulting in a high computational threshold and time consumption, which is not conducive to rapid, high-throughput prediction.

[0005] Patent CN202311324827.3 discloses a method for predicting the atmospheric oxidation reaction rate constant of volatile organic compounds (VOCs) based on graph neural networks. The process involves converting the SMILES encoding of VOCs and oxides into molecular graphs, inputting them into a graph neural network, and outputting predicted reaction rate constant values. The drawbacks of this method are: 1. Feature extraction lacks physical meaning, and its "black box" nature leads to opaque mechanisms; 2. The characterization of microscopic electronic effects is vague, limiting its ability to identify complex isomers. Summary of the Invention

[0006] To address the problems existing in the background art, this invention provides an intelligent prediction method and system for atmospheric reaction dynamics parameters that integrates multi-level quantum chemical characterization, solving the technical problems of high cost due to reliance on experimental measurement, limited characterization of microscopic electronic effects, and insufficient interpretability of model mechanisms in the background art.

[0007] The technical solution adopted in this invention is: I. An intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization: S1. Collect experimental data on the reaction of organic compounds with hydroxyl radicals, and preprocess the experimental data to obtain preprocessed experimental data.

[0008] S2. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data. Obtain the descriptor table dataset based on the SMILES structural formula and the reaction rate constant dataset based on the reaction rate constant.

[0009] S3. Extract the molecular structure formula of each organic compound from the pretreated experimental data, and calculate multiple quantum chemical descriptors corresponding to each organic compound based on the molecular structure formula.

[0010] The quantum chemical descriptor includes the local electron attachment energy, the maximum electrophilic Fukui index, and the frontier orbital gap.

[0011] S4. After using multiple machine learning algorithms to perform feature selection on the descriptor table dataset, a feature subset corresponding to each machine learning algorithm is obtained. Then, all quantum chemical descriptors are added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm.

[0012] S5. Using each new feature set as input and the reaction rate constant dataset as output, train each corresponding machine learning model to obtain the optimal machine learning model, and obtain the predicted feature set based on the optimal machine learning model.

[0013] S6. Based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, further obtain the value of each feature in the prediction feature set of the organic compound to be tested, and input all feature values ​​into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested.

[0014] S7. During the prediction process, the SHAP interpreter is called to calculate the SHAP value of each feature in the feature set, and the features are sorted in descending order according to the size of the SHAP value to obtain the importance of each feature.

[0015] The preprocessing includes cleaning up invalid values ​​and filling in missing values; cleaning up invalid values ​​involves finding invalid values ​​in the descriptor data and replacing them with null values; filling in missing values ​​involves filling in the missing values ​​in the experimental measurement data with null values.

[0016] Step S2 specifically involves: S21. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data.

[0017] S22. Generate corresponding descriptor data for the SMILES structure of each organic compound, and arrange the descriptor data corresponding to each organic compound in rows to obtain a preliminary descriptor table dataset.

[0018] S23. Perform dimensionality reduction and Z-score standardization on the initial descriptor table dataset to obtain the descriptor table dataset.

[0019] S24. After performing a logarithmic transformation on the reaction rate constants of each organic compound, the reaction rate constants dataset is obtained.

[0020] The descriptor data for each organic compound includes several molecular descriptors, each of which is a feature; each column of the descriptor table dataset corresponds to one molecular descriptor.

[0021] The dimensionality reduction process involves first deleting all-zero columns and columns containing null values ​​from the descriptor table dataset, and then sequentially processing the initial descriptor data using methods to remove low-variance features and high-correlation features.

[0022] Step S4 specifically involves: S41. Using the descriptor table dataset as input and the reaction rate constant dataset as output, various machine learning algorithms are employed, with each machine learning model serving as the base model for the recursive feature elimination method. Feature selection is then performed on the descriptor table dataset to obtain the feature subsets corresponding to each machine learning algorithm.

[0023] S42. Treat each quantum chemical descriptor as a single feature, and add all three features together to the feature subset corresponding to each machine learning algorithm, thereby obtaining the new feature set after adding each machine learning algorithm.

[0024] Step S5 specifically involves: S51. Using each new feature set as input and the reaction rate constant dataset as output, train the corresponding machine learning model to obtain each trained machine learning model and its corresponding evaluation index value.

[0025] S52. Select the optimal evaluation index value from all evaluation index values. The machine learning model trained with the optimal evaluation index value is taken as the optimal machine learning model.

[0026] S53. The new feature set corresponding to the optimal machine learning model is taken as the optimal new feature set, and all features included in the optimal new feature set are taken as the prediction feature set.

[0027] Step S6 specifically involves: S61. Obtain the molecular structural formula and SMILES structural formula of the organic compound to be tested.

[0028] S62. Based on the molecular structure formula and SMILES structure formula, obtain the value corresponding to each feature in the predicted feature set of the organic compound to be tested, thereby obtaining the feature value set of the organic compound to be tested.

[0029] S63. Input the set of feature values ​​corresponding to the organic compound to be tested into the optimal machine learning model for prediction to obtain the corresponding reaction rate constant.

[0030] II. An intelligent prediction system for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization: The data acquisition and preprocessing module collects experimental measurement data of the reaction between organic compounds and hydroxyl radicals, and preprocesses the experimental measurement data to obtain preprocessed experimental measurement data.

[0031] The dataset construction module extracts the SMILES structure and reaction rate constant of each organic compound from the preprocessed experimental data, obtains the descriptor table dataset based on the SMILES structure, and obtains the reaction rate constant dataset based on the reaction rate constant.

[0032] The supplementary descriptor acquisition module extracts the molecular structure formula of each organic compound from the preprocessed experimental measurement data, and calculates multiple quantum chemical descriptors corresponding to each organic compound based on the molecular structure formula.

[0033] The new feature set acquisition module uses multiple machine learning algorithms to select features from the descriptor table dataset, obtaining a feature subset corresponding to each machine learning algorithm. Then, all quantum chemical descriptors are added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm.

[0034] The model training and prediction feature set acquisition module takes each new feature set as input and the reaction rate constant dataset as output to train each corresponding machine learning model, thereby obtaining the optimal machine learning model and acquiring the prediction feature set based on the optimal machine learning model.

[0035] The prediction deployment module further obtains the value of each feature of the organic compound to be tested in the prediction feature set based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, and inputs the values ​​of all features into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested.

[0036] The interpretability module calls the SHAP interpreter during the prediction process to calculate the SHAP value of each feature in the feature set, and sorts the features in descending order according to the size of the SHAP value, thereby obtaining the importance of each feature.

[0037] The beneficial effects of this invention are: 1. Synergistic Enhancement Mechanism of Quantum Chemical Descriptors: By introducing the maximum electrophilic Fukui exponent, frontier orbital gap, and local electron attachment energy (introduced for the first time in this invention), a multi-level three-dimensional characterization system for electronic reactivity, ranging from global electronic stability to local electronic microenvironment and then to specific reaction sites, is constructed, overcoming the shortcomings of existing technologies in characterizing microscopic electronic effects. The synergistic effect of these three elements significantly enhances the model's ability to identify complex substituted derivatives and structurally similar molecules. While improving prediction accuracy and robustness, it reveals the microscopic mechanism of the reaction at the quantum chemical level, endowing the model with stronger physical interpretability.

[0038] 2. Breakthrough in Intelligent Modeling of High-Dimensional Chemical Data: By integrating multiple decision trees and training them in parallel, the complex nonlinear relationship between molecular structure parameters and reaction kinetics is effectively analyzed. Compared to traditional regression models, this significantly improves the ability to analyze high-dimensional data such as quantum chemical descriptors, greatly reducing computational complexity while maintaining prediction accuracy.

[0039] 3. Innovation in stability prediction of heterogeneous systems: Develop a dynamic weighted system based on molecular structure characteristics to overcome the problem of predicting fluctuations among different categories of compounds such as alkanes, aromatics, and oxygen-containing VOCs using traditional models.

[0040] 4. Revolutionary Full-Process Automated Feature Engineering: A dual-driven intelligent screening system based on machine learning and physicochemical mechanisms is created, overcoming the subjective limitations of manual feature selection. Employing a collaborative mechanism of recursive feature elimination and interpretability analysis, it accurately captures key influencing factors from massive amounts of descriptors, establishing an efficient and reliable feature engineering paradigm.

[0041] 5. Multi-scenario adaptive modeling: Construct a transfer learning framework that couples environmental parameters, and through machine model optimization, effectively solve the adaptability problem of traditional models in multi-physics field scenarios such as high-temperature combustion and normal temperature atmosphere, significantly expanding the engineering application boundaries of the technology. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention.

[0043] Figure 2 The curves show the comparison of prediction results of the optimal machine learning model XGBoost on different feature sets in this embodiment of the invention.

[0044] Figure 3 This is a schematic diagram illustrating the importance of various features that affect the reaction of all organic compounds with hydroxyl radicals in the entire experimental measurement data of this invention.

[0045] Figure 4 This is a schematic diagram illustrating the degree of importance of various features that affect the reaction between the tested organic compound and hydroxyl radicals in embodiments of the present invention.

[0046] Figure 5 This is a molecular structural diagram of vanillin, the organic compound to be tested in this embodiment of the invention. Detailed Implementation

[0047] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0048] Explanation of the name: (1) SMILES is a simplified line input notation system for representing the structure of chemical molecules. It is a method of representing chemical molecules in the form of text strings and is widely used in the fields of cheminformatics and computational chemistry. SMILES can concisely and accurately describe the structure of organic molecules, including key information such as atoms, chemical bonds, branches and ring structures.

[0049] (2) Low-variance feature removal is a commonly used feature selection method to remove features with low variance from descriptor data. These features usually contain little information, contribute little to the model's predictive ability, and may even introduce noise, affecting the model's performance. By removing low-variance features, the dataset can be simplified, and the efficiency and accuracy of the model can be improved.

[0050] (3) The highly correlated feature removal method is a feature selection method used to identify and remove highly correlated features in descriptor data. Highly correlated features usually provide similar information, and retaining all of these features may lead to model overfitting, computational inefficiency, and even affect the interpretability of the model. By removing highly correlated features, the dataset can be simplified, and the performance and interpretability of the model can be improved.

[0051] (4) Recursive feature elimination is a model-based feature selection method. It gradually removes features that contribute the least to the model’s predictive ability through recursion until a preset number of features is reached, and finally retains a set of the most important features.

[0052] (5) Z-score standardization, also known as standard deviation standardization, is a commonly used data standardization method that transforms data into a distribution with a mean of 0 and a standard deviation of 1. This method eliminates the influence of dimensions, allowing data with different characteristics to be compared and analyzed on the same scale.

[0053] (6) SHAP interpreter is a tool for interpreting the output of machine learning models. It can help us understand the model’s prediction results and the contribution of each feature to the prediction results. SHAP is based on Shapley values ​​in game theory and can assign an important value to each feature in a specific prediction to explain the influence of the feature on the model’s prediction results.

[0054] (7) Random Forest algorithm, which belongs to the bagging method in ensemble learning. It is an algorithm that uses multiple decision trees to train and predict samples. It is also an algorithm that contains multiple decision trees. For classification problems, the output class is determined by the mode of the classes output by multiple decision trees. For regression problems, the simple averaging method is usually used to obtain the final model output by arithmetically averaging the regression results obtained from multiple decision trees.

[0055] (8) GBDT, Gradient Boosting Tree, is a boosting method in ensemble learning. It is a classification and regression algorithm based on decision trees. The underlying algorithm is based on the gradient descent algorithm of regression tree and function space. In addition to the advantages of tree model, such as strong interpretability, effective handling of mixed type features, scalability invariance, and robustness to missing values, it also has the advantages of strong predictive ability and good stability.

[0056] (9) XGBoost, Extreme Gradient Boosting Tree, is a boosting method in ensemble learning. It is an efficient, flexible, and highly optimized gradient boosting framework, a machine learning algorithm designed for large-scale data. It builds strong learners by integrating multiple weak learners (usually decision trees), and performs exceptionally well in tasks such as classification, regression, and ranking.

[0057] like Figure 1 As shown in this embodiment, an intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization includes the following steps: S1. Establish an experimental data acquisition channel to collect experimental data on the reaction of organic compounds with hydroxyl radicals, and preprocess the experimental data to obtain preprocessed experimental data.

[0058] Preprocessing includes cleaning outlier and invalid values ​​and filling missing values. Cleaning outlier and invalid values ​​involves finding outlier and invalid values ​​in the descriptor data and replacing them with null values. Filling missing values ​​involves filling missing values ​​in the experimental data with null values.

[0059] Each data record in the experimental measurements includes the name of the organic compound, its molecular structural formula, the structural formula of SMILES, experimental conditions, reaction rate constant, and measurement error range. Specifically, the organic compounds include alkanes, aromatic hydrocarbons, and oxygen-containing VOCs.

[0060] In this embodiment, experimental data on the reaction of organic compounds with hydroxyl radicals were obtained from AtmVOCkin-2.1.0, an authoritative publicly available database in the field of atmospheric chemical kinetics. The data used were derived from the CompoundList Info file of this database, containing 1366 experimental data points on the reaction of different categories of compounds (such as alkanes, aromatic hydrocarbons, and oxygen-containing VOCs) with hydroxyl radicals. Each record includes the compound name, SMILES structural formula, experimental conditions, reaction rate constant, and its error range. To ensure data quality, the experimental data were preprocessed to remove meaningless and missing data items.

[0061] S2. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data. Obtain the descriptor table dataset based on the SMILES structural formula and the reaction rate constant dataset based on the reaction rate constant.

[0062] Step S2 is as follows: S21. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data.

[0063] S22. Use tools such as RDKit or Mordred to generate corresponding descriptor data for the SMILES structure of each organic compound, and arrange the descriptor data corresponding to each organic compound in rows to obtain a preliminary descriptor table dataset.

[0064] The descriptor data for each organic compound includes several molecular descriptors, each of which is a feature; each column of the descriptor table dataset corresponds to a molecular descriptor (i.e., each column corresponds to a feature), such as molecular weight, lipid-water partition coefficient, various topological indices, etc.

[0065] S23. Perform dimensionality reduction and Z-score standardization on the initial descriptor table dataset to obtain the descriptor table dataset.

[0066] The dimensionality reduction process involves first deleting columns containing all zeros and null values ​​from the descriptor table dataset, and then sequentially processing the initial descriptor data using methods to remove low-variance features and high-correlation features.

[0067] In practice, the low-variance feature removal method eliminates descriptor data with low variance. The high-correlation feature removal method identifies and removes highly correlated descriptor data. Z-score normalization is used to eliminate the influence of dimensions.

[0068] In practice, the formula for calculating the correlation matrix between features using the method of removing highly correlated features is as follows: R ij=(∑ k=1 n ((X ik -X i - (X) jk -X j - ))) / ((∑ k=1 n (X ik -X i - ) 2 ) 1 / 2 (∑ k=1 n (X jk -X j - ) 2 ) 1 / 2 ) Among them, R ij The value representing the correlation between features, where n is the number of descriptors, and X... ik It is feature X i The value of X on sample k i - It is feature X i The average value, X jk It is feature X j The value of X on sample k j - It is feature X j The average value.

[0069] S24. After performing a logarithmic transformation on the reaction rate constants of each organic compound, the reaction rate constants dataset is obtained.

[0070] Because the reaction rate constant is very small, directly predicting data may span multiple orders of magnitude, making it difficult to optimize the model due to large numerical differences. Therefore, performing a logarithmic transformation on the reaction rate constant can expand the numerical range, make the data closer to a normal distribution, and reduce the impact of extreme values, which helps to improve the model's performance and predictive ability.

[0071] In practice, the order of organic compounds in each row of the descriptor table dataset corresponds one-to-one with the order of each organic compound in the reaction rate constant dataset, which facilitates the training of subsequent machine learning models.

[0072] S3. Extract the molecular structure formula of each organic compound from the pretreated experimental data, and calculate the local electron attachment energy LEAE and the maximum electrophilic Fukui index f(-) for each organic compound based on the molecular structure formula. x and frontline orbital energy gap E GAP These three quantum chemical descriptors.

[0073] Descriptor data generated based on SMILES structures are essentially two-dimensional topological parameters, limiting their ability to capture subtle electronic behaviors within molecules caused by differences in spatial structure or substituent positions. In contrast, the local electron attachment energy, maximum electrophilic Fukui exponent, and frontier orbital gap introduced in this invention are real physical quantities calculated from molecular structures. They can directly quantify the global electronic stability of a reaction, the local electronic microenvironment, and the electronic reactivity of specific reaction sites at the microscopic level. This effectively fills the gaps in traditional methods, elevating molecular characterization from a simple statistical analysis of the "static framework" to the level of "dynamic electronic mechanisms."

[0074] S4. After employing multiple machine learning algorithms and using the reaction rate constant dataset to perform feature selection on the descriptor table dataset obtained in step S23, a feature subset corresponding to each machine learning algorithm is obtained. Then, the local electron attachment energy LEAE and the maximum electrophilic Fukui exponent f(-) are selected. x and frontline orbital energy gap E GAP Each feature is added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm.

[0075] In practice, feature selection involves selecting and retaining features (molecular descriptors) from all columns of the descriptor table dataset, thereby obtaining a subset of the descriptor table dataset, i.e., a feature subset.

[0076] For a single descriptor table dataset: The descriptor table dataset is equivalent to a feature set containing multiple molecular descriptors. After feature selection (deleting some features), a feature subset is obtained. Then, the local electron attachment energy LEAE and the maximum electrophilic Fukui index f(-) are added to the feature subset. x and frontline orbital energy gap E GAP These three features are then used to obtain a new feature set.

[0077] Step S4 is as follows: S41. Using the descriptor table dataset obtained in step S23 as input and the reaction rate constant dataset as output, various machine learning algorithms are used, and each machine learning model is used as the base model of recursive feature elimination (RFE) to perform feature selection on the descriptor table dataset, thereby obtaining the feature subsets corresponding to each machine learning algorithm.

[0078] Machine learning models include RandomForest, Gradient Boosting Tree (GBDT), and Extreme Gradient Boosting Tree (XGBoost).

[0079] The specific implementation steps for obtaining the feature subset of a single machine learning model in step S41 are as follows: Suppose we have the following model inputs: a descriptor table dataset X. The target variable set for the model is Y: a reaction rate constant dataset.

[0080] The machine learning model is M, which is used to evaluate the importance of features. The number of features to be selected is k (usually less than m, where m is the number of features in the descriptor table dataset X).

[0081] Step 1: Set the current descriptor table dataset X as the initial feature set X current =X.

[0082] Step 2: Using feature set X current Train a machine learning model M with the target variable set Y.

[0083] Step 3: Obtain the importance weight w of each feature using the machine learning model M. i =split_count i w i This refers to the number of times a certain feature is used for node splitting across all decision trees. Features that are frequently used for splitting in machine learning models are more important.

[0084] Step 4: Based on importance weight w i Sort all features in descending order.

[0085] Step 5: Remove the bottom p features from the ranking comparison. The number of features p removed each time is dynamically determined, depending on the difference between the current number of features and the target number of features. Update the feature set X. current .

[0086] Step Six: If the feature set X current If the number of features in the subset equals k, then stop and obtain the final feature subset; otherwise, return to step two.

[0087] Assuming the XGBoost model is used as the base model, the specific process of feature selection for the processed descriptor data using the Recursive Feature Elimination (RFE) method is as follows: First, Recursive Feature Elimination (RFE) combined with XGBoost is used as the base model for initial feature selection. While adhering to physicochemical rationale, key features significantly impacting the target variable set Y are automatically identified. Based on this, the XGBoost model is further trained. During training, the weights are automatically and dynamically adjusted according to the contribution of features in each round of the decision tree structure; that is, the higher the importance of a feature, the more frequently it is used for node splitting, and the greater its weight. Through this feature importance-driven dynamic weighting mechanism, combined with multiple rounds of iterative training and evaluation, the XGBoost model structure and feature combinations are continuously optimized, enabling the XGBoost model to better adapt to the structural differences of different types of compounds and improving the overall stability and accuracy of predictions.

[0088] S42. Local electron attachment energy, maximum electrophilic Fukui index, and frontier orbital gap are respectively used as single-column features (molecular descriptors). All three columns of features are added to the feature subset corresponding to each machine learning algorithm, thereby obtaining the new feature set after the addition of each machine learning algorithm.

[0089] S5. Using each new feature set as input and the reaction rate constant dataset as output, train each corresponding machine learning model to obtain the optimal machine learning model, and obtain the predicted feature set based on the optimal machine learning model.

[0090] Step S5 is as follows: S51. Using each new feature set as input and the reaction rate constant dataset as output, train the corresponding machine learning model to obtain each trained machine learning model and its corresponding evaluation index value.

[0091] The evaluation index is a weighted sum of the root mean square error (MSE), goodness-of-fit R², and mean absolute error (MAE). In this embodiment, the evaluation index is set as goodness-of-fit R² - mean absolute error (MAE) - root mean square error (MSE).

[0092] The root mean square error (MSE), goodness-of-fit R², and mean absolute error (MAE) are set according to the following formulas: MSE=(1 / n)∑ i=1 n (y i -y i ^ ) 2 R 2 =1-((∑ i=1 n (y i -y i ^ ) 2 ) / (∑ i=1n (y i -y i - ) 2 )) MAE=(1 / n)∑ i=1 n |y i -y i ^ | Where n is the total number of samples, y i For the actual value, y i ^ For the predicted value, y i - This is the sample mean.

[0093] During the training of the machine learning model, the feature values ​​of each organic compound in each row of the new feature set are used as input, and the reaction rate constants of the corresponding organic compounds in the reaction rate constant dataset are used as output, thereby achieving the training of the machine learning model.

[0094] Specifically, during the training process, the new feature set is randomly divided, with 80% of the data allocated to the training set and 20% allocated to the test set.

[0095] On the training set, 5-fold cross-validation combined with GridSearchCV was used for automatic hyperparameter tuning, and the cross-validation results were used to evaluate the model's stability and generalization ability. Using the selected optimal parameters, the final model was trained on the entire training set. Through multiple validations on different partitions, the mean of the evaluation metric values ​​for each fold was calculated as the final evaluation metric value for each machine learning algorithm. After training, the model was tested and validated on the test set.

[0096] S52. Select the optimal evaluation index value from all evaluation index values. The machine learning model trained with the optimal evaluation index value is taken as the optimal machine learning model.

[0097] In this embodiment, the algorithm was trained on RandomForest, Gradient Boosting Tree (GBDT), and Extreme Gradient Boosting Tree (XGBoost). The obtained evaluation metrics are shown in Table 1 below: Table 1. Evaluation metrics for each algorithm: algorithm <![CDATA[R 2 ]]> MAE MSE RF 0.889 0.312 0.171 GBDT 0.911 0.240 0.119 XGBoost 0.925 0.216 0.104 As shown in Table 1, the XGBoost machine learning model is selected as the optimal machine learning model because it has the best evaluation index value.

[0098] like Figure 2As shown, the horizontal axis represents the actual value of the reaction rate constant determined experimentally, and the vertical axis represents the predicted value of the reaction rate constant obtained by the XGBoost machine learning model.

[0099] Line 1 represents the experimentally determined true value of the reaction rate constant. Since the experimentally determined true value of the reaction rate constant was not predicted by the XGBoost machine learning model, the values ​​on both the horizontal and vertical axes are equal. Line 2 is the line formed by the predicted values ​​obtained from processing the descriptor data of only the optimal new feature set using the XGBoost machine learning model and their corresponding true values. Line 3 is the line formed by processing the descriptor data before feature selection and without adding quantum chemical descriptors using the XGBoost machine learning model and its corresponding true values.

[0100] thus Figure 2 As can be seen, the second line is closer to the first line, which indicates that the predicted value of the reaction rate constant obtained by the method of the present invention is closer to the true value of the experimentally determined reaction rate constant. This shows that feature selection and the addition of multiple quantum chemical descriptors have a significant improvement on the prediction results.

[0101] Table 2. The optimal new feature set and the meaning of each feature:

[0102] Table 2 shows the 21 molecular descriptors (features) included in the optimal new feature set corresponding to the optimal machine learning model XGBoost. The first 18 are the features retained by the XGBoost machine learning model after feature selection, and the last 3 are newly added features.

[0103] During the training of the optimal machine learning model XGBoost, the SHAP interpreter is invoked to calculate the SHAP value of each column feature of each organic compound in the new feature set corresponding to the machine learning model. The SHAP values ​​of each column feature of all organic compounds are summed and averaged to obtain the average SHAP value of each column feature. These average SHAP values ​​are then sorted from largest to smallest to determine the importance of each column feature. Figure 3 As shown. According to Figure 3 It is evident that the three quantum chemical descriptors proposed in this invention play a crucial role in predicting the reaction rate constant.

[0104] S53. The new feature set corresponding to the optimal machine learning model is taken as the optimal new feature set, and all features included in the optimal new feature set are taken as the prediction feature set.

[0105] S6. Based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, further obtain the value of each feature in the prediction feature set of the organic compound to be tested, and input all feature values ​​into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested.

[0106] Step S6 is as follows: S61. Obtain the molecular structural formula and SMILES structural formula of the organic compound to be tested.

[0107] In this embodiment, the name of the organic compound to be tested is vanillin, and the structural formula of SMILES is COCl=C(C=CC(=C1)C=O)O, and the molecular structural formula is as follows. Figure 5 As shown.

[0108] S62. Based on the molecular structure formula and SMILES structure formula, obtain the value corresponding to each feature in the predicted feature set of the organic compound to be tested, thereby obtaining the feature value set of the organic compound to be tested.

[0109] S63. Input the set of feature values ​​corresponding to the organic compound to be tested into the optimal machine learning model for prediction to obtain the corresponding reaction rate constant.

[0110] S7. During the prediction process, the SHAP interpreter is called to calculate the SHAP value of each feature in the feature set, and the features are sorted in descending order according to their SHAP values ​​to obtain the importance of each feature. Figure 4 As shown.

[0111] Figure 4 The results presented and Figure 3 The results on the average SHAP value are generally consistent, according to Figure 4 This further demonstrates that the three quantum chemical descriptors proposed in this invention play a crucial role in predicting the reaction rate constant.

[0112] This embodiment also provides an intelligent prediction system for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization, including: a data acquisition and preprocessing module, which acquires experimental measurement data of the reaction between organic compounds and hydroxyl radicals, and preprocesses the experimental measurement data to obtain preprocessed experimental measurement data; a dataset construction module, which extracts the SMILES structural formula and reaction rate constant of each organic compound from the preprocessed experimental measurement data, obtains a descriptor table dataset based on the SMILES structural formula, and obtains a reaction rate constant dataset based on the reaction rate constant; a supplementary descriptor acquisition module, which extracts the molecular structural formula of each organic compound from the preprocessed experimental measurement data, and calculates multiple quantum chemical descriptors corresponding to each organic compound based on the molecular structural formula; a new feature set acquisition module, which uses multiple machine learning algorithms to perform feature selection on the descriptor table dataset, obtains a feature subset corresponding to each machine learning algorithm, and then adds all quantum chemical descriptors to each feature subset to obtain a new feature set corresponding to each machine learning algorithm; and a model training and prediction feature set acquisition module, which uses each new feature set as input and the reaction rate constant dataset as output to train each corresponding machine learning model, thereby obtaining the optimal machine learning model, and obtaining the prediction feature set based on the optimal machine learning model. The prediction deployment module, based on the obtained molecular and SMILES structural formulas of the target organic compound, further obtains the value of each feature in the prediction feature set for the target organic compound. All feature values ​​are then input into the optimal machine learning model to predict the reaction rate constant of the target organic compound. The interpretability module, during the prediction process, calls the SHAP interpreter to calculate the SHAP value of each feature in the feature value set. The features are then sorted in descending order according to their SHAP values ​​to determine the importance of each feature.

[0113] By introducing local electron attachment energy, the maximum electrophilic Fukui exponent, and the frontier orbital gap, a multi-level, three-dimensional characterization system for electronic reactivity, ranging from global electronic stability to local electronic microenvironments and specific reaction sites, is constructed, overcoming the shortcomings of existing techniques in characterizing microscopic electronic effects. The synergistic effect of these three elements significantly enhances the model's ability to identify complex substituted derivatives and structurally similar molecules. While improving prediction accuracy and robustness, it also reveals the microscopic mechanisms of reactions at the quantum chemical level, endowing the model with stronger physical interpretability.

[0114] Comparative Example 1: Using the same data as in the example, the specific steps are as follows: D1. Collect experimental data on the reaction of organic compounds with hydroxyl radicals, and preprocess the experimental data to obtain preprocessed experimental data.

[0115] D2. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data. Obtain the descriptor table dataset based on the SMILES structural formula and the reaction rate constant dataset based on the reaction rate constant.

[0116] D3. After using multiple machine learning algorithms to perform feature selection on the descriptor table dataset, the feature subset corresponding to each machine learning algorithm is obtained.

[0117] D4. Using each feature subset as input and the reaction rate constant dataset as output, train the corresponding machine learning model to obtain the optimal machine learning model, and obtain the predicted feature set based on the optimal machine learning model.

[0118] D5. Based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, further obtain the value of each feature in the prediction feature set of the organic compound to be tested, and input all feature values ​​into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested.

[0119] D6. During the prediction process, the SHAP interpreter is called to calculate the SHAP value of each feature in the feature set, and the features are sorted in descending order according to the size of the SHAP value to obtain the importance of each feature.

[0120] Compared to Example 1, Comparative Example 1 directly trains the corresponding machine learning models using the feature subsets obtained after feature selection, without adding the three features of local electron attachment energy, maximum electrophilic Fukui exponent, and frontier orbital gap to the feature subsets. The results obtained by training on RandomForest, Gradient Boosting Tree (GBDT), and Extreme Gradient Boosting Tree (XGBoost) are shown in Table 3 below: Table 3. Evaluation metrics for each algorithm in Comparative Example 1: algorithm <![CDATA[R 2 ]]> MAE MSE RF 0.874 0.295 0.168 GBDT 0.903 0.258 0.130 XGBoost 0.916 0.230 0.113 Based on the comparison of the results in Table 3 of this comparative example and Table 1 of the embodiments, it can be seen that the method in the embodiments of the present invention is optimal, thus proving that the three quantum chemical descriptors added in the present invention—local electron attachment energy, maximum electrophilic Fukui exponent, and frontier orbital gap—have a significant advantage in contributing to the accuracy of predicting the reaction rate constant.

[0121] Comparative Example 2: H1. Collect experimental data on the reaction of organic compounds with hydroxyl radicals, and preprocess the experimental data to obtain preprocessed experimental data.

[0122] H2. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data. Obtain the descriptor table dataset based on the SMILES structural formula and the reaction rate constant dataset based on the reaction rate constant.

[0123] H3. Extract the molecular structure formula of each organic compound from the pretreated experimental data, and calculate the set of quantum chemical descriptors corresponding to each organic compound based on the molecular structure formula.

[0124] H4. After using multiple machine learning algorithms to perform feature selection on the descriptor table dataset, a feature subset corresponding to each machine learning algorithm is obtained. Then, the quantum chemical descriptor set is added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm.

[0125] H5. Using each new feature set as input and the reaction rate constant dataset as output, train each corresponding machine learning model, thereby obtaining the optimal machine learning model, and obtaining the predicted feature set based on the optimal machine learning model.

[0126] H6. Based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, further obtain the value of each feature in the prediction feature set of the organic compound to be tested, and input all feature values ​​into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested.

[0127] H7. During the prediction process, the SHAP interpreter is called to calculate the SHAP value of each feature in the feature set, and the features are sorted in descending order according to the size of the SHAP value to obtain the importance of each feature.

[0128] In Comparative Example 2, compared to Example 1, a new feature set is obtained by adding a set of quantum chemical descriptors after obtaining the feature subset. When the sets of quantum chemical descriptors are LEAE and f(-) respectively... x E GAP LEAE+f(-) x LEAE+E GAP and f(-) x +E GAP The results obtained from training on RandomForest, Gradient Boosting Tree (GBDT), and Extreme Gradient Boosting Tree (XGBoost) are shown in Table 4 below: Table 4. Ablation Experiment Results:

[0129] Based on the comparison of the results in Table 4 of this comparative example and Table 1 of the embodiments, it can be seen that the method in the embodiments of the present invention is optimal, thus proving that the three quantum chemical descriptors added in the present invention—local electron attachment energy, maximum electrophilic Fukui exponent, and frontier orbital gap—have a significant advantage in contributing to the accuracy of predicting the reaction rate constant.

[0130] By introducing local electron attachment energy, the maximum electrophilic Fukui exponent, and the frontier orbital gap, a multi-level, three-dimensional characterization system for electronic reactivity, ranging from global electronic stability to local electronic microenvironments and specific reaction sites, is constructed, overcoming the shortcomings of existing techniques in characterizing microscopic electronic effects. The synergistic effect of these three elements significantly enhances the model's ability to identify complex substituted derivatives and structurally similar molecules. While improving prediction accuracy and robustness, it also reveals the microscopic mechanisms of reactions at the quantum chemical level, endowing the model with stronger physical interpretability.

[0131] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. An intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization, characterized in that, Includes the following steps: S1. Collect and obtain experimental data on the reaction of organic compounds with hydroxyl radicals, and preprocess the experimental data to obtain preprocessed experimental data. S2. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data. Obtain the descriptor table dataset based on the SMILES structural formula and the reaction rate constant dataset based on the reaction rate constant. S3. Extract the molecular structure formula of each organic compound from the pretreated experimental data, and calculate multiple quantum chemical descriptors corresponding to each organic compound based on the molecular structure formula. The quantum chemical descriptor includes the local electron attachment energy, the maximum electrophilic Fukui index, and the frontier orbital gap; S4. After using multiple machine learning algorithms to perform feature selection on the descriptor table dataset, a feature subset corresponding to each machine learning algorithm is obtained. Then, all quantum chemical descriptors are added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm. S5. Using each new feature set as input and the reaction rate constant dataset as output, train each corresponding machine learning model, thereby obtaining the optimal machine learning model, and obtaining the predicted feature set based on the optimal machine learning model. S6. Based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, further obtain the value of each feature of the organic compound to be tested in the prediction feature set, and input all feature values ​​into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested. S7. During the prediction process, the SHAP interpreter is called to calculate the SHAP value of each feature in the feature set, and the features are sorted in descending order according to the size of the SHAP value to obtain the importance of each feature.

2. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 1, characterized in that: The preprocessing includes cleaning up invalid values ​​and filling in missing values; cleaning up invalid values ​​involves finding invalid values ​​in the descriptor data and replacing them with null values; filling in missing values ​​involves filling in the missing values ​​in the experimental measurement data with null values.

3. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization as described in claim 1, characterized in that, Step S2 specifically involves: S21. Extract the SMILES structural formula and reaction rate constant of each organic compound from the pretreated experimental data; S22. Generate corresponding descriptor data for the SMILES structure formula of each organic compound, and arrange the descriptor data corresponding to each organic compound in rows to obtain a preliminary descriptor table dataset. S23. Perform dimensionality reduction and Z-score standardization on the initial descriptor table dataset to obtain the descriptor table dataset. S24. After performing a logarithmic transformation on the reaction rate constants of each organic compound, the reaction rate constants dataset is obtained.

4. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 3, characterized in that: The descriptor data for each organic compound includes several molecular descriptors, each of which is a feature; each column of the descriptor table dataset corresponds to one molecular descriptor.

5. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 3, characterized in that: The dimensionality reduction process involves first deleting all-zero columns and columns containing null values ​​from the descriptor table dataset, and then sequentially processing the initial descriptor data using methods to remove low-variance features and high-correlation features.

6. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 1, characterized in that, Step S4 specifically involves: S41. Using the descriptor table dataset as input and the reaction rate constant dataset as output, multiple machine learning algorithms are employed, with each machine learning model serving as the base model for the recursive feature elimination method. Feature selection is performed on the descriptor table dataset to obtain the feature subsets corresponding to each machine learning algorithm. S42. Treat each quantum chemical descriptor as a single column of features, and add all three columns of features together to the feature subset corresponding to each machine learning algorithm, thereby obtaining the new feature set after the addition of each machine learning algorithm.

7. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 1, characterized in that, Step S5 specifically involves: S51. Using each new feature set as input and the reaction rate constant dataset as output, train the corresponding machine learning model to obtain each trained machine learning model and the corresponding evaluation index value. S52. Select the optimal evaluation index value from all evaluation index values. The machine learning model corresponding to the optimal evaluation index value is taken as the optimal machine learning model. S53. The new feature set corresponding to the optimal machine learning model is taken as the optimal new feature set, and all features included in the optimal new feature set are taken as the prediction feature set.

8. The intelligent prediction method for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization according to claim 1, characterized in that, Step S6 specifically involves: S61. Obtain the molecular structural formula and SMILES structural formula of the organic compound to be tested; S62. Based on the molecular structural formula and SMILES structural formula, obtain the value corresponding to each feature in the predicted feature set of the organic compound to be tested, thereby obtaining the feature value set of the organic compound to be tested. S63. Input the set of feature values ​​corresponding to the organic compound to be tested into the optimal machine learning model for prediction to obtain the corresponding reaction rate constant.

9. An intelligent prediction system for atmospheric reaction kinetic parameters integrating multi-level quantum chemical characterization, employing the method described in any one of claims 1-8, comprising: The data acquisition and preprocessing module acquires experimental measurement data of the reaction between organic compounds and hydroxyl radicals, and preprocesses the experimental measurement data to obtain preprocessed experimental measurement data. The dataset construction module extracts the SMILES structural formula and reaction rate constant of each organic compound from the preprocessed experimental data, obtains the descriptor table dataset based on the SMILES structural formula, and obtains the reaction rate constant dataset based on the reaction rate constant. The supplementary descriptor acquisition module extracts the molecular structure formula of each organic compound from the preprocessed experimental measurement data, and calculates multiple quantum chemical descriptors corresponding to each organic compound based on the molecular structure formula; The new feature set acquisition module uses multiple machine learning algorithms to select features from the descriptor table dataset, obtaining a feature subset corresponding to each machine learning algorithm. Then, all quantum chemical descriptors are added to each feature subset to obtain a new feature set corresponding to each machine learning algorithm. The model training and prediction feature set acquisition module takes each new feature set as input and the reaction rate constant dataset as output to train each corresponding machine learning model, thereby obtaining the optimal machine learning model and acquiring the prediction feature set based on the optimal machine learning model. The prediction deployment module further obtains the value of each feature of the organic compound to be tested in the prediction feature set based on the obtained molecular structure formula and SMILES structure formula of the organic compound to be tested, and inputs the values ​​of all features into the optimal machine learning model to predict the reaction rate constant of the organic compound to be tested. The interpretability module calls the SHAP interpreter during the prediction process to calculate the SHAP value of each feature in the feature set, and sorts the features in descending order according to the size of the SHAP value, thereby obtaining the importance of each feature.

Images