Method for predicting metallocene polypropylene synthesis catalytic performance

Abstract

The invention discloses a metallocene polypropylene synthesis catalytic performance prediction method, which comprises the following steps: data collection: collecting experimental data of metallocene catalyst synthesis polypropylene, and establishing a data set containing input variables and output variables; data preprocessing: cleaning, standardizing and coding experimental data; feature engineering: carrying out feature correlation analysis on the preprocessed data; model construction: constructing independent prediction models for predicting catalyst activity, number-average molecular weight and molecular weight distribution by using a plurality of machine learning algorithms respectively; performing hyper-parameter optimization: performing hyper-parameter optimization on the prediction model by adopting a Bayesian optimization algorithm to obtain an optimal model parameter; model evaluation: evaluating the performance of the prediction model by adopting a cross validation method; and predicting: inputting process parameters and catalyst ligand parameters input by a user into the optimized and evaluated optimal prediction model, and outputting predicted values of catalyst activity, number-average molecular weight and molecular weight distribution.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration technology, and in particular to a method for predicting the catalytic performance of metallocene polypropylene synthesis. Background Technology

[0002] Polypropylene (PP) is one of the world's largest-produced and most widely used general-purpose plastics. Due to its excellent mechanical properties, good chemical stability, ease of processing, and relatively low cost, it has become a fundamental material in modern industrial production and daily life. With the increasing demand for high-end polyolefin materials, metallocene catalysts have gradually become a key technology for producing high-performance polypropylene. Compared with traditional Ziegler-Natta catalysts, metallocene catalysts have the characteristic of a single active site, enabling precise control of the polymer chain structure. This allows for the production of polypropylene products with narrow molecular weight distribution, high transparency, and excellent mechanical and optical properties, meeting the stringent requirements of high-end applications for material consistency and performance. In the industrial process of producing polypropylene using metallocene catalysts, catalyst activity, number-average molecular weight (Mn), and molecular weight distribution (PDI) are the core indicators determining the quality and performance of the final product. Accurate and timely monitoring and prediction of these key parameters are crucial for process optimization and precise control.

[0003] However, metallocene catalysts exhibit complex and diverse ligand structures and significantly varying catalytic performances. Furthermore, the polymerization process involves complex characteristics such as strong coupling of multiple parameters and nonlinearity. Traditional research methods heavily rely on trial-and-error experiments, resulting in long development cycles, high resource consumption, and difficulty in systematically and profoundly revealing the intrinsic relationship between catalyst structure and performance. In recent years, machine learning, as a core component of artificial intelligence, has demonstrated powerful data mining and complex relationship modeling capabilities in multiple scientific fields, providing a novel approach for efficiently and accurately establishing the mapping relationship between catalyst structure and polymerization performance. Currently, there is no efficient method for predicting the activity, number-average molecular weight, and molecular weight distribution of metallocene polypropylene catalysts. Therefore, this paper proposes a method for predicting the catalytic performance of metallocene polypropylene synthesis. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for predicting the catalytic performance of metallocene polypropylene synthesis.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the catalytic performance of metallocene polypropylene synthesis includes the following steps: Data collection: Experimental data on the synthesis of polypropylene using metallocene catalysts were collected, and a dataset containing input and output variables was established. The input variables included catalyst dosage, the ratio of co-catalyst dosage to catalyst dosage, polymerization temperature, polymerization time, ligand R1 buried volume percentage, ligand R2 buried volume percentage, metal center, and ligand R3. The output variables included catalyst activity, number-average molecular weight, and molecular weight distribution. Data preprocessing: The experimental data are cleaned, standardized and encoded, wherein numerical features are Z-score standardized and categorical features are one-hot encoded; Feature engineering: Perform feature correlation analysis on the preprocessed data, calculate the Pearson correlation coefficient between the features and the output variables, and use SHAP analysis to evaluate the importance of the features; Model Construction: Independent prediction models for predicting catalyst activity, number-average molecular weight, and molecular weight distribution were constructed using a variety of machine learning algorithms, including gradient boosting regression, AdaBoost regression, bagging regression, support vector regression, random forest regression, extreme tree regression, K-nearest neighbor regression, and neural networks. Hyperparameter optimization: The prediction model is tuned using the Bayesian optimization algorithm to obtain the optimal model parameters; Model Evaluation: The performance of the prediction model was evaluated using cross-validation, with mean absolute error, root mean square error, and coefficient of determination used as evaluation metrics. Prediction: Input the process parameters and catalyst ligand parameters input by the user into the optimal prediction model after optimization and evaluation, and output the predicted values ​​of catalyst activity, number-average molecular weight and molecular weight distribution.

[0006] The above further includes: Furthermore, in the data collection, the metal center includes zirconium, titanium, and hafnium.

[0007] Furthermore, in feature engineering, the feature correlation analysis includes drawing a correlation heatmap, generating a correlation heatmap between feature variables and output variables using plotting software, quantifying the linear correlation strength between each feature and catalyst activity, number-average molecular weight, and molecular weight distribution by calculating the Pearson correlation coefficient, and performing SHAP analysis on catalyst activity, number-average molecular weight, and molecular weight distribution to quantify the feature contribution.

[0008] Furthermore, in the model construction, the neural network model adopts a dual-branch structure, which includes a first branch for processing numerical features and a second branch for processing categorical features. The first branch and the second branch contain multiple hidden layers for extracting high-dimensional features from different types of data, respectively. Finally, a shared fusion layer integrates and interacts the features extracted by the first branch and the second branch.

[0009] Furthermore, the hyperparameter optimization is implemented by calling the scikit-learn optimization library in the Python programming environment to implement the Bayesian optimization algorithm. The Bayesian optimization algorithm targets various machine learning algorithms, uses the evaluation index as the optimization target, and performs multiple rounds of iterative search in the preset parameter space to find the globally optimal hyperparameter combination for each machine learning algorithm structure and each prediction task.

[0010] Furthermore, the cross-validation method used in the model evaluation step employs a differentiated dataset partitioning strategy based on the different prediction targets; For the catalyst activity prediction model, the dataset is divided into training, validation and test sets in a ratio of 7:2:1. For the number-average molecular weight prediction model and the molecular weight distribution prediction model, the training set and the test set were divided in a ratio of 8:2.

[0011] Furthermore, the prediction is accomplished by deploying an interactive graphical user interface system developed based on the Python framework. After the user inputs all or part of the input variables, including polymerization temperature, catalyst dosage, and ligand structure parameters, according to the prompts in the user interface, the interactive graphical user interface system will automatically call the pre-trained, hyperparameter-optimized, and performance-verified optimal machine learning model, perform calculations, and display the numerical prediction results of catalyst activity, number-average molecular weight, and molecular weight distribution on the user interface.

[0012] Furthermore, after the prediction output is completed, based on the prediction results and combined with the feature importance ranking of SHAP analysis, the influence trend of each process parameter and ligand structure parameter on catalytic performance is quantitatively analyzed.

[0013] The present invention has the following beneficial effects: 1. In this invention, by constructing three independent end-to-end prediction models, catalyst activity, number-average molecular weight, and molecular weight distribution with clear engineering value can be directly output respectively. Through comparison of various machine learning algorithms and optimization using Bayesian hyperparameters, the constructed optimal model has excellent generalization performance and prediction accuracy for catalyst activity, number-average molecular weight, and molecular weight distribution. This enables multi-dimensional, high-precision parallel prediction of the core performance of the polymerization process, greatly enhancing the practical value of the technology in industrial applications.

[0014] 2. In this invention, by integrating SHAP analysis, not only can predicted values ​​be output, but the contribution of each input feature to the prediction results can also be quantitatively revealed, providing clear and quantitative guidance for catalyst molecular design and process optimization.

[0015] 3. In this invention, a pure data-driven paradigm is adopted. The model has undergone rigorous cross-validation and exhibits good stability and generalization ability. It can be extended to various catalyst systems such as post-transition metals, and has wider applicability and forward-looking nature. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a method for predicting the catalytic performance of metallocene polypropylene synthesis proposed in this invention. Figure 2 This is a graph showing the fitting of predicted and actual values ​​of the catalyst activity model in this invention. Figure 3 This is a graph showing the fitting of predicted and actual values ​​of the number-average molecular weight model in this invention. Figure 4 This is a graph showing the fitting of predicted and actual values ​​of the molecular weight distribution model in this invention. Figure 5 This is a thermogram showing the correlation coefficients of catalyst activity in this invention. Figure 6 This is a SHAP analysis diagram of the catalyst activity in this invention; Figure 7 This is a heatmap of the correlation coefficient of the number-average molecular weight in this invention; Figure 8 This is a SHAP analysis graph of the number-average molecular weight in this invention; Figure 9 This is a heatmap of the correlation coefficients of molecular weight distribution in this invention; Figure 10 This is a SHAP analysis diagram of the molecular weight distribution in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1-10 As shown, this invention provides a method for predicting the catalytic performance of metallocene polypropylene synthesis, comprising the following steps: Data Collection: Experimental data on the synthesis of polypropylene using metallocene catalysts were collected, and a dataset containing input and output variables was established. The input variables included catalyst dosage (Cat), the ratio of co-catalyst dosage to catalyst dosage ([Al] / [M]), polymerization temperature (T), polymerization time (t), ligand R1 buried volume percentage, ligand R2 buried volume percentage, metal center, and ligand R3. The output variables included catalyst activity, number-average molecular weight, and molecular weight distribution. Data preprocessing: The experimental data are cleaned, standardized, and encoded to eliminate the influence of units and missing values. Numerical features are Z-score standardized, and categorical features are one-hot encoded. Feature engineering: Perform feature correlation analysis on the preprocessed data, calculate the Pearson correlation coefficient between the features and the output variables, and use SHAP analysis to evaluate the importance of features in order to eliminate redundant features; Model Construction: Independent prediction models for predicting catalyst activity, number-average molecular weight, and molecular weight distribution were constructed using a variety of machine learning algorithms, including gradient boosting regression, AdaBoost regression, bagging regression, support vector regression, random forest regression, extreme tree regression, K-nearest neighbor regression, and neural networks. Hyperparameter optimization: The prediction model is tuned using the Bayesian optimization algorithm to obtain the optimal model parameters; Model evaluation: The performance of the prediction model was evaluated using cross-validation, with mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) used as evaluation metrics. Prediction: Input the process parameters and catalyst ligand parameters input by the user into the optimal prediction model after optimization and evaluation, and output the predicted values ​​of catalyst activity, number-average molecular weight and molecular weight distribution.

[0019] In one embodiment, the data collection process specifically draws from publicly available academic and patent literature both domestically and internationally. The dataset contains at least 207 sets of experimental data to ensure data diversity and model robustness. The metal centers include zirconium (Zr), titanium (Ti), and hafnium (Hf), three common metallocene catalyst center atoms. By integrating data from these different metal centers, the constructed prediction model can adapt to a wider range of catalyst types.

[0020] In one embodiment, in feature engineering, the feature correlation analysis includes drawing a correlation heatmap, generating a correlation heatmap between feature variables and output variables using plotting software, quantifying the linear correlation strength between each feature and catalyst activity, number-average molecular weight, and molecular weight distribution by calculating the Pearson correlation coefficient, and performing SHAP analysis on catalyst activity, number-average molecular weight, and molecular weight distribution to quantify the feature contribution.

[0021] In this embodiment: Correlation heatmap analysis: Correlation heatmaps were plotted using the correlation heatmap module in Origin software to analyze the correlation relationships between features. Results showed that for catalyst activity, number-average molecular weight, and molecular weight distribution models, the correlation coefficients between catalyst dosage (Cat) and polymerization time (t) were the highest, reaching 0.82, 0.83, and 0.83, respectively.

[0022] SHAP analysis was performed. SHAP is a game theory-based method used to interpret the prediction results of any machine learning model. By utilizing the SHapley Additive Explanations (SHAP) analysis method, the contribution of each feature parameter to the predicted value was analyzed in depth. The results show that for the catalyst activity prediction model, the metal center, catalyst dosage, and polymerization time have the greatest impact on the predicted catalyst activity; for the number-average molecular weight prediction model, the catalyst dosage, metal center, and polymerization time have the greatest impact on the predicted number-average molecular weight; for the molecular weight distribution prediction model, ligand R1, ligand R2, catalyst dosage, and polymerization time have the greatest impact on the predicted molecular weight distribution.

[0023] In one embodiment, the neural network model adopts a dual-branch structure in model construction. The dual-branch structure includes a first branch for processing numerical features and a second branch for processing categorical features. The first branch and the second branch contain multiple hidden layers for extracting high-dimensional features from different types of data, respectively. Finally, a shared fusion layer integrates and interacts the features extracted by the first branch and the second branch, thereby more effectively capturing the complex nonlinear coupling relationship between process conditions and catalyst structure during polymerization.

[0024] In one embodiment, the hyperparameter optimization is implemented by calling the scikit-learn optimization library in the Python programming environment to implement the Bayesian optimization algorithm. The Bayesian optimization algorithm targets various machine learning algorithms (gradient boosting regression, random forest regression, K-nearest neighbor regression, and neural networks with different hidden layer depths, etc.), uses the evaluation index as the optimization target, and performs multiple rounds of iterative search in the preset parameter space to find the globally optimal hyperparameter combination for each machine learning algorithm structure and each prediction task.

[0025] In one embodiment, the cross-validation method used in the model evaluation step employs a differentiated dataset partitioning strategy based on the different prediction targets. For catalyst activity prediction models, the dataset is divided into training, validation and test sets in a ratio of 7:2:1 to facilitate effective model selection and early stopping during training to prevent overfitting. For the number-average molecular weight prediction model and the molecular weight distribution prediction model, the training set and the test set are divided in a ratio of 8:2 to ensure the reliability of the model performance evaluation on unknown data.

[0026] In one embodiment, the prediction is accomplished by deploying an interactive graphical user interface system developed based on the Python framework. After the user inputs all or part of the input variables, including polymerization temperature, catalyst dosage, and ligand structure parameters, according to the prompts in the user interface, the interactive graphical user interface system will automatically call the pre-trained, hyperparameter-optimized, and performance-verified optimal machine learning model, perform calculations, and display the numerical prediction results of catalyst activity, number-average molecular weight, and molecular weight distribution on the user interface.

[0027] In one embodiment, after the prediction output is completed, based on the prediction results and combined with the feature importance ranking of SHAP analysis, the influence trend of each process parameter and ligand structure parameter on catalytic performance is quantitatively analyzed. This provides users with specific and actionable technical suggestions on how to adjust catalyst molecular design or how to optimize polymerization reaction operating conditions, thereby assisting in the precise control of polypropylene product performance.

[0028] In this embodiment: Data Collection: Experimental data on the synthesis of polypropylene using metallocene catalysts were collected from published domestic and international literature and patents, resulting in a dataset containing 207 sets of data. The dataset includes parameters related to the operating conditions during the polypropylene synthesis process and the ligand structure parameters of the metallocene catalysts used.

[0029] Input variable X includes numerical features: catalyst dosage (Cat, unit: μmol), co-catalyst dosage / catalyst dosage ratio ([Al] / [M], unit: mol), polymerization temperature (T, unit: °C), polymerization time (t, unit: min), ligand R1 buried volume percentage, ligand R2 buried volume percentage; and categorical features: metal center. , , ligand R3; Output variable Y: catalyst activity, number-average molecular weight, molecular weight distribution; The collected data is cleaned to eliminate model training errors or performance degradation caused by missing values. Then, the data is standardized and one-hot encoded to eliminate the influence of different feature units and improve model training efficiency and stability. Data cleaning and quality assurance: The integrity of the 207 sets of raw data samples was checked and cleaned, and samples containing serious missing values ​​or obvious recording errors were removed to ensure the reliability and consistency of the dataset and lay a solid foundation for model training. Feature type identification and partitioning: Based on data characteristics, input variables are explicitly divided into two main categories: numerical features and categorical features, and different preprocessing strategies are adopted accordingly. Numerical features include catalyst dosage (Cat), co-catalyst / catalyst ratio ([Al] / [M]), polymerization temperature (T), polymerization time (t), ligand R1 buried volume percentage, and ligand R2 buried volume percentage. These features are Z-score standardized to transform them into a distribution with a mean of 0 and a standard deviation of 1, thereby eliminating dimensional differences, accelerating model convergence, and improving stability.

[0030] Category-type characteristics: including metal center ( , , The features are categorical and R3 ligand types. One-hot encoding is used to convert each categorical variable into a binary vector, preventing the model from misjudging numerical relationships between categories.

[0031] Feature correlation analysis: Correlation heatmaps are used to quantify the processed features and calculate the Pearson correlation coefficient between them and the predicted targets (such as catalyst activity and number-average molecular weight). This is used to verify the strength of the linear association between the features and the target variable, providing statistical basis for the rationality of the feature set.

[0032] Structured dataset construction: All features that have undergone the above preprocessing and encoding are combined to construct a structured feature matrix (X). The feature matrix (X) and the corresponding labels (Y, i.e., the actual measured values ​​of catalyst activity, number-average molecular weight, and molecular weight distribution) together form the final dataset, which is divided into training set, validation set, and test set according to the proportions for model training, optimization, and evaluation. Interpretability analysis and feature importance verification: After model training, SHAP (Shape-Up Analysis), a game theory-based method, is used to quantitatively evaluate the contribution of each input feature to a single prediction result. This verifies the effectiveness of feature engineering, reveals key influencing factors, and enhances the model's transparency and credibility.

[0033] Model selection: Eight machine learning algorithms were selected, including Gradient Boosting Regression, AdaBoost Regression, Bagging Regression, Support Vector Regression, Random Forest Regression, ExtraTrees Regression, K-Nearest Neighbors Regressor, and Neural Network. These algorithms can handle high-dimensional, nonlinear, and complex data relationships and are suitable for predicting the catalyst activity, number-average molecular weight, and molecular weight distribution of metallocene polypropylene.

[0034] Model training: The catalyst activity prediction model divides the dataset into training set: validation set: test set in a 7:2:1 ratio; the number mean molecular weight and molecular weight distribution prediction model divides the dataset into training set and test set in an 8:2 ratio, using train_test_split from the sklearn library to split the data, and uses the training set to train the model.

[0035] Hyperparameter optimization: Bayesian optimization algorithm is used to tune model parameters and find the optimal parameters to build the optimal model. Bayesian optimization is implemented using the scikit-learn library in Python, and the optimization results are as follows:

[0036] Catalyst activity prediction model: AdaBoost regression model: It is 159. It is 0.0812.

[0037] Bagging regression model: It is 69. It is 0.6388. The value is 0.7588, and bootstrap is False. It is True.

[0038] Extreme Tree Regression (ETR) Model: It is 160. It is 3. It is 3. The value is 3.

[0039] Gradient Boosting (GBR) Regression Model: It is 72. =0.1155, The value is 3.

[0040] K-nearest neighbor regression model: The value is 9, weights is distance, and p is 1.

[0041] Support Vector Machine (SVR) regression model: C = 60, gamma = 0.1433, epsilon = 0.1548.

[0042] Random Forest Regression Model (RFR): It is 181. It is 3. It is 3. The value is 2.

[0043] A neural network model with two branches, two hidden layers on each side, and one fusion layer (2+2+1): the optimal hyperparameters are the swish function, and the activation function is the swish function. It is 65. It is 0.1456. It is 0.1719. It is 0.1056. It is 0.1744. The value is 127, and l2_reg is 4.3872e-5. It is 0.000459. It is 82. It is 112.

[0044] A neural network model with two branches, each with three hidden layers and one fusion layer (3+3+1): the optimal hyperparameters are the swish function as the activation function. It is 55. It is 0.3301. It is 0.5513. It is 0.1179. It is 0.1314. It is 0.2184. It is 0.1241. The value is 277, and l2_reg is 4.0011e-5. It is 0.00076. It is 99. The value is 117.

[0045] A neural network model with two branches, each with four hidden layers and one fusion layer (4+4+1): The optimal hyperparameters are the Leaky_ReLU activation function. It is 45. It is 0.2558. It is 0.2395. It is 0.4. It is 0.4 It is 0.4. It is 0.2. It is 0.1. It is 0.2341. The value is 163, and l2_reg is 1e-4. It is 0.00127. It is 79. It is 135.

[0046] Number-mean molecular weight prediction model: AdaBoost regression model: It is 32. It is 0.0271. It is 9.

[0047] Bagging regression model: It is 73. It is 0.6410. The value is 0.8796, and bootstrap is False. It is False.

[0048] Extreme Tree Regression (ETR) Model: It is 36. It is 6. It is 7. The value is 2.

[0049] Gradient Boosting (GBR) Regression Model: It is 52. It is 0.0418. The value is 3.

[0050] K-nearest neighbor regression model: The value is 6, weights are uniform, and p is 1.

[0051] Support Vector Machine (SVR) regression model: C = 7.190450, gamma = 0.1453, epsilon = 0.01.

[0052] Random Forest Regression Model (RFR): It is 71. It is 5. It is 3. The value is 2.

[0053] A neural network model with two branches, two hidden layers on each side, and one fusion layer (2+2+1): the optimal hyperparameters are the swish function, and the activation function is the swish function. It is 34. t is 0.0590. It is 0.1568. It is 0.1548. It is 0.2484. The value is 87, and l2_reg is 7.9945e-5. It is 0.000418. It is 91. It is 84.

[0054] A neural network model with two branches, each with three hidden layers and one fusion layer (3+3+1): the optimal hyperparameters are the swish function as the activation function. It is 57. It is 0.2682. It is 0.2530. It is 0.1028. It is 0.1605. It is 0.1064. It is 0.1873. The value is 92, and l2_reg is 5e-5. It is 0.0008. It is 80. It is 86.

[0055] A neural network model with two branches, each with four hidden layers and one fusion layer (4+4+1): The optimal hyperparameters are the swish function as the activation function. It is 39. It is 0.2491. It is 0.2760. It is 0.2745. It is 0.2872. It is 0.1850. It is 0.1435. It is 0.1468. It is 0.1209. The value is 121, and l2_reg is 1.45e-5. It is 0.00067. It is 126. It is 68.

[0056] Molecular weight distribution prediction model: AdaBoost regression model: It is 102. It is 0.1511. It is 9.

[0057] Bagging regression model: It is 84. It is 0.9644. The value is 0.7751, and bootstrap is False. It is True.

[0058] Extreme Tree Regression (ETR) Model: It is 48. It is 6. It is 2. The value is 1.

[0059] Gradient Boosting (GBR) Regression Model: It is 90. It is 0.3. The value is 3.

[0060] K-nearest neighbor regression model: The value is 5, weights is distance, and p is 1.

[0061] Support Vector Machine (SVR) regression model: C = 9.905087, gamma = 0.9957, epsilon = 0.0633.

[0062] Random Forest Regression Model (RFR): It is 50. It is 7. It is 3. The value is 1.

[0063] A neural network model with two branches, two hidden layers on each side, and one fusion layer (2+2+1): the optimal hyperparameters are the ELU activation function. It is 34. It is 0.0573. It is 0.1119. It is 0.2818. It is 0.1861. The value is 81, and l2_reg is 2.7288e-5. It is 0.000478. It is 152. It is 88.

[0064] A neural network model with two branches, three hidden layers on each side, and one fusion layer (3+3+1): the optimal hyperparameters are the ELU activation function. It is 34. It is 0.2907. It is 0.1700. It is 0.2629. It is 0.2594. It is 0.1594. It is 0.1497. The value is 98, and l2_reg is 2.11e-5. It is 0.00076. It is 94. The value is 113.

[0065] A neural network model with two branches, four hidden layers on each side, and one fusion layer (4+4+1): the optimal hyperparameters are the ELU activation function. It is 40. It is 0.2636. It is 0.2860. It is 0.2550. It is 0.2401. It is 0.1794. It is 0.1336. It is 0.2837. It is 0.2297. The value is 120, and l2_reg is 2.54e-5. It is 0.00051. It is 41. It is 64.

[0066] Cross-validation is used to evaluate the model to ensure its robustness and generalization ability. The model's mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are calculated to verify the model's accuracy and generalization ability. Catalyst activity: AdaBoost regression model: R² = 0.8835, MAE = 0.2599; Bagging regression model: R² = 0.9084, MAE = 0.2320; Extreme Tree Regression (ETR) model: R² = 0.8999, MAE = 0.2492; Gradient boosting (GBR) regression model: R² = 0.8984, MAE = 0.2258; K-nearest neighbor regression model: R² = 0.9126, MAE = 0.1912; Support Vector Machine (SVR) regression model: R² = 0.8706, MAE = 0.2579; Random Forest Regression Model (RFR): R² = 0.7660, MAE = 0.3309; Neural network model: R² = 0.9201, MAE = 0.1932; Number average molecular weight: AdaBoost regression model: R² = 0.8540, MAE = 0.2461; Bagging regression model: R² = 0.8089, MAE = 0.2956; Extreme Tree Regression (ETR) model: R² = 0.8007, MAE = 0.2486; Gradient boosting (GBR) regression model: R² = 0.8093, MAE = 0.2921; K-nearest neighbor regression model: R² = 0.8146, MAE = 0.2565; Support Vector Machine (SVR) regression model: R² = 0.8407, MAE = 0.2469; Random Forest Regression (RFR) model: R² = 0.8427, MAE = 0.2442; Neural network model: R² = 0.9133, MAE = 0.2196; Molecular weight distribution: AdaBoost regression model: R² = 0.9678, MAE = 0.1480; Bagging regression model: R² = 0.6176, MAE = 0.2366; Extreme Tree Regression (ETR) model: R² = 0.9753, MAE = 0.1238; Gradient boosting (GBR) regression model: R² = 0.6090, MAE = 0.2509; K-nearest neighbor regression model: R² = 0.9766, MAE = 0.1181; Support Vector Machine (SVR) regression model: R² = 0.9628, MAE = 0.1431; Random Forest Regression Model (RFR): R² = 0.5635, MAE = 0.2565; Neural network model: R² = 0.0058, MAE = 0.4128; Develop a prediction system based on Python. After the user inputs the process parameters of the polypropylene polymerization process, the system calls a pre-trained machine learning model to quickly output the predicted results of catalyst activity, number-average molecular weight, and molecular weight distribution.

[0067] A graphical interface is used to visualize parameter input and results, which improves the interactivity and ease of use of the system.

[0068] Based on the prediction results, the process parameters in actual production can be adjusted to optimize catalyst activity, number-average molecular weight and molecular weight distribution, thereby improving product quality and production process control.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the performance of a synthetic catalyst for a metallocene polypropylene, characterized by, The method comprises the following steps: Data collection: collecting experimental data of metallocene catalyst for synthesizing polypropylene to establish a data set containing input variables and output variables; the input variables include catalyst dosage, cocatalyst dosage, polymerization temperature, polymerization time, ligand R1 buried volume percentage, ligand R2 buried volume percentage, metal center and ligand R3; the output variables include catalyst activity, number average molecular weight and molecular weight distribution; Data preprocessing: cleaning, standardizing and encoding the experimental data, wherein the numerical features are standardized by Z-score, and the categorical features are one-hot encoded; Feature engineering: performing feature correlation analysis on the preprocessed data, calculating the Pearson correlation coefficient between the features and the output variables, and evaluating the feature importance using SHAP analysis; Model construction: using multiple machine learning algorithms to construct independent prediction models for predicting catalyst activity, number average molecular weight and molecular weight distribution, the multiple machine learning algorithms including gradient boosting regression, AdaBoost regression, bagging regression, support vector regression, random forest regression, extreme tree regression, K-nearest neighbor regression and neural network; Hyperparameter optimization: using Bayesian optimization algorithm to optimize the hyperparameters of the prediction models to obtain the optimal model parameters; Model evaluation: evaluating the performance of the prediction models using cross-validation method, and using mean absolute error, root mean square error and determination coefficient as evaluation indicators; Prediction: inputting the user input process parameters and catalyst ligand parameters into the optimized and evaluated optimal prediction model to output the predicted values of catalyst activity, number average molecular weight and molecular weight distribution.

2. A method of predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, In data collection, the metal center includes zirconium, titanium and hafnium.

3. The method for predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, In feature engineering, the feature correlation analysis includes drawing a correlation heat map, using drawing software to generate a correlation heat map between feature variables and output variables, quantifying the linear correlation strength between each feature and catalyst activity, number average molecular weight and molecular weight distribution by calculating the Pearson correlation coefficient, and performing SHAP analysis on catalyst activity, number average molecular weight and molecular weight distribution to quantify the feature contribution.

4. The method for predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, In model construction, the neural network model adopts a double-branch structure, which includes a first branch for processing numerical features and a second branch for processing categorical features, and the first branch and the second branch include multiple hidden layers for extracting high-dimensional features from different types of data, respectively, and finally a shared fusion layer is used to integrate and interact the features extracted by the first branch and the second branch.

5. The method for predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, The hyperparameter optimization realizes the Bayesian optimization algorithm by calling the scikit-learn optimization library in the Python programming environment, the Bayesian optimization algorithm is used for multiple machine learning algorithms, the evaluation indicators are used as optimization objectives, and multiple rounds of iterative search are performed in the preset parameter space to find the global performance optimal hyperparameter combination for each machine learning algorithm structure and each prediction task.

6. The method for predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, The cross-validation method used in the model evaluation step adopts different data set division strategies according to different prediction targets; For the catalyst activity prediction model, the data set is divided into training set, validation set and test set in the ratio of 7:2:1; For the number average molecular weight prediction model and the molecular weight distribution prediction model, the training set and the test set are divided in the ratio of 8:

2.

7. The method of predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, The prediction is completed by deploying an interactive graphical user interface system developed based on a Python framework. After the user inputs all or part of the input variables such as the aggregation temperature, the catalyst dosage and the ligand structure parameters in the user interface according to the prompts, the interactive graphical user interface system will automatically call the optimal machine learning model that has been pre-trained, optimized for hyperparameters and verified for performance, and perform calculations to display the numerical prediction results of the catalyst activity, the number average molecular weight and the molecular weight distribution on the user interface.

8. The method of predicting the catalytic performance of a synthetic metallocene polypropylene according to claim 1, characterized in that, After the prediction output is completed, based on the prediction results and the feature importance ranking of SHAP analysis, the influence trend of each process parameter and ligand structure parameter on the catalytic performance is quantitatively analyzed.

Images