A Multi-Objective Product Process Control Method and System Based on Machine Learning

Abstract

This invention provides a multi-objective product process control method and system based on machine learning, belonging to the field of machine learning technology. The control method classifies the original dataset into three categories—MG, EG, and ET—based on yield results, and improves the accuracy of model classification prediction through the SMOTE improvement method. Based on the SHAP method and partial dependency analysis, the relative importance of different input feature results, as well as their positive and negative correlation effects and trends on different products in the DMO hydrogenation process, are explored. The classification results obtained from the input feature data are returned to APR. 2 - In the ML regressor, APR is used 2 The ML model, combined with a genetic algorithm, optimizes the yield of different target products, improves catalyst characteristic parameters and reaction conditions, and can screen out the optimal catalyst. By applying this model to search the global optimization space, it is possible to find the best-performing novel catalysts for the hydrogenation of DMO to MG, EG, and ET, thus realizing multi-objective optimization of catalysts for multi-target product processes.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and specifically to a multi-objective product process control method and system based on machine learning. Background Technology

[0002] Catalyst development has historically relied on experimental trial and error, requiring significant experimental time and resources. However, with the help of machine learning, catalyst screening and optimization have become much more efficient. Machine learning can handle complex, multi-dimensional data, quickly identifying promising candidates by predicting and simulating the performance of numerous different catalyst combinations. This approach is rapidly gaining traction in materials science, catalysis research, and other fields. Traditional machine learning-based catalyst optimization and screening typically involves five major steps: data collection and preprocessing, feature selection and extraction, model selection and training, model validation and testing, and catalyst performance prediction and screening.

[0003] Zahrte et al. (2019) proposed a computer-driven machine learning workflow for predicting highly selective catalysts and argued that machine learning methods will provide chemists with a new paradigm for selecting and optimizing catalysts, moving from experience-guided to math-guided approaches.

[0004] Maleye et al. (2020) calculated 14 descriptors of the molecular structure of 105 Cr(P,N) catalysts, used density functional theory to calculate the transition state energy difference between 1-hexene and 1-octene in the catalytic oligomerization of ethylene, and established a random forest regression model to predict the efficiency and selectivity of the ethylene oligomerization reaction to produce specific products, and guided the design and optimization of the catalyst for this reaction.

[0005] Chandanae et al. (2023) developed a mathematical modeling (ML) framework for modeling and predicting the catalyst activity for the direct conversion of CO2 to low-carbon olefins (C2–C4) based on structural composition operating parameters. This framework includes 16 input parameters and catalyst activity as an output parameter. Eight ML models were developed, including linear, tree-based, kernel-based, and artificial neural network (ANN) models, and these models were used to predict catalyst activity.

[0006] However, most machine learning (ML) models used for catalyst development are designed for a single product and cannot be adapted to the design of multi-target catalysts for processes such as the selective hydrogenation of dimethyl oxalate (DMO). In addition, traditional ML models mostly use regression prediction based on the yield of the target product, making it difficult to predict and classify multiple target products based on catalyst characteristics and reaction condition parameters. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a multi-objective product process control method and system based on machine learning, which solves the problem of difficulty in achieving multi-objective optimization of catalysts in multi-objective product processes.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A multi-objective product process control method based on machine learning, the method comprising the following steps:

[0010] S1. Collect relevant data on catalyst properties and reaction conditions for multi-target product processes, study the data distribution, and evaluate the data quality through data preprocessing and correlation analysis between characteristics before inputting the data into APR. 2 In ML models;

[0011] S2. Evaluate APR using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML classifier model is used to automatically filter the desired target products, and then adaptive product data is automatically returned to APR via automated feedback. 2 In the ML model regressor, the APR of the root mean square error is used after cross-validation, based on the coefficient of determination. 2 -Evaluate and improve the hyperparameters of the ML regressor;

[0012] S3. Exploring the role of SHAP-based feature importance analysis and partial dependency graphs in APR 2 -Analyze the degree and trend of the influence of input on output characteristics under ML, and analyze which characteristic parameters play an important role in automatic adjustment;

[0013] S4. Collect the multi-target product process catalyst product data that has been automatically classified and screened, and automatically feed it back to APR. 2 In the regressor of -ML, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as catalyst properties and reaction conditions, and automatically generate parameter data to maximize the yield of the final screened product.

[0014] S5, APR 2 -ML models are used in multi-objective optimization screening and prediction of novel catalysts for multi-objective product processes, automatically searching for potential candidate catalysts with better catalytic performance from the global optimization space.

[0015] Preferably, the multi-target product process is set as a DMO hydrogenation process.

[0016] Preferably, the APR 2 - The classifier for the ML model uses either the RF model or the XGB model.

[0017] Preferably, the test set accuracy of the RF model is 0.85, and the recall rates on MG, EG, and ET are 0.80, 0.86, and 0.86, respectively.

[0018] The XGB model has a test set accuracy of 0.84 and recalls of 0.80, 0.85, and 0.86 on MG, EG, and ET, respectively.

[0019] Preferably, the APR 2 - The regressor for the ML model uses the XGB model.

[0020] Preferably, the XGB model has a coefficient of determination of 0.86 on the test set, and has the best predictive ability for the yield of product MG.

[0021] Preferably, the optimal catalyst for the hydrogenation of DMO to MG, as determined by the TOPSIS decision method, is Ag-Ni / SBA-15 with 7.97 wt% Ag loaded; the optimal catalyst for the hydrogenation of DMO to EG, is Cu-Au / SBA-15 with 5.32 wt% Cu loaded; and the optimal catalyst for the hydrogenation of DMO to ET, is Cu-HZ38 / SiO2 with 12.55 wt% Cu loaded.

[0022] Preferably, step S5 includes the following specific steps:

[0023] APR 2 - The ML model is combined with the NSGA-II algorithm to find and predict the novel catalysts with optimal performance for the hydrogenation of DMO to MG, EG and ET in the global optimization space of catalyst properties and reaction conditions, in order to maximize the yield and minimize the active component loading.

[0024] The NSGA-II algorithm with a group size of 1000 and a maximum number of evolutions of 100 was used to search the feature space and serve as the APR. 2 - The input to the ML model is used to compute the function value of the objective, and the mutation probability and crossover probability are set to 0.2 and 0.8, respectively;

[0025] The optimization model was run 100 times, and the results of the 100 runs were compared to ensure that it was the global optimum and to avoid getting trapped in local minima during the space search process of the NSGA-II algorithm.

[0026] Preferably, based on APR combined with the NSGA-II algorithm, 2The ML model predicted six novel catalysts for the hydrogenation of DMO to prepare MG, EG, and ET. Among them, two Ag-based catalysts were used for the preparation of MG, and four Cu-based catalysts were used for the preparation of EG and ET. They are Ag-B2O3 / HAP, Ag-CuO / KCC-1, Cu-Au / SNT, Cu-Al2O3 / SiO2, Cu-B / SNT, and Cu-Ni-foam / RGO.

[0027] A multi-objective product process control system based on machine learning, the system comprising: a data collection and evaluation module, a model selection and optimization module, an interpretability analysis module, a catalyst optimization module, and a model application module;

[0028] The data collection and evaluation module is used to collect relevant data on catalyst properties and reaction conditions for multi-target product processes, and to study the data distribution. After evaluating the data quality through data preprocessing and correlation analysis between features, the data is input into the APR (Automatic Processing Unit). 2 In ML models;

[0029] The model selection and tuning module is used to evaluate APR using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML classifier model is used to automatically filter the desired target products, and then adaptive product data is automatically returned to APR via automated feedback. 2 In the ML model regressor, the APR of the root mean square error is used after cross-validation, based on the coefficient of determination. 2 -Evaluate and improve the hyperparameters of the ML regressor;

[0030] The interpretability analysis module is used to explore APR using SHAP-based feature importance analysis and partial dependency graphs. 2 -Analyze the degree and trend of the influence of input on output characteristics under ML, and analyze which characteristic parameters play an important role in automatic adjustment;

[0031] The catalyst optimization module is used to collect product data of multi-objective product process catalysts that have been automatically classified and screened, and automatically feed them back to APR. 2 In the regressor of -ML, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as catalyst properties and reaction conditions, and automatically generate parameter data to maximize the yield of the final screened product.

[0032] The model application module is used to apply APR. 2 -ML models are used in multi-objective optimization screening and prediction of novel catalysts for multi-objective product processes, automatically searching for potential candidate catalysts with better catalytic performance from the global optimization space.

[0033] This invention provides a multi-objective product process control method and system based on machine learning. Compared with existing technologies, it has the following advantages:

[0034] In this invention, the original dataset is classified into three categories—methyl glycolate (MG), ethylene glycol (EG), and ethanol (ET)—based on the different yield results. The SMOTE improvement method is used to enhance the accuracy of the model's classification prediction, making the machine learning model more accurate in classifying results based on input feature data even without experimental testing. Based on the SHAP method and partial dependency analysis, the relative importance of different input feature results and their positive and negative correlations and trends on different products of the DMO hydrogenation process are explored. The classification results obtained from the input feature data are returned to the APR (Automatic Partial Regression Process). 2 - In the ML regressor, APR is used 2 The ML model, combined with a genetic algorithm, optimizes the yield of different target products, improves catalyst characteristic parameters and reaction conditions, and can screen out the optimal catalyst. By applying this model to search the global optimization space, it is possible to find the best-performing novel catalysts for the hydrogenation of DMO to MG, EG, and ET, thus realizing multi-objective optimization of catalysts for multi-target product processes. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the control method in an embodiment of the present invention;

[0037] Figure 2 A comparison of the confusion matrices of the four models SVM(a), DNN(b), RF(c), and XGB(d) in Embodiment 1 of the present invention;

[0038] Figure 3 In Embodiment 2 of this invention, the SVM, DNN, RF, and XGB algorithms are represented as APRs using MG. 2 - Comparison of the predictive performance of regressors in ML models;

[0039] Figure 4 The Pareto set results for the multi-objective optimization of the DMO hydrogenation catalyst in Example 3 of this invention;

[0040] Figure 5 The APR in Embodiment 4 of the present invention 2- Comparison of six novel optimal catalyst parameters predicted by ML models combined with genetic algorithms. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0042] This application provides a multi-objective product process control method and system based on machine learning, which solves the problem that it is difficult to achieve multi-objective optimization of catalysts in multi-objective product processes.

[0043] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0044] Example:

[0045] In this embodiment, the multi-target product process is exemplified by the DMO hydrogenation process.

[0046] like Figure 1 As shown, this invention provides a multi-objective product process control method based on machine learning, the method comprising the following steps:

[0047] S1. Data Collection and Evaluation

[0048] Data on catalyst properties (CP) and reaction conditions (RC) for the DMO hydrogenation process were collected from numerous experiments and literature studies. The distribution of the data was investigated, and the data quality was assessed through data preprocessing and correlation analysis between characteristics before being input into the APR (Automatic Processing) database. 2 In ML models;

[0049] S2. Model Selection and Tuning

[0050] APR was evaluated using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML classifier model is used to automatically filter the desired target products, and then adaptive product data is automatically returned to APR via automated feedback. 2 In the ML model regressor, the coefficient of determination (R²) after cross-validation is used. 2 ), Root Mean Square Error (RMSE) versus APR 2 - Evaluate and improve the hyperparameters of the ML regressor to enhance APR. 2 - The adaptive adjustment capability of ML models when facing different target products;

[0051] S3, Interpretability Analysis

[0052] Explore the role of SHAP-based feature importance analysis and partial dependency graphs in APR 2 -Analyze the degree and trend of the influence of input on output characteristics under ML, and analyze which characteristic parameters play an important role in automatic adjustment in order to better adapt to product needs;

[0053] S4. Catalyst Optimization

[0054] Data on DMO hydrogenation catalyst products that have been automatically categorized and filtered will be collected and automatically fed back to APR. 2 In the regression analyzer of -ML, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as CP and RC, and automatically generate parameter data to maximize the yield of the finally screened DMO hydrogenation product.

[0055] S5, APR 2 Applications of ML models

[0056] APR 2 - The ML model is used in the multi-objective optimization screening and prediction of novel DMO hydrogenation catalysts, automatically searching for potential candidate DMO hydrogenation catalysts with better catalytic performance from the global optimization space.

[0057] This invention provides a multi-objective product process control system based on machine learning. The system includes: a data collection and evaluation module, a model selection and optimization module, an interpretability analysis module, a catalyst optimization module, and a model application module.

[0058] The data collection and evaluation module is used to collect relevant data on catalyst properties and reaction conditions for multi-target product processes, and to study the data distribution. After evaluating the data quality through data preprocessing and correlation analysis between features, the data is input into the APR (Automatic Processing Unit). 2 In ML models;

[0059] The model selection and tuning module is used to evaluate APR using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML classifier model is used to automatically filter the desired target products, and then adaptive product data is automatically returned to APR via automated feedback. 2 In the ML model regressor, the APR of the root mean square error is used after cross-validation, based on the coefficient of determination. 2 - Evaluate and improve the hyperparameters of the ML regressor to enhance APR. 2 - The adaptive adjustment capability of ML models when facing different target products;

[0060] The interpretability analysis module is used to explore APR using SHAP-based feature importance analysis and partial dependency graphs. 2 -Analyze the degree and trend of the influence of input on output characteristics under ML, and analyze which characteristic parameters play an important role in automatic adjustment;

[0061] The catalyst optimization module is used to collect product data of multi-objective product process catalysts that have been automatically classified and screened, and automatically feed them back to APR. 2 In the regressor of -ML, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as catalyst properties and reaction conditions, and automatically generate parameter data to maximize the yield of the final screened product.

[0062] The model application module is used to apply APR. 2 -ML models are used in multi-objective optimization screening and prediction of novel catalysts for multi-objective product processes, automatically searching for potential candidate catalysts with better catalytic performance from the global optimization space.

[0063] Example 1:

[0064] The test set Acy values ​​for the RF and XGB models were 0.85 and 0.84, respectively, under the structures of n_estimators=71, max_depth=20, min_samples_split=2, min_samples_leaf=1 and n_estimators=86, max_depth=15, learning_rate=0.25, min_child_weight=8. The mean Acy values ​​for the two models under 5x cross-validation were also as high as 0.84 and 0.83.

[0065] The SVM and DNN models performed poorly on this type of dataset, with Acy scores of 0.8 and 0.81 respectively on the test set.

[0066] Therefore, based on the overall model accuracy (Acy) metric, it was initially determined that RF and XGB are more suitable for APR compared to SVR and DNN ML models. 2 - Classifier for ML models.

[0067] like Figure 2 As shown, the confusion matrix provides a clearer view of the accuracy of the four ML models under three different classification metrics: MG, EG, and ET. The darker value in the middle represents recall (R), and the higher the value, the more accurate the model classification.

[0068] observe Figure 2(ad) It can be found that the R-values ​​of the four ML models under the confusion matrix are significantly different. The R-values ​​of RF and XGB are higher than those of SVR and DNN models under each metric. Their R-values ​​for MG, EG, and ET are 0.80, 0.86, 0.86 and 0.80, 0.85, 0.86, respectively. In contrast, the highest R-values ​​of SVR and DNN models for the three target predictions are 0.82 and 0.81, respectively, which are achieved on the classification task of predicting the ET target. This shows that the RF and XGB models have a significant advantage in capturing different class samples in this type of dataset.

[0069] Example 2:

[0070] like Figure 3 As shown, the prediction performance of four models—SVM, DNN, RF, and XGB—under DMO conversion rate is compared. Figure 3 (a) shows the comparison results between the model predictions and the actual values. It can be found that all models can fit the actual values ​​well in predicting the results, and the errors on the test set are mostly stable within 10%, indicating that the four models have a certain accuracy in predicting the results.

[0071] However, according to Figure 3 (c) The results show that XGB has the highest R-value in the final prediction. 2 Its result on the test set is 0.86, so its yield Y of product MG is... MG The R-squared model has the best predictive power, followed by the RF model. 2 The R value was 0.84, while DNN and SVR did not show a significant advantage over the two ensemble models on this type of data, with an R value of 0.84 on the test set. 2 They are 0.80 and 0.83 respectively.

[0072] pass Figure 3 As shown in (d), the error distribution density reveals that the XGB model has a denser distribution near the error level of 0, with the highest frequency approaching 0.6. Therefore, XGB possesses the best predictive ability for this dataset. Based on the above analysis, XGB is ultimately determined to be the best predictive model for APR. 2 - The optimal regressor for ML models.

[0073] Example 3:

[0074] like Figure 4 As shown, we use the TOPSIS (Order Preference Technique of Similarity to an Ideal Solution) decision method to select the optimal value on the boundary. In Ag-based catalysts, such as... Figure 4As shown in (ac), the TOPSIS method identifies Ag-Ni / SBA-15 (82.60%, 7.97 wt%), Ag-Ti / KCC-1 (71.73%, 9.95 wt%), and Ag-B2O3 / SiO2 (76.94%, 12.62 wt%) as the optimal catalysts for the hydrogenation of DMO to MG, EG, and ET, respectively. It can be observed that Ag-based catalysts have significant advantages in the hydrogenation of DMO to MG, achieving an optimized MG yield of 82.60%, higher than the 71.73% and 76.94% for EG and ET, respectively, while requiring only a lower active component loading of 7.97 wt%. For Cu-based catalysts, as... Figure 3 As shown in (df), the TOPSIS method identifies Cu-Al₂O₃ / ZrO₂ (77.59%, 13.15 wt%), Cu-Au / SBA-15 (83.20%, 5.32 wt%), and Cu-HZ₃₈ / SiO₂ (82.60%, 12.55 wt%) as the optimal catalysts for the hydrogenation of DMO to MG, EG, and ET, respectively. Cu-based catalysts exhibit good performance in EG preparation, outperforming MG and ET in both yield (83.20%) and active component loading (5.32 wt%). Cu-based catalysts also show excellent performance in ET, with TOPSIS indicating that (82.60%, 12.55 wt%) is a relatively advantageous choice. Overall, Ag catalysts are more suitable for the hydrogenation of DMO to MG, while Cu-based catalysts are more advantageous in the preparation of EG and ET.

[0075] Example 4:

[0076] like Figure 5 As shown, APR is used. 2 - By combining the ML model with the NSGA-II algorithm, in the global optimization space of catalyst properties and reaction conditions, we can find and predict novel catalysts with optimal performance for the hydrogenation of DMO to MG, EG, and ET, in order to maximize yield and minimize active component loading.

[0077] The NSGA-II algorithm with a group size of 1000 and a maximum number of evolutions of 100 was used to search the feature space and serve as the APR. 2 - The input to the ML model is used to calculate the function value of the objective, and the mutation probability and crossover probability are set to 0.2 and 0.8, respectively.

[0078] The optimization model was run 100 times, and the results of the 100 runs were compared to ensure that it was the global optimum and to avoid getting trapped in local minima during the space search process of the NSGA-II algorithm.

[0079] Finally, based on the combination with the NSGA-II algorithm, APR 2 The ML model successfully predicted six novel catalysts for the hydrogenation of DMO to produce MG, EG, and ET. Two of these are Ag-based catalysts for MG production, and four are Cu-based catalysts for EG and ET production. These are Ag-B₂O₃ / HAP, Ag-CuO / KCC-1, Cu-Au / SNT, Cu-Al₂O₃ / SiO₂, Cu-B / SNT, Cu-Ni-foam / RGO, and Cu-Ni-foam / RGO. Figure 5 Their catalytic performance was compared in (a)-(b).

[0080] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0081] In this embodiment of the invention, the original dataset is classified into three categories—methyl glycolate (MG), ethylene glycol (EG), and ethanol (ET)—based on the different yield results. The SMOTE improvement method is used to enhance the accuracy of the model's classification prediction, making the machine learning model more accurate in classifying results based on input feature data even without experimental testing. Based on the SHAP method and partial dependency analysis, the relative importance of different input feature results and their positive and negative correlation effects and trends on different products of the DMO hydrogenation process are explored. The classification results obtained through the input feature data are returned to the APR (Automatic Partial Regression Process). 2 - In the ML regressor, APR is used 2 The ML model, combined with a genetic algorithm, optimizes the yield of different target products, improves catalyst characteristic parameters and reaction conditions, and can screen out the optimal catalyst. By applying this model to search the global optimization space, it is possible to find the best-performing novel catalysts for the hydrogenation of DMO to MG, EG, and ET, thus realizing multi-objective optimization of catalysts for multi-target product processes.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective product process control method based on machine learning, characterized in that, The method includes the following steps: S1. Collect relevant data on catalyst properties and reaction conditions for multi-target product processes, study the data distribution, and evaluate the data quality through data preprocessing and correlation analysis between characteristics before inputting the data into APR. 2 In ML models; APR 2 - The classifier for the ML model uses either the RF model or the XGB model; APR 2 - The regressor for the ML model uses the XGB model; S2. Evaluate APR using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML model classifier automatically filters the desired target products, and then automatically returns adaptive product data to APR via automated feedback. 2 In the regressor of the ML model, the root mean square error (RMSE) is used to evaluate the APR (Advanced Percentage Regression). 2 -Evaluate and improve the hyperparameters of the regressor in the ML model; S3. Exploring the role of SHAP-based feature importance analysis and partial dependency graphs in APR 2 -Analyze the degree and trend of the influence of input on output features in the ML model, and analyze which feature parameters play an important role in automatic adjustment; S4. Collect the multi-target product process catalyst product data that has been automatically classified and screened, and automatically feed it back to APR. 2 In the regressor of the ML model, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as catalyst properties and reaction conditions, and to automatically generate parameter data to maximize the yield of the final screened product. S5, APR 2 -ML models are used in multi-objective optimization screening and prediction of novel catalysts for multi-objective product processes, automatically searching for potential candidate catalysts with better catalytic performance from the global optimization space.

2. The multi-objective product process control method based on machine learning as described in claim 1, characterized in that, The multi-target product process is set as a DMO hydrogenation process.

3. The multi-objective product process control method based on machine learning as described in claim 1, characterized in that, The test set accuracy of the RF model is 0.85, and the recall rates on MG, EG, and ET are 0.80, 0.86, and 0.86, respectively. The XGB model has a test set accuracy of 0.84 and recalls of 0.80, 0.85, and 0.86 on MG, EG, and ET, respectively.

4. The multi-objective product process control method based on machine learning as described in claim 1, characterized in that, The XGB model has a coefficient of determination of 0.86 on the test set, and has the best predictive ability for the yield of product MG.

5. The multi-objective product process control method based on machine learning as described in claim 2, characterized in that, The optimal catalysts for the hydrogenation of DMO to MG were determined using the TOPSIS decision method: Ag-Ni / SBA-15 with 7.97 wt% Ag loading; Cu-Au / SBA-15 with 5.32 wt% Cu loading; and Cu-HZ38 / SiO2 with 12.55 wt% Cu loading.

6. The multi-objective product process control method based on machine learning as described in claim 2, characterized in that, S5 includes the following specific steps: APR 2 - The ML model is combined with the NSGA-II algorithm to find and predict the novel catalysts with optimal performance for the hydrogenation of DMO to MG, EG and ET in the global optimization space of catalyst properties and reaction conditions, in order to maximize the yield and minimize the active component loading. The NSGA-II algorithm with a group size of 1000 and a maximum number of evolutions of 100 was used to search the feature space and serve as the APR. 2 - The input to the ML model is used to compute the function value of the objective, and the mutation probability and crossover probability are set to 0.2 and 0.8, respectively; The optimization model was run 100 times, and the results of the 100 runs were compared to ensure that it was the global optimum and to avoid getting trapped in local minima during the space search process of the NSGA-II algorithm.

7. The multi-objective product process control method based on machine learning as described in claim 6, characterized in that, Based on the combination with the NSGA-II algorithm, APR 2 The ML model predicted six novel catalysts for the hydrogenation of DMO to prepare MG, EG, and ET. Among them, two Ag-based catalysts were used for the preparation of MG, and four Cu-based catalysts were used for the preparation of EG and ET. They are Ag-B2O3 / HAP, Ag-CuO / KCC-1, Cu-Au / SNT, Cu-Al2O3 / SiO2, Cu-B / SNT, and Cu-Ni-foam / RGO.

8. A multi-objective product process control system based on machine learning, characterized in that, The system includes: a data collection and evaluation module, a model selection and optimization module, an interpretability analysis module, a catalyst optimization module, and a model application module; The data collection and evaluation module is used to collect relevant data on catalyst properties and reaction conditions for multi-target product processes, and to study the data distribution. After evaluating the data quality through data preprocessing and correlation analysis between features, the data is input into the APR (Automatic Processing Unit). 2 In ML models; APR 2 - The classifier for the ML model uses either the RF model or the XGB model; APR 2 - The regressor for the ML model uses the XGB model; The model selection and tuning module is used to evaluate APR using micro-average, macro-average, and AUC values ​​under the ROC curve. 2 - An ML model classifier automatically filters the desired target products, and then automatically returns adaptive product data to APR via automated feedback. 2 In the regressor of the ML model, the root mean square error (RMSE) is used to calculate the average regression ratio (APR) after cross-validation based on the coefficient of determination. 2 -Evaluate and improve the hyperparameters of the regressor in the ML model; The interpretability analysis module is used to explore APR using SHAP-based feature importance analysis and partial dependency graphs. 2 -Analyze the degree and trend of the influence of input on output features in the ML model, and analyze which feature parameters play an important role in automatic adjustment; The catalyst optimization module is used to collect product data of multi-objective product process catalysts that have been automatically classified and screened, and automatically feed them back to APR. 2 In the regressor of the ML model, the optimization algorithm is combined to perform multi-parameter coordinated control and optimization of data such as catalyst properties and reaction conditions, and to automatically generate parameter data to maximize the yield of the final screened product. The model application module is used to apply APR. 2 -ML models are used in multi-objective optimization screening and prediction of novel catalysts for multi-objective product processes, automatically searching for potential candidate catalysts with better catalytic performance from the global optimization space.

Images