Catalyst process design and optimization method and device

Abstract

The invention discloses a catalyst process design and optimization method and device, and belongs to the field of machine learning. The method comprises the steps of obtaining original process data in a catalyst production process, performing data cleaning and feature engineering processing, and generating a preprocessing data set capable of being used for modeling; screening a plurality of candidate machine learning models by using an AutoML technology, and performing hyper-parameter optimization on the baseline model; an incremental learning technology is adopted to update the optimized model online in real time so as to adapt to data fluctuation in the production process; and finally, the online updated model is applied to quality prediction and process parameter adjustment, so that closed-loop optimization of the production process is realized, and the product consistency and the production efficiency are improved. According to the method, the automation level and the intelligent quality control capability in catalyst production can be effectively improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of machine learning, and in particular to a catalyst process design and optimization method and device. BACKGROUND

[0002] As a core material in key fields such as chemical industry and energy, the stability and controllability of the quality of catalysts are the key to improving industrial competitiveness and breaking through technical blockades. In the traditional production process of catalysts, quality control often relies on post-detection and trial-and-error experience. This way of relying on manual and experimental verification not only is inefficient, but also is difficult to respond in real time to quality problems caused by multivariate coupling fluctuations in high-intensity and fast-paced production processes, resulting in problems such as yield fluctuation and poor product consistency in catalyst production. These problems have seriously restricted the industrialized production and high-end development of catalysts, especially in large-scale production, it is difficult to achieve high-precision and high-efficiency quality optimization.

[0003] Response surface methodology (RSM) as a classic experimental design and modeling tool is widely used in process optimization of catalyst production. However, the traditional RSM method has exposed a series of problems in high-dimensional and high-complexity production processes. First, RSM relies on polynomial fitting, and when facing high-dimensional and complex nonlinear mapping, the prediction accuracy of the model will decrease exponentially with the increase of the dimension, which makes it difficult for RSM to accurately predict and optimize the coupling effect of multiple variables. Second, the traditional RSM modeling process needs to periodically interrupt production for experimental data collection, which is not practical in a high-intensity production environment. The frequent changes and diversity of process parameters in production lead to an explosive increase in the cost of model reconstruction of RSM with the increase in the number of variables, which cannot meet the demand for real-time and efficient model updating in modern industrial environments.

[0004] In addition, the traditional quality optimization method relies on manual experience to adjust process parameters, lacks effective intelligent decision support systems, and cannot quickly respond to unforeseen changes in the production process. This makes it not only impossible to identify quality fluctuations in production in a timely manner, but also can lead to huge waste and accumulation of quality problems in the production process.

[0005] Under this background, the existing traditional modeling methods and quality optimization means cannot meet the demand for real-time, efficient and intelligent quality optimization in the era of Industry 4.0. Therefore, there is an urgent need for a new technical solution that can break through the limitations of traditional methods and solve the quality design and optimization problems in the production process of catalysts through automated and intelligent means. SUMMARY

[0006] The embodiment of the application provides a catalyst process design and optimization method and device, which can solve the problems of complex optimization process and low optimization efficiency in the related art when a multi-lingual optimization is performed in an artificial manner. The technical solution is as follows: In one aspect, a catalyst process design and optimization method is provided, which comprises the following steps: Obtaining original process data in a catalyst production process; Performing data cleaning and feature engineering on the original process data to generate a pretreatment data set that can be used for modeling; Filtering a plurality of candidate machine learning models based on the pretreatment data set by using an automatic machine learning (AutoML) technology; Optimizing hyperparameters of a selected baseline model based on a generative intelligent process system; Performing online updating on the optimized baseline model by using an incremental learning technology; Applying the online updated baseline model to quality prediction and process parameter adjustment in the catalyst production process.

[0007] Optionally, the original process data in the catalyst production process is obtained by: Collecting parameter data of raw material groups, process groups and equipment groups in the catalyst production process; Obtaining real-time environmental data in the catalyst production process; Collecting the original process data in real time according to the collection frequency of a data collection module.

[0008] Optionally, the original process data is cleaned and feature engineered to generate a pretreatment data set that can be used for modeling, comprising: Performing abnormality detection on the original process data by using a 3 sigma rule to remove abnormal values; Filling in missing values in the original process data by using an interpolation method; Filtering features in the original process data according to Pearson correlation coefficients and Spearman correlation coefficients to filter out features with a correlation greater than 0.3 with a quality index; Performing dimension reduction on the original process data by using a principal component analysis (PCA) method to retain principal components with a higher contribution rate.

[0009] Optionally, the plurality of candidate machine learning models are filtered based on the pretreatment data set by using an automatic machine learning (AutoML) technology, comprising: According to the characteristics of the catalyst process data, a suitable learning algorithm is automatically selected and a candidate model set is constructed; The candidate models are evaluated using cross-validation to obtain performance evaluation metrics for each candidate model. Based on the performance evaluation metrics, the baseline model with the best performance is selected.

[0010] Optionally, the generative intelligent process system optimizes the hyperparameters of the selected baseline model, including: Define the hyperparameter search space for the optimal baseline model; Hyperparameter optimization based on Gaussian process regression in generative intelligent process systems; The optimal combination of hyperparameters is selected by maximizing the expected value improvement criterion.

[0011] Optionally, the application of incremental learning techniques to update the optimized baseline model online includes: The system receives new production data from the catalyst production process in real time and processes it in batches. The new production data is divided into batches using a sliding window technique, and the parameters of the optimal baseline model are updated online based on the data. The optimal baseline model, updated online, is validated to ensure it can adapt to new data in the production process.

[0012] Optionally, applying the online-updated baseline model to quality prediction and process parameter adjustment in catalyst production includes: The online-updated optimal baseline model is applied to predict quality indicators in the catalyst production process in real time. Based on the quality prediction results, key process parameters in the catalyst production process are adjusted in real time. The model is corrected based on actual quality data during the production process to achieve closed-loop optimization in the production process.

[0013] On the other hand, a catalyst process design and optimization apparatus is provided, the apparatus comprising: The data acquisition module is used to acquire raw process data during catalyst production. The preprocessing module is used to perform data cleaning and feature engineering on the raw process data to generate a preprocessed dataset that can be used for modeling. The model selection module is used to select multiple candidate machine learning models based on the preprocessed dataset using Automated Machine Learning (AutoML) technology. The parameter optimization module is used to optimize the hyperparameters of the selected baseline model based on the generative intelligent process system. The model update module is used to apply incremental learning techniques to update the optimized baseline model online. The parameter adjustment module is used to apply the online updated baseline model to quality prediction and process parameter adjustment in the catalyst production process.

[0014] On the other hand, a parameter optimization device is provided, the device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being executed by the processor to implement the catalyst process design and optimization method as described above.

[0015] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, the at least one program, the code set, or the instruction set is executed by the processor to implement the catalyst process design and optimization method as described above.

[0016] The beneficial effects of the technical solution provided in this application include at least the following: a catalyst process design and optimization method and apparatus, comprising acquiring raw process data during catalyst production, performing data cleaning and feature engineering to generate a preprocessed dataset that can be used for modeling; using AutoML technology to screen multiple candidate machine learning models and optimizing the hyperparameters of the baseline model; employing incremental learning technology to update the optimized model online in real time to adapt to data fluctuations during production; and finally applying the online-updated model to quality prediction and process parameter adjustment to achieve closed-loop optimization of the production process, thereby improving product consistency and production efficiency. This method can effectively improve the automation level and intelligent quality control capabilities in catalyst production. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This illustration shows a schematic flow chart of a catalyst process design and optimization method according to an embodiment of this application; Figure 2 This illustration shows a schematic diagram of a technology stack structure provided in one embodiment of this application; Figure 3 This illustration shows a schematic diagram of a 3σ discrepancy rule provided in one embodiment of this application; Figure 4 A comparison chart of three correlation coefficient formulas provided in one embodiment of this application is shown; Figure 5 A comparative diagram of two standardized methods provided in one embodiment of this application is shown; Figure 6 A model evaluation index diagram provided in one embodiment of this application is shown; Figure 7 This is a schematic diagram illustrating the principle of separating sample optimization and sample background noise processes; Figure 8 A block diagram of a catalyst process design and optimization apparatus provided in one embodiment of this application is shown; Figure 9 A schematic diagram of the structure of a parameter optimization device provided in one embodiment of this application is shown; Figure 10 This illustration shows a schematic diagram of principal component contributions provided in one embodiment of this application; Figure 11 This illustration shows a principal component analysis diagram for objective 1 provided in one embodiment of this application; Figure 12 This illustration shows a principal component analysis diagram for objective 2 provided in one embodiment of this application; Figure 13 This paper illustrates the BO-SVR optimization process and model performance diagram of objective 1 provided in one embodiment of this application; Figure 14 This paper illustrates the model hyperparameter iteration process of objective 1 provided in one embodiment of the present application; Figure 15 The BO-SVR optimization process and model performance diagram of objective 2 provided in one embodiment of this application are shown; Figure 16 This paper illustrates the model hyperparameter iteration process of objective 2 provided in one embodiment of this application; Figure 17 The diagram illustrates the performance of a Lasso regression model for two target values ​​according to an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0020] Before introducing specific embodiments, let me briefly describe the research background of this patent.

[0021] In the current context of profound adjustments in the global manufacturing landscape, the fourth paradigm (data-intensive scientific discovery) combining big data and models, along with the AI4S (AI for Science) technology wave, is reshaping the competitive logic of the catalyst manufacturing industry. With the intensification of anti-globalization trends, catalysts, as core materials in key fields such as chemicals and energy, face critical barriers to overcoming technological blockades and building industrial competitiveness due to their performance stability and quality controllability. Under traditional manufacturing models, quality control relies on post-production testing and trial-and-error, making it difficult to respond promptly to multi-variable coupled fluctuations in high-intensity, fast-paced production environments. This leads to poor product consistency and fluctuating yield rates, severely hindering the localization of high-end catalysts. Modern quality theory points out that quality is essentially a systematic output of the design and manufacturing process, not the result of end-of-pipe testing. Small fluctuations in variables such as material properties, process parameters, and equipment status during production accumulate and amplify through multi-physics coupling, ultimately becoming the root cause of quality problems. Among key technologies for quality design, Response Surface Methodology (RSM), as a classic experimental design and modeling tool, has long been used to analyze the mapping relationship between process parameters and quality indicators. However, in industrial practice, this method has revealed significant limitations: First, high-intensity production scenarios involve process parameters of tens or even hundreds of dimensions. Traditional RSM, which relies on multinomial fitting, struggles to characterize complex nonlinear mappings in high-dimensional spaces, leading to an exponential decrease in prediction accuracy as the dimensionality of variables increases. Second, fast-paced production demands real-time model updates, but RSM modeling requires periodic production interruptions to obtain experimental data, and the cost of model reconstruction increases exponentially with the number of variables, failing to meet the dynamic quality control requirements of "production and optimization" in the Industry 4.0 era. Therefore, overcoming the dimensional bottlenecks and real-time limitations of traditional modeling methods has become a common technical challenge that urgently needs to be addressed in the field of catalyst manufacturing quality design.

[0022] In the field of industrial catalysis, quality design and process optimization technologies are undergoing a paradigm shift from experience-driven to data-driven approaches, including improvements and expansions of RSM, deep applications of machine learning, and the integration of parameter optimization and online learning technologies.

[0023] Existing research reveals that the application of response surface methodology (RSM) in industrial quality design primarily focuses on catalyst application and waste catalyst recycling, with limited research on catalyst industrial production, remaining largely at the laboratory preparation level. Its application in the industrial manufacturing of catalysts is rarely addressed. Furthermore, while traditional mathematical modeling methods are employed in catalyst manufacturing and application processes, the high dimensionality and nonlinearity of variables in the process-oriented industry of catalysts highlight the limitations of traditional mathematical methods in characterization. In addition, parameter fluctuations in fast-paced production leading to decreased model generalization performance is another significant challenge.

[0024] Machine learning technology has been used in industrial quality prediction for nearly 20 years, evolving from early support vector machines to XGBoost, LSTM5, and deep learning, and further to the recent emergence of Automated Machine Learning (AutoML). With the widespread application of big data across various fields, more and more stakeholders are looking to fully utilize their data, driving researchers to explore the possibilities of AutoML. AutoML tools aim to enable non-machine learning experts (domain experts) to use machine learning, improve its efficiency, and accelerate research. Catalysts, as a branch of materials science, exhibit significant characteristics of small to medium data volumes. For machine learning with small data, scholars in the medical and biological fields were among the first to propose validation strategies for machine learning algorithms with small sample sizes, providing insights for materials science. Subsequently, scholars have optimized casting processes by combining phase-field simulation, addressed data scarcity issues by combining Latin hypercube sampling with Bayesian optimization, and solved the problem of insufficient physical interpretability using SHAP value analysis. Although machine learning has become a tool for accelerating materials design and discovery, the problem of data scarcity remains severe. To address this, some scholars have developed high-throughput computing methods by extracting literature and constructing materials databases, providing algorithms that are friendly to small data volumes. Furthermore, given the significant focus on model accuracy in materials science, some scholars have reduced prediction errors by incorporating coarse estimates into the feature space. Support Vector Machine (SVM) regression, as a mature algorithm, possesses high-dimensional processing capabilities, strong robustness, and good generalization performance, making it commonly used in industrial quality prediction and having implemented incremental learning relatively early. However, considering the continuous nature of its parameters, there are currently no effective techniques for optimizing model parameters, thus increasing the possibility of getting trapped in local optima.

[0025] Traditional parameter optimization methods employ parametric grids, which rely on setting combinations of hyperparameters. While efficient and accurate for baseline model selection, they face limitations when dealing with continuous variables. To address this, some researchers have used genetic algorithms and particle swarm optimization to optimize kernel function parameters, achieving results superior to parametric grids. To further improve efficiency and accuracy, researchers have proposed parameter optimization methods based on parallel mechanisms, employing heuristic algorithms to significantly improve the efficiency of small-sample machine learning. With technological advancements, researchers have begun experimenting with Bayesian optimization algorithms for machine learning parameter optimization (XGboost), showing superior classification performance compared to manual tuning and exhaustive methods (parametric grids). Other approaches utilize Bayesian optimization for Bagging, decision trees, random forests, and fully connected neural networks, demonstrating good results for single-objective outputs. However, these techniques are based on offline machine learning and do not consider model iteration. To address the issues of data updates and model iteration, an online restrained learning method based on batch incremental learning has emerged. This method generates offline models that can be updated at appropriate times as needed to handle constantly changing and newly added data. There are also LSSVM models developed using offline models combined with incremental learning, which effectively solve the problems of offline learning and achieve accurate online predictions. These studies demonstrate the application prospects of Bayesian optimization and incremental learning in parameter optimization and model iteration, providing useful references for catalyst process design and optimization.

[0026] To further improve the efficiency of process modeling and control factor tuning, this paper proposes a technical approach based on automated machine learning and multivariate statistical regression, combined with incremental learning to achieve online model iteration. This approach is validated in practical process design and optimization. The technology stack is as follows: Figure 2 As shown.

[0027] In catalyst manufacturing quality design, data-driven modeling methods require efficient extreme model screening as a prerequisite. This section constructs a model screening framework based on AutoML, and builds a modeling method through industrial data preprocessing, multi-model comparison and evaluation, and multi-dimensional statistical optimization.

[0028] Example 1 like Figure 1 As shown, a method for catalyst process design and optimization is provided, the method including: Step 101: Obtain the raw process data during catalyst production.

[0029] Raw data collected from industrial sites commonly suffers from noise interference, the curse of dimensionality, and collinearity of features, requiring systematic preprocessing to improve data quality.

[0030] Therefore, it is necessary to perform data cleaning and anomaly handling on the original process data, which is accomplished through the following steps.

[0031] Step 102: Perform data cleaning and feature engineering on the original process data to generate a preprocessed dataset that can be used for modeling.

[0032] The 3σ principle SPC outlier detection rule is used for data preprocessing, such as screening outliers in the original process data (raw material group, process group, equipment group parameters, etc.) of catalyst synthesis, and filling missing values ​​for jump points using interpolation. Figure 3 The figure shows a schematic diagram of the 3σ outlier rule, where the horizontal axis represents the number of sample groups and the vertical axis represents the value of the parameter.

[0033] Furthermore, feature engineering and dimensionality reduction are employed. Based on descriptive statistical results, the Pearson correlation coefficient and Spearman correlation coefficient are used to examine the correlation between the independent and dependent variables, such as... Figure 4 The diagram shows a comparison of the three correlation coefficient formulas. Dimensionality reduction is achieved by filtering out correlated parameters. For complex processes such as multi-stage heating and distributed loads, time-series features and process cumulative features are constructed to enhance the data's ability to characterize the quality formation process. PCA dimensionality reduction is used for high-dimensional data, retaining principal components with high contribution rates to balance model complexity and information loss.

[0034] Furthermore, data standardization and normalization are crucial. Z-score standardization is applied to continuous features to ensure the training stability of distance-based models such as SVR. Standardization and normalization bring data of different dimensions and distributions to the same scale, effectively improving model convergence speed and prediction accuracy. After the aforementioned data cleaning, feature engineering, and dimensionality reduction operations, different types of features still exhibit significant differences in data distribution and dimensions. To further eliminate the negative impact of these differences on model training, targeted standardization and normalization techniques are needed to optimize different types of features. The numerical fluctuation range and distribution of continuous features directly affect the model's training efficiency and prediction accuracy, while the encoding method of categorical features is crucial for the model's accurate understanding of discrete information. To address the differences in characteristics between continuous and categorical features, appropriate data transformation strategies must be adopted, such as… Figure 5 The figure shown is a comparison chart of the two standardization methods.

[0035] Step 103: Using Automated Machine Learning (AutoML) technology, select multiple candidate machine learning models based on the preprocessed dataset.

[0036] A dictionary of algorithms and typical hyperparameters is constructed by integrating 12 algorithms. Using parametric grid technology, RMSE, MAE, and R2 are used as evaluation metrics. Figure 6The diagram illustrates a model evaluation metric provided in one embodiment of this application, which selects the candidate machine learning model with the best overall performance as a baseline model for further optimization.

[0037] Step 104: Optimize the hyperparameters of the selected baseline model based on the generative intelligent process system.

[0038] Through its independently developed "Generative Intelligent Process System," the system integrates 12 mainstream machine learning algorithms and multivariate statistical methods to quickly determine the hyperparameter combination of the model using a three-step method: "baseline model screening - parameter optimization - fine-tuning."

[0039] The first step is hyperparameter construction. Based on the baseline model selected in the previous section, the available hyperparameters of the model algorithm are constructed and combined, which involves the construction of the range of continuous parameters.

[0040] The second step is hyperparameter optimization. As shown in Table 1, discontinuous parameters are combined, and continuous parameters are fitted and tuned using the central composite method, Box-Behnken design, and Bayesian optimization, respectively. Forty samples were randomly selected from the production process for comparison, and it was found that the Bayesian optimization model had higher performance metrics (MSE, R²).

[0041] Table 1 The third step is hyperparameter fine-tuning. Based on the best model selected in the second step, and combined with the experience of algorithm engineers and process engineers, a final optimization is performed. The main purpose is to avoid overfitting while ensuring the model's prediction accuracy.

[0042] The fourth step is model standardization. The optimized model is saved locally for production control. To address the complexity of industrial scenarios, data augmentation is performed in edge scenarios, supplementing data from long-tail scenarios through transfer learning. Regarding the model itself, an online learning mechanism is introduced to update parameters and address concept drift, mitigating the impact of equipment aging on quality during long-term operation. Regularization is employed to enhance the model's robustness. The model is deployed on the PLC to reduce communication latency.

[0043] Based on production needs, a central value and fluctuation range for the target value were set. Step sizes were set according to the settable range of the equipment and the controllable range of the process parameters, generating 90,000 parameter combinations. This is a small sample dataset, and the local computing power (NVIDIA A100) is sufficient to handle this scale of computation. Therefore, a full-factor combination approach was adopted to call the best model selected in the previous step for prediction. The minimum Euclidean distance from the predicted results to the set central value was calculated, and the top 10 parameter combinations with the smallest distances were selected as the process parameters for industrial trials. Parallel trials were conducted with manually adjusted process parameters. The results in the first 5 parallel trials showed that the prediction accuracy of the machine learning model was higher than that of manual parameter tuning, and the predicted pass rate was >99.5%, meeting the company's requirement for first-pass yield. Therefore, the process parameters optimized by automatic machine learning and multivariate statistical methods were standardized into a Standard Operating Procedure (SOP) to guide the production of this batch of products.

[0044] Step 105: Apply incremental learning techniques to update the optimized baseline model online.

[0045] Step 106: Apply the updated baseline model online to the quality prediction and process parameter adjustment in the catalyst production process.

[0046] Furthermore, in steps 105 and 106, the online quality optimization closed loop is further completed through incremental learning.

[0047] Based on the need to control the dynamic fluctuations of production parameters and the implementation of quality indicators in catalyst manufacturing, this section constructs an online optimization closed-loop framework based on incremental learning to solve the problems after the iteration of traditional offline models and realize an adaptive cycle of "real-time data acquisition - dynamic model update - quality prediction feedback".

[0048] Regarding this, traditional batch learning for industrial applications faces bottlenecks. Its iteration cycle is long, and offline models rely on periodic batch data reconstruction, making them unable to respond to hourly or even minute-level drifts in process parameters. Secondly, data utilization is low; each training run loads the entire historical dataset, and computational efficiency decreases as the data volume increases. Furthermore, repeated computations lead to wasted computing power and increased energy consumption. In addition, there is insufficient robustness; this full-scale computation faces gradual changes such as equipment aging and sensor accuracy degradation, causing the cumulative increase in prediction errors of offline models to show an upward trend.

[0049] Therefore, an incremental learning mechanism is needed to modify the offline model. One approach is to process the data in batches, dividing the production data into fixed-size micro-batches in real time. A sliding window technique is used to maintain a cache of recent data, setting the number of samples to retain while balancing model memory and finer-grained flexibility. For dynamic parameter updates, taking BOSVR as an example, a gradient descent algorithm is used to optimize the kernel matrix. When new samples are added, only the relevant support vectors are updated, avoiding full retraining to reduce the time consumption of online updates. Simultaneously, SPC is used to monitor the residual distribution of the model, and a supplementary training mechanism is triggered using outlier rules to selectively update affected feature dimensions.

[0050] The adaptive optimization architecture is implemented in three layers. At the edge computing layer, a lightweight data preprocessing module is deployed to perform real-time standardization and outlier filtering, transforming raw production data into model-recognizable feature vectors. At the incremental learning layer, offline training models such as BOSVR are used as output parameters. Batch data drives model updates, outputting real-time predicted quality indicators and calculating prediction confidence. At the closed-loop control layer, a mapping rule between predicted values ​​and process parameters is established. Adjustment commands are sent to the PLC via the OPCUA protocol, or the DCS system is automatically adjusted, forming a prediction-control-feedback process with a second-level response time.

[0051] To enhance model robustness, the impact of data noise must first be reduced. Therefore, during incremental updates, exponential weighting and moving averages need to be introduced, with decreasing weights assigned to high-frequency noisy data to suppress the cumulative effect of random measurement errors. Simultaneously, the possibility of model degradation must be considered. A historical best-in-class model library needs to be maintained, and a threshold should be manually set to control the retention range of the best model. If the model's continuous prediction error exceeds the threshold, a model rollback mechanism should be triggered to prevent continuous performance degradation due to data drift. Furthermore, balancing settings based on local computing power are necessary. The update frequency can be dynamically adjusted according to the production rhythm, using longer update intervals during stable production phases and increasing the update frequency during trial runs and grade changeovers. It is also crucial to ensure that the CPU utilization of edge devices is kept within a reasonable range to avoid increasing the burden on the DCS system.

[0052] Considering the limitations of catalyst manufacturing equipment, incremental learning algorithm optimization necessitates the use of a sparse update mechanism. Taking the BO-SVR model as an example, only support vectors that play a crucial role in prediction are retained. This allows for conditional checks on whether new samples support the vector set, maintaining the model's small size. An asynchronous update mechanism decouples the model training and prediction processes. The prediction module sets its response level based on production intensity, while the training module asynchronously processes micro-batch data in the background. Queuing mechanisms and multithreading ensure no data loss or duplication. Finally, a streamlined incremental learning framework is deployed on edge devices to reduce code size and adapt to mainstream PLCs and other hardware.

[0053] Example 2 In one possible implementation, step 101 includes the following.

[0054] Step S11: Collect parameter data of the raw material group, process group and equipment group in the catalyst production process.

[0055] In catalyst production, the first step is to systematically collect various parameters related to raw materials, processes, and equipment. These parameters cover the properties of the raw materials, each step in the process flow, and the operating status of the equipment. Data is collected periodically or in real-time using sensors and data acquisition modules, and transmitted to the central control system via a data interface. This system can process and store large amounts of real-time data, ensuring data integrity and accuracy, and supports high-frequency data acquisition.

[0056] Accurate and comprehensive data acquisition enables full monitoring of the catalyst production process, providing a foundation for subsequent data analysis, anomaly detection, and process optimization. Timely acquisition of parameter data helps identify potential quality issues, ensuring improved product consistency and reduced quality fluctuations.

[0057] Step S12: Obtain real-time environmental data during the catalyst production process.

[0058] Environmental data includes external conditions affecting the catalyst production process, such as temperature, humidity, and air pressure. This data is collected in real-time by environmental sensors installed in the production environment and transmitted along with other production data to a central data platform. The real-time nature of this environmental data ensures timely understanding of the potential impact of environmental changes on the production process, further improving the accuracy and reliability of production data.

[0059] Real-time acquisition of environmental data helps production systems adjust production parameters promptly to cope with changes in the external environment and reduce quality problems caused by environmental fluctuations. By integrating environmental data, production systems can more accurately predict product quality, thereby optimizing production efficiency and product consistency.

[0060] Step S13: Collect raw process data in real time according to the acquisition frequency of the data acquisition module.

[0061] In one possible implementation, step 102 includes the following.

[0062] Step S21: Apply the 3σ anomaly detection rule to detect anomalies in the original process data and remove outliers.

[0063] The 3σ outlier rule is a statistical method used to identify outliers in a dataset. Based on the mean and standard deviation of the data, any data point exceeding ±3σ is considered an outlier. The system automatically labels and excludes these outliers to ensure the quality and accuracy of the dataset.

[0064] By applying the 3σ outlier rule, outlier data can be effectively removed, ensuring the accuracy of subsequent analysis and modeling. This helps avoid interference from outlier data in model training and process analysis, improving the reliability of prediction results.

[0065] Step S22: Use interpolation to fill in the missing values ​​in the original process data.

[0066] Interpolation is used to fill in missing values ​​in data. It uses linear or polynomial interpolation methods to estimate the value of missing data based on trends in existing data. The interpolation algorithm needs to select an appropriate interpolation model based on the characteristics of the data to reduce errors introduced during the filling process.

[0067] By filling in missing values, the integrity of the data is restored, avoiding analytical errors or model training interruptions caused by missing data. Precise interpolation ensures the smooth progress of subsequent data analysis and modeling, improving the accuracy of process optimization and quality prediction.

[0068] Step S23: Based on the Pearson correlation coefficient and the Spearman correlation coefficient, the features in the original process data are screened to select features with a correlation greater than 0.3 with the quality indicators.

[0069] Pearson and Spearman correlation coefficients are used to measure the linear or nonlinear correlation between different process data features and quality indicators. By calculating the correlation coefficient between each feature and the target quality indicator, features highly correlated with quality are selected. This process helps eliminate redundant features and improves the efficiency and accuracy of the model.

[0070] By selecting features that are highly correlated with quality metrics, the efficiency and prediction accuracy of subsequent modeling can be significantly improved. Removing irrelevant features shortens model training time, avoids the impact of multicollinearity on model performance, and improves prediction accuracy and the model's generalization ability.

[0071] Step S24: Apply principal component analysis (PCA) to the original process data to reduce dimensionality and retain the principal components with higher contribution rates.

[0072] Principal component analysis (PCA) is a statistical method commonly used for dimensionality reduction. It reduces the dimensionality of data by extracting the direction of maximum variance. In catalyst production, PCA can help process high-dimensional data by reducing the number of features and thus simplifying the model's complexity by preserving principal components.

[0073] By reducing dimensionality, PCA can reduce data redundancy and improve the training efficiency of subsequent models. By preserving information about important features, PCA not only reduces the dimensionality of the data but also retains most of the data's variability, thereby enhancing the model's predictive ability.

[0074] In one possible implementation, step 103 includes the following.

[0075] Step S31: Based on the characteristics of the catalyst process data, automatically select a suitable learning algorithm and construct a candidate model set.

[0076] Based on data characteristics and process requirements, the system automatically selects the most suitable machine learning algorithm (such as support vector regression, random forest, etc.) and builds multiple candidate models. This process typically relies on automated tools for algorithm selection and model generation (such as AutoML) to reduce human intervention and ensure the diversity and adaptability of the models.

[0077] Automatic selection of learning algorithms can significantly improve modeling efficiency and avoid the tedious process of manually adjusting algorithms. By constructing multiple candidate models, it is possible to ensure that the optimal model is selected under different production scenarios, thereby improving the accuracy and stability of quality prediction.

[0078] Step S32: Use cross-validation to evaluate the performance of candidate models and obtain performance evaluation metrics for each candidate model.

[0079] Cross-validation is a technique used to evaluate the performance of machine learning models. It involves dividing the data into multiple subsets, validating the model sequentially with each subset, and using the remaining subsets to train the model. A common cross-validation method is K-fold cross-validation. This process avoids overfitting and provides an assessment of the model's generalization ability across different datasets.

[0080] Cross-validation effectively evaluates the performance of candidate models on real-world production data, ensuring that the selected model has high stability and generalization ability. This helps ensure that the model operates effectively under various conditions and maintains good prediction accuracy in different scenarios.

[0081] Step S33: Select the baseline model with the best performance based on the performance evaluation indicators.

[0082] By calculating model evaluation metrics (such as RMSE, MAE, R², etc.) after cross-validation, the baseline model with the best performance is selected. This model will serve as a reference benchmark in subsequent optimizations, and the performance of other models will be compared with it.

[0083] By selecting the best-performing baseline model, we can ensure that subsequent model optimization work starts from the optimal foundation, thus improving optimization efficiency. The selected model is usually the most stable in performance and has the smallest error, providing reliable support for real-time quality prediction in the production process.

[0084] In one possible implementation, step 104 includes the following.

[0085] Step S41: Define the hyperparameter search space for the optimal baseline model.

[0086] In hyperparameter optimization, the first step is to define the search space for hyperparameters. This includes important parameters such as the model's learning rate, regularization coefficient, and kernel function type. By specifying a reasonable search range, we can ensure that the optimal combination of hyperparameters is found within a suitable scope.

[0087] Setting a reasonable hyperparameter search space can avoid wasting computational resources due to an excessively large search range, while ensuring the efficiency and effectiveness of the hyperparameter optimization process. This step is crucial for improving model performance and provides a solid foundation for further optimization.

[0088] Step S42: Perform hyperparameter optimization based on Gaussian process regression in the generative intelligent process system.

[0089] Gaussian Process Regression (GPR) is a machine learning method based on Bayesian inference that can perform efficient optimization in continuous space. This method constructs a surrogate model to predict unknown data using known data and optimizes hyperparameters using the Expected Improvement (EI) criterion.

[0090] Optimizing hyperparameters through Gaussian process regression can find the optimal combination of hyperparameters with fewer experiments, avoiding redundant computations in traditional methods. This method is particularly effective in high-dimensional nonlinear problems, significantly improving the model's prediction accuracy and reducing computational costs.

[0091] Step S43: Select the optimal combination of hyperparameters by maximizing the expected improvement criterion.

[0092] The Expected Improvement (EI) criterion is a collection function used in Bayesian optimization to measure the potential of the current model in the parameter space. Maximizing the EI finds the hyperparameter combinations that are likely to bring the most significant performance improvement. By selecting the hyperparameter configuration that maximizes the expected improvement in each iteration, the system can efficiently search for optimal parameters.

[0093] Maximizing the expected value improvement criterion helps to quickly find the optimal combination of hyperparameters and improve the model's prediction accuracy in production environments. This process not only optimizes model performance but also effectively avoids computational redundancy during the search process, improving overall efficiency.

[0094] In one possible implementation, step 105 includes the following.

[0095] Step S51: Receive new production data in real time during the catalyst production process and process it in batches.

[0096] New production data is integrated into the system in real time. By monitoring various production parameters in real time, the system ensures that predictions and optimizations can be made based on the latest production data. The data is processed according to pre-defined batches to accommodate subsequent online updates and optimization needs.

[0097] Processing new data in batches effectively reduces the real-time computational burden on the system, while ensuring that the data processed in each batch is updated in a timely manner, and the optimization process is not limited by excessive data volume. By receiving new data in real time, it is possible to ensure that the quality control of each batch during the production process keeps up with the latest changes in process parameters.

[0098] Step S52: Using the sliding window technique, the new production data is divided into batches, and the parameters of the optimal baseline model are updated online based on the data.

[0099] The sliding window technique is used to dynamically adjust the size of a data window as new data arrives, thereby enabling real-time updates of historical data. The latest data within the window is used to update the parameters of the optimal baseline model to adapt to changes in data during production.

[0100] The sliding window technique ensures that model parameters are updated in a timely manner as new data flows in during production, improving the model's adaptability to production fluctuations. The online-updated model reflects the current production status, avoiding a decrease in prediction accuracy due to data drift or equipment aging.

[0101] Step S53: Validate the optimal baseline model after online updates to ensure that it can adapt to new data in the production process.

[0102] After updating the model, a validation step is required to confirm its performance on the new data. Validation methods typically include re-performing cross-validation or testing directly on real production data. By comparing the model's prediction accuracy before and after the update, it is ensured that the model is well-suited to its new data.

[0103] The validation process ensures that the updated model can accurately predict quality indicators in the new data, avoiding performance degradation due to improper updates. It also ensures the model can adapt to constantly changing process parameters during production, guaranteeing the stability and accuracy of quality predictions.

[0104] In one possible implementation, step 106 includes the following.

[0105] Step S61: Apply the online updated optimal baseline model to predict quality indicators in the catalyst production process in real time.

[0106] By deploying the optimal baseline model, updated online, to the production site, the system predicts quality indicators in the production process in real time. The system inputs real-time process data into the model, outputs corresponding quality prediction results, and issues timely alerts when quality deviates from standards.

[0107] Real-time forecasting can significantly improve quality control capabilities in the production process. By dynamically adjusting process parameters, the system can accurately predict quality during production, avoiding product quality fluctuations and the generation of defective products, thereby improving overall production efficiency.

[0108] Step S62: Based on the quality prediction results, adjust the key process parameters in the catalyst production process in real time.

[0109] Based on the quality predictions provided by the model, the system automatically or manually adjusts key process parameters, such as temperature, pressure, and reaction time. This adjustment process is dynamically optimized based on the prediction results to ensure that product quality always meets predetermined standards.

[0110] Real-time adjustment of process parameters ensures that quality remains within predetermined ranges throughout production, preventing product quality issues caused by process fluctuations. Precise control of each process parameter significantly improves production efficiency and product consistency.

[0111] Step S63: Feedback correction of the model based on actual quality data during the production process to achieve closed-loop optimization in the production process.

[0112] During actual production, the system adjusts the model's predictions based on a comparison of actual production data and quality indicators. This feedback mechanism ensures that the model can self-correct according to actual conditions, thereby adapting to long-term production changes.

[0113] The feedback correction mechanism enables closed-loop optimization, ensuring that the model continuously self-optimizes during production, thus improving long-term prediction accuracy and adaptability. Through continuous feedback, the system can cope with the impact of factors such as equipment aging and sensor changes on quality prediction, maintaining the model's stability and accuracy.

[0114] Example 3 After completing the above description of the embodiments, the applicant also conducted experimental verification and result analysis.

[0115] Taking the mother liquor compounding process of a certain catalyst as an example, there are a total of 8 potential influencing factors and 2 target values ​​identified in production. The traditional compounding process relies entirely on post-production testing to determine the target values, requiring laboratory analysts to follow up and analyze the content before operation. Even so, there are still significant deviations, requiring engineers to make further manual adjustments based on production experience. The entire production process exhibits serious lag and experience-driven problems.

[0116] To this end, 190 sets of historical data produced according to the same formula were collected for preliminary analysis. These variables were found to have a normal distribution characteristic; therefore, the Z-score method was used for standardization during data cleaning. Figure 7 The table shows a histogram of influencing factors and objectives. Histogram analysis reveals that, except for factors 4 and 7, the distribution largely conforms to a normal distribution; therefore, the correlation using the Pearson correlation coefficient calculator is relatively reliable.

[0117] like Figure 8 The diagram shown is a Pearson correlation coefficient matrix. Based on this, the principal components between the eight potential influencing factors and the two objectives are further calculated, as follows: Figure 9 The principal component matrix diagram shown Figure 10 The diagram showing the principal component contribution is shown below. Figure 11 The diagram shown illustrates the principal component analysis for objective 1. Figure 12 The principal component analysis diagram for objective 2 shows a certain consistency with the correlation coefficient calculation results. Although factors 4 and 7 have a small contribution to the objective, their contribution can still reach above 0.3 under different calculation methods. Considering the model's representational ability, dimensionality reduction is not performed, and all eight potential influencing factors are included as independent variables in the modeling process.

[0118] Furthermore, the following explanation is provided regarding the model performance evaluation metrics.

[0119] After determining the feature engineering, RSM and BO-SVR were compared, and an anomaly was found in the RSM model constructed using this data structure, such as... Figure 13 The BO-SVR optimization process and model performance diagram for objective 1 are shown below. Figure 14 The diagram shows the iterative process of the hyperparameters of the model for objective 1. Figure 15 The BO-SVR optimization process and model performance diagram for objective 2 are shown below. Figure 16 The diagram shows the hyperparameter iteration process of the model for objective 2. We first fitted and analyzed the BO-SVR model, and it can be seen that the fitting effect of both objective values ​​is relatively good.

[0120] In the past, major problems encountered in production processes included insufficient generalization ability of RSM for unknown data, low efficiency and poor representativeness due to the need for extensive interactive testing when handling high-dimensional models, and significant differences in prediction accuracy across different processes. Production is a complex and comprehensive process; defects at each stage affect the final product's pass rate. The first-pass yield can be represented by the product of the pass rates of each process, and the process capability index Cpk is calculated based on the first-pass yield. For high-cost, high-value catalyst production enterprises, a Cpk of at least ≥1.37 is required for significant benefits. Therefore, AutoML technology is used to automatically match the most suitable machine learning model for each process to address the insufficient generalization ability of a single model. To address parameter fluctuations during production, noise and outliers are proactively added during model training for adversarial training to enhance model robustness and prevent concept drift. For different production scenarios, universal data collection and baseline model construction are performed for common processes in the early stages. During product design, only the baseline models for each process need to be combined and rearranged to carry out process design and optimization. Taking the mother liquor compounding process as an example, considering the high requirement for a one-time molding pass rate (≥99.5%) in the production process, in order to find the most suitable model for this process, in addition to BO-SVR, 12 other algorithms were introduced for parameter grid screening. The optimal process obtained was Lasso regression, such as... Figure 17 The figure shows the model performance of Lasso regression for the two target values. Comparing the parametric grid and the BO-SVR algorithm, it is found that the BO-SVR algorithm has higher model performance. Although both the parametric grid and BO-SVR achieve the prediction accuracy required for production, considering the impact of the pass rate of each other process step on the total pass rate, the more accurate BO-SVR model should be used. See Table 2 for details.

[0121] Table 2 In terms of efficiency, processing 60 sets of data using BO-SVR took 8.2 seconds, while parametric grid search took 13.7 seconds. The difference is negligible and both are sufficient to meet the high-intensity, fast-paced demands of actual production. To verify the model's generalization ability, 16 sets of data were extracted from 130 batches using the LHS method as a test set, and the BO-SVR model was used for prediction. See Table 3 for the test set target and true value pairs.

[0122] Table 3 To address the dynamic characteristics of industrial catalyst production, an incremental learning framework based on micro-batch updates was designed. Core components include data caching and a sliding window mechanism. A circular buffer maintains the latest 1000 production data entries (covering approximately 5 complete production batches). When new batch data arrives, it automatically replaces the oldest batch data. The sliding window width parameter λ (set to 200 entries / window) controls the balance between model memory and adaptability. In platinum group catalyst production, this setting reduces the model's response delay to feedstock fluctuations from 2 hours in traditional batch processing to 15 minutes. The incremental update algorithm, based on SVR, employs a variant of the SMO (Sequential Minimal Optimization) algorithm. When a new sample x is added: the kernel function value of x and existing support vectors is calculated; KKT conditions are used to determine whether x becomes a new support vector; if x is accepted, the weight vector α and bias b are updated. .

[0123] Where η is the learning rate (set to 0.01), dynamically adjusted through an exponential decay mechanism (decreasing by a factor of 0.95 every 1000 updates). Concept drift detection and response: The mean and standard deviation of the model residuals are monitored through statistical process control (SPC), triggering an alarm when the residuals exceed ±3σ for three consecutive windows. When the content of factor 1 suddenly increases, the incremental learning model reduces the RMSE from 0.042 to 0.015 within three update cycles. The single-sample update time is stable between 18-25ms, far lower than the sampling cycle (5 seconds) of industrial control systems. Memory usage is maintained in the range of 80-120MB, only 1 / 10 of Batch Training, meeting the resource constraints of edge computing devices. Through the asynchronous update mechanism, the peak CPU utilization is controlled below 45%, ensuring the stability of the control system. In 7 days of continuous testing, the incremental learning model has an average RMSE of 0.012 and a standard deviation of 0.003, demonstrating good stability.

[0124] Experimental results show that Bayesian optimization (BO) has a significant advantage in improving the performance of automated machine learning models. Taking catalyst mother liquor compounding process data as an example, in the modeling scenario of 8-dimensional features and 2-dimensional target values, the BO-SVR model exhibits higher prediction accuracy than the Lasso regression model obtained by parametric grid search: the RMSE of target 1 decreased from 0.5307 to 0.0102, and R² increased from 0.9992 to 0.9999; the RMSE of target 2 decreased from 0.3467 to 0.0124, and R² increased from 0.9997 to 0.9999 (see Table 2). This result verifies the global optimization capability of Bayesian optimization in continuous hyperparameter space. By modeling Gaussian processes and improving the acquisition function by expectation, it avoids the defect of grid search getting trapped in local optima, and is particularly suitable for the accurate representation of high-dimensional nonlinear relationships. In terms of optimization efficiency, BO-SVR takes 8.2 seconds to process 60 sets of data, which is not much different from the 13.7 seconds of parametric grid search, both of which can meet the real-time requirements of industry. Further analysis of the hyperparameter iteration process reveals that BO converges to the optimal solution within 15 iterations, while traditional grid search requires traversing over 200 parameter combinations, resulting in an efficiency improvement of approximately 90%. This characteristic is crucial for catalyst multi-variety, small-batch production scenarios. Specifically, when formula changes lead to variations in data distribution, BOSVR can quickly reconstruct model parameters, saving 85% of the time cost compared to manual parameter tuning.

[0125] Therefore, to address the quality design challenges in the highly dynamic production environment of catalyst manufacturing, this application proposes a closed-loop quality design method integrating Automated Machine Learning (AutoML), Bayesian optimization, and incremental learning. This method systematically solves the modeling bottleneck problem of traditional Response Surface Method (RSM) and constructs a full-process technical framework of "data preprocessing - baseline model selection - hyperparameter optimization - online iteration." AutoML selects Support Vector Regression (SVR) as the baseline model from 12 algorithms. Bayesian optimization solves the problem of continuous space optimization of SVR hyperparameters. Incremental learning enables real-time model updates, forming a quality closed loop of "design-manufacturing-optimization." This overcomes the limitations of traditional RSM in terms of low prediction accuracy and poor modeling efficiency in high-dimensional nonlinear scenarios. Performance breakthrough. Industrial experiments have demonstrated that this method significantly improves prediction accuracy in catalyst mother liquor blending processes. Specifically, the RMSE for objective 1 decreased from 0.5307 in the parameter grid search to 0.0102, with an R² of 0.9999. When faced with ±3% fluctuations in raw material parameters, the incremental learning model completes parameter adjustments within 15 minutes, shortening the iteration cycle by over 95% compared to offline models. Furthermore, the model's memory footprint was reduced from 1.2GB to 80MB, meeting the real-time requirements of edge computing devices. Engineering Value: This method has been successfully applied to the industrial production of catalysts, improving batch quality consistency by 57%, increasing the pass rate from 89% to 98.3%, and saving over 800,000 yuan in quality costs annually. Its technical system adapts to the dynamic control requirements of Industry 4.0's "production-while-optimization" approach, providing core technical support for the localization of high-end catalysts under the pressure of deglobalization competition.

[0126] This application also provides a computer-readable medium storing at least one instruction, which is loaded and executed by the processor to implement a catalyst process design and optimization method as described in the above embodiments.

[0127] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for catalyst process design and optimization, characterized in that, The method includes: Obtain raw process data during catalyst production; The original process data is cleaned and feature-engineered to generate a preprocessed dataset that can be used for modeling. Using automated machine learning techniques, multiple candidate machine learning models are selected based on the preprocessed dataset; Based on the generative intelligent process system, the hyperparameters of the selected baseline model are optimized; Incremental learning techniques are applied to update the optimized baseline model online. The updated baseline model is then applied to quality prediction and process parameter adjustment in catalyst production.

2. The method according to claim 1, characterized in that, The acquisition of raw process data during catalyst production includes: Collect parameter data of the raw material group, process group and equipment group in the catalyst production process; Acquire real-time environmental data during the catalyst production process; The raw process data is collected in real time according to the acquisition frequency of the data acquisition module.

3. The method according to claim 1, characterized in that, The process of cleaning and feature engineering the raw process data to generate a preprocessed dataset that can be used for modeling includes: The original process data is anomaly detected and outliers are removed by applying the 3σ outlier detection rule. The missing values ​​in the original process data were filled using interpolation. Features in the original process data were filtered based on Pearson correlation coefficient and Spearman correlation coefficient, and features with a correlation greater than 0.3 with the quality indicators were selected. Principal component analysis (PCA) was applied to the original process data to reduce dimensionality, retaining the principal components with higher contribution rates.

4. The method according to claim 1, characterized in that, The method of using automated machine learning technology to select multiple candidate machine learning models based on the preprocessed dataset includes: Based on the characteristics of the catalyst process data, an appropriate learning algorithm is automatically selected and a candidate model set is constructed. The candidate models are evaluated using cross-validation to obtain performance evaluation metrics for each candidate model. Based on the performance evaluation metrics, the baseline model with the best performance is selected.

5. The method according to claim 1, characterized in that, The generative intelligent process system optimizes the hyperparameters of the selected baseline model, including: Define the hyperparameter search space for the optimal baseline model; Hyperparameter optimization based on Gaussian process regression in generative intelligent process systems; The optimal combination of hyperparameters is selected by maximizing the expected value improvement criterion.

6. The method according to claim 1, characterized in that, The application of incremental learning technology to update the optimized baseline model online includes: The system receives new production data from the catalyst production process in real time and processes it in batches. The new production data is divided into batches using a sliding window technique, and the parameters of the optimal baseline model are updated online based on the data. The optimal baseline model, updated online, is validated to ensure it can adapt to new data in the production process.

7. The method according to claim 1, characterized in that, The step of applying the online updated baseline model to quality prediction and process parameter adjustment in catalyst production includes: The online-updated optimal baseline model is applied to predict quality indicators in the catalyst production process in real time. Based on the quality prediction results, key process parameters in the catalyst production process are adjusted in real time. The model is corrected based on actual quality data during the production process to achieve closed-loop optimization in the production process.

8. A catalyst process design and optimization device, characterized in that, The device includes: The data acquisition module is used to acquire raw process data during catalyst production. The preprocessing module is used to perform data cleaning and feature engineering on the raw process data to generate a preprocessed dataset that can be used for modeling. The model selection module is used to select multiple candidate machine learning models based on the preprocessed dataset using automated machine learning techniques. The parameter optimization module is used to optimize the hyperparameters of the selected baseline model based on the generative intelligent process system. The model update module is used to apply incremental learning techniques to update the optimized baseline model online. The parameter adjustment module is used to apply the online updated baseline model to quality prediction and process parameter adjustment in the catalyst production process.

9. A parameter optimization device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being executed by the processor to implement the catalyst process design and optimization method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is executed by a processor to implement the catalyst process design and optimization method as described in any one of claims 1 to 7.

Images