Active learning and uncertainty evaluation-based zinc salt modified ruthenium catalyst performance optimization system and method

Abstract

The invention relates to a zinc salt modified ruthenium catalyst performance optimization system and a zinc salt modified ruthenium catalyst performance optimization method based on active learning and uncertainty evaluation, and aims to overcome the defects of low efficiency, experience dependence and lack of unknown scheme prediction capability when a traditional trial and error method is used for screening a zinc salt modifier. A candidate zinc salt scheme is digitally represented, a standardized feature library is constructed, and a Gaussian process model capable of quantitatively predicting uncertainty is established in combination with historical data; the model and an active learning strategy are utilized to intelligently evaluate the potential and uncertainty of unexperimented schemes, and the most valuable experiment scheme is automatically recommended; and finally, through a closed-loop process of experimental verification, data feedback and model iteration updating, the purpose of efficiently finding out the combination of the optimal zinc salt type and addition amount with the least experiment times is achieved. According to the method, traditional blind screening is converted into data-driven intelligent optimization, and the efficiency and systematicness of catalyst performance optimization are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial catalysis and artificial intelligence, specifically to a system and method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment. Background Technology

[0002] The selective hydrogenation of benzene to cyclohexene is a key reaction process in the chemical industry, especially in the nylon industrial chain. Ruthenium catalysts are widely used in this reaction due to their high activity, but their inherent reaction characteristics often lead to unsatisfactory selectivity for the target product, cyclohexene. To improve selectivity, industrial practice and research commonly introduce zinc salts as modifiers into the reaction system, utilizing the chemical properties of different zinc salts to adjust the reaction pathway. However, the selection of zinc salts itself constitutes a complex optimization problem: on the one hand, there are numerous candidate zinc salts, covering both inorganic and organic acid salts, with significant differences in their chemical properties; on the other hand, the amount of zinc salt added (i.e., the Zn / Ru molar ratio) is also a key variable, and it exhibits complex interaction effects with the type of zinc salt. The core challenge in improving catalyst performance is how to quickly and economically find the optimal combination from the high-dimensional, discrete-continuous mixed parameter space constituted by "zinc salt type" and "addition amount".

[0003] Currently, solutions to this optimization problem primarily rely on traditional trial-and-error methods or grid search based on limited experience. Researchers must pre-select a few zinc salts and limited dosages based on chemical intuition, conducting numerous experiments. This process is not only time-consuming and costly but also heavily dependent on the personal experience of experts. More importantly, existing methods are essentially a "passive" selection process, lacking the ability to predict the performance of untried solutions and failing to intelligently assess the exploration value of different regions (such as high-potential or high-uncertainty regions) during experiments, thus failing to proactively plan the most efficient experimental sequence. This results in a considerable degree of blindness in the research and development process, with a large amount of experimental resources potentially consumed in parameter regions with mediocre performance, while truly outstanding "dark horse" solutions may never be considered. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a system and method for optimizing the performance of zinc salt modified ruthenium catalysts based on active learning and uncertainty assessment, which can actively learn, make intelligent decisions, and efficiently traverse the chemical space in a data-driven manner to systematically discover the optimal zinc salt modification scheme.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment, comprising the following steps:

[0007] S1: Perform digital characterization on the candidate zinc salt modification schemes, convert the zinc salt types and Zn / Ru molar ratios into digital vectors of uniform dimension, and generate a standardized feature library containing standard feature vectors of all candidate schemes;

[0008] S2: Match and align the historical data list obtained from historical experimental records with the standardized feature library, selectively pair the historical schemes with their standard feature vectors and measured cyclohexene, and generate the initial training set and the current best performance benchmark.

[0009] S3: Perform Gaussian process regression training on the initial training set, fit the mapping relationship between the standard feature vector and cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output.

[0010] S4: Input the unexperimented candidate solutions in the standardized feature library and the current best performance benchmark into the uncertainty prediction model for active learning and scoring, calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them to generate an experimental recommendation list containing specific solutions.

[0011] S5: Perform experimental verification and model iteration based on the experimental recommendation list. Execute the experiments in the list, add the new measured data to the training set, re-train the uncertainty prediction model, update the historical experimental records, and update the uncertainty prediction model and the current best performance benchmark.

[0012] In one embodiment, S1 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0013] S11: Perform feature encoding on the zinc salt types and Zn / Ru molar ratio in the candidate zinc salt modification schemes, convert the SMILES string of organic acid zinc salts into Morgan fingerprint vectors, and convert the chemical formula of inorganic zinc salts into anion descriptor vectors containing electronegativity, ionic radius and pKa value, and generate basic feature data for each scheme.

[0014] S12: The basic feature data and the corresponding Zn / Ru molar ratio are concatenated and standardized. The Morgan fingerprint vector and anion descriptor vector are connected with the molar ratio values ​​of the basic feature data by dimension and scale normalization is performed to generate a standard feature vector.

[0015] S13: Store and index the standard feature vectors of all candidate schemes, establish the mapping relationship between scheme identifiers and feature vectors, and generate a standardized feature library for unified querying.

[0016] In one embodiment, S2 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0017] S21: Analyze and extract information from historical experimental records to extract the types of zinc salts, Zn / Ru molar ratios, and corresponding measured values ​​of cyclohexene selectivity of completed experiments, and generate a structured historical data list.

[0018] S22: For each scheme in the historical data list, perform feature query and pairing processing in the standardized feature library based on zinc salt type and Zn / Ru molar ratio, match the corresponding standard feature vector and associate it with the measured selectivity value to generate a feature-performance paired dataset.

[0019] S23: Integrate and statistically process the feature-performance paired dataset, summarize all data pairs to form the initial training set, and select the maximum value from all measured selectivity to generate the current best performance benchmark.

[0020] In one embodiment, S3 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0021] S31: Perform input-output separation processing on the initial training set, extract all standard feature vectors as the training input matrix, extract the corresponding measured values ​​of cyclohexene selectivity as the training target vector, and generate model training data.

[0022] S32: Based on the Matrn kernel function, Gaussian process modeling is performed on the model training data, the covariance function form and initial hyperparameters are configured, and the length scale and noise variance parameters are optimized by maximizing the marginal likelihood function to generate the trained regression model;

[0023] S33: Perform interface encapsulation processing on the trained regression model so that it can output the predicted mean and predicted variance of cyclohexene selectivity from the input feature vector, generating an uncertainty prediction model.

[0024] In one embodiment, S4 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0025] S41: Perform set difference analysis on the standardized feature library and the initial training set, screen out all candidate solutions that did not appear in the initial training set and the standard feature vectors corresponding to the candidate solutions, and generate a set of solutions to be evaluated;

[0026] S42: Perform Gaussian process posterior inference on each scheme in the set of schemes to be evaluated and the input uncertainty prediction model, calculate the predicted distribution parameters of cyclohexene selectivity based on the covariance relationship obtained from training, and generate the prediction results of candidate schemes containing the predicted mean and the predicted variance.

[0027] S43: Perform expectation enhancement and uncertainty fusion scoring on the prediction results of candidate schemes and the current best performance benchmark. Calculate the comprehensive score of each scheme based on the preset exploration factor and balance coefficient, sort them, and generate an experimental recommendation list.

[0028] In one embodiment, S5 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0029] S51: Perform catalytic experiments and data acquisition and processing for each scheme in the experimental recommendation list. Based on the zinc salt type and Zn / Ru molar ratio specified in the scheme, carry out benzene hydrogenation reaction, measure and record cyclohexene selectivity, and generate new measured data.

[0030] S52: Perform data integration and performance benchmark update processing on the newly added measured data and the initial training set. Selectively pair the standard feature vector corresponding to each new experimental scheme with the measured data and add it to the training set. At the same time, extract the highest value from the updated measured value set to generate an updated training set and an updated current best performance benchmark.

[0031] S53: The updated training set is used as new training data input into the Gaussian process regression framework to optimize the model parameters and generate an updated uncertainty prediction model.

[0032] Secondly, the present invention provides a performance optimization system for zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment, the system comprising the following modules:

[0033] The zinc salt scheme characterization module is used to digitally characterize candidate zinc salt modification schemes, converting zinc salt types and Zn / Ru molar ratios into digital vectors of uniform dimensions, and generating a standardized feature library containing standard feature vectors of all candidate schemes.

[0034] The data matching and alignment module is used to match and align the historical data list obtained from historical experimental records with the standardized feature library. It selectively pairs historical schemes with their standard feature vectors and measured cyclohexene to generate an initial training set and the current best performance benchmark.

[0035] The Gaussian regression modeling module is used to train the initial training set using Gaussian process regression, fit the mapping relationship between the standard feature vector and the cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output.

[0036] The active learning recommendation module is used to input unexperimented candidate solutions and the current best performance benchmark from the standardized feature library into the uncertainty prediction model for active learning scoring, calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them, and generate an experimental recommendation list containing specific solutions.

[0037] The experimental iteration optimization module is used to perform experimental verification and model iteration based on the experimental recommendation list. It executes the experiments in the list, adds the new measured data to the training set, re-trains the uncertainty prediction model, updates the historical experimental records, and updates the uncertainty prediction model and the current best performance benchmark.

[0038] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned methods for optimizing the performance of zinc salt modified ruthenium catalysts based on active learning and uncertainty assessment.

[0039] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned methods for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment.

[0040] In summary, this application provides a method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment. By constructing a standardized feature library, it transforms diverse zinc salt modification schemes into machine-processable feature vectors, establishing a digital representation foundation for standardized comparison and efficient retrieval among schemes. Subsequently, historical experimental data is matched with the feature library to generate an initial training set and performance benchmark, enabling the rapid establishment of a predictive model starting point from existing knowledge. Through Gaussian process regression training, this method establishes a nonlinear mapping relationship between features and performance and simultaneously quantifies predictive uncertainty, achieving reliable estimation and risk measurement of the selectivity of unknown zinc salt schemes. Furthermore, by combining an active learning strategy to intelligently score and rank unexperimented schemes, it proactively recommends the most valuable or potentially promising experimental schemes, replacing the traditional experience-based blind trial and error. Finally, through a closed-loop process of experimental verification, data feedback, and model iteration, the model continuously evolves with the experimental progress, constantly approaching the optimal solution. This method transforms the traditional, inefficient, and passive catalyst screening model into a highly efficient, targeted, and data-driven intelligent optimization process. It significantly reduces the over-reliance on expert experience and a large number of ineffective experiments, and can systematically discover high-performance zinc salt modification schemes within limited experimental resources, thereby accelerating the catalyst development process and reducing development costs.

[0041] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0042] Figure 1 A flowchart illustrating a method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment, provided for embodiments of this application;

[0043] Figure 2 This is a schematic diagram of a zinc salt modified ruthenium catalyst performance optimization system based on active learning and uncertainty assessment, provided as another embodiment of this application. Detailed Implementation

[0044] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0046] In one embodiment, such as Figure 1 As shown, a method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0047] S1: Perform digital characterization on candidate zinc salt modification schemes, convert the zinc salt types and Zn / Ru molar ratios into digital vectors of uniform dimension, and generate a standardized feature library containing standard feature vectors of all candidate schemes.

[0048] Specifically, the system screens zinc salts covering different chemical structures and classifies them into organic and inorganic zinc salts based on differences in chemical composition and bonding methods. This ensures that the screened zinc salts can cover potential modifiers with different anion types and molecular structural characteristics, providing a foundation for subsequent comprehensive chemical space exploration. For organic zinc salts, the system obtains corresponding standard structural information from authoritative chemical databases and uses cheminformatics tools to convert the structural information into molecular fingerprints. This conversion process can encode the atomic connection patterns and functional group distribution characteristics within a specific range of the molecule. At the same time, the system calculates key molecular descriptors for organic zinc salts. These descriptors can reflect the basic physicochemical properties of the molecule. Subsequently, the system concatenates the molecular fingerprints and molecular descriptors to form the basic characteristic expression of organic zinc salts.

[0049] For inorganic zinc salts, the system uses their chemical formula as the core identifier, extracts the anionic components, and then calculates multiple physicochemical descriptors for the anions. These descriptors cover key parameters reflecting the electronic properties, spatial properties, and reactivity of the anions. All descriptors are standardized to eliminate dimensional differences, forming the basic characteristic expression of inorganic zinc salts. The system normalizes the Zn / Ru molar ratio as a continuous variable and integrates it with the basic characteristic expressions of both organic and inorganic zinc salts. By using a unified dimension filling method, the feature vectors of the two types of zinc salts reach the same dimension, forming standardized feature vectors. The system constructs a structured standardized feature library. Each record contains core fields such as a unique scheme identifier, basic information related to the zinc salt, the original and normalized Zn / Ru molar ratio, the standard feature vector, and experimental status identifiers. Simultaneously, the system performs data validation on the feature library, removing abnormal data with inconsistent dimensions or exceeding reasonable ranges to ensure the standardization and validity of all data in the feature library.

[0050] S2: Match and align the historical data list obtained from historical experimental records with the standardized feature library, selectively pair the historical schemes with their standard feature vectors and measured cyclohexene, and generate the initial training set and the current best performance benchmark.

[0051] Specifically, the system acquires multi-source historical experimental data, including experimental records accumulated in the laboratory and relevant data from publicly available literature. All acquired historical data must include key information such as zinc salt name, structural identification information, Zn / Ru molar ratio, reaction conditions, measured values ​​of cyclohexene selectivity, and relevant records of parallel experiments. The system preprocesses the historical data, first filtering out data whose reaction conditions are consistent with preset standard reaction conditions, and removing data that deviates from the preset range. Simultaneously, data whose parallel experimental results show dispersion exceeding an acceptable range are also removed to ensure the reliability and comparability of subsequent modeling data. The system uses the zinc salt name and Zn / Ru molar ratio as joint search criteria to find the corresponding standard feature vector in a standardized feature library. For zinc salts with inconsistent names but identical structures, the system achieves precise matching by comparing their structural identification information, ensuring a unique mapping relationship between historical experimental schemes and corresponding schemes in the standardized feature library. The system pairs successfully matched standard feature vectors with corresponding measured values ​​of cyclohexene selectivity, forming a structured set of data pairs. The system then divides this set into an initial training set and a validation set according to a specific ratio. During this division, the system ensures that the distribution ratio of organic and inorganic zinc salts in both sets remains consistent with the original data pair set, avoiding the impact of imbalanced data distribution on model training. The initial training set is stored as a feature matrix and label vectors for easy model retrieval. Simultaneously, the system iterates through all measured selectivity values ​​in the initial training set, extracts the maximum value as the current optimal performance benchmark, and records the zinc salt name, Zn / Ru molar ratio, and other scheme information corresponding to this benchmark value.

[0052] S3: Perform Gaussian process regression training on the initial training set, fit the mapping relationship between the standard feature vector and cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output.

[0053] Specifically, the system employs a Gaussian process regression model as the core model for performance prediction. Based on a Bayesian inference framework, this model can effectively fit nonlinear mapping relationships in scenarios with small sample data, and also has the ability to output quantitative information on the uncertainty of prediction results, aligning with the practical need for gradual data accumulation during the zinc salt modification optimization process. The system selects the Matrn kernel function as the model's kernel function. This kernel function can capture the smooth nonlinear relationship between feature vectors and cyclohexene selectivity through parameter adjustment, and its mathematical expression can reflect the impact of the similarity between different feature vectors on the prediction results. The system optimizes the model's hyperparameters through grid search. The hyperparameter search range is determined based on model characteristics and data features, with the optimization objective being to minimize the prediction error on the validation set. The hyperparameters determined in this way enable the model to achieve the best fitting effect under the current data conditions.

[0054] The system inputs the feature matrix and label vectors of the initial training set into the model for training. During training, a fixed number of iterations and a convergence threshold are set. Training stops when the model reaches the maximum number of iterations or the difference between the objective function values ​​of two adjacent iterations is less than the convergence threshold, completing the initial model fitting. The system uses a validation set to evaluate the performance of the trained model. Evaluation metrics include the coefficient of determination, root mean square error, and mean absolute error. These metrics comprehensively reflect the degree of agreement between the model's predictions and the measured values, verifying the model's predictive reliability. When making predictions, for any input standard feature vector, the model outputs the corresponding predicted mean and predicted variance. The predicted mean serves as the predicted value for cyclohexene selectivity, and the predicted variance serves as the uncertainty quantification result. The variance reflects the model's sufficient understanding of the performance of the input scheme. The system stores the trained model in a specific format, including metadata such as model structure, optimal hyperparameters, kernel function configuration, training set data, and validation set evaluation results. This ensures the model's traceability and reusability, providing a foundation for subsequent model iterations and calls.

[0055] S4: Input the unexperimented candidate solutions in the standardized feature library and the current best performance benchmark into the uncertainty prediction model for active learning and scoring. Calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them to generate an experimental recommendation list containing specific solutions.

[0056] Specifically, the system selects candidate solutions marked as "not yet tested" from a standardized feature library, extracts standard feature vectors from these solutions, constructs a feature matrix to be scored, and inputs the current best performance benchmark as key reference data into the uncertainty prediction model. The system calculates the predicted mean and variance for each candidate solution using the model, and further calculates the expected improvement value based on these results. The expected improvement value is calculated based on probability and statistics principles, quantifying the potential probability and magnitude of a candidate solution's performance exceeding the current best performance benchmark. The magnitude of this value directly reflects the performance improvement potential of the candidate solution. To balance performance improvement potential with the model's need to explore unknown areas, the system introduces an uncertainty comprehensive scoring mechanism. This mechanism sets fixed weight coefficients and weights the expected improvement value and the normalized predicted variance to obtain a comprehensive score for each candidate solution. The weight coefficients are determined based on the priority of the optimization objectives, ensuring that the scoring results highlight high-potential solutions while also considering the exploration of areas where the model's knowledge is limited.

[0057] Furthermore, the system sorts all unexperimented candidate schemes according to their comprehensive scores from high to low, forming a preliminary ranking list. To avoid the model getting trapped in local optima, the system introduces a cross-zinc salt diversity constraint mechanism. By analyzing the zinc salt chemical structure types corresponding to the candidate schemes, it ensures a balanced distribution of zinc salt schemes with different structural types in the ranking results, avoiding excessive concentration of schemes of a single type. The system then selects a specified number of candidate schemes from the ranked list. These schemes possess both high comprehensive scores and cover different types of zinc salts. Based on these selected schemes, the system generates a standardized experimental recommendation list. The list includes fields such as recommendation ranking, scheme unique identifier, zinc salt name, original Zn / Ru molar ratio, predicted mean, predicted variance, expected improvement, comprehensive score, and recommendation level. Each field provides key reference information for experimental execution. The recommendation level is determined based on the range of the comprehensive score, providing an intuitive basis for prioritizing experiments. Throughout the process, the system ensures the efficiency and diversity of recommended schemes through multi-dimensional scoring and screening.

[0058] S5: Perform experimental verification and model iteration based on the experimental recommendation list. Execute the experiments in the list, add the new measured data to the training set, re-train the uncertainty prediction model, update the historical experimental records, and update the uncertainty prediction model and the current best performance benchmark.

[0059] Specifically, the system establishes standardized experimental execution procedures and conditions. All recommended experiments are conducted according to this standard procedure to ensure consistency and repeatability. During the experiment, the system records key parameters of the reaction process, including reaction temperature, pressure, time, and stirring rate. Multiple parallel experiments are also set up to ensure the reliability of the experimental data. After the experiment, the system uses specified detection instruments and methods to perform component analysis on the reaction products. Quantitative calculations yield the measured cyclohexene selectivity for each recommended scheme. The system performs statistical analysis on the measured values ​​of parallel experiments to ensure that the dispersion of the measured values ​​is within an acceptable range. This statistically analyzed measured value is then used as the final performance data for that scheme. The system supplements the initial training set with the standard feature vector of the new experimental scheme and the final performance data according to the pre-defined data pairing rules, forming an updated training set. Simultaneously, the system updates the experimental status identifier of the corresponding scheme in the standardized feature library, updating "not yet experimented" to "experimented," and enters the final performance data.

[0060] The system synchronously updates the historical experimental record database, recording complete information such as the new experiment's scheme details, reaction conditions, measured data, and parallel experiment results, forming a coherent chain of experimental data traceability. The system employs an incremental training mode to iteratively update the uncertainty prediction model. Using the updated training set as input, it retains the original kernel function type and hyperparameter optimization method, preserving the hyperparameters from the previous round as initial values ​​for iterative optimization. Reasonable iteration counts and convergence thresholds are set to improve iteration efficiency while ensuring model prediction accuracy. After iteration, the system uses a validation set to evaluate the updated model's performance. By comparing evaluation metrics before and after iteration, the system verifies the improvement in model prediction accuracy and stability. The system compares the final performance data of the new experiment with the current best performance benchmark. If a scheme has final performance data higher than the current best performance benchmark, the system will update the current best performance benchmark with that higher value. The system determines whether the iteration termination condition is met. This condition is determined based on the improvement of the best performance benchmark and the range of model prediction deviation in two consecutive iterations. If the termination condition is met, the iteration stops and the optimal solution is output. If the termination condition is not met, the system returns to the active learning scoring stage and starts the next round of solution recommendation and experimental verification process until the termination condition is met, thus completing the entire optimization process.

[0061] In summary, this application provides a method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment. By constructing a standardized feature library, it transforms diverse zinc salt modification schemes into machine-processable feature vectors, establishing a digital representation foundation for standardized comparison and efficient retrieval among schemes. Subsequently, historical experimental data is matched with the feature library to generate an initial training set and performance benchmark, enabling the rapid establishment of a predictive model starting point from existing knowledge. Through Gaussian process regression training, this method establishes a nonlinear mapping relationship between features and performance and simultaneously quantifies predictive uncertainty, achieving reliable estimation and risk measurement of the selectivity of unknown zinc salt schemes. Furthermore, by combining an active learning strategy to intelligently score and rank unexperimented schemes, it proactively recommends the most valuable or potentially promising experimental schemes, replacing the traditional experience-based blind trial and error. Finally, through a closed-loop process of experimental verification, data feedback, and model iteration, the model continuously evolves with the experimental progress, constantly approaching the optimal solution. This method transforms the traditional, inefficient, and passive catalyst screening model into a highly efficient, targeted, and data-driven intelligent optimization process. It significantly reduces the over-reliance on expert experience and a large number of ineffective experiments, and can systematically discover high-performance zinc salt modification schemes within limited experimental resources, thereby accelerating the catalyst development process and reducing development costs.

[0062] In one embodiment, S1 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0063] S11: Perform feature encoding on the zinc salt types and Zn / Ru molar ratios in the candidate zinc salt modification schemes, convert the SMILES strings of organic acid zinc salts into Morgan fingerprint vectors, and convert the chemical formulas of inorganic zinc salts into anion descriptor vectors containing electronegativity, ionic radius and pKa value, generating basic feature data for each scheme.

[0064] Specifically, for zinc organic acid salts among the candidate solutions, the system first obtains the corresponding SMILES strings for each type of zinc organic acid salt. These strings are standard identifiers reflecting the atomic connections, functional group distribution, and molecular structural framework of organic molecules. The system obtains these strings through a professional chemical data retrieval system to ensure that the strings accurately and completely reflect the molecular structural information of the zinc organic acid salts. Furthermore, the system uses cheminformatics processing tools to transform the obtained SMILES strings. By traversing the atomic nodes and chemical bond connections in the molecular structure, it captures local structural features and functional group-related information, transforming the one-dimensional SMILES strings into multi-dimensional Morgan fingerprint vectors. These vectors numerically represent the molecular structural characteristics of the zinc organic acid salts, which can be directly identified and utilized in subsequent data processing and model training.

[0065] For inorganic zinc salts in the candidate solutions, the system extracts the corresponding chemical formulas of various inorganic zinc salts. Using these formulas as the core basis, the anionic components of the inorganic zinc salts are separated, clarifying their composition and structural characteristics. Further, the system calculates the electronegativity of the separated anionic components, identifying the constituent atoms and the number of each atom. Electronegativity scaling values ​​for each constituent atom are obtained from a standard chemical database. Based on the bonding mode and electron cloud distribution of the atoms in the anion, the weighting coefficients of each atom are determined, and the overall electronegativity of the anion is calculated through a weighted average. Further, the system analyzes the coordination environment of the anion in the reaction system, determines the corresponding coordination number, and searches a standard crystal radius database based on the chemical composition and structural type of the anion. Ionic radius data matching the coordination number are selected, and this data is directly used as the ionic radius parameter of the anion. Simultaneously, the system determines the pKa value of the anion, identifies the corresponding conjugate acid, and queries the dissociation constant of the conjugate acid under standard conditions from a thermodynamic database. Using the mathematical conversion relationship between the dissociation constant and pKa value, the system calculates the pKa value of the anion. The system integrates these three types of parameters in a preset order to form an anion descriptor vector with fixed dimensions. This vector comprehensively reflects the core physicochemical characteristics of inorganic zinc salt anions, providing data support for subsequent analysis of the correlation between inorganic zinc salts and catalyst performance. Finally, the system uses the Morgan fingerprint vector corresponding to organic acid zinc salts and the anion descriptor vector corresponding to inorganic zinc salts as the basic characteristic data for the corresponding candidate zinc salt modification schemes, completing the generation of basic characteristic data for all candidate schemes and ensuring that each candidate scheme has corresponding basic characteristic data that reflects its core material properties.

[0066] S12: The basic feature data and the corresponding Zn / Ru molar ratio are concatenated and standardized. The Morgan fingerprint vector and anion descriptor vector are connected to the molar ratio values ​​of the basic feature data by dimension and scale normalized to generate a standard feature vector.

[0067] Specifically, the system extracts the basic feature data corresponding to each candidate scheme, which is either a Morgan fingerprint vector or an anion descriptor vector. Simultaneously, the system extracts the Zn / Ru molar ratio value corresponding to the candidate scheme, a key process parameter reflecting the amount of zinc salt added. Following preset dimensional connection rules, the system integrates the Morgan fingerprint vector or anion descriptor vector with the corresponding Zn / Ru molar ratio value at the dimensional level. For the connection between the Morgan fingerprint vector and the Zn / Ru molar ratio value, the system adds the molar ratio value as an independent dimension to the specified dimensional position of the Morgan fingerprint vector. Similarly, for the connection between the anion descriptor vector and the Zn / Ru molar ratio value, the system adds the molar ratio value as an independent dimension to the specified dimensional position of the anion descriptor vector, ensuring consistency in the vector structure after concatenating the two different types of basic feature data with the molar ratio value.

[0068] Furthermore, the system performs scale normalization on the concatenated combined vectors. A unified scale normalization method is used to process each dimension of the combined vectors individually, transforming each dimension into a uniform numerical range to ensure the comparability and standardization of the processed data. The system performs scale normalization on the concatenated combined vectors of all candidate schemes. After processing, a standard feature vector is generated for each candidate scheme. This vector is a structured high-dimensional numerical vector that can completely and accurately reflect the core characteristics of zinc salt types and Zn / Ru molar ratios in the candidate schemes.

[0069] S13: Store and index the standard feature vectors of all candidate schemes, establish the mapping relationship between scheme identifiers and feature vectors, and generate a standardized feature library for unified querying.

[0070] Specifically, the system employs a structured data storage format to batch store the standard feature vectors of each candidate solution. During storage, it ensures that every dimension of the standard feature vector is completely recorded, preventing data loss or distortion. Simultaneously, the system assigns a unique solution identifier to each candidate zinc salt modification solution. This identifier serves as a unique identity for the candidate solution and can be used for subsequent solution queries and data matching. The system performs indexing, establishing a one-to-one mapping between solution identifiers and their corresponding standard feature vectors. This mapping enables quick retrieval of the corresponding standard feature vector using the solution identifier, and also allows tracing back to the corresponding solution identifier using the standard feature vector, providing efficient indexing support for subsequent data query and retrieval processes. During the mapping process, each set of mapping data is validated to ensure the accuracy of the correspondence between solution identifiers and standard feature vectors, avoiding one-to-many or many-to-one erroneous mappings, thus guaranteeing the integrity and accuracy of the indexing system. Based on the completed standard feature vector storage and the mapping relationship between scheme identifiers and feature vectors, the system generates a standardized feature library for unified querying. This feature library contains core content such as scheme identifiers, corresponding standard feature vectors, zinc salt type information of candidate schemes, and Zn / Ru molar ratio information, forming a structured database system.

[0071] In one embodiment, S2 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0072] S21: Analyze and extract information from historical experimental records to extract the types of zinc salts, Zn / Ru molar ratios, and corresponding measured values ​​of cyclohexene selectivity of completed experiments, and generate a structured historical data list.

[0073] Specifically, the system acquires multi-source historical experimental records, including experimental archives of past reactions conducted in the laboratory, relevant experimental data disclosed in publicly available academic literature, and similar experimental records shared within the industry. All records contain information related to zinc salt modification involved in the reaction and reaction performance test results. The system first performs format recognition and reading processing on the acquired historical experimental records, supporting the parsing of various common record formats to ensure the complete reading of key information such as text and numerical values. Subsequently, the system initiates a key information recognition process, locating descriptive information related to the type of zinc salt in the records through preset field recognition rules. This information may exist in the form of names, chemical identifiers, etc. The system accurately extracts the type of zinc salt used in each completed experiment through text matching and semantic recognition technology.

[0074] Furthermore, the system identifies numerical information related to the Zn / Ru molar ratio in the records. This information is a core process parameter reflecting the amount of zinc salt added. The system uses a numerical extraction algorithm to separate the corresponding molar ratio-related data from the process parameter descriptions in the records. In addition, the system focuses on the measured results descriptions of cyclohexene selectivity in the records, extracting the corresponding measured values. These values ​​are core indicators reflecting catalyst performance. During the extraction process, the system performs preliminary verification on all extracted information, removing obviously illogical anomalies, such as content inconsistent with the conventional descriptions of zinc salt types, unreasonable descriptions of molar ratio-related data, and abnormal records of measured values. The system then organizes the extracted zinc salt types, Zn / Ru molar ratios, and cyclohexene selectivity measured values ​​into fixed fields. Each field corresponds to a unified information type, ensuring data format consistency. Finally, a structured historical data list is generated, where each record uniquely corresponds to a completed experiment, fully preserving the core characteristics and performance results of the experiment.

[0075] S22: For each scheme in the historical data list, perform feature query and pairing processing in the standardized feature library based on the zinc salt type and Zn / Ru molar ratio, match the corresponding standard feature vector and associate it with the measured selectivity value to generate a feature-performance paired dataset.

[0076] Specifically, for each experimental protocol in the historical data list, the system extracts two key pieces of information: the type of zinc salt and the Zn / Ru molar ratio. These are used as joint query conditions to perform a retrieval operation in the standardized feature library. During the retrieval process, the system performs precise matching based on the zinc salt type identifier information with relevant records in the feature library. Simultaneously, it combines this with relevant Zn / Ru molar ratio data to locate the standard feature vector that perfectly corresponds to the experimental protocol. During the matching process, the system employs a multi-verification mechanism. First, it verifies the consistency of the zinc salt type identifier to ensure that the zinc salt corresponding to the matched feature vector is completely consistent with the zinc salt used in the historical experimental protocols. Then, it verifies the correspondence of Zn / Ru molar ratio data to avoid matching errors caused by differences in molar ratios. If multiple potential matching results exist under a single query condition during the retrieval process, the system further compares the precise correspondence between other auxiliary identifier information of the zinc salt and the molar ratio to select the unique and correct standard feature vector. If no corresponding feature vector is found, the system marks the historical experimental record as pending verification and does not include it in the subsequent processing flow.

[0077] After matching the standard feature vector, the system associates and binds the vector with the corresponding measured cyclohexene selectivity value in the historical data list, forming a feature-performance data pair for a single experimental scheme. The system performs the above query, pairing, and association operations on all historical experimental schemes, processing each record one by one. Finally, it integrates and aggregates all the validly associated feature-performance data pairs to generate a feature-performance paired dataset. This dataset completely preserves the standard feature vector and the corresponding actual performance results for each completed experiment.

[0078] S23: Integrate and statistically process the feature-performance paired dataset, summarize all data pairs to form the initial training set, and select the maximum value from all measured selectivity to generate the current best performance benchmark.

[0079] Specifically, the system performs integrity checks on all feature-performance data pairs in the paired dataset, verifying that each data pair simultaneously contains a complete standard feature vector and the corresponding measured value of cyclohexene selectivity. Data pairs with missing data, abnormal vector dimensions, or non-standard representation of measured values ​​are removed, ensuring that all data included in subsequent processing is valid and standardized. The system then aggregates and integrates the validated valid data pairs, storing them in batches according to a unified data format to form a structured dataset set, which serves as the initial training set. The structure of the initial training set maintains a one-to-one correspondence between feature vectors and measured performance values, facilitating direct retrieval and access during subsequent model training.

[0080] Furthermore, the system performs statistical traversal processing on the measured values ​​of cyclohexene selectivity in all valid data pairs. The system reads each measured value sequentially, records the maximum value during the current traversal through comparison operations, and continuously updates the recorded maximum value as the traversal progresses until all measured values ​​have been traversed. After the traversal is complete, the final maximum value recorded by the system is the optimal selectivity result among all completed experiments. The system officially generates the current optimal performance benchmark from this result, and simultaneously records the relevant information of the experimental scheme corresponding to this benchmark, including key features such as the corresponding zinc salt type and Zn / Ru molar ratio, so that it can serve as a core reference for performance comparison during subsequent active learning.

[0081] In one embodiment, S3 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0082] S31: Perform input-output separation processing on the initial training set, extract all standard feature vectors as the training input matrix, extract the corresponding measured values ​​of cyclohexene selectivity as the training target vector, and generate model training data.

[0083] Specifically, the system iterates through each data record in the initial training set, checking whether it simultaneously contains the standard feature vector and the corresponding measured value of cyclohexene selectivity. For records missing any core data item, the system filters and removes them, retaining only complete and valid records for subsequent processing. After verification, the system performs input-output separation. The system extracts the standard feature vectors corresponding to all valid records and integrates them in an ordered manner according to the experimental record's chronological order or scheme number order to form the training input matrix. The mathematical expression of the training input matrix is:

[0084]

[0085] in, To train the input matrix, Let represent the standard feature vector of the i-th valid record, N represent the number of valid records, and D represent the dimension of the standard feature vector. The number of columns in the matrix is ​​consistent with D. Subsequently, the system extracts the measured cyclohexene selectivity value corresponding to each valid record, and arranges them in a one-dimensional ordered manner according to the order that completely corresponds to the rows of the training input matrix, forming the training target vector, whose mathematical expression is:

[0086]

[0087] in, To train the target vector, This represents the measured value of cyclohexene selectivity for the i-th valid record. The vector length perfectly matches the number of rows N in the training input matrix, ensuring that the measured value at each position is accurately associated with the standard feature vector of the corresponding row. The system performs a unified format conversion on all numerical data in the training input matrix and the training target vector, transforming all types of numerical values ​​into floating-point numbers that can be directly recognized and processed by the Gaussian process model, eliminating modeling interference caused by different data storage formats. Finally, the system associates and stores the training input matrix and the training target vector to form complete model training data. It also records key information such as the execution time of the separation process, the number of valid data records, and the dimension of the training input matrix, ensuring the traceability and standardization of the model training data.

[0088] S32: Based on the Matrn kernel function, Gaussian process modeling is performed on the model training data, the covariance function form and initial hyperparameters are configured, and the length scale and noise variance parameters are optimized by maximizing the marginal likelihood function to generate the trained regression model.

[0089] Specifically, the system determines the form of the covariance function and uses the Matrn kernel function to construct the covariance structure, which is mathematically expressed as follows:

[0090]

[0091] in, and These represent the i-th and j-th feature vectors in the training input matrix, respectively. Let be the Euclidean distance between the two vectors. The length scale parameter is used to adjust the sensitivity of the feature space. The smoothing parameters determine the model's ability to fit nonlinear relationships. For gamma function, for The modified Bessel function of order 1. Based on this kernel function, the system constructs the covariance matrix of the training data, which is mathematically expressed as: Matrix elements ,in The noise variance parameter is used to fit the random error in the experimental data, where $$\delta_{ij}$$ is the Kronecker function, i.e., 1 when i=j and 0 otherwise. The system then configures and initializes hyperparameters, pre-setting initial value ranges according to hyperparameter type, and assigning initial values ​​to the length scaling parameter, smoothing parameter, and noise variance parameter. The system optimizes hyperparameters by maximizing the marginal likelihood function, which is mathematically expressed as:

[0092]

[0093] in, To train the target vector, To train the input matrix, This is the determinant of the covariance matrix. The system employs a numerical optimization algorithm for iterative solution, continuously adjusting the length scale and noise variance parameters, and simultaneously calculating the corresponding marginal likelihood function values ​​until the function values ​​reach the convergence condition (i.e., the difference between function values ​​in two adjacent iterations is less than a preset threshold), thus obtaining the optimal parameter combination. Substituting the optimal parameters into the model, the system completes the training and fitting of the regression model based on the training input matrix and the training target vector, constructing a mapping relationship between feature information and cyclohexene selectivity. After training, the system performs basic fitting performance verification on the model by calculating the correlation index between the model's fitted values ​​and the measured values, validating the fitting effect. The model is then stored in a unified binary format.

[0094] S33: Perform interface encapsulation processing on the trained regression model so that it can output the predicted mean and predicted variance of cyclohexene selectivity from the input feature vector, generating an uncertainty prediction model.

[0095] Specifically, the system defines a unified specification for the model interface, specifying that the input object is a standard feature vector, and the output objects are the predicted mean and variance of cyclohexene selectivity. A standardized model call interface framework is constructed, integrating data format validation logic to perform dimensionality consistency and data format standardization checks on the input standard feature vectors. Dimensionality consistency checks confirm that the dimension of the input feature vector matches the number of columns in the training input matrix, while data format standardization checks confirm that the values ​​of the input vectors are all valid floating-point numbers. The interface framework also integrates data conversion logic, converting the compliant input standard feature vectors into a tensor format that the trained regression model can directly recognize and process. The converted data is then sequentially input into the regression model, triggering the model's internal calculation process. Preferably, the model calculation process follows the Gaussian process prediction formula, with the mathematical expressions for the predicted mean and predicted variance being:

[0096]

[0097]

[0098] in, The input is the feature vector to be predicted. Let the covariance vector of the feature vector to be predicted and each feature vector in the training input matrix have the following elements: , The covariance of the feature vector to be predicted is itself. , The covariance matrix of the training data, This serves as the training target vector. After the model calculation is complete, the results are passed to the interface's output processing module. This module converts the results into clear predicted mean and predicted variance according to a preset format and outputs them back. The predicted mean reflects the core predicted result of the cyclohexene selectivity of the candidate scheme, while the predicted variance reflects the degree of uncertainty of the predicted result.

[0099] During the encapsulation process, the system incorporates robust exception handling logic to capture real-time anomalies such as computational overflow, data transmission interruption, and model loading failure that may occur during model invocation. It also executes exception logging and process termination operations to prevent the entire optimization process from being interrupted by a single call anomaly. After encapsulation, the system conducts comprehensive functional testing on the model. By inputting multiple sets of different standard feature vectors, it verifies whether the interface can accurately and stably output the corresponding predicted mean and predicted variance, while also verifying the effectiveness of the exception handling logic. Upon successful testing, the system formally designates the model with standardized input / output interfaces as an uncertainty prediction model, assigns it a unique number and stores it, and records key information such as interface specification details, test result summaries, and encapsulation completion time during the model encapsulation process. This ensures that the uncertainty prediction model can be directly invoked by subsequent active learning scoring steps.

[0100] In one embodiment, S4 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0101] S41: Perform set difference analysis on the standardized feature library and the initial training set, screen out all candidate schemes that did not appear in the initial training set and the standard feature vectors corresponding to the candidate schemes, and generate a set of schemes to be evaluated.

[0102] Specifically, the system extracts the core identifier information of all candidate solutions from the standardized feature library. The core identifier information is the standard feature vector or unique solution identifier corresponding to the solution. Simultaneously, it extracts the corresponding core identifier information of all experimental solutions in the initial training set, forming two independent identifier sets. The system uses set difference operations to perform difference analysis. The mathematical expression of set difference operations is:

[0103]

[0104] in, The set of identifiers for the proposed solutions. This is the set of identifiers for all schemes in the standardized feature library. This is the set of identifiers for the experimental schemes in the initial training set. Through this operation, the system filters out all schemes that exist only in... Not existing The system uses the identified information to perform precise matching queries in a standardized feature library, extracting candidate solution details and standard feature vectors for each identified information. The candidate solution details include the type of zinc salt and the Zn / Ru molar ratio.

[0105] Furthermore, the system performs integrity checks on the extracted candidate solutions and standard feature vectors to ensure that each selected solution has complete feature data and solution information, eliminating solutions with missing feature data or incomplete information. Finally, the system summarizes all qualified unexperimented candidate solutions and their corresponding standard feature vectors to generate a set of solutions to be evaluated. The set of solutions to be evaluated is stored in a structured format, establishing a one-to-one mapping relationship between solution details and standard feature vectors, and recording key information about the generation of the set, including the generation time, the number of solutions in the set, and the dimension of the feature vectors.

[0106] S42: Perform Gaussian process posterior inference on each scheme in the set of schemes to be evaluated and the input uncertainty prediction model, calculate the predicted distribution parameters of cyclohexene selectivity based on the covariance relationship obtained from training, and generate candidate scheme prediction results containing the predicted mean and predicted variance.

[0107] Specifically, the system sequentially extracts the standard feature vector corresponding to each candidate scheme in the set of schemes to be evaluated. This feature vector is then input into the uncertainty prediction model through a standardized interface, triggering the model's posterior inference calculation. The posterior inference is based on the covariance relationship determined during the training phase. Its core is calculating the predicted distribution parameters of the cyclohexene selectivity of the candidate schemes. The predicted distribution follows a normal distribution, with parameters being the predicted mean and the predicted variance. The core formula for posterior inference follows the Gaussian process prediction formula, i.e., the predicted mean... With prediction variance .in Let be the standard feature vector of the scheme to be evaluated. The feature vectors of the scheme to be evaluated and the training input matrix The covariance vector of each eigenvector in the equation. The covariance matrix of the training data, To train the target vector, Let be the covariance of the eigenvectors of the scheme to be evaluated.

[0108] During the computation process, the system synchronously calls upon the optimal hyperparameters obtained during the model training phase, including the length scale parameter. Smoothing parameters and noise variance parameter To ensure the accuracy of covariance calculation remains consistent with that during the training phase, the system executes the aforementioned posterior inference process for each scheme in the set of schemes to be evaluated, obtaining the corresponding predicted mean and predicted variance one by one. The system verifies the reasonableness of the prediction results for each scheme, confirming that the predicted mean is within a reasonable performance range and the predicted variance is non-negative, and discards schemes with abnormal prediction results. Finally, the system correlates and summarizes all qualified candidate schemes and their corresponding predicted means and predicted variances to generate candidate scheme prediction results. The prediction results are arranged in scheme order, establishing a one-to-one mapping relationship between scheme identifiers and prediction parameters.

[0109] S43: Perform expectation enhancement and uncertainty fusion scoring on the prediction results of candidate schemes and the current best performance benchmark. Calculate the comprehensive score of each scheme based on the preset exploration factor and balance coefficient, sort them, and generate an experimental recommendation list.

[0110] Specifically, after obtaining the prediction results of candidate solutions, the system compares and analyzes them with the current best performance benchmark to determine the current best performance benchmark. The benchmark is the maximum measured value of cyclohexene selectivity in the initial training set. The system's prediction mean is based on the candidate schemes. With prediction variance Calculate the expected improvement value for each solution. The mathematical expression for the expected improvement value is:

[0111]

[0112] in, The cumulative distribution function of the standard normal distribution. Let be the probability density function of the standard normal distribution. The expected improvement is the square root of the predicted standard deviation, i.e., the predicted variance. It quantifies the expected product of the probability that a candidate solution outperforms the current best benchmark and the magnitude of that improvement. Preferably, the system calculates an uncertainty score, which uses the normalized result of the predicted standard deviation. The calculation formula is as follows:

[0113]

[0114] in, The maximum predicted standard deviation for all candidate solutions is used, and the uncertainty score reflects the ambiguity of the model's understanding of the candidate solutions' performance. The system uses a weighted summation method for fusion scoring, incorporating a pre-defined exploration factor. As a weighting factor in the uncertainty score, the balance coefficient As a weighting factor for the expected improvement value, the comprehensive scoring formula is as follows:

[0115]

[0116] in, The normalized result for the expected improvement value ranges from 0 to 1. The system sorts all candidate schemes in descending order of comprehensive score and selects the top-ranked schemes as recommended experimental schemes. The system summarizes the details, prediction parameters, and comprehensive score of the recommended schemes to generate an experimental recommendation list. The list includes fields such as recommendation number, scheme details, prediction mean, prediction variance, expected improvement value, and comprehensive score.

[0117] In one embodiment, S5 of the performance optimization method for zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment provided by the present invention specifically includes the following steps:

[0118] S51: Perform catalytic experiments and data acquisition and processing for each scheme in the experimental recommendation list. Conduct benzene hydrogenation reaction according to the zinc salt type and Zn / Ru molar ratio specified in the scheme, measure and record the cyclohexene selectivity, and generate new measured data.

[0119] Specifically, the system extracts the core information of each scheme from the recommended experiment list in sequence, including the type of zinc salt and the Zn / Ru molar ratio. This information is then converted into standardized experimental operation instructions, which include control logic for key operational parameters such as raw material ratios and reaction condition control. Based on these instructions, the system performs experimental preparation, calculates the required amount of zinc salt according to the Zn / Ru molar ratio, and uniformly mixes the calculated amount of zinc salt with the ruthenium catalyst. Subsequently, the mixed catalyst system, along with benzene and the reaction medium, is added to the high-pressure reactor in a preset ratio. The system performs airtightness testing and environmental pretreatment of the reactor, replacing the air inside the reactor with inert gas to ensure that the reaction environment meets the atmospheric requirements for the benzene hydrogenation reaction.

[0120] After pretreatment, the system initiates the reaction process according to preset reaction conditions, monitoring and adjusting the reaction temperature, pressure, and stirring rate in real time to ensure that all parameters remain stable within the set range during the reaction. Once the preset reaction time is reached, the system automatically terminates the reaction process and initiates a cooling and depressurization procedure. After the temperature and pressure inside the device drop to a safe range, the reaction product is sampled. The system sends the extracted reaction product to the product analysis module, using gas chromatography to separate and detect the product components. The cyclohexene selectivity value is calculated using the standard curve method. The system verifies the validity of the detected values, eliminating abnormal data caused by operational deviations, recording qualified values, and associating them with corresponding experimental protocol information, including zinc salt type, Zn / Ru molar ratio, and unique experimental identifier. Finally, the system summarizes the qualified cyclohexene selectivity values ​​and associated information for all recommended protocols, generating structured new measured data. This new measured data establishes a one-to-one mapping relationship between experimental protocols and measured performance.

[0121] S52: Perform data integration and performance benchmark update processing on the newly added measured data and the initial training set. Selectively pair the standard feature vector corresponding to each new experimental scheme with the measured data and add it to the training set. At the same time, extract the highest value from the updated measured value set to generate an updated training set and an updated current best performance benchmark.

[0122] Specifically, the system performs comprehensive data validation on newly added measured data. Validation includes data format completeness, numerical rationality, and consistency with experimental scheme information. This ensures that the new data has no missing fields, that the values ​​are within a reasonable performance range, and that they accurately correspond to the experimental scheme information. Data that fails validation is marked and excluded, and only valid data is retained for integration. Based on the experimental scheme information in the newly added valid data—namely, the type of zinc salt and the Zn / Ru molar ratio—the system precisely matches the standard feature vector corresponding to each scheme from the standardized feature library. The matching logic follows the dual-keyword matching rule of step S2 to ensure a unique correspondence between the feature vector and the experimental scheme.

[0123] Furthermore, the system pairs the matched standard feature vectors with the corresponding newly added measured cyclohexene selectivity values ​​to form new feature-performance data pairs. The format of the new data pairs is completely consistent with that of the data pairs in the initial training set. The system integrates all the new feature-performance data pairs into the initial training set, performing a data integration operation. During the integration process, the integrity and order of the data pairs are maintained. The integrated data is reordered according to the experimental time sequence to generate an updated training set. The updated training set contains all valid data from the initial training set and the newly added valid feature-performance data pairs, using the same structured storage format as the initial training set. It also records key information about the data integration, including the integration time, the amount of new data, and the total amount of data in the updated training set. Subsequently, the system initiates the update process for the current best performance benchmark. The system extracts the measured cyclohexene selectivity values ​​corresponding to all feature-performance data pairs in the updated training set, forming a set of measured values. By traversing and comparing all values ​​in the set, the system selects the maximum value and determines this maximum value as the updated current best performance benchmark. The system records the experimental scheme information and determination basis corresponding to the updated best performance benchmark, and stores the updated training set and the updated current best performance benchmark together to ensure the traceability of both, providing updated basic data and performance reference standards for subsequent model parameter optimization.

[0124] S53: The updated training set is used as new training data input into the Gaussian process regression framework to optimize the model parameters and generate an updated uncertainty prediction model.

[0125] Specifically, the system performs an input-output separation operation on the updated training set. The separation logic follows step S3, extracting the standard feature vectors of all feature-performance data pairs in the updated training set, integrating them sequentially to form a new training input matrix, and extracting the corresponding measured values ​​of cyclohexene selectivity to form a new training target vector. The mathematical expression of the new training input matrix is ​​as follows:

[0126]

[0127] in, The training input matrix for the initial training set. The matrix is ​​formed by integrating the standard feature vectors of the newly added feature-performance data pairs. N is the initial training set size, M is the amount of newly added valid data, and D is the dimension of the standard feature vectors. The mathematical expression for the new training target vector is:

[0128]

[0129] in, The training target vector for the initial training set. This is a vector formed from newly added measured data. The system inputs the new training input matrix and the new training target vector into the Gaussian process regression framework, using the Matrn kernel function determined in step S3 as the covariance function, maintaining the kernel function form and core parameter types unchanged. The system calls the hyperparameters of the original uncertainty prediction model as initial values ​​and initiates the hyperparameter optimization process. The optimization objective remains maximizing the marginal likelihood function, and the mathematical expression of the marginal likelihood function uses the following formula:

[0130]

[0131] in, The covariance matrix is ​​constructed based on the new training data. The matrix elements are calculated using the Matrn kernel function and a noise variance parameter is added. The system uses a numerical optimization algorithm to iteratively adjust the hyperparameters until the marginal likelihood function reaches the convergence condition, resulting in an optimized new combination of hyperparameters, including updated length scaling parameters, smoothing parameters, and noise variance parameters.

[0132] Furthermore, the system substitutes the new hyperparameter combination into the Gaussian process regression model, retrains the model based on the new training input matrix and the new training target vector, and constructs a mapping relationship between the updated feature information and cyclohexene selectivity. After training, the system verifies the fitting performance of the updated model by calculating the correlation index between the model's fitted values ​​and the measured values ​​on the updated training set, thus evaluating the model's performance improvement. After successful verification, the system performs interface encapsulation processing on the updated model, following the S3 step to ensure that the model has standardized input and output capabilities, able to receive standard feature vectors and output the predicted mean and variance of cyclohexene selectivity. Finally, the system identifies the encapsulated model as the updated uncertainty prediction model, assigns it a unified number and stores it, overwriting the original model, and records key information about the model update, including update time, amount of new data, hyperparameter changes, and model performance indicators, ensuring that the updated uncertainty prediction model can be directly called by subsequent iterative optimization processes.

[0133] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0134] Based on the same inventive concept, this application also provides a system for optimizing the performance of zinc-salt-modified ruthenium catalysts based on active learning and uncertainty assessment, used to implement the aforementioned method for optimizing the performance of zinc-salt-modified ruthenium catalysts based on active learning and uncertainty assessment. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for optimizing the performance of zinc-salt-modified ruthenium catalysts based on active learning and uncertainty assessment provided below can be found in the limitations of the method for optimizing the performance of zinc-salt-modified ruthenium catalysts based on active learning and uncertainty assessment described above, and will not be repeated here.

[0135] Preferably, such as Figure 2 As shown, this invention provides a performance optimization system 600 for zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment. This system is configured with the following modules:

[0136] The zinc salt scheme characterization module 610 is used to perform digital characterization processing on candidate zinc salt modification schemes, converting zinc salt types and Zn / Ru molar ratios into digital vectors of a unified dimension, and generating a standardized feature library containing standard feature vectors of all candidate schemes.

[0137] The data matching and alignment module 620 is used to match and align the historical data list obtained from historical experimental records with the standardized feature library, selectively pairing historical schemes with their standard feature vectors and measured cyclohexene to generate an initial training set and the current best performance benchmark.

[0138] The Gaussian regression modeling module 630 is used to train the initial training set using Gaussian process regression, fit the mapping relationship between the standard feature vector and the cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output.

[0139] The active learning recommendation module 640 is used to input unexperimented candidate solutions and the current best performance benchmark from the standardized feature library into the uncertainty prediction model for active learning scoring, calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them, and generate an experimental recommendation list containing specific solutions.

[0140] The experimental iteration optimization module 650 is used to perform experimental verification and model iteration processing based on the experimental recommendation list. It executes the experiments in the list, adds the new measured data obtained to the training set, re-trains the uncertainty prediction model, updates the historical experimental records, and updates the uncertainty prediction model and the current best performance benchmark.

[0141] Preferably, the zinc salt characterization module 610 provided in this application is configured with the following units:

[0142] The zinc salt feature encoding unit is used to perform feature encoding processing on the zinc salt type and Zn / Ru molar ratio in the candidate zinc salt modification schemes. It converts the SMILES string of organic acid zinc salts into Morgan fingerprint vectors and converts the chemical formula of inorganic zinc salts into anion descriptor vectors containing electronegativity, ionic radius and pKa value, generating basic feature data for each scheme.

[0143] The feature splicing and standardization unit is used to splice and standardize the basic feature data and the corresponding Zn / Ru molar ratio. It connects the Morgan fingerprint vector and the anion descriptor vector with the molar ratio values ​​of the basic feature data by dimension and performs scale normalization to generate a standard feature vector.

[0144] The feature library storage index unit is used to store and index the standard feature vectors of all candidate schemes, establish the mapping relationship between scheme identifiers and feature vectors, and generate a standardized feature library for unified querying.

[0145] Preferably, the data matching and alignment module 620 provided in this application is configured with the following units:

[0146] The historical data parsing unit is used to parse and extract information from historical experimental records, extracting the types of zinc salts, Zn / Ru molar ratios, and corresponding measured values ​​of cyclohexene selectivity of completed experiments, and generating a structured historical data list.

[0147] The feature-performance pairing unit is used to perform feature querying and pairing processing on each scheme in the historical data list based on zinc salt type and Zn / Ru molar ratio in a standardized feature library, match the corresponding standard feature vector and associate it with the measured selectivity value to generate a feature-performance pairing dataset.

[0148] The training set construction unit is used to integrate and statistically process the feature-performance paired dataset, summarize all data pairs to form the initial training set, and select the maximum value from all measured selectivity to generate the current best performance benchmark.

[0149] Preferably, the Gaussian regression modeling module 630 provided in this application is configured with the following units:

[0150] The training data separation unit is used to perform input-output separation processing on the initial training set, extract all standard feature vectors as the training input matrix, extract the corresponding measured values ​​of cyclohexene selectivity as the training target vector, and generate model training data.

[0151] The Gaussian process modeling unit is used to perform Gaussian process modeling on the model training data based on the Matrn kernel function, configure the covariance function form and initialize the hyperparameters, optimize the length scale and noise variance parameters by maximizing the marginal likelihood function, and generate the trained regression model.

[0152] The model interface encapsulation unit is used to encapsulate the trained regression model, enabling it to output the cyclohexene selectivity prediction mean and prediction variance from the input feature vector, thus generating an uncertainty prediction model.

[0153] Preferably, the active learning recommendation module 640 provided in this application is configured with the following units:

[0154] The scheme selection unit is used to perform set difference analysis between the standardized feature library and the initial training set, select all candidate schemes that do not appear in the initial training set and the standard feature vectors corresponding to the candidate schemes, and generate a set of schemes to be evaluated.

[0155] The scheme prediction and inference unit is used to perform Gaussian process posterior inference processing on each scheme in the set of schemes to be evaluated and the input uncertainty prediction model, calculate the predicted distribution parameters of cyclohexene selectivity based on the covariance relationship obtained from training, and generate candidate scheme prediction results including the predicted mean and the predicted variance.

[0156] The experimental scheme ranking unit is used to perform expected improvement and uncertainty fusion scoring on the prediction results of candidate schemes and the current best performance benchmark. It calculates the comprehensive score of each scheme based on the preset exploration factor and balance coefficient, and ranks them to generate an experimental recommendation list.

[0157] Preferably, the experimental iterative optimization module 650 provided in this application is configured with the following units:

[0158] The experimental data acquisition unit is used to perform catalytic experiments and data acquisition and processing for each scheme in the experimental recommendation list. Based on the zinc salt type and Zn / Ru molar ratio specified in the scheme, the benzene hydrogenation reaction is carried out, the cyclohexene selectivity is measured and recorded, and new measured data is generated.

[0159] The training set benchmark update unit is used to integrate the newly added measured data with the initial training set and update the performance benchmark. It selectively pairs the standard feature vector corresponding to each new experimental scheme with the measured data and adds it to the training set. At the same time, it extracts the highest value from the updated set of measured values ​​to generate an updated training set and an updated current best performance benchmark.

[0160] The prediction model iteration unit is used to input the updated training set as new training data into the Gaussian process regression framework to optimize the model parameters and generate an updated uncertainty prediction model.

[0161] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for optimizing the performance of zinc salt modified ruthenium catalyst based on active learning and uncertainty assessment.

[0162] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for optimizing the performance of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment.

[0163] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0164] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0165] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for performance optimization of zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment, characterized in that, Includes the following steps: S1: Perform digital characterization on the candidate zinc salt modification schemes, convert the zinc salt types and Zn / Ru molar ratios into digital vectors of uniform dimension, and generate a standardized feature library containing standard feature vectors of all candidate schemes; S2: Match and align the historical data list obtained from historical experimental records with the standardized feature library, selectively pair the historical schemes with their standard feature vectors and measured cyclohexene, and generate an initial training set and the current best performance benchmark. S3: Perform Gaussian process regression training on the initial training set, fit the mapping relationship between the standard feature vector and cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output. S4: Input the unexperimented candidate solutions in the standardized feature library and the current best performance benchmark into the uncertainty prediction model for active learning and scoring, calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them to generate an experimental recommendation list containing specific solutions. S5: Based on the experimental recommendation list, perform experimental verification and model iteration processing, execute the experiments in the list, add the new measured data obtained to the training set, re-train the uncertainty prediction model, update the historical experimental records, and update the uncertainty prediction model and the current best performance benchmark.

2. The method according to claim 1, characterized in that, S1 includes: S11: Perform feature encoding on the zinc salt types and Zn / Ru molar ratio in the candidate zinc salt modification schemes, convert the SMILES string of organic acid zinc salts into Morgan fingerprint vectors, and convert the chemical formula of inorganic zinc salts into anion descriptor vectors containing electronegativity, ionic radius and pKa value, and generate basic feature data for each scheme. S12: The basic feature data and the corresponding Zn / Ru molar ratio are concatenated and standardized. The Morgan fingerprint vector and the anion descriptor vector are connected with the molar ratio values ​​of the basic feature data according to the dimension, and scale normalization is performed to generate a standard feature vector. S13: Store and index the standard feature vectors of all candidate schemes, establish the mapping relationship between scheme identifiers and feature vectors, and generate a standardized feature library for unified query.

3. The method according to claim 1, characterized in that, S2 includes: S21: Analyze and extract information from historical experimental records to extract the types of zinc salts, Zn / Ru molar ratios, and corresponding measured values ​​of cyclohexene selectivity of completed experiments, and generate a structured historical data list. S22: For each scheme in the historical data list, based on the zinc salt type and Zn / Ru molar ratio, perform feature query and pairing processing in the standardized feature library, match the corresponding standard feature vector and associate it with the measured selectivity value to generate a feature-performance paired dataset. S23: Integrate and statistically process the feature-performance pairing dataset, summarize all data pairs to form the initial training set, and select the maximum value from all measured selectivity to generate the current best performance benchmark.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform input-output separation processing on the initial training set, extract all the standard feature vectors as the training input matrix, extract the corresponding measured values ​​of cyclohexene selectivity as the training target vector, and generate model training data. S32: Based on the Matrn kernel function, Gaussian process modeling is performed on the training data of the model, the covariance function form and initial hyperparameters are configured, and the length scale and noise variance parameters are optimized by maximizing the marginal likelihood function to generate the trained regression model; S33: Perform interface encapsulation processing on the trained regression model so that it can output the predicted mean and predicted variance of cyclohexene selectivity for the input feature vector, thereby generating an uncertainty prediction model.

5. The method according to claim 1, characterized in that, S4 includes: S41: Perform set difference analysis on the standardized feature library and the initial training set, filter out all candidate schemes that do not appear in the initial training set and the standard feature vectors corresponding to the candidate schemes, and generate a set of schemes to be evaluated; S42: Perform Gaussian process posterior inference processing on each scheme in the set of schemes to be evaluated and the input uncertainty prediction model, calculate the predicted distribution parameters of cyclohexene selectivity based on the covariance relationship obtained from training, and generate candidate scheme prediction results containing the predicted mean and predicted variance. S43: Perform expectation enhancement and uncertainty fusion scoring on the predicted results of the candidate schemes and the current best performance benchmark, calculate the comprehensive score of each scheme according to the preset exploration factor and balance coefficient, sort them, and generate the experimental recommendation list.

6. The method according to any one of claims 1-5, characterized in that, S5 includes: S51: Perform catalytic experiments and data acquisition and processing for each scheme in the experimental recommendation list, carry out benzene hydrogenation reaction according to the zinc salt type and Zn / Ru molar ratio specified in the scheme, measure and record cyclohexene selectivity, and generate new measured data. S52: Perform data integration and performance benchmark update processing on the newly added measured data and the initial training set. Selectively pair the standard feature vector corresponding to each new experimental scheme with the measured data and add it to the training set. At the same time, extract the highest value from the updated measured value set to generate an updated training set and an updated current best performance benchmark. S53: The updated training set is used as new training data to input into the Gaussian process regression framework to optimize the model parameters and generate an updated uncertainty prediction model.

7. A performance optimization system for zinc salt-modified ruthenium catalysts based on active learning and uncertainty assessment, characterized in that, The system includes: The zinc salt scheme characterization module is used to digitally characterize candidate zinc salt modification schemes, converting zinc salt types and Zn / Ru molar ratios into digital vectors of uniform dimensions, and generating a standardized feature library containing standard feature vectors of all candidate schemes. The data matching and alignment module is used to match and align the historical data list obtained from historical experimental records with the standardized feature library, selectively pairing the historical schemes with their standard feature vectors and measured cyclohexene to generate an initial training set and the current best performance benchmark. The Gaussian regression modeling module is used to perform Gaussian process regression training on the initial training set, fit the mapping relationship between the standard feature vector and cyclohexene selectivity, quantify the uncertainty of the model prediction results, and generate an uncertainty prediction model with the ability to predict the mean and variance output. The active learning recommendation module is used to input the unexperimented candidate solutions in the standardized feature library and the current best performance benchmark into the uncertainty prediction model for active learning scoring, calculate the expected improvement value and uncertainty comprehensive score of each candidate solution and sort them, and generate an experimental recommendation list containing specific solutions. The experimental iteration optimization module is used to perform experimental verification and model iteration processing based on the experimental recommendation list, execute the experiments in the list, add the new measured data obtained to the training set, re-train the uncertainty prediction model, update the historical experimental records, and update the uncertainty prediction model and the current best performance benchmark.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images