Bayesian optimization method for nitration reaction migration based on substrate characteristic path planning

Abstract

This invention discloses a migration Bayesian optimization method for nitration reactions based on substrate feature path planning. The method acquires the substrate features of each substrate within the reaction domain; calculates the similarity relationships between substrates, constructs a substrate path graph, and determines the starting substrate; selects the next substrate to be explored based on the feature similarity between explored and unexplored substrates, forming a cross-substrate experimental path; for the current substrate, the substrate features and reaction condition variables are input into the migration Bayesian optimization model, and historical data is used to model the current substrate reaction output; the next experimental point is obtained through a data acquisition function and search; the experiment is executed, and new experimental outputs are obtained; the model is updated and iteration continues until the set termination condition is reached and the optimization result is output. This invention enables knowledge expansion from near to far within multi-substrate reaction domains, improves the efficiency of utilizing historical experimental information, and reduces the number of experiments required for new substrates to reach optimal conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization and automated experimental technology for chemical reactions, specifically to a Bayesian optimization method for nitration reaction migration based on substrate feature path planning. Background Technology

[0002] In the development of chemical reactions, the research object is often not a single isolated substrate, but a reaction domain composed of multiple structurally related substrates. Taking aromatic nitration and ester nitration reactions as examples, researchers usually need to find optimal reaction conditions for each substrate in a group with similar skeletons or similar functional group distributions. Traditional approaches often treat each substrate as an independent task, designing and optimizing conditions separately. This results in the inability to fully reuse existing substrate experimental information, high experimental costs, and long development cycles.

[0003] In recent years, Bayesian optimization has been widely used for optimizing chemical reaction conditions due to its advantages such as high sample efficiency, applicability to black-box objective functions, and ability to handle multi-objective problems. Existing methods typically model the relationship between experimental conditions and reaction performance using surrogate models such as Gaussian processes, and select the next experimental point based on a sampling function to search for the optimal region within a limited experimental budget. However, most Bayesian optimization methods primarily model single-substrate tasks. Even when transfer learning or multi-task learning is introduced, they usually assume that historical task data is already given, without further considering how to actively plan the experimental sequence within a reaction domain composed of multiple related substrates to maximize the efficiency of subsequent knowledge transfer.

[0004] On the other hand, to achieve cross-substrate transfer, existing methods often use molecular fingerprints, molecular descriptors, or task tags as additional inputs to Gaussian process models to characterize the similarity between different substrates. Although such methods can improve the optimization efficiency of new substrates to some extent, their substrate features are mostly used only in the surrogate model modeling stage, and not in the higher-level experimental planning stage. Therefore, it is difficult to organize cross-substrate experimental paths from a global perspective, and the value of the similarity structure between substrates in the knowledge acquisition order cannot be fully reflected.

[0005] Therefore, there is an urgent need to develop an optimization method for multi-substrate reaction domains, so that substrate features can not only participate in the construction of the transfer Bayesian optimization model, but also be used to plan cross-substrate experimental paths. This would enable a knowledge acquisition method that expands gradually from near to far and from similar substrates, thereby improving the overall efficiency of experimental information utilization and accelerating the condition optimization process for new substrates. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a migration Bayesian optimization method for nitration reactions based on substrate feature path planning. This method addresses the problems of independent optimization of each substrate, insufficient utilization of historical experimental information, and low optimization efficiency for new substrates in existing multi-substrate nitration reaction optimization processes. This invention utilizes substrate features to simultaneously characterize the similarity relationships between substrates and the task similarity in the migration Bayesian optimization model. Within the same reaction domain, experimental objects are progressively expanded according to substrate feature similarity. Bayesian optimization is then performed on the current substrate in conjunction with historical experimental data from previously explored substrates. This achieves a unification of experimental path planning and migration optimization modeling, improving the efficiency of experimental information utilization and the speed of developing new substrate conditions in multi-substrate reaction optimization processes.

[0007] To achieve the above objectives, the present invention proposes the following technical solution:

[0008] This invention first provides a migration Bayesian optimization method based on substrate feature path planning, which includes the following steps:

[0009] S1: Determine the target reaction domain, which includes multiple nitration reaction substrates, and obtain the substrate characteristics corresponding to each substrate; calculate the feature similarity between substrates based on the substrate characteristics of each substrate, construct a substrate path graph, and determine the starting substrate;

[0010] S2: For the starting substrate, the Latin hypercube sampling method is used to select multiple initial experimental points in the reaction condition space to carry out nitration reaction experiments, measure the experimental output, obtain the initial experimental dataset, and use the starting substrate as the explored substrate.

[0011] S3: Based on the feature similarity between explored and unexplored substrates, select the next substrate to be explored from the unexplored substrates;

[0012] S4: For the current substrate to be explored, input the substrate characteristics and reaction condition variables into the Gaussian process-based migration Bayesian optimization model, and combine the historical experimental data of the explored substrates to model the experimental output of the current substrate;

[0013] S5: Construct the acquisition function and search the candidate response condition space of the current substrate to obtain the next experimental point; execute the experiment corresponding to the next experimental point, obtain the new experimental output, and use the new experimental output to update the migration Bayesian optimization model; repeat S5 until the optimization process of the current substrate to be explored reaches the set termination condition, at which point the current substrate to be explored is regarded as the explored substrate.

[0014] S6: Repeat S3 to S5 until the optimization process of all the substrates to be explored is completed, and the optimization results of the nitration reaction are obtained.

[0015] According to a preferred embodiment of the present invention, the substrate features include one or more of molecular descriptors, structural parameters, empirical parameters, quantum chemical parameters, or molecular fingerprints; when the substrate is an aromatic compound, the substrate features include one or more of Hammett constant, Sterimol parameter, HOMO level, LUMO level, or electrostatic potential parameter.

[0016] According to a preferred embodiment of the present invention, the feature similarity between the substrates in S1 is calculated using a radial basis function kernel; the nodes in the substrate path graph represent each substrate, the edges between the nodes represent the similarity relationship between the substrates, and the edge weights are determined by the feature similarity between the substrates.

[0017] According to a preferred embodiment of the present invention, determining the starting substrate in S1 includes: for each substrate in the reaction domain, calculating its feature similarity with the remaining substrates, and obtaining the average feature similarity of the substrate relative to the remaining substrates; selecting the substrate with the largest average feature similarity as the starting substrate.

[0018] According to a preferred embodiment of the present invention, the next substrate to be explored in S3 is determined by the following formula:

[0019] ;

[0020] in, This indicates that the substrate set has been explored. Indicates the set of unexplored substrates. Indicates the next substrate to be explored. It indicates the similarity of features between substrates.

[0021] According to a preferred embodiment of the present invention, the migration Bayesian optimization model in S4 uses a Gaussian process as a surrogate model. The input of the Gaussian process includes the reaction condition variable and the substrate feature vector, and its covariance function is expressed as:

[0022] ;

[0023] in, As a conditional variable for the reaction, This is the current feature vector of the substrate. The kernel function corresponding to the reaction condition variable. is the kernel function corresponding to the substrate characteristics.

[0024] According to a preferred embodiment of the present invention, the experimental output includes conversion rate and selectivity; in S4, Gaussian process sub-models are established for conversion rate and selectivity respectively and combined to obtain a combined Gaussian process model, and the model is jointly trained based on the existing experimental data of the current substrate and the historical experimental data of the explored substrate.

[0025] According to a preferred embodiment of the present invention, in S5, a data acquisition function for multi-objective optimization is constructed based on the predicted mean and predicted variance of the outputs of each sub-model in the combined Gaussian process model. The data acquisition function is used to evaluate the expected improvement of candidate experimental points relative to the current Pareto front. The data acquisition function is the expected hypervolume improvement (EHVI).

[0026] In the candidate reaction condition space of the current substrate to be explored, the experimental condition combination that maximizes the acquisition function value is obtained by searching the algorithm as the next experimental point. After obtaining the next experimental point, the corresponding experiment is executed and the conversion rate and selectivity data of the current substrate under the experimental conditions are obtained. Then, the data is incorporated into the current substrate dataset, and the combined Gaussian process model is retrained or updated, and the acquisition function is calculated again to generate the next experimental point.

[0027] According to a preferred embodiment of the present invention, the termination conditions set in S5 include: the current substrate has reached the set experimental round, or the acquisition function value is lower than the threshold, or the Pareto front improvement is lower than the threshold.

[0028] The present invention also provides a nitration reaction migration Bayesian optimization system for implementing the aforementioned method, which includes a substrate feature acquisition module, a path planning module, an experimental data acquisition module, a migration optimization module, a data update module, and a result output module;

[0029] The substrate feature acquisition module is used to acquire the substrate features of the substrate within the target reaction domain;

[0030] The path planning module is used to calculate the feature similarity between substrates based on substrate features and construct a substrate path map to determine the starting substrate and the next substrate to be explored.

[0031] The experimental data acquisition module is used to acquire experimental data of the current substrate under different reaction conditions;

[0032] The migration optimization module is used to construct a migration Bayesian optimization model based on substrate features and response condition variables, and to search for the next experimental point based on the acquisition function;

[0033] The data update module is used to incorporate new experimental data into the historical dataset and update the migration Bayesian optimization model;

[0034] The result output module is used to output the optimization result after the termination condition is met. Through the above technical solution, the present invention has the following significant advantages:

[0035] 1) This invention takes the reaction domain composed of multiple related unexplored substrates as the overall optimization object, and uses the experimental data of explored substrates to assist the current substrate optimization, thereby improving the utilization rate of cross-substrate experimental information and reducing repeated experiments.

[0036] 2) This invention uses the same set of substrate features for both experimental path planning and transfer Bayesian optimization modeling, so that the experimental order selection is consistent with the substrate similarity judgment in the surrogate model, thereby improving the rationality of cross-substrate data transfer.

[0037] 3) This invention selects the next experimental substrate based on the similarity of substrate characteristics, so that the experimental process is carried out in the order of similar substrates to substrates with large differences, thereby reducing the uncertainty of cross-substrate migration and improving the efficiency of subsequent optimization.

[0038] 4) This invention incorporates historical experimental data of explored substrates into the current substrate optimization process and selects the next experimental conditions through transfer Bayesian optimization. Compared to methods that optimize each substrate independently, this invention can reduce the number of experiments required to achieve optimal reaction conditions for new substrates.

[0039] 5) This invention is applicable to multi-objective reaction optimization problems such as conversion rate and selectivity, and can output the Pareto front, providing a basis for the selection of conditions among different objectives. Attached Figure Description

[0040] Figure 1 This is a flowchart of a migration Bayesian optimization method based on substrate feature path planning in one embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the characteristic similarity path of multiple substrates within the reaction domain in one embodiment of the present invention.

[0042] Figure 3 This is a boundary evaluation result diagram of the efficiency improvement of migration Bayesian optimization based on leave-one-out verification in one embodiment of the present invention.

[0043] Figure 4 shows the migration Bayesian optimization effect after path planning in this invention. Detailed Implementation

[0044] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0045] The following is a specific embodiment from the aromatic nitration reaction domain, and with reference to Figure 1 This paper describes the Bayesian optimization method for nitration reaction migration based on substrate feature path planning as described in this invention. The method of this invention mainly includes the following steps:

[0046] S1: Determine the target reaction domain, which includes multiple nitration reaction substrates, and obtain the substrate characteristics corresponding to each substrate; calculate the feature similarity between substrates based on the substrate characteristics of each substrate, construct a substrate path graph, and determine the starting substrate;

[0047] S2: For the starting substrate, the Latin hypercube sampling method is used to select multiple initial experimental points in the reaction condition space to carry out nitration reaction experiments, measure the experimental output (conversion and selectivity), obtain the initial experimental dataset, and use the starting substrate as the explored substrate;

[0048] S3: Based on the feature similarity between explored and unexplored substrates, select the next substrate to be explored from the unexplored substrates;

[0049] S4: For the current substrate to be explored, input the substrate characteristics and reaction condition variables into the Gaussian process-based migration Bayesian optimization model, and combine the historical experimental data of the explored substrates to model the experimental output of the current substrate;

[0050] S5: Construct the acquisition function and search the candidate response condition space of the current substrate to obtain the next experimental point; execute the experiment corresponding to the next experimental point, obtain the new experimental output, and use the new experimental output to update the migration Bayesian optimization model; repeat S5 until the optimization process of the current substrate to be explored reaches the set termination condition, at which point the current substrate to be explored is regarded as the explored substrate.

[0051] S6: Repeat S3 to S5 until the optimization process of all the substrates to be explored is completed, and the optimization results of the nitration reaction are obtained.

[0052] The target reaction domain of this invention can be any substrate capable of nitration, such as aromatic compounds, heteroaromatic compounds, ester compounds, etc. Nitration is a reaction in which hydrogen atoms in organic compound molecules are replaced by nitro groups. It is an important type of organic reaction for the preparation of energetic materials, pharmaceuticals, pesticides, dyes, etc.

[0053] In a specific embodiment of the present invention, the target reaction domain is selected as a substrate domain for the nitration of a group of aromatic compounds. The substrate domain includes multiple aromatic substrates with similar aromatic skeletons but different substituents, including chlorobenzene, fluorobenzene, trifluorotoluene, and trifluoromethoxybenzene. Conditions for each aromatic substrate are optimized under the same type of aromatic nitration reaction system, with the optimization objective being to improve substrate conversion and target product selectivity.

[0054] The following describes the process of the present invention in detail with reference to the aforementioned method steps and target reaction domain. In this embodiment, substrate features are first extracted for each substrate in the aromatic substrate domain. These substrate features are used to characterize the differences in electronic effects and steric hindrance between different substrates, and preferably include the Hammett constant and Sterimol parameter. The substrate feature parameters corresponding to each substrate are combined to form a substrate feature vector, and the feature vector is normalized to eliminate the influence of differences in the dimensions of different features on subsequent similarity calculations.

[0055] Suppose there are N aromatic substrates in the substrate domain, and the eigenvector of the i-th substrate is denoted as . The correlation between substrates is calculated based on the similarity between their feature vectors, preferably represented by a radial basis function kernel:

[0056]

[0057] in, This represents the feature similarity between substrate i and substrate j. is the length scale parameter of the substrate feature kernel. The larger the value, the closer the two substrates are in the feature space; The smaller the value, the greater the difference between the two substrates in the feature space.

[0058] Furthermore, a substrate path graph is constructed based on the feature similarity between the substrates. Nodes in the substrate path graph represent aromatic substrates, and edges between nodes represent the similarity relationships between substrates. The edge weights are determined by the feature similarity between the substrates. Through this substrate path graph, multiple aromatic nitrification substrates can be uniformly organized within the same feature space, such as... Figure 2 As shown.

[0059] In this embodiment, the starting substrate is first determined in the substrate path diagram. Specifically, for each substrate within the reaction domain, its feature similarity with the remaining substrates is calculated based on the radial basis function kernel, and the average feature similarity of the substrate relative to the remaining substrates is obtained; the substrate with the highest average feature similarity is selected as the starting substrate. This method ensures that the selected starting substrate is positioned relatively centrally in the substrate feature space, thereby improving the transfer support capability of the initial experimental data for the optimization of multiple subsequent substrates. In the aromatic nitration substrate domain of this embodiment, chlorobenzene is found to have the highest average feature similarity relative to the remaining substrates, therefore chlorobenzene is selected as the starting substrate. After the starting substrate is determined, several initial experiments are conducted on chlorobenzene to obtain experimental data for chlorobenzene under different reaction conditions.

[0060] In this embodiment, the reaction condition variables for the aromatic nitration reaction include reaction temperature and residence time. Preferably, the reaction temperature range is set to -5°C to 10°C, and the residence time range is set within a preset interval according to specific process conditions. For chlorobenzene, multiple initial experimental points are selected within the reaction condition variable space using the Latin hypercube sampling method, with each experimental point corresponding to a set of reaction condition parameters. Experiments are conducted on each initial experimental point, and the corresponding conversion rate and selectivity are measured to form an initial experimental dataset for chlorobenzene.

[0061] The main products in the chlorobenzene nitration reaction include ortho-nitrochlorobenzene, meta-nitrochlorobenzene, and para-nitrochlorobenzene. Preferably, the chlorobenzene conversion rate and the selectivity of the mononitrochlorobenzene product are used as optimization targets, where the selectivity is calculated based on the mole fractions of the ortho, meta, and para products. By conducting initial experiments on para-chlorobenzene under different temperature and residence time conditions, the basic response distribution of chlorobenzene in the target reaction space can be obtained.

[0062] After completing the chlorobenzene experiment, based on the characteristic similarity between the explored and unexplored substrates, the substrate with the greatest similarity to the explored substrates is selected as the next substrate to be explored from the unexplored substrates. Let the set of explored substrates be denoted as . The unexplored substrate set is The next substrate to be explored is preferably determined in the following way:

[0063]

[0064] in, This indicates the next substrate to be explored. In this way, the experimental subjects are expanded gradually from strong to weak similarity in the substrate feature space, ensuring that information obtained from the chlorobenzene experiment is preferentially propagated to substrates with similar characteristics to chlorobenzene.

[0065] After identifying the current substrate to be explored, migration Bayesian optimization is performed on it. Specifically, the substrate feature vector and reaction condition variables (including but not limited to reaction temperature, residence time, and mixed acid dosage) are input into a Gaussian process-based migration Bayesian optimization model. Let the reaction condition variables be... The current substrate feature vector is The input is then represented as The covariance function of a Gaussian process can be expressed as:

[0066]

[0067] in, The kernel function corresponding to the reaction condition variable. This is the kernel function corresponding to the substrate features. Preferably, A radial basis function kernel is employed to characterize the similarity relationships between different substrates. Through this combined kernel function, the model can learn not only the effects of reaction condition variables such as temperature and residence time on conversion and selectivity, but also the similarity relationships between different aromatic substrates, thereby achieving cross-substrate migration.

[0068] In this embodiment, Gaussian process sub-models are established for conversion and selectivity respectively, and then combined to obtain a combined Gaussian process model. The model is then jointly trained based on existing experimental data of the current substrate and historical experimental data of the explored substrates. After training, an acquisition function for multi-objective optimization is constructed based on the predicted mean and predicted variance output by each sub-model in the combined Gaussian process model. Since the optimization objectives in this embodiment are both conversion and selectivity, the expected hypervolume improvement (EHVI) is preferably used as the acquisition function to evaluate the expected improvement of candidate experimental points relative to the current Pareto front.

[0069] In the candidate response condition space of the current substrate to be explored, the experimental condition combination that maximizes the acquisition function value is obtained through a search algorithm as the next experimental point. Preferably, a genetic algorithm or a multi-starting point search algorithm can be used to optimize the acquisition function. After obtaining the next experimental point, the corresponding experiment is executed to obtain the conversion rate and selectivity data of the current substrate under the experimental conditions; then, this data is incorporated into the current substrate dataset, and the combined Gaussian process model is retrained or updated to recalculate the acquisition function to generate the next experimental point.

[0070] In this embodiment, when the current substrate reaches a set number of experimental rounds, the acquisition function value is lower than a threshold, or the Pareto front improvement is lower than a threshold, the optimization process for the current substrate is considered complete, and the substrate is added to the set of explored substrates. Subsequently, based on the feature similarity between the updated set of explored substrates and the set of unexplored substrates, the next substrate to be explored is selected, and the above-described transfer Bayesian optimization process is repeated.

[0071] Through the above steps, this embodiment establishes a cross-substrate experimental path in the aromatic nitration substrate domain, starting with chlorobenzene and based on the similarity of the radial basis function kernel of substrate features. Transfer Bayesian optimization is then performed progressively along this path. Since substrate features are used simultaneously for substrate path graph construction and kernel function input for the Gaussian process model, the expansion relationships in the experimental path remain consistent with the task similarity relationships in the surrogate model, thereby enhancing the auxiliary role of historical chlorobenzene experimental data in the subsequent optimization of new substrates.

[0072] In this embodiment, chlorobenzene is used as the starting substrate. Initial experiments are conducted within a preset temperature and residence time range, and a historical dataset is established. Subsequently, based on the radial basis function kernel similarity of the substrate characteristics, the unexplored aromatic substrates most similar to chlorobenzene and the currently explored substrates are selected as subsequent optimization targets. For each new substrate, the historical experimental data of chlorobenzene and other explored substrates are input into the migration Gaussian process model to optimize the conditions of the new substrate to be optimized. Figure 3 To evaluate the potential of cross-substrate information migration in this invention to improve optimization efficiency using leave-one-out validation (i.e., each time one substrate is used as a new substrate to be optimized, and the remaining substrates are used as historical data sources). Figure 4 The performance of the complete method described in this invention (combining path planning and transfer optimization) was compared with that of ordinary Bayesian optimization. The results show that, compared to methods that treat each aromatic substrate as a completely independent task and perform Bayesian optimization separately, this invention, through leave-one-out verification, evaluates the potential for improving optimization efficiency through cross-substrate information transfer. The results show that the average efficiency improvement boundary of transfer Bayesian optimization can reach 60%. Based on this, a path planning strategy starting with chlorobenzene was further combined to conduct actual transfer optimization. Compared to ordinary Bayesian optimization, the efficiency of transfer Bayesian optimization after path planning was improved by 25%, and it can construct the Pareto front of the target substrate more quickly.

[0073] Therefore, the migration Bayesian optimization method based on substrate feature path planning described in this invention can efficiently organize and reuse experimental information of multiple related substrates in the aromatic nitration reaction domain with chlorobenzene as the starting substrate, reduce the number of experiments required for new substrate optimization, and improve the efficiency of multi-objective reaction optimization.

[0074] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of this invention. For those skilled in the art, various modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of this invention.

Claims

1. A Bayesian optimization method for nitration reaction migration based on substrate feature path planning, characterized in that, Includes the following steps: S1: Determine the target reaction domain, which includes multiple nitration reaction substrates, and obtain the substrate characteristics corresponding to each substrate; calculate the feature similarity between substrates based on the substrate characteristics of each substrate, construct a substrate path graph, and determine the starting substrate; the substrate characteristics include one or more of molecular descriptors, structural parameters, empirical parameters, quantum chemical parameters, or molecular fingerprints; the feature similarity between substrates in S1 is calculated using radial basis function kernels; the nodes in the substrate path graph represent each substrate, the edges between nodes represent the similarity relationship between substrates, and the edge weights are determined by the feature similarity between substrates; determining the starting substrate in S1 includes: for each substrate in the reaction domain, calculating its feature similarity with the other substrates, and obtaining the average feature similarity of the substrate relative to the other substrates; selecting the substrate with the largest average feature similarity as the starting substrate; S2: For the starting substrate, the Latin hypercube sampling method is used to select multiple initial experimental points in the reaction condition space to carry out nitration reaction experiments, measure the experimental output, obtain the initial experimental dataset, and use the starting substrate as the explored substrate. S3: Based on the feature similarity between explored and unexplored substrates, select the next substrate to be explored from the unexplored substrates; the next substrate to be explored in S3 is determined by the following formula: ； in, This indicates that the substrate set has been explored. Indicates the set of unexplored substrates. Indicates the next substrate to be explored. Indicates the characteristic similarity between substrates; S4: For the current substrate to be explored, the substrate features and reaction condition variables are jointly input into a Gaussian process-based transfer Bayesian optimization model, and the experimental output of the current substrate is modeled by combining historical experimental data of the explored substrates; the transfer Bayesian optimization model in S4 uses a Gaussian process as a surrogate model. The input of this Gaussian process includes the reaction condition variables and the substrate feature vector, and its covariance function is expressed as: ； in, As a conditional variable for the reaction, This is the current feature vector of the substrate. The kernel function corresponding to the reaction condition variable. The kernel function corresponding to the substrate characteristics; Experimental outputs include conversion rate and selectivity. Gaussian process sub-models are established for conversion rate and selectivity respectively and combined to obtain a combined Gaussian process model. The model is jointly trained based on existing experimental data of the current substrate and historical experimental data of the explored substrate. S5: Construct the acquisition function and search the candidate response condition space of the current substrate to obtain the next experimental point; execute the experiment corresponding to the next experimental point, obtain the new experimental output, and use the new experimental output to update the migration Bayesian optimization model; repeat S5 until the optimization process of the current substrate to be explored reaches the set termination condition, at which point the current substrate to be explored is regarded as the explored substrate. In S5, based on the predicted mean and predicted variance of each sub-model output in the combined Gaussian process model, a data acquisition function for multi-objective optimization is constructed. The data acquisition function is used to evaluate the expected improvement of candidate experimental points relative to the current Pareto front. The data acquisition function is the expected hypervolume improvement (EHVI). In the candidate response condition space of the current substrate to be explored, the experimental condition combination that maximizes the acquisition function value is obtained through a search algorithm as the next experimental point. After obtaining the next experimental point, the corresponding experiment is executed and the conversion rate and selectivity data of the current substrate under the experimental conditions are obtained. Then, the data is incorporated into the current substrate dataset, and the combined Gaussian process model is retrained or updated, and the acquisition function is calculated again to generate the next experimental point. S6: Repeat S3 to S5 until the optimization process of all the substrates to be explored is completed, and the optimization results of the nitration reaction are obtained.

2. The method according to claim 1, characterized in that, When the substrate is an aromatic compound, the substrate characteristics include one or more of the following: Hammett constant, Sterimol parameter, HOMO level, LUMO level, or electrostatic potential parameter.

3. The method according to claim 1, characterized in that, The termination conditions set in S5 include: the current substrate has reached the set experimental round, or the acquisition function value is lower than the threshold, or the Pareto front improvement is lower than the threshold.

4. A migration Bayesian optimization system for implementing the method according to any one of claims 1-3, characterized in that, It includes a substrate feature acquisition module, a path planning module, an experimental data acquisition module, a migration optimization module, a data update module, and a result output module; The substrate feature acquisition module is used to acquire the substrate features of the substrate within the target reaction domain; The path planning module is used to calculate the feature similarity between substrates based on substrate features and construct a substrate path map to determine the starting substrate and the next substrate to be explored. The experimental data acquisition module is used to acquire experimental data of the current substrate under different reaction conditions; The migration optimization module is used to construct a migration Bayesian optimization model based on substrate features and response condition variables, and to search for the next experimental point based on the acquisition function; The data update module is used to incorporate new experimental data into the historical dataset and update the migration Bayesian optimization model; The result output module is used to output the optimization result after the termination condition is met.

Images