A method for automatically optimizing a continuous intelligent microreactor platform

Abstract

The present invention provides a method for automatically optimizing a continuous intelligent microreactor platform, which relates to the technical fields of chemistry, chemical engineering, and pharmaceutical synthesis processes. The method comprises the following steps: S1, initializing input parameters of an actual design space; S2, obtaining an initial design by maximizing minimum distance Latin hypercube design according to sample size and variable dimension, and mapping the initial design to the actual design space as a sample matrix by reflection transformation; S3, iteratively calculating the current sample entropy of the sample matrix, constructing a crossover-mutation coevolution framework, and adaptively adjusting parameters according to the current sample entropy to obtain a non-dominated solution set; and S4, performing a chemical reaction on the continuous intelligent microreactor platform using process parameters corresponding to the non-dominated solution set to obtain a final product. The method solves the problems of excessively long automatic optimization time and excessive reagent consumption costs of existing chemical reactions, effectively reduces computing resource consumption, reaction automatic optimization time, and reagent consumption costs, and reduces environmental pollution.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical engineering and pharmaceutical synthesis processes, and in particular to a method for automatically optimizing a continuous intelligent microreactor platform. Background Art

[0002] With the deep integration of automation and artificial intelligence technologies, the field of chemical synthesis is undergoing a paradigm shift from traditional intermittent operations to automated continuous flow synthesis. This technological innovation achieves intelligent optimization and real-time regulation of the synthesis process by integrating automation systems with continuous flow reactors, significantly improving reaction efficiency, selectivity, and operational safety. The core advantage of automatic optimization of continuous synthesis technology lies in the organic integration of continuous flow systems and automated intelligent control. The former breaks through the bottleneck of traditional processes by enhancing mass transfer and heat transfer and precise parameter control, while the latter dynamically optimizes reaction pathways based on experimental feedback data. The synergistic effect of the two drives the evolution of the synthesis process towards full automation and intelligence.

[0003] Chemical reaction optimization inherently presents multiple complex challenges: reaction performance metrics (such as yield, selectivity, and environmental factors) often exhibit highly nonlinear relationships with process parameters (such as temperature, flow rate, and concentration), and multi-dimensional objectives such as economic efficiency, environmental friendliness, and product quality must be simultaneously balanced. While traditional optimization algorithms (such as the Simplex Method and SNOBFIT) demonstrate rapid convergence in single-objective scenarios, they are unable to effectively reveal trade-offs between multiple objectives, making them inadequate for the flexible needs of practical process design. For example, when combining multiple objectives through scalarization methods (such as weighted summation), the pre-set weights rely on subjective experience, potentially introducing bias. Traditional Latin Hypercube Design (LHD) suffers from issues such as insufficient sample distribution uniformity and limited global exploration capabilities in complex optimization problems, resulting in low computational efficiency. While genetic algorithms (such as NSGA-II) can generate Pareto frontiers, they require extensive experimental evaluation (typically hundreds of iterations), making them difficult to apply to costly real-world scenarios. Furthermore, these extensive experiments can result in excessive emissions of hazardous substances and waste, potentially polluting the environment.

[0004] Therefore, the existing automatic optimization continuous synthesis technology is unable to effectively reveal the trade-off relationship between multiple objectives, making it difficult to meet the flexible needs of actual process design. In addition, its low computational efficiency leads to excessive consumption of computing resources, long reaction automatic optimization time, and high reagent consumption costs, which can cause environmental pollution and other problems. There is an urgent need for a method for automatically optimizing a continuous intelligent microreactor platform to solve the above technical problems. Summary of the Invention

[0005] In order to overcome the deficiencies in the background technology, the present invention provides a method for automatically optimizing a continuous intelligent microreactor platform.

[0006] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for automatically optimizing a continuous intelligent microreactor platform, wherein the continuous intelligent microreactor platform includes chemical reaction raw materials, a continuous flow reactor, a solvent absorption device, and a control system; the control system includes a TS-EMO monitoring system and a ReactIR real-time monitoring system, and comprises the following steps:

[0008] S1. Initialize the input parameters of the actual design space; the input parameters include sample size n, variable dimension d ,Design space boundary[ lb , ub ], maximum number of iterations T , convergence threshold ε ;

[0009] S2, according to the sample size n and Variable dimensions d Obtain the initial design using the maximized minimum distance Latin hypercube design X , and mapped it to the actual design space as a sample matrix through reflection transformation X 0 ;

[0010] S3, sample matrix X 0 Iteratively calculate the current sample entropy H ( X t ), build a cross-mutation co-evolution framework and according to the current sample entropy H ( X t ) to perform adaptive parameter adjustment and reach the maximum or entropy change at the number of iterations t ΔH Greater than the convergence threshold ε Output the current sample X t As a non-dominated solution set X opt ; Among them, the current sample X t express X 0 No. t Iterations, t <T;

[0011] The crossover-mutation co-evolution framework is constructed based on formula (2):

[0012] (2),

[0013] in, θ c is the crossover rate; θ m is the mutation rate; the CROSSOVER function represents the crossover operation; the MUTATE function represents the mutation operation;

[0014] S4, the continuous intelligent microreactor platform adopts a non-dominated solution set X opt The corresponding process parameters are used for chemical reaction to obtain the final product; specifically including:

[0015] The chemical reaction raw materials are input into the continuous flow reactor, and the TS-EMO monitoring system adopts the non-dominated solution set X opt The corresponding process parameters are combined with the ReactIR real-time monitoring system to carry out a chemical reaction, and the product of the chemical reaction is purified through a solvent absorption device to obtain the final product.

[0016] Specifically, step S2 includes the following steps:

[0017] S21, according to the sample size n and Variable dimensions d Generate candidate sample sets that satisfy Latin hypercube constraints X candidate =L(n,d);

[0018] S22, based on the candidate sample set X candidate Solve the maximin optimization problem of formula (1) by tabu search and obtain the initial design X ;

[0019] (1),

[0020] in, L(n,d) express n The sample points are d The set of Latin hypercubes in dimensional space; x i and x j All are sample points; for x i and x j The Euclidean distance between

[0021] S23, through the affine transformation formula X 0 = X ⊗( ub - lb ) +lb Initial design X Mapping to the actual design space to obtain the sample matrix X 0 .

[0022] Specifically, step S3 includes the following steps:

[0023] S31. Calculate the current sample X t The current sample entropy of H ( X t ); X t Represents the sample matrix X 0 No. t iterations;

[0024] S32, according to the current sample entropy H ( X t ) Based on formula (4)-(5) for the crossover rate θ c and mutation rate θ m Make adjustments:

[0025] θ c = θ base + k 1 *(1 - H ( X t ) / H max ) (4),

[0026] θ m = θ base + k 2*( H ( X t ) / H max ) (5),

[0027] in, H max is the theoretical maximum value of population distribution entropy, H max = d ·log n, n is the sample size, d is the variable dimension, representing the entropy value when it is completely uniformly distributed; θ baseIt is the benchmark value of crossover rate and mutation rate, generally ranging from 0.1 to 0.3;

[0028] S33, according to formula (6) for the current sample X t Perform a directed crossover operation to obtain a crossover sample set X cross :

[0029] X cross = ArithmeticCrossover( X t , θ c )(6),

[0030] in, ArithmeticCrossover is the directional crossover function, corresponding to the CROSSOVER function in formula (2);

[0031] S34, according to formula (8) for the cross sample set X cross Perform the guided mutation operation to obtain the cross-mutation sample set X mut :

[0032] X mut = GuidedMutation( X cross , θ m )(8),

[0033] Among them, GuidedMutation is the guided mutation function, corresponding to the MUTATE function of formula (2);

[0034] S35. For the current sample X t and cross-mutation sample set X mut Perform selection operations to generate next generation samples X t+1 ;

[0035] S36. Calculate the entropy change according to formula (12) ΔH, And increase the number of iterations t by 1 ， If the number of iterations t reaches the maximum or the entropy change ΔH Greater than the convergence threshold ε Output the current sample X t As a non-dominated solution set X opt; Otherwise, return to step S31 and continue iteration; Formula (12) is as follows:

[0036] ΔH = | H ( X t+1 ) - H ( X t )|(12).

[0037] Specifically, step S31 calculates the current sample according to formula (3) X t The current sample entropy of H ( X t ):

[0038] (3),

[0039] in, x ij Refers to the current sample X t No. i Individuals in j The value of the dimension; p ( x ij ) indicates the current sample X t In the individual i The probability that a value in dimension falls into the jth dimension.

[0040] Specifically, the ArithmeticCrossover function in step S33 adopts an arithmetic crossover strategy based on space segmentation to select the current sample X t The parent sample points in x p and x q Perform linear combination to generate offspring x child , so that it is located in the middle area of ​​the parent line, which is expressed as shown in formula (7):

[0041] (7),

[0042] in, α is the weight coefficient.

[0043] Specifically, the GuidedMutation function in step S34 is implemented based on formula (9):

[0044] (9),

[0045] in, represents the individual after mutation; x i Indicates the current sample X t The i individual; δ is the step size factor; x elite For the current elite sample set; Indicates that the mean is 0 and the covariance matrix is Multidimensional Gaussian distribution.

[0046] Specifically, step S35 calculates the current sample according to formula (11): X t and cross-mutation sample set X mut Make a selection:

[0047] X t+1 = ParetoFront( X t ∪ X mut )(11),

[0048] Among them, ParetoFront function is the Pareto front surface function.

[0049] Specifically, the cross rate control coefficient in step S32 k 1 The value range is [0.1, 0.3]; the variation rate control coefficient k 2 The value range is [0.2,0.5].

[0050] Specifically, the current elite sample set x elite To sort from the current population by non-dominated sorting X t Select the top 10% of solutions.

[0051] Specifically, the weight coefficient α =1- θ c .

[0052] Specifically, the step size factor is calculated by formula (10): δ Make dynamic adjustments:

[0053] (10),

[0054] in, is the step size factor of the tth generation;δ 0 is the initial step size, δ 0=(0.1~0.2)*||ub-lb||2; δ T is the termination step length, δ T =(0.001~0.01)* δ 0; τ is the decay constant.

[0055] The present invention provides a method for automatically optimizing a continuous intelligent microreactor platform. Addressing the problems of traditional Latin Hypercube Design (LHD) in complex optimization problems, such as insufficient sample distribution uniformity, limited global exploration capabilities, low computational efficiency leading to excessive consumption of computing resources, prolonged reaction automatic optimization time, and excessive reagent costs, an improved TS-EMO algorithm based on an evolutionary multi-objective optimization framework is proposed. By introducing a dynamic parameter adjustment mechanism and a hybrid evolutionary operator, the initial sample generation strategy and iterative optimization process are improved. The maximum and minimum distance criterion is used to generate the initial Latin Hypercube (LHC) sample, maximize the minimum distance between sample points, and make the sample points evenly distributed in the input space, thereby reducing the sampling bias in the initial design, and improving the quality of the initial sample through a more efficient search strategy of taboo search, thereby increasing the probability of finding the global optimal solution and the efficiency of the algorithm, and reducing the consumption of computing resources; the crossover operation mixes the existing samples to generate new sample points, explores more design spaces, improves sample diversity, and helps the algorithm to more comprehensively explore potential solutions in the search space; mutation is an operation that randomly changes the values ​​of certain sample points to avoid falling into local optimality, while increasing the randomness of the search, and also helps the algorithm to more comprehensively explore potential solutions in the search space, effectively revealing the trade-off relationship between multiple objectives, meeting the flexible needs of actual process design, and obtaining the optimal solution (non-dominated solution set X opt ) The corresponding process parameters are more suitable for chemical reactions on a continuous intelligent microreactor platform, which enables the continuous intelligent microreactor platform to operate efficiently and stably and produce high-quality products, significantly reducing the reaction automatic optimization time and reagent consumption costs. It also effectively solves the problem that a large number of experimental evaluations (usually hundreds of iterations) are difficult to apply to actual scenarios with high experimental costs, and has robustness and engineering applicability.

[0056] The present invention effectively reduces the number of experiments in the automatic optimization process, thereby correspondingly reducing the energy consumption and chemical reagent consumption required for the experiment, reducing carbon emissions, reducing the emission of harmful substances and the generation of waste, alleviating pollution to the environment, and promoting the development of green experiments.

[0057] In addition, the present invention monitors the sample distribution status in real time through sample entropy, and dynamically adjusts the crossover rate and mutation rate based on the sample entropy to achieve a dynamic balance between exploration and development capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 2 is a schematic diagram of a method for automatically optimizing a continuous intelligent microreactor platform according to an embodiment of the present invention;

[0060] Figure 2 1 is a schematic diagram of a process of automatically optimizing a continuous intelligent microreactor platform according to an embodiment of the present invention;

[0061] Figure 3 2 is a schematic diagram of the results of automatic optimization of the propylene high-temperature chlorination reaction before the algorithm optimization and improvement provided by an embodiment of the present invention;

[0062] Figure 4 2. Schematic diagram of automatic optimization of propylene high-temperature chlorination reaction after algorithm optimization and improvement provided by an embodiment of the present invention;

[0063] Figure 5 2. Schematic diagram of automatic optimization of propylene low-temperature chlorination reaction before algorithm optimization and improvement according to an embodiment of the present invention;

[0064] Figure 6 1 is a schematic diagram of automatic optimization of the low-temperature chlorination reaction of propylene after the algorithm optimization and improvement provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The present invention can be explained in detail through the following embodiments. The purpose of providing the present invention is to protect all technical improvements within the scope of the present invention. In the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right", etc. to indicate directions or positional relationships, they only correspond to the drawings of this application for the convenience of describing the present invention, and do not indicate or imply that the device or element referred to must have a specific direction.

[0066] Example 1

[0067] refer to Figure 1 This embodiment provides a method for automatically optimizing a continuous intelligent microreactor platform, comprising the following steps:

[0068] S1. Initialize the input parameters of the actual design space; the input parameters include sample size n, variable dimension d ,Design space boundary[ lb , ub ], maximum number of iterations T , convergence threshold ε ;

[0069] Design space boundaries[ lb , ub ], lb and ub represent the global lower bound (Lower Bound) and global upper bound (Upper Bound) of all design variables (input parameters), respectively;

[0070] The maximum number of iterations in this embodiment T The value range is [50,500], which can be selected according to actual needs;

[0071] The convergence threshold in this embodiment ε The value range is [1e-04, 1e-06], which can be selected according to actual needs;

[0072] S2. Improve LHC initialization: According to the sample size n and Variable dimensions d Obtain the initial design using the maximized minimum distance Latin hypercube design X , and mapped it to the actual design space as a sample matrix through reflection transformation X 0 ;

[0073] The specific steps include:

[0074] S21, according to the sample size n and Variable dimensions d Generate candidate sample sets that satisfy Latin hypercube constraints X candidate =L(n,d);

[0075] The Latin Hypercube (LHC) is an efficient method for sampling multivariate parametric distributions. Its core approach is to improve sample uniformity and representativeness through a stratified strategy. Combining the characteristics of stratified sampling with randomization, it holds significant value in fields such as experimental design and model optimization.

[0076] The core of this method lies in its stratified sampling implementation. Each dimension is divided into several strata (e.g., n strata correspond to n samples), with values ​​randomly selected within each stratum, but with strict non-overlap between strata. This stratification strategy ensures that the sample projections on each dimension cover all subintervals, eliminating the local gaps that can occur with random sampling. For example, when the 0-1 interval is divided into five strata, each sample point must come from a different 0.2-width interval, ensuring low-bias global coverage.

[0077] The construction of the Latin hypercube is divided into three steps:

[0078] 1) Stratification: Divide the value range of each variable (usually standardized to 0-1) into n sub-intervals, where n is the total sample size.

[0079] 2) Random sampling within the stratum: A value is randomly selected within each subinterval to ensure that there is only one sample in each stratum.

[0080] 3) Dimensional Permutation and Combination: Randomly permutate and combine the sampling results of each dimension to avoid interference from correlations between variables. For example, in a two-dimensional Latin hypercube, the subinterval samples are arranged sequentially in the first dimension, while the second dimension is shuffled and then matched, forming an irregular spatial distribution.

[0081] Candidate sample set X candidate express n The sample points are d The Latin hypercube set in the dimensional space is a large sample set, in which each sample point satisfies the Latin hypercube constraint, that is, in each dimension, the sample points are evenly distributed in equally spaced intervals, and each interval has only one sample point;

[0082] S22, based on the candidate sample set X candidate Solve the maximin optimization problem of formula (1) by tabu search and obtain the initial design X ;

[0083] In the initial sample generation stage, the optimized Latin hypercube design (Optimal LHD) that maximizes the minimum distance (maximin criterion) is adopted to determine the optimal initial distribution by solving the maximin optimization problem of the discrete combination shown in formula (1) through tabu search:

[0084] (1),

[0085] in, L(n,d) express n The sample points are d The set of Latin hypercubes in dimensional space; xi and x j All are sample points; for x i and x j The purpose of the formula is to find the set of points that maximizes the minimum distance between any two points in all possible Latin hypercube designs. X， The point set X is used as the initial design, and the tabu search in step S22 is a tabu search algorithm.

[0086] The maximum-minimum distance criterion is used to generate the initial Latin Hypercube (LHC) samples, maximizing the minimum distance between sample points so that the sample points are evenly distributed in the input space, thereby reducing the sampling bias in the initial design.

[0087] In this improved LHC initialization, the tabu search algorithm is used to optimize the initial sample distribution of the Latin Hypercube Design (LHD). The goal is to maximize the minimum distance between sample points, namely the maximin criterion. That is, the quality of the initial samples is improved through a more efficient search strategy of tabu search, further solving the problem that the traditional LHD generation method may have an uneven sample distribution.

[0088] Tabu Search (TS), also known as the tabu search method, is a modern heuristic algorithm proposed by Fred Glover, a professor at the University of Colorado, around 1986. It is used to escape local optima. It first creates an initial solution and then "moves" to a neighboring solution based on it. Through repeated moves, the quality of the solution is improved.

[0089] The tabu search algorithm uses a tabu list to avoid repeated searches and dynamically generates new candidate solution sets through a neighborhood movement strategy. These candidate solution sets expand the algorithm's search space and help find better solutions.

[0090] This example also provides a dynamic domain tabu algorithm based on an improved tabu search algorithm, which can be used to replace the tabu search algorithm in step S22. The dynamic domain tabu algorithm can improve the tabu algorithm in the following aspects:

[0091] 1. Improved Neighborhood Structure: Traditional tabu search neighborhoods may involve swapping a dimension between two sample points or performing small perturbations. The improved tabu search introduces a more complex neighborhood structure, dynamically adjusts the neighborhood size, and combines multiple neighborhood operations (such as swapping and perturbations) to increase search diversity.

[0092] 2. Dynamic tabu list management: Traditional tabu lists record recent operations to avoid loops, but improved algorithms may adopt adaptive tabu deadlines and dynamically adjust the length of the tabu list according to the search process to balance exploration and exploitation.

[0093] 3. Optimization of the amnesty rule: When a forbidden move can bring significant improvement, the amnesty rule allows the removal of the taboo. The improved algorithm may introduce dynamic amnesty conditions based on the improvement of the objective function, rather than just fixed rules.

[0094] The dynamic adjustment strategy of the neighborhood structure can effectively balance computational efficiency and solution quality, ensure the uniform distribution of initial samples in the input space, and significantly reduce the spatial coverage deviation that may exist in the traditional random Latin Hypercube (LHC) method.

[0095] S23, through the affine transformation formula X 0 = X ⊗( ub - lb ) + lb Initial design X Mapping to the actual design space to obtain the sample matrix X 0 .

[0096] Affine transformation is the result of mathematical optimization X 0 The "bridge" of lossless conversion to engineering-usable designs ensures that the superiority of the LHC is retained in practical problems. This is existing technology and will not be described in detail here.

[0097] This step uses the maximum-minimum distance criterion to generate the initial Latin Hypercube (LHC) samples, maximizing the minimum distance between sample points so that the sample points are evenly distributed in the input space, thereby reducing the sampling bias in the initial design.

[0098] S3, sample matrix X 0 Iteratively calculate the current sample entropy H ( X t ), build a cross-mutation co-evolution framework and according to the current sample entropy H ( X t ) to perform adaptive parameter adjustment and reach the maximum or entropy change at the number of iterations t ΔH Greater than the convergence threshold ε Output the current sample X t As a non-dominated solution set Xopt ; Among them, the current sample X t express X 0 No. t Iterations, t <T;

[0099] In the iterative optimization stage, a crossover-mutation co-evolution framework is constructed, and its mathematical expression is shown in formula (2):

[0100] (2),

[0101] in ​ c and ​ m are the adaptive control parameters of crossover rate and mutation rate, X t For the t The population or solution set of generations represents the current sample; each crossover operation (CROSSOVER function) generates offspring individuals x child , these offspring individuals x child Together with the individuals after the mutation operation (MUTATE function), they form the intermediate population, which is the X mut , and then generate a new generation of population through selection operation X t+1 .

[0102] Specifically, step S3 includes:

[0103] S31. Calculate the current sample according to formula (3) X t The current sample entropy of H ( X t ), X t Represents the sample matrix X 0 No. t iterations;

[0104] (3),

[0105] in, x ij Refers to the current sample X t No. i Individuals in j The value of the dimension. p ( x ij ) indicates the current sample Xt (Population matrix) individuals in i The probability that the value in dimension falls into the jth dimension;

[0106] S32, according to the current sample entropy H ( X t ) Based on formula (4)-(5) for the crossover rate ​ c and mutation rate ​ m Make adjustments:

[0107] ​ c = ​ base + k 1 *(1 - H ( X t ) / H max ) (4),

[0108] ​ m = ​ base + k 2*( H ( X t ) / H max ) (5),

[0109] in, H max is the theoretical maximum value of population distribution entropy, H max = ​ represents the entropy value when the distribution is completely uniform (i.e. the diversity is the highest), n is the sample size, d is the variable dimension; ​ base It is the benchmark value of crossover rate and mutation rate, generally ranging from 0.1 to 0.3; k 1 is the crossover rate control coefficient, ; k 2 is the variation rate control coefficient, ;

[0110] Mutation is an operation that randomly changes the values ​​of certain sample points to avoid falling into local optimality. It also increases the randomness of the search and helps the algorithm explore potential solutions more comprehensively in the search space.

[0111] whenH ( X t ) approaches 0, indicating that the sample aggregation is high and the diversity is insufficient, and the exploration ability needs to be enhanced. The crossover rate is automatically increased through formula (4) ​ c , automatically reduce the mutation rate through formula (5) ​ m To enhance population diversity;

[0112] when H ( X t ) approaches the theoretical maximum value H max = d log n When , it means that the diversity is high, and the crossover rate is reduced by formula (4) ​ c , increase the mutation rate through formula (5) ​ m To enhance local search intensity;

[0113] This embodiment monitors the sample distribution status in real time through sample entropy, and dynamically and adaptively adjusts the crossover rate and mutation rate based on the sample entropy according to formulas (4)-(5), thereby achieving a dynamic balance between exploration and development capabilities.

[0114] S33, according to formula (6) for the current sample X t Perform a directed crossover operation to obtain a crossover sample set X cross :

[0115] X cross = ArithmeticCrossover( X t , ​ c )(6),

[0116] in, is the cross sample set; ArithmeticCrossover is the directional cross function, corresponding to the CROSSOVER function of formula (2); ArithmeticCrossover function adopts the arithmetic cross strategy based on space segmentation, and selects X t The parent sample points in x p and x q Perform linear combination to generate offspring x child , so that it is located in the middle area of ​​the parent line, which is expressed as shown in formula (7):

[0117] (7),

[0118] This strategy maintains the Latin hypercube structure characteristics through the weight coefficient α =1- ​ c (Depend on ​ c Dynamic adjustment, when the population diversity decreases, ​ c Increase, possibly expand α The value range of is [0.3, 0.7] to enhance the exploration ability, and the random perturbation corresponding to formula (4) realizes the directional exploration of the design space; in this embodiment α The value range is [0.4,0.6].

[0119] It is understandable that X t The parent sample set in is X t-1， x p and x q is the parent sample set X t-1 Two sample points in ;

[0120] S34, according to formula (8) for the cross sample set X cross Perform the guided mutation operation to obtain the cross-mutation sample set X mut :

[0121] X mut = GuidedMutation( X cross , ​ m )(8),

[0122] Among them, GuidedMutation is the guided mutation function, corresponding to the MUTATE function of formula (2); the mutation operation is designed as a Gaussian perturbation model with directional guidance, which is implemented by the GuidedMutation function, as shown in formula (9):

[0123] (9),

[0124] in, represents the individual after mutation; x i Represents the current sample (current population) Xt The i Individuals (sample points) are the target individuals of the mutation operation; ​ is the step size factor, and a dynamic adjustment strategy is adopted; x elite is the current elite sample set, from the current population X t The individual with the best fitness is selected; Indicates that the mean is 0 and the covariance matrix is The multi-dimensional Gaussian distribution controls the randomness and amplitude of the disturbance; this mechanism effectively balances the needs of local development and global exploration.

[0125] Specifically, the step size factor is calculated by formula (10): ​ Make dynamic adjustments:

[0126] (10),

[0127] in, is the t-th generation step factor, that is, the step factor ​ Dynamic adjustment is performed according to the number of iterations t; ​ 0 is the initial step size, ​ 0=(0.1~0.2)*||ub-lb||2; ​ T is the termination step length, ​ T =(0.001~0.01)* ​ 0; ​ is the decay constant, which is determined by the maximum number of iterations T Decide.

[0128] Furthermore, the current elite sample set x elite Through non-dominated sorting (non-dominated sorting is a key technology in multi-objective optimization, used to hierarchically sort solutions according to their dominance relationship and construct the Pareto frontier) from the current population X t The top 10% of solutions are selected to guide the mutation direction.

[0129] S35, according to formula (11) for the current sample X t and cross-mutation sample set X mut Perform selection operations to generate next generation samples X t+1 ;

[0130] X t+1 = ParetoFront( Xt ∪ X mut )(11),

[0131] The ParetoFront function is the Pareto frontier function, a core concept in multi-objective optimization problems. It is also called the Pareto optimal frontier. Its essence is to describe the solution set where it is impossible to further optimize a certain objective without compromising other objectives under multi-objective conditions.

[0132] S36. Calculate the entropy change according to formula (12) ​ And increase the number of iterations t by 1 ， If the number of iterations t reaches the maximum or the entropy change ​ Greater than the convergence threshold ​ Output the current sample X t As a non-dominated solution set X opt ; Otherwise, return to step S31 and continue iteration; Formula (12) is as follows:

[0133] ​ = | H ( X t+1 ) - H ( X t )|(12).

[0134] S4, the continuous intelligent microreactor platform adopts a non-dominated solution set X opt The corresponding process parameters are used for chemical reaction to obtain the final product;

[0135] The TS-EMO algorithm is an efficient multi-objective optimization algorithm based on Thompson sampling. It is designed to solve global multi-objective optimization problems involving black-box functions with high evaluation costs. The algorithm uses intelligent sampling strategies to reduce unnecessary computation, improve optimization efficiency, and conduct extensive searches in the solution space to avoid being trapped in local optima.

[0136] The working principle of the TS-EMO algorithm includes the following steps:

[0137] Initialization: Randomly generate a set of initial solutions as the starting point of the algorithm;

[0138] Sampling: Generate new candidate solutions in the solution space using the Thompson sampling strategy;

[0139] Evaluation: Calculate the performance of the newly generated candidate solutions on various objective functions;

[0140] Selection: Select the optimal solution set from the candidate solutions based on a certain selection mechanism (such as Pareto dominance relation);

[0141] Iteration: Repeat the above steps until the termination condition is met (such as reaching the maximum number of iterations or the solution set no longer changes significantly).

[0142] This example provides a method for automatically optimizing a continuous intelligent microreactor platform. To address the challenges of traditional Latin Hypercube Design (LHD) in complex optimization problems, including insufficient sample distribution uniformity, limited global exploration capabilities, low computational efficiency leading to excessive computational resource consumption, prolonged reaction automatic optimization time, and excessive reagent costs, an improved TS-EMO algorithm based on an evolutionary multi-objective optimization framework is proposed. By introducing a dynamic parameter adjustment mechanism and a hybrid evolutionary operator, the algorithm improves the initial sample generation strategy and iterative optimization process. The maximum and minimum distance criterion is used to generate the initial Latin Hypercube (LHC) sample, maximize the minimum distance between sample points, and make the sample points evenly distributed in the input space, thereby reducing the sampling bias in the initial design, and improving the quality of the initial sample through a more efficient search strategy of taboo search, thereby increasing the probability of finding the global optimal solution and the efficiency of the algorithm, and reducing the consumption of computing resources; the crossover operation mixes the existing samples to generate new sample points, explores more design spaces, improves sample diversity, and helps the algorithm to more comprehensively explore potential solutions in the search space; mutation is an operation that randomly changes the values ​​of certain sample points to avoid falling into local optimality, while increasing the randomness of the search, and also helps the algorithm to more comprehensively explore potential solutions in the search space, effectively revealing the trade-off relationship between multiple objectives, meeting the flexible needs of actual process design, and obtaining the optimal solution (non-dominated solution set X opt ) The corresponding process parameters are more suitable for chemical reactions on a continuous intelligent microreactor platform, which enables the continuous intelligent microreactor platform to operate efficiently and stably and produce high-quality products, significantly reducing the reaction automatic optimization time and reagent consumption costs. It also effectively solves the problem that a large number of experimental evaluations (usually hundreds of iterations) are difficult to apply to actual scenarios with high experimental costs, and has robustness and engineering applicability.

[0143] Example 2

[0144] Based on the method of automatically optimizing the continuous intelligent microreactor platform described in Example 1, a ​ The continuous intelligent microreactor platform shown is used to conduct dual-objective collaborative automatic optimization research on the high-temperature chlorination reaction of propylene.

[0145] The continuous intelligent microreactor platform is a highly integrated and automated chemical reaction system. ​ middle:

[0146] The continuous intelligent microreactor platform includes chemical reaction raw materials, a continuous flow reactor, a solvent absorption device and a control system; the control system includes a TS-EMO monitoring system and a ReactIR real-time monitoring system;

[0147] Chlorine (Cl 2 ) and propane (C 3 H 6 ) as inputs, these two components, the main raw materials for the chemical reaction, are transported to the continuous flow reactor through a dedicated piping system. Chlorine is one of the reactants, and propane is the other. They mix and react in the reactor.

[0148] The continuous flow reactor is the core of the entire system, providing the environment for the chemical reaction. Here, chlorine and propane react under specific conditions (such as temperature and pressure) to produce the desired products. The design of the continuous flow reactor ensures efficient reaction and stable product output.

[0149] After the reaction products are generated in the continuous flow reactor, they are transported to the solvent absorption unit. This unit uses a specific solvent to absorb and separate the reaction products, removing impurities and waste, and improving the purity and quality of the products. The solvent absorption unit is a key step in product purification.

[0150] The control system (TS-EMO and ReactIR) is the "brain" of the entire continuous intelligent microreactor platform. It includes the TS-EMO monitoring system and the ReactIR real-time monitoring system. The TS-EMO monitoring system is responsible for monitoring and controlling reaction conditions such as temperature, pressure, and flow rate to ensure that the chemical reaction proceeds under optimal conditions.

[0151] The ReactIR real-time monitoring system is based on Fourier transform infrared spectroscopy (FT-IR) technology. By detecting changes in the vibration frequency of chemical bonds in a chemical reaction system, it can obtain real-time concentration information of reactants, intermediates and products, as well as the formation and breaking of chemical bonds during the reaction process, provide information on reaction kinetics and mechanisms, and provide data support for the optimization of reaction conditions.

[0152] These modules are closely connected through pipes and control systems to form a highly automated and integrated chemical reaction system.

[0153] The chemical reaction raw materials (chlorine and propane) are fed into the continuous flow reactor, and the TS-EMO monitoring system adopts the non-dominated solution set Xopt The corresponding process parameters are combined with the ReactIR real-time monitoring system to carry out a chemical reaction, and the product of the chemical reaction is purified through a solvent absorption device to obtain the final product.

[0154] The input of chlorine and propane provides the raw materials for the reaction. The continuous flow reactor is the primary site of the reaction. The solvent absorption device purifies the reaction products to obtain the final product. The control system uses the process parameters optimized by the method described in Example 1 to ensure the stability and safety of the entire reaction process. This design enables the continuous intelligent microreactor platform to operate efficiently and stably, produce high-quality products, and reduce computing resource consumption, reaction automatic optimization time, and reagent consumption costs.

[0155] When using the initial TS-EMO algorithm, the initial LHC sample size is n =32, 30 sets of non-dominated solutions were successfully collected through the adaptive evolution strategy to construct the initial Pareto frontier. ​ As shown in the figure, after 72 iterative experiments, the algorithm converged to a high-density Pareto frontier consisting of 40 non-dominated solutions (convergence threshold ε=1e-5), and its hypervolume index (Hypervolume) reached 0.852±0.015.

[0156] According to the improvement of TS-EMO algorithm described above, the cross-mutation optimization TS-EMO algorithm with dynamic parameter adjustment mechanism is used to reduce the initial LHC sample size to n =15. ​ As shown, the improved algorithm converged to a Pareto front consisting of 31 non-dominated solutions after 45 iterations, achieving a 37.5% convergence rate compared to the original algorithm. The improved frontier hypervolume index reached 0.849±0.014, showing no significant difference from the original algorithm, but with improved algorithm iteration efficiency. The results demonstrate that the improved parameter adjustment mechanism significantly reduces computational resource consumption while maintaining solution quality, validating the robustness and engineering applicability of the algorithm framework. This method effectively reduces sample size and the number of iterative experiments, significantly reducing computational resource consumption, reaction optimization time, and reagent costs.

[0157] Example 3

[0158] Based on the method of automatically optimizing the continuous intelligent microreactor platform described in Example 1, a ​ The continuous intelligent microreactor platform shown in the figure carries out dual-objective collaborative optimization analysis of the propylene low-temperature chlorination reaction system.

[0159] When using the initial TS-EMO algorithm, set the LHC initial sample size n=16, a high-density Pareto front containing 30 non-dominated solutions is constructed through a systematic sampling strategy. ​ As shown in the figure, after 56 iterative experiments, the algorithm converged to a Pareto frontier consisting of 21 non-dominated solutions (convergence threshold ε=1e-5), and its hypervolume index (Hypervolume) was 0.768±0.021.

[0160] Improved optimized TS-EMO algorithm reduces the initial LHC sample size to n =14. ​ As shown, after 44 iterations, the algorithm converged to a Pareto front consisting of 18 non-dominated solutions, achieving a 21.4% convergence rate compared to the initial algorithm. The improved frontier hypervolume index was 0.761±0.019, showing no statistical difference from the initial algorithm. These results demonstrate that the improved parameter optimization mechanism significantly improves algorithm efficiency while maintaining the quality of the Pareto solution set, validating its robustness in low-temperature reaction systems. This method effectively reduces the sample size and the number of iterative experiments, significantly reducing computational resource consumption, reaction optimization time, and reagent costs.

[0161] ​ In the horizontal axis f 1 represents objective function 1 (main product selectivity); the vertical axis f 2 Represents objective function 2 (the mass ratio of by-products to main products).

[0162] In summary, the results of the automatic optimization of the propylene chlorination reaction after the improvement of the TS-EMO algorithm show that the initial sample size is significantly reduced, and the optimization results also achieve good results. After the algorithm is optimized through crossover and mutation operations, the sample points generated by the initial LHC are further improved. The crossover operation helps to increase the diversity of the sample points, while the mutation operation increases the randomness of the sample points to a certain extent, which helps the algorithm to escape the local optimal solution and find the global optimal solution. In addition, the function added to the algorithm can apply crossover and mutation operations more frequently by increasing the number of iterations, improve the quality and diversity of the sample points, explore the solution space more deeply, and increase the probability of finding the global optimal solution and the efficiency of the algorithm. Therefore, the improved TS-EMO algorithm will be more suitable for the automatic optimization of continuous flow reactor platforms, which will significantly reduce the reaction automatic optimization time and reagent consumption costs.

[0163] The present invention effectively reduces the number of experiments in the automatic optimization process, thereby correspondingly reducing the energy consumption and chemical reagent consumption required for the experiment, reducing carbon emissions, reducing the emission of harmful substances and waste generation, reducing pollution to the environment, and promoting the development of green experiments.

[0164] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and it is intended that all changes that fall within the meaning and scope of equivalent elements are included in the present invention.

Claims

1. A method for automatically optimizing a continuous intelligent microreactor platform, characterized in that: The continuous intelligent microreactor platform includes chemical reaction raw materials, a continuous flow reactor, a solvent absorption device and a control system; the control system includes a TS-EMO monitoring system and a ReactIR real-time monitoring system, and includes the following steps: S1. Initialize the input parameters of the actual design space; the input parameters include sample size n, variable dimension d ,Design space boundary[ lb , ub ], maximum number of iterations T , convergence threshold ε ; S2, according to the sample size n and Variable dimensions d Obtain the initial design using the maximized minimum distance Latin hypercube design X , and mapped it to the actual design space as a sample matrix through reflection transformation X 0 ; S3, sample matrix X 0 Iteratively calculate the current sample entropy H ( X t ), build a cross-mutation co-evolution framework and according to the current sample entropy H ( X t ) to perform adaptive parameter adjustment and reach the maximum or entropy change at the number of iterations t ΔH Greater than the convergence threshold ε Output the current sample X t As a non-dominated solution set X opt ; Among them, the current sample X t express X 0 No. t Iterations, t <T; The crossover-mutation co-evolution framework is constructed based on formula (2): （2）， in, θ c is the crossover rate; θ m is the mutation rate; the CROSSOVER function represents the crossover operation; the MUTATE function represents the mutation operation; S4, the continuous intelligent microreactor platform adopts a non-dominated solution set X opt The corresponding process parameters are used for chemical reaction to obtain the final product; specifically including: The chemical reaction raw materials are input into the continuous flow reactor, and the TS-EMO monitoring system adopts the non-dominated solution set X opt The corresponding process parameters are combined with the ReactIR real-time monitoring system to carry out a chemical reaction, and the product of the chemical reaction is purified through a solvent absorption device to obtain the final product.

2. The method according to claim 1, characterized in that Step S2 specifically includes the following steps: S21, according to the sample size n and Variable dimensions d Generate candidate sample sets that satisfy Latin hypercube constraints X candidate =L(n,d); S22, based on the candidate sample set X candidate Solve the maximin optimization problem of formula (1) by tabu search and obtain the initial design X ; （1）， in, L(n,d) express n The sample points are d The set of Latin hypercubes in dimensional space; x i and x j All are sample points; for x i and x j The Euclidean distance between S23, through the affine transformation formula X 0 = X ⊗( ub - lb ) + lb Initial design X Mapping to the actual design space to obtain the sample matrix X 0 .

3. The method according to claim 1, characterized in that Step S3 specifically includes the following steps: S31. Calculate the current sample X t The current sample entropy of H ( X t ); X t Represents the sample matrix X 0 No. t iterations; S32, according to the current sample entropy H ( X t ) Based on formula (4)-(5) for crossover rate θ c and mutation rate θ m Make adjustments: θ c = θ base + k 1 *(1 - H ( X t ) / H max ) （4）， θ m = θ base + k 2*( H ( X t ) / H max ) （5）， in, H max is the theoretical maximum value of population distribution entropy, H max = d ·log n , n is the sample size, d is the variable dimension, representing the entropy value when it is completely uniformly distributed; θ base is the benchmark value of crossover rate and mutation rate; k 1 is the crossover rate control coefficient; k 2 is the variation rate control coefficient; S33, according to formula (6) for the current sample X t Perform a directed crossover operation to obtain a crossover sample set X cross : X cross = ArithmeticCrossover( X t , θ c )(6), Among them, ArithmeticCrossover is the directional crossover function, corresponding to the CROSSOVER function of formula (2); S34, according to formula (8) for the cross sample set X cross Perform the guided mutation operation to obtain the cross-mutation sample set X mut : X mut = GuidedMutation( X cross , θ m )(8), Among them, GuidedMutation is the guided mutation function, corresponding to the MUTATE function of formula (2); S35. For the current sample X t and cross-mutation sample set X mut Perform selection operations to generate next generation samples X t+1 ; S36. Calculate the entropy change according to formula (12) ΔH, And increase the number of iterations t by 1 ， If the number of iterations t reaches the maximum or the entropy change ΔH Greater than the convergence threshold ε Output the current sample X t As a non-dominated solution set X opt ; Otherwise, return to step S31 and continue iteration; Formula (12) is as follows: ΔH = | H ( X t+1 ) - H ( X t )|（12）。 4. The method according to claim 3, characterized in that Step S31 calculates the current sample according to formula (3) X t The current sample entropy of H ( X t ): （3）， in, x ij Refers to the current sample X t No. i Individuals in j The value of the dimension; p ( x ij ) indicates the current sample X t In the individual i The probability that a value in dimension falls into the jth dimension.

5. The method according to claim 3, characterized in that ,Step S33ArithmeticCrossover function adopts arithmetic crossover strategy based on space segmentation, selects the current sample X t The parent sample points in x p and x q Perform linear combination to generate offspring x child , so that it is located in the middle area of ​​the parent line, which is expressed as shown in formula (7): （7）， in, α is the weight coefficient.

6. The method according to claim 3, characterized in that The GuidedMutation function in step S34 is implemented based on formula (9): （9）， in, represents the individual after mutation; x i Indicates the current sample X t The i individual; δ is the step size factor; x elite For the current elite sample set; Indicates that the mean is 0 and the covariance matrix is Multidimensional Gaussian distribution.

7. The method according to claim 3, characterized in that Step S35 calculates the current sample according to formula (11): X t and cross-mutation sample set X mut Make a selection: X t+1 = ParetoFront( X t ∪ X mut )(11), Among them, ParetoFront function is the Pareto front surface function.

8. The method according to claim 5, characterized in that The current elite sample set in formula (7) x elite To sort from the current population by non-dominated sorting X t Select the top 10% of solutions.

9. The method according to claim 5, characterized in that The weight coefficient α =1- θ c .

10. The method according to claim 6, characterized in that The step size factor is calculated by formula (10): δ Make dynamic adjustments: （10）， in, is the step size factor of the tth generation; δ 0 is the initial step size; δ T is the termination step length; τ is the decay constant.

Images