Chemical reaction condition closed-loop optimization method based on interpretable Bayesian optimization algorithm

Abstract

The invention provides a chemical reaction condition closed-loop optimization method based on an interpretable Bayesian optimization algorithm. The method comprises the following steps: firstly, collecting initial experimental data, and carrying out noise reduction and distribution mapping preprocessing on a reaction independent variable through quantile conversion; then constructing a symbol regression model, accurately fitting a complex nonlinear relationship between a reaction condition and an experimental result, and evaluating uncertainty; reconstructing a symbolic regression model by using a probability agent model, defining an acquisition function, performing optimization in a parameter space by using L-BFGS-B, differential evolution and other optimizers, balancing local development and global exploration, and intelligently recommending a next group of experimental conditions for maximizing target benefits; and finally, real experiment feedback is introduced into the model for closed-loop iteration. The embodiment of the invention shows that the method can obviously improve the product yield in the scenes of photocatalytic synthesis of hydrogen peroxide, flavone and the like by using extremely few experimental batches.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and chemical engineering, and proposes a closed-loop optimization method for chemical reaction conditions based on an interpretable Bayesian optimization algorithm. Background Technology

[0002] In fields such as chemical synthesis, materials preparation, and bioengineering, optimizing reaction conditions is crucial for improving product yield and selectivity. Traditional experimental optimization mainly employs the "one-factor experiment" (OFAT) or the "design of experiments" (DoE) based on the coordinated response surface methodology (RSM). However, the OFAT method neglects the potential interactions between multiple variables and requires significant time and reagent consumption; while RSM, when dealing with highly nonlinear relationships in complex reaction systems, is often limited by simple polynomial models, resulting in insufficient fitting ability and prediction accuracy.

[0003] To further improve experimental design efficiency, machine learning methods are increasingly being introduced into response condition optimization. Among these, Bayesian optimization (BO) has attracted significant attention due to its superior performance with small sample sizes and high evaluation costs. Traditional methods often get stuck in local optima or require massive amounts of experimentation when exploring high-dimensional variables. Bayesian optimization, however, can effectively capture the complex nonlinear relationship between experimental conditions and response performance using probabilistic surrogate models (such as Gaussian process regression), balancing the "exploration" of unknown regions with the "utilization" of high-yield regions, thereby quickly identifying the optimal response conditions within a very small number of experimental batches. However, Gaussian process regression in Bayesian optimization suffers from poor interpretability. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a closed-loop optimization method for chemical reaction conditions based on an interpretable Bayesian optimization algorithm. By introducing quantile transformation preprocessing, a symbolic regression algorithm, and an efficient acquisition function optimization mechanism, this invention significantly reduces the number of experimental trials and the research and development costs.

[0005] The specific technical solution adopted in this invention includes the following steps:

[0006] Step 1: Collect initial experimental sample data using OFAT or DoE. The experimental data includes independent variables of reaction conditions (such as reaction time, temperature, catalyst concentration, etc.) and corresponding dependent variables of experimental results (such as yield, conversion rate). The collected independent variable data are then preprocessed using the quantile transformation method.

[0007] Step 2: Train and fit the preprocessed response conditional independent variables and corresponding dependent variable labels into a Kolmogorov-Arnold network (KAN). Use the high-precision prediction space (or explicit mathematical expression) output by the trained KAN model as a benchmark and define it as the "target function" for the subsequent Bayesian optimization process.

[0008] Step 3: For the KAN objective function constructed in Step 2, establish a Gaussian Process (GP) regression model for infinite approximation fitting. Based on the predicted mean and uncertainty output by the surrogate model, define an acquisition function, balancing the exploration of high uncertainty regions with the utilization of high prediction value regions. Use optimizers such as L-BFGS-B or Differential Evolution to globally maximize the acquisition function, outputting the next set of recommended response conditions that maximize the utility of the acquisition function (Next try).

[0009] Step 4: Perform a reverse quantile transformation on the recommended reaction conditions output in Step 3 to restore the actual physical operating space. Perform a real chemical synthesis experiment based on these recommended conditions to obtain the actual dependent variable results. Add the newly obtained experimental conditions and results to the original dataset, update the dataset, and return to Steps 2 and 3 to retrain the model and recommend new experimental conditions. This closed-loop process iterates continuously until the experimental results meet the preset yield / yield target or the maximum number of iterations is reached.

[0010] The beneficial effects of this invention are:

[0011] Breaking through the bottleneck of small-sample nonlinear fitting and providing extremely high prediction accuracy: This invention utilizes the KAN model to break through the bottleneck of complex nonlinear fitting of small samples, and constructs an extremely high-precision global objective function base for the optimization process.

[0012] Significantly reduces experimental trial-and-error costs and achieves efficient global optimization: This invention uses a Bayesian surrogate model to infinitely approximate and globally optimize the objective function, which effectively avoids getting trapped in local optima while significantly reducing experimental trial-and-error costs.

[0013] Significantly improves actual reaction yield and synthesis efficiency: The closed-loop optimization mechanism of this invention can accurately lock the optimal reaction conditions with very few experimental iterations, thereby significantly improving the final yield of the target product in complex chemical synthesis. Attached Figure Description

[0014] Figure 1This is a schematic diagram of OFAT and DoE data collection and quantile conversion preprocessing provided by the present invention.

[0015] Figure 2 This diagram illustrates the optimization of the acquisition function and the recommendation of the next set of experimental conditions (Next best guess) for this invention. In this diagram, a is the Bayesian optimized surrogate model, and b is the Bayesian optimized acquisition function.

[0016] Figure 3 The diagram shows a comparison of the Bayesian optimization iteration effects in Implementation Case 1 (hydrogen peroxide synthesis optimization) of the present invention. In the diagram, a is a comparison of the models in the first iteration, b is a comparison of the results in the first iteration, c is a comparison of the models in the second iteration, and d is a comparison of the results in the second iteration.

[0017] Figure 4 The diagram shows a comparison of the Bayesian optimization iteration effects in Implementation Case 2 (flavonoid synthesis optimization) of the present invention. In the diagram, a is the model comparison diagram in the first iteration, b is the result comparison diagram in the first iteration, c is the model comparison diagram in the second iteration, d is the result comparison diagram in the second iteration, e is the model comparison diagram in the third iteration, and f is the result comparison diagram in the third iteration.

[0018] Figure 5 This is a flowchart of the present invention. Detailed Implementation

[0019] A closed-loop optimization method for chemical reaction conditions based on an interpretable Bayesian optimization algorithm, characterized by the following steps:

[0020] Step 1: Collect an initial dataset containing the independent variables of the reaction conditions and the dependent variables of the experimental results, and preprocess the data of the independent variables;

[0021] Step 2: Construct a Kolmogorov-Arnold network (KAN) model, train the KAN model using the preprocessed dataset, fit the nonlinear mapping relationship between the response conditions and the experimental results with high accuracy, and define the output space of the trained KAN model as the objective function of Bayesian optimization.

[0022] Step 3: Construct a probabilistic surrogate model and use it to infinitely approximate the KAN objective function in Step 2, outputting the predicted mean and prediction uncertainty; based on this, define the acquisition function, and within the set independent variable constraint space, use an optimizer to perform global maximization optimization on the acquisition function, outputting the next set of recommended optimal experimental response conditions.

[0023] Step 4: Conduct actual verification experiments based on the reaction conditions recommended in Step 3, obtain real results and update them to the initial dataset, and repeat Steps 2 to 3 in a closed loop until the preset optimization goal is met or the maximum number of iterations is reached.

[0024] The preprocessing described in step 1 specifically employs the quantile transformation method, which maps the independent variable data to a uniform or standard normal distribution interval to eliminate the influence of outliers and unify the scale of multi-dimensional features.

[0025] The KAN model described in step 2 is mainly used to perform continuous, high-precision nonlinear base fitting on sparse discrete real experimental data, replacing the traditional response surface model (RSM) as the objective function benchmark.

[0026] The probabilistic proxy model mentioned in step 3 is a Gaussian process regression model, which is used to perform a probabilistic infinite approximation fit on the deterministic target space characterized by the KAN model, thereby quantifying the uncertainty boundary of the search space.

[0027] The optimizer described in step 3 includes the finite-memory Broyden-Fletcher-Goldfarb-Shanno algorithm (L-BFGS-B) with boundary constraints and the Differential Evolution algorithm, which are used to escape local optima in the multidimensional reaction condition space to find the maximum value of the acquisition function.

[0028] The acquisition function aims to balance the exploration of regions with high uncertainty in response parameters with the exploitation of regions with high predictive returns obtained by infinite approximation of surrogate models.

[0029] The method is applied to the optimization of process parameters in the fields of chemical engineering and materials science, including but not limited to the optimization of conditions for the photocatalytic synthesis of hydrogen peroxide and the optimization of conditions for the photocatalytic synthesis of flavonoids.

[0030] Tab. 1 Data collection for hydrogen peroxide synthesis example and comparison of Bayesian optimization recommendations with experimental values.

[0031]

[0032] After two rounds of closed-loop iteration of "KAN fitting real data - BO surrogate model infinite approximation", the actual yield of H2O2 increased significantly from the initial highest of 626 mmol / L to 1006 mmol / L, achieving a significant improvement of 60.7%.

[0033] Table 2. Data collection for flavonoid synthesis examples and comparison of Bayesian optimization recommendations with experimental values.

[0034]

[0035] After three rounds of iteration, the flavonoid yield increased from the initial 91.1% to 95.8% (an absolute increase of nearly 5%), demonstrating strong generalization ability and optimization accuracy.

Claims

1. A closed-loop optimization method for chemical reaction conditions based on an interpretable Bayesian optimization algorithm, characterized in that, Includes the following steps: Step 1: Collect an initial dataset containing the independent variables of the reaction conditions and the dependent variables of the experimental results, and preprocess the data of the independent variables; Step 2: Construct the Kolmogorov-Arnold network model, train the KAN model using the preprocessed dataset, fit the nonlinear mapping relationship between the response conditions and the experimental results with high precision, and define the output space of the trained KAN model as the objective function of Bayesian optimization. Step 3: Construct a probabilistic surrogate model and use it to infinitely approximate the KAN objective function in Step 2, outputting the predicted mean and prediction uncertainty; based on this, define the acquisition function, and within the set independent variable constraint space, use an optimizer to perform global maximization optimization on the acquisition function, outputting the next set of recommended optimal experimental response conditions. Step 4: Conduct actual verification experiments based on the reaction conditions recommended in Step 3, obtain real results and update them to the initial dataset, and repeat Steps 2 to 3 in a closed loop until the preset optimization goal is met or the maximum number of iterations is reached.

2. The method according to claim 1, characterized in that, The preprocessing described in step 1 specifically employs the quantile transformation method to map the independent variable data to a uniform or standard normal distribution interval, thereby eliminating the influence of outliers and unifying the scale of multi-dimensional features.

3. The method according to claim 1, characterized in that, The KAN model described in step 2 is mainly used to perform continuous, high-precision nonlinear base fitting on sparse discrete real experimental data, replacing the traditional response surface model as the objective function benchmark.

4. The method according to claim 1, characterized in that, The probabilistic surrogate model mentioned in step 3 is a Gaussian process regression model, which is used to perform a probabilistic infinite approximation fit on the deterministic target space characterized by the KAN model, thereby quantifying the uncertainty boundary of the search space.

5. The method according to claim 1, characterized in that, The optimizer described in step 3 includes the finite-memory Broyden-Fletcher-Goldfarb-Shanno algorithm with boundary constraints and the differential evolution algorithm, which are used to escape local optima in the multidimensional reaction condition space to find the maximum value of the acquisition function.

6. The method according to claim 1, characterized in that, The acquisition function aims to balance the exploration of regions with high uncertainty in response parameters with the utilization of regions with high predictive returns obtained by infinitely approximating the surrogate model.

7. The application of the method according to any one of claims 1 to 6, characterized in that, The method is applied to the optimization of process parameters in the fields of chemical engineering and materials science, including but not limited to the optimization of conditions for the photocatalytic synthesis of hydrogen peroxide and the optimization of conditions for the photocatalytic synthesis of flavonoids.

Images