Method for optimizing synthesis path of polycaprolactone polyol based on genetic algorithm

Abstract

The application relates to the technical field of data processing, in particular to a polycaprolactone polyol synthesis path optimization method based on a genetic algorithm, which comprises the following steps: constructing an adversarial framework with the genetic algorithm as a generator and a process stability discriminator; using a failure path deduction model to generate virtual failure data in a directional mode to solve a data sparsity problem and train the discriminator; quantifying the robustness risk of candidate process parameters through a virtual disturbance unit and taking the risk as a key part of a genetic algorithm fitness function; and simultaneously using the explainability analysis result of the discriminator to guide the mutation direction of the genetic algorithm. Through the above method, a process path with high performance and high robustness can be found, and the industrial practicability and decision efficiency of the optimization result are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a polycaprolactone polyol synthesis path optimization method based on a genetic algorithm. BACKGROUND

[0002] Heuristic search models such as genetic algorithms are key computing tools for solving complex engineering optimization problems. For example, in the synthesis path optimization of high-performance materials such as polycaprolactone polyols, how to apply these computing models to deal with real-world systems with uncertainty and sparse failure data, and ensure the robustness and practicality of the optimization results, is a core challenge in the field of artificial intelligence.

[0003] Existing computing methods usually couple genetic algorithms with physical models or data-driven surrogate models. The former is inefficient due to high evaluation cost. The latter generally uses models such as artificial neural networks as surrogates, but as a black box model, the prediction logic is difficult to explain, and its training effect is highly dependent on data quality; in industrial applications such as polycaprolactone synthesis, failure condition data representing unqualified product quality is extremely sparse, resulting in poor model generalization ability and inability to accurately assess risks. Although existing research has attempted to use generative adversarial networks to enhance data, conventional generative adversarial networks have problems such as unstable training and pattern collapse when dealing with high-dimensional and complex process data, making it difficult to generate high-fidelity virtual samples that meet physical laws. More importantly, regardless of the coupling method, existing technologies mostly find isolated optimal points, lack mechanisms to directly quantify and optimize the robustness of solutions to parameter perturbations at the algorithm level, and lack feedback loops that use internal insights from the model to intelligently guide evolutionary search.

[0004] Therefore, there is an urgent need in the art for a new computing optimization method to solve the core algorithmic challenges encountered in the optimization of complex processes such as polycaprolactone synthesis: how to create high-quality virtual data through a failure path deduction model to train an accurate process stability discriminator when negative sample data is sparse; and how to directly quantify and optimize the robustness of solutions to parameter perturbations in evolutionary search.

[0005] To this end, a polycaprolactone polyol synthesis path optimization method based on a genetic algorithm is proposed. SUMMARY

[0006] The present application aims to provide a polycaprolactone polyol synthesis path optimization method based on a genetic algorithm, to solve the technical problem that the production decision-making has great uncertainty due to the sparse production failure case data in the prior art, which cannot construct an accurate risk prediction model. The present application constructs an adversarial optimization framework containing a generator and a discriminator, generates high-quality virtual failure data using a failure path deduction model to enhance the training of the process stability discriminator, and combines the robustness risk of the virtual disturbance unit quantifying the process parameters under real fluctuations to finally include the risk index into the fitness function of the genetic algorithm for iterative optimization, thereby providing the manager with a production strategy with the highest business return on investment under the premise of balancing product performance and production stability.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] The polycaprolactone polyol synthesis path optimization method based on a genetic algorithm comprises a genetic algorithm as a generator and a process stability discriminator.

[0009] The failure path deduction model is trained using historical synthesis data to learn the conditional probability distribution of synthesis parameters composed of reaction temperature, reaction time, monomer to initiator molar ratio, and catalyst to initiator molar ratio under the condition of a given failure mode label; and virtual failure data is generated by inputting the failure mode label into the model.

[0010] For the candidate synthesis parameters generated by the genetic algorithm, parameter clouds are generated in the neighborhood of the candidate synthesis parameters by a virtual disturbance unit; the failure risk values of the disturbed parameter points in the parameter clouds are evaluated by the process stability discriminator to construct a strategy volatility risk matrix; and the strategy volatility risk matrix is aggregated to generate a robustness risk index representing the overall failure risk of the candidate synthesis parameters.

[0011] The robustness risk index is input into the fitness function of the genetic algorithm, and the results of the explainability analysis of the process stability discriminator are used to guide the mutation operator of the genetic algorithm to accelerate the convergence of the optimization process to the robustness optimal solution.

[0012] Preferably, the training step of the process stability discriminator comprises:

[0013] The historical synthesis data and the virtual failure data are merged to form an enhanced training set together.

[0014] The process stability discriminator is supervised learning trained by using the enhanced training set, so that the discriminator can input a set of synthesis parameters composed of reaction temperature, reaction time, monomer to initiator molar ratio and catalyst to initiator molar ratio, and output the corresponding failure risk probability.

[0015] Preferably, the explainability analysis result is the quantification of the positive and / or negative contribution of each candidate synthesis parameter to the overall failure risk by applying the SHAP analysis method to the process stability discriminator; and the mutation operator specifically dynamically adjusts the mutation probability and mutation direction of the corresponding parameter gene according to the contribution, applies reverse mutation to the parameters contributing to the positive risk, and increases the mutation probability of the parameters.

[0016] Preferably, the failure path inference model comprises:

[0017] a forward diffusion unit that gradually adds Gaussian noise to the real synthesis parameters in the historical synthesis data through a Markov chain until the distribution is standard normal distribution;

[0018] a conditional denoising unit, which is a trained neural network, receives the noisy synthesis parameters at any time step of the forward diffusion process and the failure mode label as common input, and predicts the noise added to the noisy synthesis parameters;

[0019] and a sampling generation unit for sampling an initial noise from a standard normal distribution and iteratively calling the conditional denoising unit step by step to denoise under the guidance of the failure mode label to reversely reconstruct the virtual failure data.

[0020] Preferably, the virtual disturbance unit is a parameterized noise generation module:

[0021] For each parameter in the candidate synthesis parameters, multiple random samplings are independently performed from a probability distribution according to a process tolerance range matched with the physical characteristics of the parameter to generate the parameter cloud, wherein the process tolerance range is set based on statistical fluctuations of historical production data and / or expert experience.

[0022] Preferably, the generation of the robust risk indicator representing the overall failure risk of the candidate synthesis parameters comprises:

[0023] receiving a single candidate synthesis parameter generated by the genetic algorithm; the virtual perturbation unit generates a parameter cloud containing multiple perturbed parameter points around the single candidate synthesis parameter through multiple random samplings; the failure risk values corresponding to the perturbed parameter points in the parameter cloud are constructed into a strategy volatility risk matrix; and the strategy volatility risk matrix is aggregated to generate a single robustness risk indicator representing the overall failure risk of the candidate synthesis parameter.

[0024] Preferably, the robustness risk indicator is input into the fitness function of the genetic algorithm, specifically to construct a composite fitness function that combines the robustness risk indicator as a negative weighting item with a positive weighting item based on an expected product performance indicator of the candidate synthesis parameter; the mutation operator of the genetic algorithm is guided, specifically the key risk parameters that have a positive contribution to the overall failure risk are determined using the explainability analysis results, and when performing the mutation operation, the probability of the key risk parameters being selected for mutation is increased, and a directional mutation bias that reduces the key risk parameters is applied.

[0025] Preferably, the aggregation processing is specifically to select the maximum value and / or a preset high percentile value from the failure risk values of all perturbed parameter points in the parameter cloud as the robustness risk indicator to evaluate the worst possible performance under process fluctuations.

[0026] Preferably, the expected product performance indicator is calculated by a proxy model trained on the historical synthesis data for predicting product performance, which receives the candidate synthesis parameter as input and outputs performance prediction values including expected molecular weight and expected polydispersity index.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] 1. The technical problem of being difficult to construct an accurate risk model due to sparse failure case data in industrial production is solved. The present application generates high-quality and diversified virtual failure data using the failure path deduction model, significantly enhances the training set of the process stability discriminator, enables it to accurately predict the potential failure risk of different process paths, and provides reliable data support for production decisions.

[0029] 2. The industrial practicability and robustness of the process optimization result are significantly improved. The present application simulates the process fluctuations in real production through the virtual perturbation unit and incorporates the quantified robustness risk into the optimization objective, so that the final found process parameters not only have the best performance in theory, but also have a wider process window, can effectively resist disturbances in actual production, thereby reducing the batch failure rate and improving the stability of product quality.

[0030] 3. A more efficient and intelligent optimization process is achieved. By introducing explainable analysis to guide the mutation direction of the genetic algorithm, the present application changes the optimization process from a "black box" search to a "white box" evolution with clear guidance and the ability to learn from mistakes, avoiding a large amount of invalid exploration, accelerating the convergence speed to the robust optimal solution, and improving the decision-making efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0031] Fig. 1 The construction flowchart of the polycaprolactone polyol synthesis path optimization method based on a genetic algorithm proposed for the embodiments of the present application is shown in the figure.

[0032] Fig. 2 The method flowchart of the polycaprolactone polyol synthesis path optimization method based on a genetic algorithm proposed for the embodiments of the present application is shown in the figure.

[0033] Fig. 3 The model structure diagram of the failure path deduction model proposed for the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0035] Embodiment one

[0036] Please refer to Figs. 1 to 3 The present application provides a polycaprolactone polyol synthesis path optimization method based on a genetic algorithm, and the technical solutions are as follows:

[0037] The polycaprolactone polyol synthesis path optimization method based on a genetic algorithm, as shown in Fig. 1 and Fig. 2 includes a genetic algorithm as a generator and a process stability discriminator;

[0038] The failure path deduction model is trained using historical synthesis data to learn the conditional probability distribution of the synthesis parameters consisting of reaction temperature, reaction time, monomer to initiator molar ratio, and catalyst to initiator molar ratio under the condition of a given failure mode label; and by inputting the failure mode label to the model, virtual failure data is generated in a targeted manner;

[0039] For a candidate synthetic parameter generated by the genetic algorithm, a parameter cloud is generated in the neighborhood of the candidate synthetic parameter by a virtual perturbation unit; the failure risk values of the perturbed parameter points in the parameter cloud are evaluated by the process stability discriminator to construct a strategy volatility risk matrix; and the strategy volatility risk matrix is aggregated to generate a robustness risk indicator representing the overall failure risk of the candidate synthetic parameter;

[0040] The robustness risk indicator is input into the fitness function of the genetic algorithm, and the results of the explainability analysis of the process stability discriminator are used to guide the mutation operator of the genetic algorithm to accelerate the convergence of the optimization process to the robustness optimal solution.

[0041] Further, the optimization process adopts a two-stage hierarchical optimization strategy: the first stage is a coarse-grained performance optimization stage, and the genetic algorithm uses a simplified fitness function containing only the expected product performance indicators for iteration to quickly locate several high-performance regions in the entire parameter space and obtain an elite population composed of candidate synthetic parameters with excellent performance; the second stage is a fine-grained robustness tuning stage, and the elite population generated in the first stage is used as the initial population, and the iteration optimization is switched to a composite fitness function containing the robustness risk indicator, so as to concentrate the computing resources in the determined high-performance region for fine robustness optimization.

[0042] This hierarchical strategy significantly improves the overall optimization efficiency. It avoids time-consuming robustness risk evaluation of a large number of low-performance solutions in the early optimization stage, but concentrates valuable computing resources in the most promising candidate region, thereby significantly shortening the total time required to find the optimal solution and reducing the computing cost while ensuring the quality of the final solution.

[0043] Further, the training step of the process stability discriminator includes:

[0044] The historical synthesis data and the virtual failure data are merged to form an enhanced training set;

[0045] The enhanced training set is used to train the process stability discriminator through supervised learning, so that the discriminator can input a set of synthesis parameters composed of reaction temperature, reaction time, monomer to initiator molar ratio, and catalyst to initiator molar ratio, and output the corresponding failure risk probability.

[0046] By combining a small amount of real historical data with a large amount of generated virtual failure data, the key technical problem of being unable to train an effective risk model due to the sparsity of failure cases and the lack of negative samples in industrial applications is effectively solved. The balanced enhanced training set constructed significantly improves the prediction accuracy and generalization ability of the process stability discriminator, enabling it to accurately identify potential process risks and providing a solid and reliable decision-making basis for the subsequent robustness optimization process.

[0047] Further, the explainability analysis result is obtained by applying a SHAP analysis method to the process stability discriminator to quantify the positive and / or negative contribution of each parameter in the candidate synthetic parameters to the overall failure risk; and the mutation operator dynamically adjusts the mutation probability and mutation direction of the corresponding parameter gene according to the contribution, applies reverse mutation to the parameters contributing positively to the risk, and increases the mutation probability of the parameters.

[0048] For each parameter in the candidate synthetic parameters, the probability of being selected for mutation is dynamically calculated. First, the absolute value of the risk contribution of each parameter is calculated, and the absolute values of the risk contributions of all parameters are added to obtain a total sum. Then, the relative risk importance of a single parameter is obtained by dividing the absolute value of the risk contribution of the parameter by the total sum. Finally, the probability of the parameter being selected for mutation is equal to a preset base probability plus the product of the relative risk importance of the parameter and an influence factor. The greater the risk impact of a parameter, the greater the likelihood of being selected for modification, and all parameters have a basic mutation opportunity.

[0049] When a parameter is selected for mutation according to its probability, the new value is calculated as follows: First, determine whether the risk contribution of the parameter is positive or negative. Then, calculate the mutation amplitude. The amplitude is equal to a preset "base step" multiplied by "the absolute value of the risk contribution of the parameter plus one". The greater the risk contribution of a parameter, the greater the adjustment amplitude. Finally, update the parameter value according to the risk direction determined in the first step. If the risk contribution is positive, subtract the calculated mutation amplitude from the original parameter value; if the risk contribution is negative, add the calculated mutation amplitude to the original parameter value. At the same time, boundary checking needs to be performed on the calculated new parameter value to ensure that its value is within the preset reasonable process range.

[0050] The genetic algorithm is transformed from "blind trial and error" to "intelligent optimization". Through SHAP analysis, the contribution of each parameter to the failure risk is accurately located, and the algorithm can preferentially and directionally correct the parameter with the highest risk. This strategy not only significantly speeds up the search for the optimal solution, improving decision-making efficiency, but more importantly, it can find a process path with high performance and high robustness. The final solution can better resist process fluctuations in actual production, thereby reducing the failure rate and improving industrial applicability.

[0051] Further, as shown in Fig. 3 The failure path inference model comprises:

[0052] a forward diffusion unit that gradually adds Gaussian noise to the real synthetic parameters in the historical synthetic data through a Markov chain until their distribution is a standard normal distribution;

[0053] a conditional denoising unit, which is a trained neural network that receives the noisy synthetic parameters at any time step of the forward diffusion process and the failure mode label as common inputs, and predicts the noise added to the noisy synthetic parameters;

[0054] and a sampling generation unit for sampling an initial noise from a standard normal distribution and gradually denoising it under the guidance of the failure mode label by iteratively calling the conditional denoising unit to reversely reconstruct the virtual failure data.

[0055] The conditional denoising unit is a neural network based on the U-Net architecture. The network includes an encoder path for downsampling, a bottleneck layer, and a decoder path for upsampling that receives information from the encoder using skip connections. In the residual blocks of each path, self-attention modules can also be selectively added to capture the internal dependencies between parameters.

[0056] The specific structure of the neural network based on the U-Net architecture is as follows: both the encoder and decoder paths of the network include 4 levels. The number of channels of the convolutional blocks of the encoder path is [64, 128, 256, 512] in turn, each convolutional block includes two convolutional layers (convolution kernel size is 3x1, padding is 1), followed by a SiLU activation function and a batch normalization layer. A convolution with a step of 2 is used for downsampling at the end of each level. The decoder path is symmetrical to the encoder and uses transposed convolution for upsampling. The number of channels of the bottleneck layer is 1024. The self-attention module is added in the 3rd level, using the standard scaled dot-product attention mechanism.

[0057] The forward diffusion unit adopts a linear noise schedule scheme. In a Markov chain with a total of T (for example, T = 1000) time steps, the variance of the Gaussian noise added at the t-th step is linearly increased from a small initial value (for example, 0.0001) to a large end value (for example, 0.02).

[0058] The failure mode label is converted into a conditional embedding vector through an embedding layer. The vector is added to the time embedding of the current time step t, and then processed through a small feedforward network, and then applied as a bias term or scaling term after the normalization layer of each residual block in the U-Net architecture, thereby guiding the denoising process at each step.

[0059] The model can generate virtual failure data for supplementing the training set with high quality and controllability, to solve the problem of sparsity of real failure cases in industrial applications. The use of the U-Net architecture ensures that the generated virtual data is highly similar in distribution to the real process parameters, ensuring the authenticity of the data. The ingenious label embedding guidance mechanism enables the model to generate failure data of specific types according to specific requirements, achieving controllability. This high-quality and controllable virtual data can significantly enhance the training effect of the downstream process stability discriminator, making its prediction more accurate and reliable.

[0060] Further, the genetic algorithm of the present application adopts an elite reservation strategy combined with tournament selection, wherein the elite proportion is 0.1 and the tournament size is 3. The crossover operation adopts simulated binary crossover, and the crossover probability is set to 0.9; the mutation operation adopts the directional mutation operator guided by the explainability analysis result.

[0061] The population size of the genetic algorithm is set to 120, and the iteration number is 250 generations. Each individual represents a candidate poly (caprolactone) polyol synthesis path, and the parameters are optimized within the following preset ranges: reaction temperature is [120.0, 160.0] ℃, residence time is [30, 120] minutes, monomer to initiator molar ratio is [50, 200], and catalyst to initiator molar ratio is [0.001, 0.01].

[0062] Further, the virtual disturbance unit is a parameterized noise generation module:

[0063] For each of the candidate synthesis parameters, a plurality of random samplings are independently performed from a probability distribution according to a process tolerance range matching the physical characteristic of the parameter, to generate the parameter cloud, wherein the process tolerance range is set based on statistical fluctuation of historical production data and / or expert experience.

[0064] The probability distribution is specifically a truncated normal distribution. The mean of the distribution is the value of the current candidate parameter, and the standard deviation is set according to the process tolerance range, and the truncation boundary is the upper and lower limits of the process tolerance range. This can not only simulate the characteristics that fluctuations in the real process are more frequent around the center value, but also ensure that the sampling points will not exceed the range allowed by physics or experience.

[0065] The process tolerance range is quantitatively set in the following manner: for parameters with sufficient historical data, the range is set as plus or minus three times the standard deviation of the historical data mean; for parameters lacking data or specified by experts, the operating upper and lower limits given by the experts are directly used as their tolerance range.

[0066] The plurality of random samplings are specifically to generate a preset number N of parameter points, which collectively constitute the parameter cloud. The number N is determined after balancing between ensuring the stability of the evaluation results and the computational efficiency. In this embodiment, N = 100.

[0067] By using the tolerance range set based on historical data statistics and using the truncated normal distribution for sampling, the generated parameter cloud can accurately reflect the fluctuation characteristics of various parameters in the real world. This makes the robustness evaluation results of the candidate process scheme more accurate and reliable.

[0068] Further, the disturbance amplitude applied by the virtual disturbance unit is dynamically self-adaptively adjusted: in the early stage of the optimization process, a larger disturbance amplitude is used to quickly eliminate candidate synthesis parameters that are extremely sensitive to process fluctuations; in the middle and later stages of the optimization process, the average fitness change rate of the genetic algorithm population is monitored, and when the change rate is lower than a preset convergence threshold, the disturbance amplitude is automatically reduced to perform more fine local robustness search within the neighborhood of the current optimal solution; when the change rate is lower than the convergence threshold for a plurality of generations, the disturbance amplitude is temporarily increased to stress test the current optimal solution and assist the algorithm to jump out of the local optimum.

[0069] This self-adaptive adjustment mechanism realizes the intelligent balance between global exploration and local optimization of the algorithm; the large disturbance in the early stage ensures the breadth of the search, which can effectively avoid falling into the local optimum; the fine adjustment in the later stage ensures the depth of the search, which can more accurately depict the robustness boundary of the optimal solution; this dynamic strategy makes the optimization process more intelligent and efficient.

[0070] Further, the generation process of the robustness risk indicator comprises:

[0071] receiving a single candidate synthetic parameter generated by the genetic algorithm; the virtual perturbation unit generates a parameter cloud containing multiple perturbed parameter points by multiple random samplings around the single candidate synthetic parameter; the failure risk values corresponding to the perturbed parameter points in the parameter cloud are constructed into a strategy volatility risk matrix; and the strategy volatility risk matrix is aggregated to generate a single robustness risk indicator representing the overall failure risk of the candidate synthetic parameter.

[0072] The aggregation process is specifically selecting the maximum value and / or a preset high percentile value (for example, the 95th percentile value) of the failure risk values of all perturbed parameter points in the parameter cloud as the single robustness risk indicator, which is used to evaluate the worst possible performance under process fluctuations.

[0073] By selecting the maximum risk value or the high percentile risk value in the parameter cloud as the final indicator, it no longer focuses on the average performance, but focuses on the "worst possible situation" under real production fluctuations. This evaluation method is more in line with the core demand of "safety first, stability first" in industrial production. Therefore, the process parameters optimized for this purpose will naturally have stronger anti-interference ability and higher production reliability, which can effectively reduce the risk of batch failure.

[0074] Further, the robustness risk indicator is input into the fitness function of the genetic algorithm, specifically a composite fitness function is constructed, which takes the robustness risk indicator as a negative weighted item and combines it with a positive weighted item based on the expected product performance indicator of the candidate synthetic parameter; the mutation operator of the genetic algorithm is guided, specifically the key risk parameters that have a positive contribution to the overall failure risk are determined by using the explainability analysis result, and when performing the mutation operation, the probability of the key risk parameters being selected for mutation is improved, and a directional mutation bias that reduces the key risk parameters is applied.

[0075] The specific form of the composite fitness function F is: Wherein, and are the values after the expected product performance indicator and the robustness risk indicator are mapped to the [0, 1] interval by the max-min normalization method; and are preset weight coefficients representing the importance of performance and robustness respectively, and the sum of the two is 1 (for example, , ).

[0076] The specific operation of the mutation operator of the genetic algorithm is:

[0077] Using the explainability analysis results (such as SHAP values ), the mutation probability of each parameter is calculated , which can be set as a basic mutation probability plus an increment proportional to the absolute value.

[0078] When a parameter determined as a key risk parameter (i.e. ) is selected for mutation, a directional mutation bias is applied to it, specifically subtracting an adjustment amount proportional to the size of the original value. New value = original value , where is a preset learning rate or step factor.

[0079] The composite fitness function sets a clear and controllable optimization goal for the algorithm, allowing users to accurately balance the two sometimes conflicting indicators of high performance and high robustness according to actual needs. Secondly, the directional mutation operator driven by explainability analysis provides the most efficient path to achieve this goal. It can accurately identify and correct specific process parameters that cause risks, avoiding blind search. The synergistic effect of this target + path ensures that the optimization process converges quickly and stably, finding the best process solution that truly balances performance and reliability.

[0080] Further, the process stability discriminator is an ensemble learning model composed of multiple base models: by performing K-fold cross-validation on the augmented training set, K independent risk assessment base models are trained; when evaluating the failure risk of each perturbed parameter point in the parameter cloud, the parameter point is input into the K base models to obtain K independent risk prediction values; the K independent risk prediction values are processed through an aggregation function to obtain a final, more robust failure risk value; the aggregation function is specifically the average and / or maximum of the K risk prediction values.

[0081] The use of an ensemble learning model significantly enhances the accuracy and stability of risk assessment. By aggregating the prediction results of multiple independently trained base models, the risk of misjudgment due to accidental bias of a single model or randomness of training data division can be effectively reduced, making the final robustness evaluation result more reliable, thereby providing a more reliable decision basis for optimization.

[0082] Further, the aggregation processing is specifically selecting the maximum value and / or a preset high percentile value from the failure risk values of all perturbed parameter points in the parameter cloud as the robustness risk indicator to evaluate the worst possible performance under process fluctuations.

[0083] The aggregation process specifically involves selecting the maximum value from the failure risk values ​​of all perturbated parameter points in the parameter cloud as the single robustness risk index; or selecting a preset high percentile value (e.g., the 95th percentile value) as the single robustness risk index; or taking the larger of the aforementioned calculated maximum value and the high percentile value as the final single robustness risk index to assess the worst possible performance under process fluctuations.

[0084] By selecting the maximum potential risk value or the high percentile risk value in the parameter cloud as the final indicator, this method no longer focuses on average performance but rather on the "worst-case scenario" under real production fluctuations. This evaluation approach is more in line with the core requirements of "safety first, stability paramount" in industrial production. Therefore, process parameters optimized with this goal in mind will naturally have stronger anti-interference capabilities and higher production reliability, effectively reducing the risk of batch failures.

[0085] Furthermore, the expected product performance indicators are calculated using a surrogate model trained on the historical synthetic data to predict product performance. The surrogate model receives the candidate synthetic parameters as input and outputs performance prediction values ​​including the expected molecular weight and the expected polydispersity index.

[0086] The surrogate model is specifically a gradient boosting decision tree model. This model effectively captures the complex nonlinear relationship between synthetic parameters and product performance, and exhibits high prediction accuracy and generalization ability.

[0087] The multiple performance predictions output are combined into a single performance index using a preset utility function. This utility function can be a weighted sum: Single performance index = × (normalized molecular weight) - × (Normalized polydispersity index).

[0088] in, and Preset positive weights are assigned to represent the importance of molecular weight and polydispersity, respectively; the polydispersity index term is negative because a lower index represents better performance. Before combination, each performance prediction value needs to be processed by methods such as max-min normalization to eliminate dimensional differences.

[0089] This embodiment first employs a gradient boosting decision tree model, which accurately captures the complex nonlinear relationship between process parameters and product performance, ensuring the accuracy of performance prediction. Secondly, through a clearly defined weighted utility function, it integrates multiple sometimes conflicting indicators, such as molecular weight and polydispersity, into a single quantifiable performance score.

[0090] The application generates high-quality virtual failure data through a failure path deduction model, solves the problem that it is difficult to construct an accurate risk model in industrial production due to sparse failure case data, and provides reliable data support for production decision-making. By simulating process fluctuations in real production through a virtual disturbance unit, and incorporating the quantified robustness risk into the optimization objective, the found process path not only has the optimal performance, but also has a wider process window to resist interference, thereby reducing the batch failure rate. By introducing explainability analysis to guide the mutation direction of the genetic algorithm, the optimization process is changed from "black box" search to "white box" evolution with clear guidance, significantly accelerating the convergence speed to the robustness optimal solution and improving the decision-making efficiency.

[0091] Embodiment two

[0092] The application scenario of this embodiment is that an A enterprise develops a core raw material, polycaprolactone (PCL) polyol, for a special polyurethane (PU) adhesive used for electric vehicle (EV) battery pack packaging for downstream new energy vehicle customers.

[0093] The core requirement of this application scenario is that the adhesive must have excellent bonding strength, as well as extremely high heat aging resistance and electrolyte corrosion resistance, to ensure the structural integrity and safety of the battery pack under long-term harsh working conditions. Therefore, the optimization objective is to maximize the robustness of the process path while ensuring the bonding performance, and to minimize the risk of batch material aging performance not meeting the standard due to process fluctuations.

[0094] A enterprise uses its hundred-ton pilot production line to collect 180 batches of PCL polyol production data. These data are generated using the leading micro-reaction continuous flow process, including reaction temperature, residence time, monomer to initiator molar ratio, catalyst dosage, and other key synthesis parameters. Among them, 172 batches of products are verified by downstream customers and have excellent performance (labeled as "success"). However, 8 batches have a bonding strength decay exceeding the acceptable threshold after accelerated aging test, and are labeled as "aging failure resistant".

[0095] Due to the severe shortage of real data samples (8 batches) of "aging failure resistant", a failure path deduction model is first used for data enhancement. All 180 batches of data are used to train a failure path deduction model based on the U-Net architecture. After training, the model is input with the label of "aging failure resistant" to generate 500 sets of virtual synthesis parameter data that may cause this failure mode.

[0096] Merge 172 batches of success data (labeled as 0), 8 batches of real failure data (labeled as 1) and 500 batches of virtual failure data (labeled as 1), train a LightGBM classifier. This model is used to predict the probability of the final product of any set of synthetic parameters failing the aging test.

[0097] Using only 172 batches of success data, train a gradient boosting regression tree model. This model is used to input a set of synthetic parameters and quickly predict the corresponding core product performance indicator - initial peel strength (N / mm).

[0098] Set the population size of the genetic algorithm to 120 and the number of iterations to 250 generations. Each individual represents a candidate PCL polyol synthesis path. The evaluation (calculate fitness) process for each individual is as follows:

[0099] Input the individual's synthetic parameters into the trained performance proxy model to get the predicted "initial peel strength" value, which is directly used as the single performance indicator P.

[0100] For each parameter of the individual, perturb around its central value to generate a parameter cloud containing N=100 points. The perturbation range is set taking into full consideration the "precise control" feature of A enterprise's micro-reaction continuous flow process, setting a narrower process tolerance than traditional batch reactors (for example, temperature fluctuation range is ±0.5°C).

[0101] Input the 100 points in the parameter cloud one by one into the process stability discriminator to get 100 corresponding "aging failure" risk probability values.

[0102] Considering the extreme importance of battery safety, select the maximum value of these 100 risk values as the robustness risk indicator R of the individual, representing the risk performance of the process path under the worst fluctuation.

[0103] Construct a composite fitness function. For this safety-critical application, give the robustness risk term a very high weight to ensure that the optimization direction converges to "safest" first: .

[0104] where, and are the normalized values of the performance indicator and the risk indicator, respectively.

[0105] In the mutation stage of the genetic algorithm, for a selected individual to be mutated, first apply the SHAP analysis method to the process stability discriminator to quantify the contribution of each parameter to the "aging failure" risk .

[0106] For example, analysis found that the The value is significantly positive, which is the main factor leading to potential aging risk. Then, in the mutation, the probability of being selected is significantly increased, and the mutation operation performed on it will be directed and small to reduce its use.

[0107] After 250 iterations, the algorithm converges to an optimal synthesis path. Compare this result with the path found by a traditional optimization method that only maximizes the "initial peel strength". As shown in Table 1:

[0108] Table 1 Comparison of the present invention and the traditional optimization method

[0109]

[0110] The traditional optimization method found a path that achieved the highest initial adhesive strength, but its robustness risk was as high as 25%, meaning that in A's actual continuous flow production, even with slight process fluctuations, there was a high risk of batch quality failure.

[0111] The path found by the present invention has only a slight decrease in initial peel strength of about 4.6%, which is completely within the customer's acceptable range, but its robustness risk is reduced by 92%, almost to zero. This indicates that the process path has a very wide process window and very high production stability.

[0112] Example Three

[0113] The application scenario of this embodiment is that A Company develops a long-acting protective coating for marine wind turbine blades for its downstream high-end paint customers. The coating uses the special polycaprolactone polyol produced by the company as the core raw material to prepare a polyurethane dispersion.

[0114] The blade coating needs to meet two major core requirements: one is extreme weather resistance, maintaining gloss and mechanical properties for a long time under strong ultraviolet radiation; the other is excellent salt spray corrosion resistance to resist marine environment erosion. The key failure mode is "coating salt spray blistering / peeling", which is difficult to detect in the laboratory research and development stage, but is disastrous in actual application. Therefore, the optimization goal is to maximize the robustness of the process path while ensuring excellent weather resistance, and to minimize the risk of coating corrosion resistance decline due to process fluctuations.

[0115] The research and development team has accumulated corresponding historical data by synthesizing and evaluating 250 different formulations of PCL polyols in their advanced detection and analysis room. These data include synthesis parameters (reaction temperature, residence time, monomer / initiator molar ratio, catalyst dosage) and performance test results of the PUD coating prepared therefrom. Among them, 238 formulations of the coating showed excellent performance. However, 12 formulations showed slight "coating blistering" phenomenon in more than 1000 hours of accelerated salt spray testing.

[0116] Since the real failure data sample (12) of “coating blistering” is too small to effectively train the model, data augmentation is first performed using the failure path inference model. All 250 sets of data are used to train a failure path inference model based on the U-Net architecture. After training, the model is input with the label of “coating blistering” to generate 600 sets of virtual synthetic parameter data that may cause this failure mode.

[0117] The 238 sets of successful data (label 0), 12 sets of real failure data (label 1), and 600 sets of virtual failure data (label 1) are combined to train a Random Forest classifier. This model is used to predict the risk probability of “coating blistering” of any set of synthetic parameters.

[0118] Using the 238 sets of successful data, a gradient boosting regression tree model is trained. This model is used to input a set of synthetic parameters to quickly predict its core weather resistance indicator, the gloss retention rate (%) after 2000 hours of QUV accelerated aging.

[0119] The population size of the genetic algorithm is set to 150, and the number of iterations is set to 300 generations. Each individual represents a candidate PCL polyol synthesis path. The evaluation (calculate fitness) process of each individual is as follows:

[0120] The individual's synthesis parameters are input into the trained performance proxy model to obtain the predicted “gloss retention rate” value, which is directly used as the single performance indicator P.

[0121] For each parameter of the individual, perturbations are made around its central value to generate a parameter cloud containing N=100 points. The perturbation range is set according to the accuracy of the A enterprise's laboratory automation batching system. The 100 points in the parameter cloud are input one by one into the process stability discriminator to obtain 100 corresponding “coating blistering” risk probability values. For wind turbine blades, which require ultra-long service life and high reliability, the maximum value of the 100 risk values is selected as the robustness risk indicator R of the individual to evaluate the corrosion resistance stability in the worst case.

[0122] The composite fitness function is constructed. For this application, weather resistance (performance) and corrosion resistance (risk) are equally important, so balanced weights are used: .

[0123] where, and are the normalized values of the performance indicator and the risk indicator, respectively.

[0124] At the mutation stage of the genetic algorithm, for a selected individual to be mutated, the SHAP analysis method is first applied to the process stability discriminator to quantify the contribution of each parameter to the "coating blistering" risk .

[0125] For example, the analysis found that the "monomer to initiator molar ratio" was too high, which was the main factor leading to potential blistering risk. Then, at the mutation stage, the probability of this parameter being selected was significantly increased, and the mutation operation performed on it would be directed to reduce its value.

[0126] After 300 iterations, the algorithm converged to an optimal synthesis path. This result was compared with the path found by a traditional optimization method that only maximized the "gloss retention" target, as shown in Table 2:

[0127] Table 2 Comparison of the present invention and the traditional optimization method

[0128]

[0129] The traditional optimization method found a formula with extremely high weather resistance indicators, but its potential salt spray blistering risk was also relatively high (22%), which was unacceptable for marine wind power coatings that required more than 20 years of life.

[0130] The path found by the present invention had a slightly lower gloss retention (about 2.6%), but its robustness risk was reduced by more than 95%, reaching an extremely high reliability level. This indicates that the process path has a very wide process window, effectively resisting production fluctuations and ensuring that each batch of products has excellent and stable corrosion resistance.

[0131] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A genetic algorithm-based polycaprolactone polyol synthesis path optimization method, characterized by, The method comprises: a genetic algorithm as a generator and a process stability discriminator; training a failure path inference model using historical synthesis data to learn the conditional probability distribution of synthesis parameters consisting of reaction temperature, reaction time, monomer-to-initiator molar ratio, and catalyst-to-initiator molar ratio given a failure mode label; and generating virtual failure data by inputting a failure mode label into the failure path inference model; the training step of the process stability discriminator comprises: merging the historical synthesis data and the virtual failure data to form an enhanced training set; and using the enhanced training set to perform supervised learning training on the process stability discriminator, so that the process stability discriminator can input a set of synthesis parameters consisting of reaction temperature, reaction time, monomer-to-initiator molar ratio, and catalyst-to-initiator molar ratio and output the corresponding failure risk probability; the failure path inference model comprises: a forward diffusion unit that gradually adds Gaussian noise to the real synthesis parameters in the historical synthesis data through a Markov chain until the distribution is standard normal distribution; a conditional denoising unit, which is a trained neural network that receives the noisy synthesis parameters at any time step of the forward diffusion process and the failure mode label as common inputs and predicts the noise added to the noisy synthesis parameters; and a sampling generation unit for sampling an initial noise from a standard normal distribution and iteratively calling the conditional denoising unit to gradually denoise under the guidance of the failure mode label to reversely reconstruct the virtual failure data; for the candidate synthesis parameters generated by the genetic algorithm, a virtual perturbation unit generates a parameter cloud in the neighborhood of the candidate synthesis parameters; the process stability discriminator is used to evaluate the failure risk values of the perturbed parameter points in the parameter cloud to construct a strategy volatility risk matrix; and the strategy volatility risk matrix is aggregated to generate a robustness risk indicator representing the overall failure risk of the candidate synthesis parameters; the virtual perturbation unit is a parameterized noise generation module: for each parameter in the candidate synthesis parameters, multiple random samples are independently generated from a probability distribution according to a process tolerance range matching the physical characteristics of the parameter, wherein the process tolerance range is set based on statistical fluctuations of historical production data and / or expert experience; the robustness risk indicator is input into the fitness function of the genetic algorithm, and the results of the explainability analysis of the process stability discriminator are used to guide the mutation operator of the genetic algorithm to accelerate the convergence of the optimization process to the robustness optimal solution. The explainability analysis result is obtained by applying a SHAP analysis method to the process stability discriminator, and quantifying the positive and / or negative contribution of each candidate synthesis parameter to the overall failure risk.

2. The genetic algorithm-based optimization method of a synthesis pathway of polycaprolactone polyol according to claim 1, characterized by, The variation operator specifically dynamically adjusts the variation probability and variation direction of the corresponding parameter gene according to the contribution, applies reverse variation to the parameter that contributes positively to the risk, and increases the variation probability of the parameter. The generation of the robust risk indicator representing the overall failure risk of the candidate synthesis parameter includes: The virtual disturbance unit generates a parameter cloud containing a plurality of perturbed parameter points by multiple random samplings around the single candidate synthesis parameter generated by the genetic algorithm; the failure risk values of the perturbed parameter points in the parameter cloud are constructed into a strategy volatility risk matrix; and a single robust risk indicator representing the overall failure risk of the candidate synthesis parameter is generated by aggregating the strategy volatility risk matrix.

3. The polycaprolactone polyol synthesis path optimization method based on genetic algorithm according to claim 2, characterized in that:

4. The genetic algorithm based optimization method of poly (caprolactone) polyol synthesis pathway according to claim 2, characterized in that: The robust risk indicator is input into the fitness function of the genetic algorithm, specifically a composite fitness function is constructed, the composite fitness function takes the robust risk indicator as a negative weighted item and combines it with a positive weighted item based on an expected product performance indicator of the candidate synthesis parameter; the variation operator of the genetic algorithm is guided, specifically the key risk parameters that positively contribute to the overall failure risk are determined by using the explainability analysis result, and when performing the variation operation, the probability of the key risk parameters being selected for variation is increased, and a directional variation bias for reducing the key risk parameters is applied.

5. The genetic algorithm based optimization method of poly (caprolactone) polyol synthesis pathway according to claim 3, wherein: The aggregation processing specifically selects the maximum value and / or a preset high percentile value from the failure risk values of all perturbed parameter points in the parameter cloud as the robust risk indicator to evaluate the worst possible performance under process fluctuations. The expected product performance indicator is calculated by an agent model trained on the historical synthesis data for predicting product performance, the agent model receives the candidate synthesis parameter as input and outputs performance prediction values including expected molecular weight and expected polydispersity index. The variation operator specifically dynamically adjusts the variation probability and variation direction of the corresponding parameter gene according to the contribution, applies reverse variation to the parameter that contributes positively to the risk, and increases the variation probability of the parameter.

Images