Polymerization process simulation method for caprolactone using generative adversarial network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WEIBOJIE BIOMATERIALS (ZHEJIANG) CO LTD
- Filing Date
- 2025-09-11
- Publication Date
- 2026-06-02
Smart Images
Figure CN121145641B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of generative adversarial networks (GANs), specifically to a method for simulating the polymerization process of caprolactone by incorporating GANs. Background Technology
[0002] In recent years, accurately simulating the polymerization process and predicting the performance of caprolactone has been a major challenge in industry. Traditional mechanistic and empirical models, due to their inherent limitations, struggle to capture the complex nonlinear and non-idealized behavior during polymerization, resulting in insufficient prediction accuracy and generalization ability.
[0003] With the development of computer systems for biological models, especially neural network technology, new directions have been provided for solving the above challenges. Modeling and optimizing chemical processes using deep learning technology has become a research hotspot. By constructing deep neural networks, researchers can automatically extract complex features from large amounts of experimental data and establish nonlinear mapping relationships between inputs and outputs.
[0004] However, existing machine learning methods, especially traditional feedforward neural networks, still have limitations when simulating complex aggregation processes. These models often require a large amount of training data to achieve ideal performance, which is a significant bottleneck in practical industrial applications because acquiring high-quality aggregation experimental data is costly and time-consuming. Furthermore, traditional neural networks exhibit insufficient generalization ability when faced with small samples or novel operating conditions, making it difficult to guarantee the accuracy and reliability of predictions. Essentially, they are discriminative models that can only learn the mapping relationships of input data and cannot autonomously generate new simulated samples that conform to the real data distribution, thus limiting their application in process optimization and virtual experiments. Summary of the Invention
[0005] The purpose of this invention is to provide a method for simulating the polymerization process of caprolactone by incorporating generative adversarial networks (GANs), thereby simulating the polymerization process of caprolactone through GANs.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A simulation method for caprolactone polymerization using generative adversarial networks includes:
[0008] A generator network is constructed, which receives multidimensional parameters of the aggregation process as input. Its training process is constrained by a physical information verification network, which quantifies the residuals of the caprolactone chemical kinetic differential equation into differentiable physical constraint loss terms, so that the generator network generates an aggregation data sequence that conforms to physical laws.
[0009] A discriminator network is constructed, and a constraint knowledge graph is built based on the aggregated data sequence. The effectiveness of the sequence is evaluated from the perspective of reaction product characteristics and multidimensional physical constraints. The combined adversarial loss used to optimize the generator network is calculated. The combined loss function includes standard adversarial loss, physical constraint loss and dynamic process consistency loss. The weights and biases of the generator are adjusted according to the combined adversarial loss.
[0010] The design parameter-sequence adversarial mechanism includes parameter-level adversarial and sequence-level adversarial, in which the generator network simultaneously engages in adversarial activities against the discriminator network and the physical information verification network, and generates a simulated aggregated data sequence based on the adjusted generator network.
[0011] The multidimensional parameters include reaction conditions, monomer information, and a sequence of actual polymerization process data.
[0012] The generator network is specifically constructed as follows:
[0013] The generator network uses a deep neural network structure to process the multidimensional parameters, capture the regularity and nonlinear relationship in the reaction process, and generate the aggregated data sequence; a physical information verification network is constructed, which combines the caprolactone chemical kinetic equation to constrain the aggregated data sequence, and quantifies the residual of the caprolactone chemical kinetic differential equation into a differentiable physical constraint loss term;
[0014] The training process is constrained by the physical information verification network, and the specific process is as follows:
[0015] A physical information verification network is constructed to receive the aggregated data sequence generated by the generator network, and to apply physical constraints to the aggregated data sequence in conjunction with the caprolactone chemical kinetic equation, and to calculate the physical constraint loss.
[0016] The caprolactone chemical kinetic equation is used to describe the dynamic characteristics of the polymerization reaction, including reaction rate calculation, monomer conversion and polymer molecular weight relationship, polymerization rate and temperature dependence, and polydispersity index and polymerization reaction relationship.
[0017] The physical information verification network is specifically as follows:
[0018] The physical information verification network is a differentiable physical model built on a neural network. It receives the aggregated data sequence output by the generator network and uses automatic differentiation technology to calculate the derivatives of key physical quantities during the polymerization reaction process as a function of time based on the aggregated data sequence.
[0019] The aggregated data sequence and derivative values are substituted into the preset set of differential equations for caprolactone chemical kinetics to calculate the residuals of the equations. The residuals are then quantified into backpropagable physical constraint loss terms by taking the norm of the residuals and incorporated into the total loss function of the generator network.
[0020] The construction of the discriminator network specifically involves:
[0021] The discriminator network uses a recurrent neural network to receive the aggregated data sequence generated by the generator network, and at the same time constructs a constraint knowledge graph to compare the aggregated data sequence generated by the generator with the real aggregated data sequence.
[0022] The generator network and the discriminator network are alternately optimized through an adversarial training mechanism;
[0023] The specific process by which the discriminator network constructs a constrained knowledge graph based on the aggregated data sequence includes:
[0024] The discriminator network extracts features from the aggregated data sequence and performs hierarchical processing to extract data features at each time step;
[0025] The data feature discriminator network constructs a constraint knowledge graph to compare the generated data sequence with the real aggregation process data sequence, and evaluates the authenticity of the aggregated data sequence generated by the generator network. The constraint knowledge graph contains the product characteristics of the reaction process and multi-dimensional physical constraints, and maps the interdependencies between various parameters in the reaction process into graph nodes and edges.
[0026] The combined adversarial loss of the generator network is calculated based on the differences evaluated in the constrained knowledge graph. The combined adversarial loss includes adversarial loss, physical constraint loss, and dynamic process consistency loss. The adversarial loss is calculated by comparing the differences between the generated data and the real data. The physical constraint loss is calculated based on the deviation between the generated data and the physical constraints. The dynamic process consistency loss measures whether the generated data conforms to the dynamic evolution law of the aggregation process in time.
[0027] The multi-level adversarial mechanism includes parameter-level adversarial and sequence-level adversarial. The parameter-level adversarial is used for adversarial training between the generator network and the physical information verification network. The physical information verification network receives the polymerization process data sequence generated by the generator network and calculates the physical loss term based on the chemical kinetic equation and thermodynamic principle of caprolactone polymerization. The goal of the generator network is to make the physical loss term approach 0.
[0028] The sequence layer adversarial network consists of the generator network and the discriminator network. The discriminator network determines the authenticity of the input data based on the data sequence of the aggregation process and generates a simulated probability, making the simulated probability approach 1.
[0029] The discriminator network calculates the adversarial loss of the generator network based on the simulated probability, and calculates the total loss value of the generator network by combining the physical loss term and the regression loss. The weights and biases of the generator network are then adjusted based on the total loss value.
[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0031] 1. This invention constructs a generator network with an embedded physical information verification network, quantifying the residual of the chemical kinetic differential equation for caprolactone polymerization into a differentiable physical constraint loss. This directly applies "hard constraints" of physical laws to the generator during the training process, fundamentally solving the problem that traditional data-driven models may generate data that violates physical reality. This ensures that the simulated data sequence not only statistically approximates the real data, but also conforms to the inherent laws of chemical reactions in terms of mechanism. As a result, it significantly improves the accuracy and physical reliability of the polymerization process simulation and reduces the dependence on massive amounts of high-quality experimental data.
[0032] 2. This invention designs a discriminator network employing a constrained knowledge graph and introduces a combined loss function comprising standard adversarial loss, physical constraint loss, and dynamic process consistency loss. This discriminator no longer simply performs true / false judgments but can deeply and comprehensively evaluate the validity and internal logic of the generated data sequences from multiple levels, including reaction product characteristics, multidimensional physical constraints, and dynamic evolution laws. This refined evaluation mechanism provides the generator with richer and more instructive gradient information, forcing it to learn and reproduce the complex interdependencies between parameters during the aggregation process, thereby generating higher-quality and more realistic simulation results.
[0033] 3. This invention designs a multi-layered adversarial mechanism that combines parameter layers and sequence layers, subjecting the generator to dual adversarial interactions with the physical information verification network and the discriminator network. The parameter layer adversarial interaction ensures the physical correctness of the generated results, while the sequence layer adversarial interaction guarantees the overall authenticity and statistical similarity of the generated data sequence. This structured adversarial training framework decouples and synchronously optimizes the accuracy requirements of the physical mechanism and the fitting requirements of the data distribution, effectively avoiding the training instability or mode collapse problems that may be caused by a single adversarial target. This makes the training process of the entire model more stable and the convergence effect better, and the simulation method obtained at the end has both high fidelity and high robustness. Attached Figure Description
[0034] Figure 1 A schematic diagram of the simulation method for caprolactone polymerization process using a generative adversarial network;
[0035] Figure 2 A data framework diagram for simulating the polymerization process of caprolactone using a generative adversarial network. Detailed Implementation
[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0037] Example 1:
[0038] Please see Figure 1 , Figure 2 This invention provides a method for simulating the polymerization process of caprolactone by fusing generative adversarial networks, the technical solution of which is as follows:
[0039] A simulation method for caprolactone polymerization using generative adversarial networks includes:
[0040] A generator network is constructed, which receives multidimensional parameters of the aggregation process as input. Its training process is constrained by a physical information verification network, which quantifies the residuals of the caprolactone chemical kinetic differential equation into differentiable physical constraint loss terms, so that the generator network generates an aggregation data sequence that conforms to physical laws.
[0041] A discriminator network is constructed, and a constraint knowledge graph is built based on the aggregated data sequence. The effectiveness of the sequence is evaluated from the perspective of reaction product characteristics and multidimensional physical constraints. The combined adversarial loss used to optimize the generator network is calculated. The combined loss function includes standard adversarial loss, physical constraint loss and dynamic process consistency loss. The weights and biases of the generator are adjusted according to the combined adversarial loss.
[0042] The design parameter-sequence adversarial mechanism includes parameter-level adversarial and sequence-level adversarial, in which the generator network simultaneously engages in adversarial activities against the discriminator network and the physical information verification network, and generates a simulated aggregated data sequence based on the adjusted generator network.
[0043] Furthermore, the multidimensional parameters include reaction conditions, monomer information, and a sequence of actual polymerization process data;
[0044] The reaction conditions include the polymerization reaction set temperature, reaction time, catalyst type, catalyst concentration, and molar ratio of initiator to monomer. For example, the polymerization reaction set temperature is 130°C, the reaction time is 8 hours, the catalyst type is stannous octoate, the catalyst concentration is a molar ratio of monomer to catalyst of 10000:1, and 1,4-butanediol is used as the initiator with a molar ratio of 1:100 to caprolactone monomer.
[0045] The monomer information includes the initial concentration and purity of caprolactone monomers, as well as the type and proportion of comonomers. For example, if only bulk polymerization of caprolactone is used, i.e. without solvent, the concentration is the density of the caprolactone monomer itself, and there are no comonomers.
[0046] The real polymerization process data sequence is a series of key process indicators collected at multiple discrete time points in a real physical experiment corresponding to the reaction conditions and monomer information. These indicators include the monomer conversion rate, number-average molecular weight, and polydispersity index of the polymer at each time point. The real polymerization process data sequence is time-series data measured in the real experiment and is the target learned by the generator network. The monomer conversion rate is the percentage of monomers that have been converted into polymers at a certain point in the reaction, relative to the initial total monomer amount. The calculation formula is:
[0047] ;
[0048] The number-average molecular weight of the polymer is the statistical average of the molecular weights of all polymer chains in the polymer sample. It is calculated by dividing the total mass of the sample by the total number of moles of all molecular chains. The polydispersity index (PDI) is an indicator that measures the breadth of the molecular weight distribution in the polymer sample. It is defined as the ratio of the weight-average molecular weight to the number-average molecular weight. An ideal polymer that is completely homogeneous and has all molecular chains of the same length has a PDI value of 1. The closer the PDI value is to 1, the higher the degree of control of the polymerization reaction, as shown in Table 1.
[0049]
[0050] By using specific, quantifiable multidimensional parameters, including explicit reaction conditions, monomer information, and corresponding real process data sequences, as inputs and learning targets, the trained simulation model can accurately establish the mapping relationship between a complete experimental formulation and the entire dynamic reaction result. It can not only reproduce existing experiments with high fidelity but also has powerful predictive capabilities. Users can conduct virtual experiments by adjusting input parameters and efficiently predict polymerization results without the need for expensive and time-consuming physical experiments.
[0051] A generator network was constructed and trained by collecting historical multidimensional parameters of the caprolactone polymerization reaction. The collected raw data was cleaned to remove outliers and erroneous records. Each data point was organized into an independent sample-label pair, where the sample is a combination of reaction conditions and monomer information for that experiment, and the label is a sequence of real polymerization process data that completely corresponds to the conditions of that sample. Since computers cannot directly process text such as stannous octoate, all parameters were numerically encoded.
[0052] A small batch of data (e.g., 32 different historical experimental data pairs) is randomly selected from the preprocessed dataset. The input sample vectors of these 32 experiments (each vector representing a set of reaction conditions and monomer information) are fed into the feedforward neural network layer of the generator network. The feedforward neural network layer maps discrete and heterogeneous static parameters into condition vectors. The condition vectors contain a digital description of the complete setting of a specific aggregation experiment, resulting in 32 different condition vectors.
[0053] The generator network uses a deep neural network to generate time series data. The conditional vector is used as the initial hidden state. Based on the initial conditions, an aggregated data sequence covering the entire reaction time is generated step by step. The aggregated data sequence is composed of data vectors at multiple time points. Each data vector contains the predicted values of [monomer conversion rate, number-average molecular weight, polydispersity index] at that time. The aggregated data sequence is based on statistical laws learned from a large amount of real data, but there may be minor physical contradictions. For example, the conversion rate at a certain time is slightly lower than that at the previous time, or the molecular weight growth and conversion rate do not have a specific physical relationship.
[0054] By establishing a standardized data preprocessing process from historical data cleaning and structuring (sample-label pairs) to parameter numerical encoding, and by using a feedforward neural network to map complex and heterogeneous response conditions into a unified and information-dense condition vector, a sequence generation core capable of effectively understanding and responding to diverse input instructions is constructed. This ensures that the generator can fully learn and reproduce the statistical regularities and nonlinear relationships contained in supervised real experimental data, thereby producing a highly realistic simulation result.
[0055] The training process of the generator network is constrained by the physical information verification network. The specific process of this constraint is as follows:
[0056] A physical information verification network is constructed to receive the aggregated data sequence generated by the generator network, and to apply physical constraints to the aggregated data sequence in conjunction with the caprolactone chemical kinetic equation, and to calculate the physical constraint loss.
[0057] The caprolactone chemical kinetic equation is used to describe the dynamic characteristics of the polymerization reaction, including reaction rate calculation, monomer conversion and polymer molecular weight relationship, polymerization rate and temperature dependence, and polydispersity index and polymerization reaction relationship.
[0058] The reaction rate calculation refers to the rate of polymerization reaction, which usually refers to the rate of monomer consumption.
[0059] The relationship between monomer conversion rate and polymer molecular weight is that, in controlled polymerization, there is a definite linear relationship between the number-average molecular weight of the polymer and the monomer conversion rate.
[0060] The polymerization rate dependence on temperature means that the polymerization rate constant depends on the reaction temperature. This dependence is described in this embodiment using the Arrhenius equation.
[0061] The relationship between the polydispersity index (PDI) and the polymerization reaction is as follows: For the controlled polymerization of caprolactone, the reaction mechanism dictates that all polymer chains should grow at similar rates. Therefore, theoretically, the molecular weight distribution of the product should be narrow, and the PDI value should be close to 1. Throughout the reaction process, the PDI value should remain at a low level (e.g., usually less than 1.5) and should not exhibit irregular, drastic fluctuations or a continuous increasing trend.
[0062] By decomposing the complex polymerization process into multiple basic physicochemical dimensions such as reaction rate, linear relationship between molecular weight and conversion, temperature dependence and product distribution (PDI), and constructing clear mathematical constraints for each dimension, a multi-faceted and in-depth physical information verification system is established. This system allows the generative network to not only mimic the surface trends of historical data, but also forces it to learn and comply with the core mechanism of the polymerization reaction and the coupling relationship between multiple variables.
[0063] Specifically, the process constrained by the physical information verification network involves receiving a preliminary aggregated data sequence generated by the generator network as input. This preliminary aggregated data sequence contains predicted values for monomer conversion rate, polymer number-average molecular weight, and polydispersity index at multiple simulation time points. The physical information verification network utilizes the automatic differentiation function in the deep learning framework to calculate the specific rate of change of key physical quantities with simulation time. During the forward computation of the generator network, a detailed computation graph is automatically constructed. This computation graph records every basic mathematical operation performed from the input of the time variable to the final physical quantity output. When it is necessary to calculate the physical quantity over time... When the rate of change is measured, the automatic differentiation employs the chain rule, starting from the final output value and tracing back along the computation graph. In each backtracking step, the gradient of the previous step is multiplied by the local derivative of the current operation and accumulated. When the time input node is reached, the accumulated result is the precise analytical derivative of the physical quantity with respect to time at a specific moment. The automatic differentiation function ensures high accuracy in calculating the residuals of physical equations, avoids misleading model training due to derivative calculation errors, ensures the strictness and effectiveness of physical constraints, and fully automates the complex differentiation process, eliminating the need for researchers to manually derive and write derivative formulas for large neural networks.
[0064] After obtaining the rate of change of each physical quantity in the generator output sequence over time through automatic differentiation, the abstract physical laws are transformed into specific loss terms that can be used to optimize the neural network.
[0065] The generated data are substituted into a pre-defined set of chemical kinetic differential equations to calculate residuals, which represent the deviation between the generated data and the predicted values of physical laws. These calculations are performed separately for each physical law at each time point in the sequence. For example, the generated conversion rate and its rate of change are substituted into the reaction rate relationship to obtain the rate residual; simultaneously, the generated molecular weight and conversion rate are substituted into the theoretical molecular weight growth relationship to obtain the molecular weight residual; and the generated polydispersity index (PDI) is checked to determine if it exceeds a reasonable upper limit, yielding the PDI residual. This time-point residual calculation for each physical law ensures that key parameters such as reaction rate, molecular weight, and polydispersity index conform to actual chemical kinetic laws during the simulation. Through this residual feedback mechanism, the generator can adaptively adjust its output, thereby improving the accuracy and reliability of the polymerization process simulation and ensuring that the generated polymerization data has high practical application value.
[0066] After calculating the residual sequence representing each physical law at all time points, the mean square error is calculated to square each residual value in the sequence, and the average of these squared values at all time points is calculated. The overall degree of violation of each physical law in the entire simulation process is transformed into a single, non-negative value. The independent loss terms representing different physical dimensions are averaged and weighted to form the physical constraint loss, which is calculated as a key regularization term in the total loss function of the generator network. The mean square error is used to quantify the degree of violation of multiple physical laws by the generated data over time into a clear and optimizable physical constraint loss term, which improves the stability of model training and the overall fidelity of the final simulation results.
[0067] During the backpropagation process of model training, the physical constraint loss generates a corresponding gradient, which guides the parameters of the generator network to be adjusted in a direction that not only simulates real data but also satisfies all preset physical and chemical laws, thereby achieving a deep integration of data-driven and physical mechanisms.
[0068] Since the input is time series data, the discriminator network uses a recurrent neural network as its core architecture. During the training process, the discriminator learns a large number of real aggregation data sequences and builds a complex cognitive model about "what is the real aggregation process".
[0069] A constrained knowledge graph is constructed based on aggregated data sequences to evaluate the effectiveness of the sequences from the perspectives of reaction product characteristics and multidimensional physical constraints. The constrained knowledge graph includes reaction characteristic constraints and multidimensional physical constraints. When a discriminator receives a sequence, it is actually matching and comparing the sequence with its internally learned knowledge graph. If the various features and dynamic evolution laws of the input sequence are highly consistent with the knowledge graph, the discriminator tends to judge it as true; otherwise, if there are any flaws that do not conform to its internal knowledge, it is judged as false.
[0070] For any input aggregated data sequence, the discriminator will eventually output a single probability value (usually between 0 and 1), which represents the confidence level that the sequence is real data;
[0071] Fix the generator network, train the discriminator network, obtain real aggregated data sequences, and simultaneously have the current generator network generate a batch of fake aggregated data sequences; feed both batches of data into the discriminator network at the same time. The discriminator's goal is to correctly push the judgment probability of the real sequence to 1 and the judgment probability of the fake sequence to 0; fix the discriminator network, train the generator network. The generator's goal is the opposite of the discriminator's goal. It hopes that the fake data it generates will be misjudged as "real" by the discriminator, that is, the closer the judgment probability is to 1, the better.
[0072] The discriminator network not only examines the instantaneous rationality of data points, but also, based on its learned knowledge graph, comprehensively judges the intrinsic consistency and mechanistic authenticity of the entire data sequence from multiple levels, such as the characteristics of reaction products, multidimensional physical constraints, and dynamic evolution laws.
[0073] When the discriminator network receives the data sequence of the aggregation process, whether it is real or forged by the generator, the recurrent neural network will process it step by step from the first time point of the sequence. When processing the data at each time point, the network will combine the physical quantity information at the current moment with its "memory summary" of the information at all previous moments. Through its internal complex gating structure, the network will update and form a new memory state. This new memory state not only contains the information at the current moment, but also integrates the evolutionary history of the entire past sequence, forming hierarchical features.
[0074] After the network processes the last time point of the sequence, its final memory state becomes a highly condensed comprehensive feature representation of the entire aggregation reaction process. This comprehensive feature is then fed into a standard feedforward neural network layer for final decision-making. Here, the network, having been trained on a large amount of real data, has implicitly constructed a complex constraint knowledge graph in its internal parameters. This constraint knowledge graph contains a deep understanding of the characteristics of the reaction products, multidimensional physical constraints, and dynamic evolution laws. The network matches and compares the input sequence features with this internal knowledge graph and finally outputs a single probability value, representing the credibility of the sequence as real data. After the input aggregation process data sequence has undergone hierarchical processing, the network can accurately judge the authenticity of the data based on its internal knowledge graph, providing a reliable basis for the final credibility output.
[0075] After the discriminator completes its evaluation, a combined adversarial loss is calculated for the generator. This combined loss consists of three parts: the standard adversarial loss, which measures the success of the generator in deceiving the discriminator; when the discriminator assigns a low probability of confidence to a forged sequence, the generator's adversarial loss will be correspondingly higher; the dynamic process consistency loss, which is a supervised loss that directly measures the difference between the generator's output sequence and the real experimental data sequence by calculating the mean square error; and the physical constraint loss, which is independent of the discriminator and measures the generator's adherence to the first principles of physics by substituting its output into the chemical kinetic equations.
[0076] The three different loss terms are combined by averaging to form a final total loss value that guides the generator optimization. The gradient is calculated based on the combined adversarial loss to update the weights and biases within the generator network.
[0077] By designing a three-dimensional combined objective function for the generator network that integrates standard adversarial loss, dynamic process consistency loss, and physical constraint loss, the three core requirements of the simulation results—fidelity, accuracy, and physical plausibility—are transformed into specific mathematical indicators that can be optimized simultaneously. At the same time, the generator is constrained and guided in three dimensions, training a more comprehensive and robust simulation model. The final aggregated process data generated by this model is not only indistinguishable from real experiments in form and numerically accurate, but also scientifically sound, thus obtaining simulation results with both high fidelity and high credibility.
[0078] The design parameter-sequence adversarial mechanism includes parameter-level adversarial and sequence-level adversarial. The generator network is simultaneously adversarial against the discriminator network and the physical information verification network. In a single update iteration of the generator network, the generator is synchronously adversarial and constrained from different levels and angles. The whole process begins with the generator network receiving a batch of input reaction conditions and individual information, and generating a batch of corresponding aggregation process data sequences accordingly.
[0079] The parameter layer adversarial training involves the generator network and the physical information verification network training against each other. The generator network generates aggregated data sequences based on the input multidimensional parameters. The physical information verification network receives the generated aggregated data sequences and calculates the physical loss by combining the chemical kinetic equations and thermodynamic principles of caprolactone polymerization. The goal of the generator network is to minimize the physical loss term by adjusting the weights and biases, so that the generated data sequences conform as closely as possible to the actual physical and chemical laws.
[0080] The sequence layer adversarial process is performed. After the generator network generates the aggregated data sequence, it inputs the data sequence into the discriminator network. The discriminator network analyzes the input data sequence and judges its authenticity. Based on the temporal characteristics of the data sequence in the aggregation process, the discriminator network generates a simulated probability, which represents the similarity between the generated data and the real data. The goal of the generator network is to optimize and push the simulated probability close to 1, so that the generated data is considered by the discriminator network to be highly consistent with the real data.
[0081] By designing parameter-level adversarial and sequence-level adversarial mechanisms, this invention can simultaneously optimize and constrain the generator network from multiple dimensions, ensuring that the generated aggregated data sequence not only conforms to physical and chemical laws but also has a high degree of temporal consistency with real data. The parameter-level adversarial mechanism verifies the constraints of the physical information verification network, making the generated data more consistent with the dynamic characteristics of the actual reaction process and improving the physical rationality of the data. The sequence-level adversarial mechanism, through the judgment of the discriminator network, further ensures that the generated data and real data have a high degree of temporal similarity, enabling the generator to provide more accurate and reliable data sequences in the simulation of the actual aggregation process, thereby improving the accuracy and practicality of the simulation results.
[0082] This invention provides a precise simulation method for caprolactone polymerization by integrating generative adversarial networks (GANs) and physical information constraints. Under the constraints of a physical information verification network, the generator network optimizes the physical constraint loss of residual quantization to ensure that the generated data sequences strictly follow chemical kinetics, thus improving the physical plausibility of the simulation results. The discriminator network, based on a constraint knowledge graph, comprehensively considers the characteristics of reaction products and multidimensional physical constraints, effectively evaluating the authenticity and validity of the generated data, further improving data accuracy. The parameter-sequence adversarial mechanism combines parameter-level and sequence-level adversarial approaches, enhancing the generator's adaptive capability through synchronous optimization.
[0083] Example 2:
[0084] To address the challenges of high R&D costs, long development cycles, poor predictability, difficulty in optimization, and data scarcity in the development of novel polycaprolactone (PCL) materials for high-end medical implants, this embodiment proposes a simulation method for the PCL polymerization process incorporating generative adversarial networks.
[0085] First, define and input the multidimensional parameters of the specific experimental setup. The reaction conditions are: reaction temperature 130℃, reaction time 8 hours, catalyst stannous octoate, catalyst concentration, monomer to catalyst molar ratio of 10000:1, initiator 1,4-butanediol, initiator to monomer molar ratio of 1:100, monomer information is: monomer type caprolactone bulk polymerization, initial concentration, density of caprolactone monomer itself, and actual polymerization process data sequence.
[0086] The generator network is a deep neural network used to generate time series data. It receives multidimensional parameters and is trained using real polymerization process data sequences to generate polymerization process sequences. During training, it is constrained by a physical information verification network, which embeds the basic chemical kinetic differential equations of caprolactone polymerization, including reaction rate calculation, the relationship between monomer conversion rate and polymer molecular weight, and the dependence of polymerization rate on temperature. The residuals calculated after substituting the generated data into these equations are quantified as a differentiable physical constraint loss term. This loss term serves as a key penalty signal, forcing the generator to follow physical laws during the learning process, thereby ensuring that the generated polymerization data sequences not only approximate reality in terms of data but also are scientifically sound in terms of mechanism.
[0087] A discriminator network with a recurrent neural network structure is constructed. The discriminator network receives the aggregated data sequence generated by the generator and constructs a constraint knowledge graph based on this sequence. The constraint knowledge graph deeply evaluates the effectiveness of the generated sequence from two levels: reaction product characteristics (such as final molecular weight and PDI distribution) and multidimensional physical constraints (such as dynamic correlations between parameters). Rate residuals, molecular weight residuals, and PDI residuals are generated. The combined adversarial loss is calculated based on the above residuals. The combined loss function includes standard adversarial loss, physical constraint loss, and dynamic process consistency loss. The gradient is backpropagated based on the combined adversarial loss to adjust the weights and biases of the generator.
[0088] The design parameter-sequence adversarial mechanism decomposes the generator training into two parallel adversarial layers:
[0089] Parameter layer adversarial: The generator network and the physical information verification network compete against each other, with the goal of generating a sequence of parameters that can make the physical constraint loss term approach zero;
[0090] Sequence layer adversarial: The generator network and the discriminator network compete against each other, with the goal of generating a sequence of data sufficient to "deceive" the discriminator into identifying it as a real aggregation process data sequence;
[0091] By simultaneously pitting the generator network against both the discriminator network and the physical information verification network, the stability and efficiency of model training are ensured, resulting in simulation results that are both data-realistic and physically accurate.
[0092] After the above iterative training and optimization process, the adjusted generator network can be used for actual prediction. When a new set of experimental conditions is input, the network can efficiently and accurately generate the final simulated aggregated data sequence.
[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for simulating the polymerization process of caprolactone using a generative adversarial network, characterized in that, include: A generator network is constructed to receive multidimensional parameters of the polymerization process as input, including reaction conditions, monomer information, and the actual polymerization process data sequence. The aggregated data sequence is generated through training, and the training process is constrained by a physical information verification network, which quantifies the residual of the differential equation of caprolactone chemical kinetics into a differentiable physical constraint loss term. The aggregated data sequence consists of data vectors at multiple time points, and each data vector contains the predicted values of monomer conversion rate, number-average molecular weight, and polydispersity index at that time. A discriminator network is constructed, and a constraint knowledge graph is built based on the aggregated data sequence. The effectiveness of the aggregated data sequence is evaluated from the perspectives of reaction product characteristics and multidimensional physical constraints. The combined adversarial loss used to optimize the generator network is calculated. The combined adversarial loss includes standard adversarial loss, physical constraint loss, and dynamic process consistency loss. The weights and biases of the generator network are adjusted according to the combined adversarial loss. The dynamic process consistency loss measures whether the generated data conforms to the dynamic evolution law of the aggregation process in time. The design incorporates a parameter-sequence adversarial mechanism, including parameter-level adversarial and sequence-level adversarial training. The generator network simultaneously engages in adversarial training against both the discriminator network and the physical information verification network. Simulated aggregation data sequences are generated based on the adjusted generator network. The parameter-sequence adversarial mechanism involves parameter-level adversarial training between the generator network and the physical information verification network. The physical information verification network receives the aggregation process data sequences generated by the generator network and calculates a physical constraint loss term based on the chemical kinetic equations for caprolactone polymerization. The generator network aims to make the physical constraint loss term approach zero. The sequence layer adversarial network consists of the generator network and the discriminator network. The discriminator network determines the authenticity of the input data based on the data sequence of the aggregation process and generates a simulated probability, making the simulated probability approach 1. The discriminator network calculates the adversarial loss of the generator network based on the simulated probability, and calculates the total loss value of the generator network by combining the physical loss term and the regression loss. The weights and biases of the generator network are then adjusted based on the total loss value.
2. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 1, characterized in that, The multidimensional parameters include reaction conditions, monomer information, and a sequence of actual polymerization process data.
3. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 1, characterized in that, The generator network is specifically constructed as follows: The generator network uses a deep neural network structure to process the multidimensional parameters, capture the regularity and nonlinear relationship in the reaction process, and generate the aggregated data sequence. A physical information verification network is constructed, which combines the caprolactone chemical kinetic equation to constrain the aggregated data sequence, and quantifies the residual of the caprolactone chemical kinetic differential equation into a differentiable physical constraint loss term.
4. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 1, characterized in that, The training process is constrained by the physical information verification network, and the specific process is as follows: A physical information verification network is constructed to receive the aggregated data sequence generated by the generator network, and to apply physical constraints to the aggregated data sequence in conjunction with the caprolactone chemical kinetic equation, and to calculate the physical constraint loss. The caprolactone chemical kinetic equation is used to describe the dynamic characteristics of the polymerization reaction, including reaction rate calculation, monomer conversion and polymer molecular weight relationship, polymerization rate and temperature dependence, and polydispersity index and polymerization reaction relationship.
5. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 4, characterized in that, The physical information verification network is specifically as follows: The physical information verification network is a differentiable physical model built on a neural network. It receives the aggregated data sequence output by the generator network and uses automatic differentiation technology to calculate the derivatives of key physical quantities during the polymerization reaction process as a function of time based on the aggregated data sequence. The aggregated data sequence and derivative values are substituted into the preset set of differential equations for caprolactone chemical kinetics to calculate the residuals of the equations. The residuals are then quantified into backpropagable physical constraint loss terms by taking the norm of the residuals and incorporated into the total loss function of the generator network.
6. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 1, characterized in that, The construction of the discriminator network specifically involves: The discriminator network uses a recurrent neural network to receive the aggregated data sequence generated by the generator network, and at the same time constructs a constraint knowledge graph to compare the aggregated data sequence generated by the generator with the real aggregated data sequence. The generator network and the discriminator network are optimized alternately through an adversarial training mechanism.
7. The method for simulating caprolactone polymerization using a fused generative adversarial network according to claim 1, characterized in that, The specific process by which the discriminator network constructs a constrained knowledge graph based on the aggregated data sequence includes: The discriminator network extracts features from the aggregated data sequence and performs hierarchical processing to extract data features at each time step; The data feature discriminator network constructs a constraint knowledge graph to compare the generated data sequence with the real aggregation process data sequence, and evaluates the authenticity of the aggregated data sequence generated by the generator network. The constraint knowledge graph contains the product characteristics of the reaction process and multi-dimensional physical constraints, and maps the interdependencies between various parameters in the reaction process into graph nodes and edges. The combined adversarial loss of the generator network is calculated based on the differences evaluated in the constrained knowledge graph. The combined adversarial loss includes standard adversarial loss, physical constraint loss, and dynamic process consistency loss. The adversarial loss is calculated by comparing the differences between the generated data and the real data. The physical constraint loss is calculated based on the deviation between the generated data and the physical constraints. The dynamic process consistency loss measures whether the generated data conforms to the dynamic evolution law of the aggregation process in time.