High polymer material performance prediction method based on generative adversarial network

Abstract

The invention provides a high polymer material performance prediction method based on a generative adversarial network, and belongs to the field of high polymer material science and engineering. According to the method, a generative adversarial network is utilized to perform data amplification on an acquired original data set to obtain a mixed data set, a pipeline optimization tool is utilized to obtain a glass-transition temperature prediction model suitable for the mixed data set, and finally the glass-transition temperature prediction model is utilized to predict the glass-transition temperature. According to the method, the structural design and synthesis of the solution polymerized styrene-butadiene rubber are promoted, so that the prediction precision is improved, the manual intervention is reduced, and the discovery and design period of an innovative material is accelerated.

Description

Technical Field

[0001] The present invention belongs to the field of polymer material science and engineering, and specifically relates to a polymer material performance prediction method based on a generative adversarial network. Background Art

[0002] With the development of the automotive industry, the increasing scarcity of oil resources, and the increasing environmental protection requirements of various countries, people are paying more and more attention to green tires. Solution styrene butadiene rubber (SSBR), due to its narrow molecular weight distribution and few branch terminals, is primarily used in the design and production of high-performance tires. It is also widely used in other industries such as cables.

[0003] Solution-polymerized styrene-butadiene rubber (SSBR) is an elastomer formed by anionic solution polymerization of styrene and butadiene in an organic lithium-initiated manner. Several studies have examined the relationship between the structure and properties of SSBR, including: Hou et al. investigated the effects of five different brands of SSBR on the overall properties of their composites, analyzing in detail the influence of SSBR microstructure (styrene, 1,2-butadiene, cis-1,4-butadiene, and trans-1,4-butadiene content) on the overall properties of the composites [J. Appl. Polym. Sci. 2018, 135, 45749]. Hua et al. studied the effects of vinyl and phenyl groups and their contents on the vulcanization behavior, mechanical strength, fatigue resistance, heat resistance and wear resistance of HVBR and solution-polymerized styrene-butadiene rubber (SSBR). The results showed that vinyl contributed more to wear resistance and fatigue resistance than phenyl groups, but phenyl groups contributed more to mechanical strength than vinyl groups [J.Appl.Polym.Sci.2018,135(12),45975].

[0004] Although these studies have conducted qualitative analysis on the relationship between the structure and properties of solution-polymerized styrene-butadiene rubber (SSBR), no quantitative model has been established. Wang Bin et al. used the multivariate linear regression method to establish a relationship model between the wear resistance and other mechanical properties of solution-polymerized styrene-butadiene rubber (SSBR)-based composites [Synthetic Rubber Industry, 2013, 36(02): 123-126]. On this basis, Li et al. realized that there may be a nonlinear relationship between wear resistance and other properties, and successfully established a multi-layer feedforward neural network model (MLFN-3) with three nodes, with an RMS error of 0.07, showing a high prediction accuracy [2014IEEE Workshop Electron. Comput. Appl. 2014, 581-584]. However, these studies only used 23 sets of data. The small amount of data may lead to insufficient generalization ability of the model, and only explored the relationship between performance and performance, lacking guidance for practical design and development.

[0005] In summary, all of the aforementioned literature lacks quantitative research on the relationship between the structure and properties of solution-polymerized styrene-butadiene rubber (SSBR). The glass transition temperature (Tg) of SSBR is a critical physical parameter, reflecting the transition temperature from the glassy to the highly elastic state, which directly impacts the material's performance and processing properties. In addition to experimental studies, computational evaluation and prediction of Tg are increasingly important for accelerating the design and development of polymer materials.

[0006] At the same time, generative adversarial networks (GANs) have attracted increasing attention in the field of materials science due to their ability to generate the required content. The GAN proposed by Goodfellow et al. can generate artificial data and has been applied in material property prediction [In Proceedings of the 27th International Conference on Neural Information Processing Systems; NIPS'14; MIT Press: Cambridge, MA, USA, 2014; 2672–2680]. Marani et al. used TGAN to generate 6513 reasonable synthetic data, trained the machine learning model, and tested it on 810 experimental data, showing good prediction performance [Materials 2020, 13(21), 4757]. Yang et al. proposed a two-step data enhancement method based on GAN to improve the evaluation performance of the machine learning model in the prediction of high entropy alloy properties [Comput. Mater. Sci. 2023, 220, 112064]. Chen et al. combined GAN and XGBoost models to predict the crystallinity of natural rubber and achieved significant results [Langmuir 2023, 39(48), 17088–17099].

[0007] Although the above-mentioned VSG method improves the prediction accuracy of machine learning to a certain extent, using a single machine learning model as the final prediction model may have problems such as low prediction accuracy, weak generalization ability, and cumbersome hyperparameter optimization. Summary of the Invention

[0008] The purpose of the present invention is to solve the difficulties existing in the above-mentioned prior art and provide a polymer material performance prediction method based on a generative adversarial network. A new automatic machine learning framework based on the combination of a generative adversarial network (GAN) and a pipeline optimization tool (TPOT) is used to predict the glass transition temperature of solution-polymerized styrene-butadiene rubber. First, a generative adversarial network (GAN) is used to amplify the collected sample data set, and then the generated data is mixed with the original data. The pipeline optimization tool (TPOT) is used to develop an optimal machine learning prediction model suitable for the mixed data set.

[0009] The present invention is achieved through the following technical solutions:

[0010] A polymer material performance prediction method based on generative adversarial networks,

[0011] The method uses a generative adversarial network to perform data amplification on a collected original data set to obtain a mixed data set, and uses a pipeline optimization tool to obtain a glass transition temperature prediction model suitable for the mixed data set. Finally, the glass transition temperature is predicted using the glass transition temperature prediction model.

[0012] Furthermore, the method comprises:

[0013] Step 1: Collect data on seven variables of the solution-polymerized styrene-butadiene rubber structure and compile an original data set;

[0014] Step 2: Build an automated machine learning regression prediction model consisting of a generative adversarial network and pipeline optimization tools;

[0015] Step 3: Use the generative adversarial network to amplify the original dataset to form a mixed dataset;

[0016] Step 4: Use the mixed data set to train the automatic machine learning regression prediction model to obtain a glass transition temperature prediction model;

[0017] Step 5: Input the data of the seven variables of the solution-polymerized styrene-butadiene rubber to be predicted into the glass transition temperature prediction model, and the glass transition temperature prediction model outputs the polymer glass transition temperature.

[0018] Furthermore, the generative adversarial network consists of a generator and a discriminator.

[0019] Furthermore, the operation of step 3 includes:

[0020] The generator of the generative adversarial network is optimized and trained through the cosine decay scheduler and adaptive moment estimation optimizer to generate augmented data;

[0021] The augmented data and the original dataset are mixed to form a mixed dataset.

[0022] Furthermore, the operation of optimizing and training the generator of the generative adversarial network to generate augmented data includes:

[0023] Set the initial learning rate of the adaptive moment estimation optimizer and generate augmented data through the generator of the generative adversarial network;

[0024] A cosine decay scheduler is used to gradually reduce the learning rate of the adaptive moment estimation optimizer, generating new augmented data during each training iteration of the generator of the generative adversarial network;

[0025] After reaching the set number of iterations, the final expanded data is obtained;

[0026] Further, the operation of gradually reducing the learning rate of the adaptive moment estimation optimizer by using the cosine decay scheduler includes:

[0027] The learning rate of the adaptive moment estimation optimizer is dynamically adjusted according to the following formula:

[0028]

[0029] Among them, η max represents the initial learning rate, η min Indicates the minimum learning rate, set to 0, t indicates the current training iteration round, T max Indicates the total number of training iterations.

[0030] Furthermore, the operation of step 4 includes:

[0031] S41 uses pipeline optimization tools to perform regression prediction training based on the original dataset;

[0032] S42 uses pipeline optimization tools to perform regression prediction training based on mixed datasets;

[0033] S43 obtains the most suitable glass transition temperature prediction model and model parameter combination.

[0034] Furthermore, the operation of step 5 includes:

[0035] S51 performs normalization preprocessing on the solution-polymerized styrene-butadiene rubber structure variables in the original data set to form normalized solution-polymerized styrene-butadiene rubber structure variables;

[0036] S52 inputs the normalized solution-polymerized styrene-butadiene rubber structural variables and the known glass transition temperatures in the original data set into the glass transition temperature prediction model for training;

[0037] S53 inputs the seven variable data of the solution-polymerized butadiene styrene rubber structure to be predicted into the trained glass transition temperature prediction model and outputs the glass transition temperature.

[0038] Furthermore, the glass transition temperature prediction model is an extreme random tree regression model.

[0039] Compared with the existing technology, the present invention establishes a quantitative structure-property relationship (QSPR) model to predict the glass transition temperature (Tg) of solution-polymerized styrene-butadiene rubber, and on this basis establishes an automatic machine learning regression prediction model to explore the influence of seven variables of solution-polymerized styrene-butadiene rubber on the glass transition temperature, so as to promote the structural design and synthesis of solution-polymerized styrene-butadiene rubber, thereby improving the prediction accuracy and reducing manual intervention, thereby accelerating the discovery and design cycle of innovative materials.

[0040] The present invention explores the effects of seven variables of solution-polymerized styrene-butadiene rubber (SSBR): styrene (St) mass fraction, 1,2-butadiene (12Bd) mass fraction, cis-1,4-butadiene (C14Bd) mass fraction, trans-1,4-butadiene (T14Bd) mass fraction, weight-average relative molecular weight (Mw), number-average relative molecular weight (Mn) and polydispersity index (PDI) on its Tg by establishing an automatic machine learning regression prediction model. The model can effectively utilize a limited feature set and a small data set to promote the structural design and synthesis of solution-polymerized styrene-butadiene rubber (SSBR), thereby improving prediction accuracy and reducing manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Automated machine learning regression prediction model prediction flow chart.

[0042] Figure 2 The overall framework structure diagram of the generative adversarial network.

[0043] Figure 3 Cosine decay learning rate curve.

[0044] Figure 4 Feature importance analysis plot of feature descriptors for the ETR model. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below with reference to the accompanying drawings:

[0046] This study primarily analyzes the effects of seven variables in solution-polymerized styrene-butadiene rubber (SSBR): styrene (St) mass fraction, 1,2-butadiene (12Bd) mass fraction, cis-1,4-butadiene (C14Bd) mass fraction, trans-1,4-butadiene (T14Bd) mass fraction, weight-average molecular weight (Mw), number-average molecular weight (Mn), and polydispersity index (PDI) on the glass transition temperature (T) of solution-polymerized styrene-butadiene rubber (SSBR). Given the lack of comprehensive data on SSBR in current databases, a raw dataset containing 68 data points was compiled based on the collected data on the seven structural variables of SSBR. Table 1 provides a partial set of data from this dataset. The complete dataset is detailed in Table 5.

[0047] Table 1. Partial data of glass transition temperature of solution styrene butadiene rubber (SSBR)

[0048]

[0049] like Figure 1 As shown, this is a prediction flow chart of the automatic machine learning regression prediction model of the present invention, and its prediction process includes four steps.

[0050] 1. Build an automatic machine learning regression prediction model

[0051] The automatic machine learning regression prediction model is trained and generated by improving the generative adversarial network (GAN) to train and generate a glass transition temperature prediction model in the machine learning tool, namely the pipeline optimization tool (TPOT), that is, the regression model consists of the generative adversarial network (GAN) and the pipeline optimization tool (TPOT).

[0052] Among them, the Generative Adversarial Network (GAN) consists of a generator and a discriminator. Specifically, the task of the generator is to generate realistic data samples from random noise, while the task of the discriminator is to distinguish the samples generated by the generator from the original data samples, such as Figure 2 shown.

[0053] The layer configuration of the generative adversarial network and the functional objectives of each layer are shown in Table 2.

[0054] Table 2. Layer configurations of the generator and discriminator of GAN.

[0055]

[0056] 2. Amplify the original data to form a mixed data set

[0057] Generative adversarial networks are used to amplify the original dataset, generating high-quality augmented data samples. The adversarial game between the generator and the discriminator is used to optimize the training of the generative adversarial network, ultimately amplifying the original data samples to form a hybrid dataset.

[0058] Specifically, the optimization training process involves repeatedly iteratively optimizing the parameters of the generator and discriminator to achieve a distribution that approximates the real data distribution as closely as possible. During this process, the generator and discriminator gradually improve their performance through alternating optimizations, ultimately reaching a dynamic equilibrium. Ultimately, the generator is able to generate realistic data samples, making it impossible for the discriminator to accurately distinguish between the augmented data generated by the generator and the real data samples, thereby amplifying the original data samples to form a mixed dataset.

[0059] The specific process is as follows:

[0060] The present invention uses a cosine decay scheduler and an adaptive moment estimation (Adam) optimizer to optimize the training of the generator of a generative adversarial network (GAN). Specifically, the cosine decay scheduler gradually reduces the learning rate, making the training process smoother and effectively preventing overfitting. At the same time, the adaptive learning rate mechanism of the adaptive moment estimation (Adam) optimizer can quickly adapt to the gradient changes of different parameters, improving training efficiency and generation quality.

[0061] The specific implementation includes: when training a generative adversarial network (GAN), the initial learning rate is set high, and as the number of training iterations increases, the learning rate is gradually reduced through a cosine decay scheduler. This dynamic adjustment mechanism not only ensures rapid convergence of the model in the early stages, but also allows for fine-tuning in the later stages, thereby optimizing the learning rate management strategy during training and optimizing the generative adversarial network (GAN) generator through parameter updates during iterative training. Specifically, the optimization process includes defining the generator's network architecture and loss function. After each training process, the cosine decay scheduler automatically adjusts the learning rate in the optimizer based on the number of repetitions of the current training process (i.e., the number of epochs), and updates the parameters through the backpropagation algorithm.

[0062] Specifically, it is dynamically adjusted according to the following formula:

[0063]

[0064] Here, η max represents the initial learning rate, set to 0.001, η min Indicates the minimum learning rate, set to 0, t indicates the current training iteration round, T max Indicates the total number of training iterations, set to 10000.

[0065] Furthermore, the adaptive moment estimation (Adam) optimizer with a learning rate of 0.00015 and a beta_1 parameter of 0.5 helps to make the gradient estimation more accurate during the optimization process, thereby improving the training efficiency and convergence speed.

[0066] In the early stage of GAN optimization training, a higher learning rate is started to promote rapid convergence, and then the learning rate is gradually reduced to fine-tune the GAN parameters to improve its generalization ability.

[0067] Specifically, the main adjustments to the parameters in the Generative Adversarial Network (GAN) include the following aspects:

[0068] (1) Generator Parameters

[0069] Weights and Biases: The weights and biases of each generator network layer. These parameters directly affect how the generator converts input noise into realistic data samples.

[0070] Regularization Parameters: These parameters, such as those of weight decay or batch normalization layers, help control the generalization ability and training stability of the generator.

[0071] (2) Discriminator Parameters

[0072] Weights and Biases: The weights and biases of each discriminator network layer. These parameters determine how the discriminator distinguishes between real data and generated data.

[0073] Regularization Parameters: They are also factors that control the generalization ability and training stability of the discriminator.

[0074] (3) Optimizer Parameters

[0075] Learning rate: This is a key parameter used by the optimizer to adjust the step size of each parameter update. In the cosine decay scheduler, changes in the learning rate affect the convergence speed during training and the final performance of the model.

[0076] Momentum and Decay Rates: The Adam optimizer also involves parameters such as momentum and decay rate, which affect the acceleration and direction of the optimizer when updating parameters.

[0077] By fine-tuning these parameters, especially adjusting the weights of the generator and discriminator, the learning rate and momentum of the optimizer, the training effect of the generative adversarial network and the quality of the generated data are optimized.

[0078] During the training process, the batch size is set to 64. The learning rate decay curve is as follows Figure 3 As shown in the figure, the sine decay learning rate curve dynamically adjusts the learning rate during training: Initially, the learning rate is 0.001. At the beginning of training, the model uses a higher learning rate to accelerate convergence. As training progresses, the cosine decay scheduler gradually reduces the learning rate until the minimum learning rate reaches 0. This dynamic adjustment mechanism enables the model to quickly converge to a local optimal solution in the early stages of training, and to make fine adjustments as it approaches the optimal solution. Figure 3 The horizontal axis represents the number of training rounds, and the vertical axis represents the corresponding learning rate value.

[0079] During the optimized training of a generative adversarial network (GAN), the generator generates new augmented data at each iteration. These data samples change with each iteration to gradually approximate the true data distribution. This training is repeated until, after the set number of iterations, the original data set of 1568 is replaced by the final 1500 augmented data. This augmented data is then mixed with the original data to form a mixed dataset.

[0080] 3. Obtain glass transition temperature prediction model and parameter combination

[0081] The glass transition temperature prediction model and parameter combination are obtained using the Pipeline Optimization Tool (TPOT). The specific steps are as follows:

[0082] S311. Regression prediction training on the original dataset: Model training and optimization are performed using the Pipeline Optimization Tool (TPOT) based on the original dataset. This process utilizes existing real raw data to perform predictions and understand the relationships between variables and their impact on the target variable. The original dataset refers to the 68 data sets listed in Table 1, which are actual, collected data.

[0083] S312. Regression prediction training on a mixed dataset: The same process as S311 is applied to the mixed dataset using the pipeline optimization tool (TPOT). The mixed dataset contains both the generated and original datasets. By merging the generated data with the existing data, the model can more comprehensively learn and understand the patterns and features in the data. Specifically, 80% of the data is randomly selected from the mixed dataset for training, and the remaining 20% is used for model validation and testing. This ensures that the model is exposed to a sufficient number of diverse samples during training, thereby improving the model's generalization ability and predictive performance.

[0084] S313, Pipeline Optimization Tool (TPOT) automatically searches and optimizes machine learning pipelines through genetic programming, specifically including feature selection and preprocessing steps, in order to obtain the model and parameter combination that best suits the current regression prediction task.

[0085] Table 3 shows the hyperparameter settings of the pipeline optimization tool (TPOT). Hyperparameters are parameters whose values are set before the learning process begins. They include: population size, mutation rate, crossover rate, crossover rate, and cross-validation folds.

[0086] Table 3. Hyperparameters used in the TPOT regression model.

[0087]

[0088]

[0089] All the models mentioned above are implemented in Python 3.9.

[0090] Processing of the merged data set: After generating 1500 augmented samples and merging them with the 1568 original data, the pipeline optimization tool (TPOT) was used to perform regression prediction to select the optimal glass transition temperature prediction model.

[0091] Optimal model selection: The Pipeline Optimization Tool (TPOT) automatically selects and optimizes the Extreme Tree Regression (ETR) model as the glass transition temperature prediction model. This automated search and optimization process reduces human intervention and improves the scientific nature and accuracy of model selection.

[0092] 4. Make predictions using the glass transition temperature prediction model.

[0093] The Extreme Randomized Trees Regression (ETR) model constructs multiple decision trees using random subsets of features and random split points; the final regression output is then generated by averaging the predictions of these trees. Figure 4 The feature importance rankings derived from the models are shown, determined by the mean absolute value of the SHAP scores.

[0094] Among them, (a) the histogram summarizes the importance of features and ranks the features according to the average absolute value of SHAP; (b) the density scatter plot also summarizes the importance of features.

[0095] Based on the original and mixed data sets, it was observed that among the seven variables, 12Bd had the greatest effect on the glass transition temperature (Tg) of solution styrene butadiene rubber (SSBR), while the polydispersity index (PDI) of the polymer had the least effect. Figure 4 The density scatter plot in , plots 1568 sample points. Figure 4As shown in the figure, a larger area indicates a higher density of points, where the color of the points represents the size of the eigenvalue, with red indicating high values and blue indicating low values. As can be seen from the figure, T14Bd, C14Bd, and PDI show negative correlations, while other variables show positive correlations.

[0096] Generally, more flexible polymer chains are associated with lower glass transition temperatures (Tg). Increasing the 12Bd content enhances polymer chain rigidity and intermolecular forces, leading to higher glass transition temperatures (Tg). Increasing the styrene content increases polymer chain rigidity and promotes overlap and interaction of electron clouds between benzene rings. This restricts chain mobility, leading to higher Tg. An increase in the average molecular weight (Mw) divided by the number average molecular weight (Mn) generally correlates with increased polymer chain length. Longer polymer chains enhance intermolecular forces and molecular density, increasing polymer robustness and order, thereby raising the glass transition temperature (Tg). The isolated double bonds in the C14Bd structure enhance polymer chain flexibility, leading to lower glass transition temperatures (Tg) with increasing C14Bd content. In contrast, T14Bd exhibits additional branching, resulting in irregular arrangement of rubbery polymer chains. Therefore, T14Bd contributes more significantly to the glass transition temperature (Tg) than C14Bd, with higher T14Bd content corresponding to lower glass transition temperatures (Tg). A narrower PDI generally results in a higher glass transition temperature (Tg) because the consistent chain length and minimal molecular weight difference promote the formation of a more ordered structure, restricting chain mobility and raising the glass transition temperature (Tg).

[0097] The specific prediction process is:

[0098] In the first step, Min-Max normalization preprocessing was performed on the seven variables, i.e., seven features, of solution styrene butadiene rubber (SSBR) in the original data set. The values of the data features were scaled to the range [0, 1]. That is, the solution styrene butadiene rubber structural variables whose data feature values were normalized within the range [0, 1] were obtained. Mapping the new solution styrene butadiene rubber structural variables to the same scale helped to eliminate the deviations caused by different dimensions between different features.

[0099] In the second step, the normalized solution-polymerized styrene-butadiene rubber structural variables and the existing glass transition temperature data in the original data set are input into the extreme randomized tree regression (ETR) model for training. During the training process, the model learns the complex relationship between the input features, namely the seven variables, and the target variable (i.e., glass transition temperature), so as to realize the prediction of the glass transition temperature of the unknown solution-polymerized styrene-butadiene rubber structural variable data.

[0100] The third step is to use the trained extreme random tree regression (ETR) model for prediction, that is, input the seven variable data into the extreme random tree regression (ETR) model and output the value of the glass transition temperature (Tg).

[0101] The specific prediction process and effects are described below through examples.

[0102] Example 1: Example of a Generative Adversarial Network consisting of a Generator and a Discriminator

[0103] The goal of the generator is to minimize the cross entropy between the generator and the discriminator, which represents the probability that the samples generated by the generator are misclassified as false samples by the discriminator. The objective function of the generator can be expressed as:

[0104]

[0105] The goal of the discriminator is to maximize its correct classification probability for real data samples and maximize its misclassification probability for samples generated by the generator. The objective function of the discriminator can be expressed as:

[0106]

[0107] Formula (2) and Formula (3) describe the optimization process in the Generative Adversarial Network, which is a zero-sum game between the generator and the discriminator. By combining the objective functions of the generator and the discriminator, we can obtain the overall objective function of the GAN. The optimization problem of the GAN is a minimax problem, and the objective function can be described as:

[0108]

[0109] In formulas (2) to (4), G represents the parameters of the generator, D represents the parameters of the discriminator, and p dara (x) represents the distribution of real data, p z (z) represents the distribution of the latent space z, x represents the real data sample, G(z) represents the data sample generated by the generator, D(x) represents the probability that the discriminator recognizes the real data sample x as real, and D(G(z)) represents the probability that the discriminator recognizes the sample G(z) generated by the generator as real.

[0110] Example 2: Specific Example for Verifying the Prediction Effect of Automatic Machine Learning Regression Model

[0111] In this embodiment, the model performance evaluation based on the original data set is adopted. Specifically, in this embodiment, a regression model based on the pipeline optimization tool (TPOT) is used to perform regression prediction on the original data set. The original data set is selected as the basis because it represents a data sample in the real world, and comparative analysis of the performance of different regression models on the same data set can provide objective evaluation indicators. The original data set is used for regression prediction training, and the test set is 20% of the data randomly extracted from the original data set as the test set. The random search method is used for training, and the 10-fold cross-validation method is used for evaluation. In order to ensure the stability of the experiment, six methods, including the automatic machine learning regression model of the pipeline optimization tool (TPOT), were tested 10 times on different data partitions to ensure the fairness and accuracy of the comparison. The results are shown in Table 4.

[0112] Table 4. Evaluation results of six machine learning regression methods.

[0113]

[0114] Among all models, the optimized model (i.e., the model using the pipeline optimization tool) performed the best. These results show that the pipeline optimization tool (TPOT) has superior prediction accuracy and data interpretability compared to other manually selected regression models.

[0115] Table 5. Complete data on the glass transition temperature of solution styrene butadiene rubber (SSBR)

[0116]

[0117]

[0118] The above technical solution is only one embodiment of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the principles disclosed in the present invention, and it is not limited to the technical solution described in the above specific embodiments of the present invention. Therefore, the above description is only preferred and does not have a restrictive meaning.

Claims

1. A polymer material performance prediction method based on generative adversarial networks, characterized in that: The method uses a generative adversarial network to perform data amplification on a collected original data set to obtain a mixed data set, and uses a pipeline optimization tool to obtain a glass transition temperature prediction model suitable for the mixed data set. Finally, the glass transition temperature is predicted using the glass transition temperature prediction model.

2. The polymer material performance prediction method based on generative adversarial network according to claim 1, characterized in that: The method comprises: Step 1: Collect data on seven variables of the solution-polymerized styrene-butadiene rubber structure and compile an original data set; Step 2: Build an automated machine learning regression prediction model consisting of a generative adversarial network and pipeline optimization tools; Step 3: Use the generative adversarial network to amplify the original dataset to form a mixed dataset; Step 4: Use the mixed data set to train the automatic machine learning regression prediction model to obtain a glass transition temperature prediction model; Step 5: Input the data of the seven variables of the solution-polymerized styrene-butadiene rubber to be predicted into the glass transition temperature prediction model, and the glass transition temperature prediction model outputs the polymer glass transition temperature.

3. The polymer material performance prediction method based on generative adversarial network according to claim 2, characterized in that: The generative adversarial network consists of a generator and a discriminator.

4. The polymer material performance prediction method based on generative adversarial network according to claim 3, characterized in that: The operation of step 3 includes: The generator of the generative adversarial network is optimized and trained through the cosine decay scheduler and adaptive moment estimation optimizer to generate augmented data; The augmented data and the original dataset are mixed to form a mixed dataset.

5. The polymer material performance prediction method based on generative adversarial network according to claim 4, characterized in that: The operation of optimizing and training the generator of the generative adversarial network to generate augmented data includes: Set the initial learning rate of the adaptive moment estimation optimizer and generate augmented data through the generator of the generative adversarial network; A cosine decay scheduler is used to gradually reduce the learning rate of the adaptive moment estimation optimizer, generating new augmented data during each training iteration of the generator of the generative adversarial network; After reaching the set number of iterations, the final augmented data is obtained.

6. The polymer material performance prediction method based on generative adversarial network according to claim 5, characterized in that: The operation of gradually reducing the learning rate of the adaptive moment estimation optimizer by using the cosine decay scheduler includes: The learning rate of the adaptive moment estimation optimizer is dynamically adjusted according to the following formula: Among them, η max represents the initial learning rate, η min Indicates the minimum learning rate, set to 0, t indicates the current training iteration round, T max Indicates the total number of training iterations.

7. The polymer material performance prediction method based on generative adversarial network according to claim 2, characterized in that: The operation of step 4 includes: S41 uses pipeline optimization tools to perform regression prediction training based on the original dataset; S42 uses pipeline optimization tools to perform regression prediction training based on mixed datasets; S43 obtains the most suitable glass transition temperature prediction model and model parameter combination.

8. The polymer material performance prediction method based on generative adversarial network according to claim 2, characterized in that: The operation of step 5 includes: S51 performs normalization preprocessing on the solution-polymerized styrene-butadiene rubber structure variables in the original data set to form normalized solution-polymerized styrene-butadiene rubber structure variables; S52 inputs the normalized solution-polymerized styrene-butadiene rubber structural variables and the known glass transition temperatures in the original data set into the glass transition temperature prediction model for training; S53 inputs the seven variable data of the solution-polymerized butadiene styrene rubber structure to be predicted into the trained glass transition temperature prediction model and outputs the glass transition temperature.

9. The polymer material performance prediction method based on generative adversarial network according to claim 8, characterized in that: The glass transition temperature prediction model is an extreme random tree regression model.

Images