A prediction method based on the glass transition temperature of fluororubber

Abstract

The present invention provides a prediction method based on the glass transition temperature of fluororubber, which relates to the technical field of material parameter measurement. The method includes: obtaining the data of the fluororubber to be measured; inputting the data of the fluororubber to be measured into the final glass transition temperature prediction model to obtain the prediction result of the transition temperature. The present invention solves the problems of too high cost of measuring Tg in the prior art and limited methods for directly predicting the material Tg by using basic formulation and process parameters, with low prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of material parameter measurement, and particularly to a prediction method based on the glass transition temperature of fluororubber. Background Art

[0002] FKM is a type of high molecular elastomer polymerized from fluoroolefins. Its unique comprehensive properties such as high temperature resistance, oil resistance, aging resistance, acid and alkali resistance make it play an indispensable role in fields such as aerospace, electronic devices, and chemical engineering. However, due to its chemical structure, FKM has poor low temperature resistance. In practical applications, the brittle temperature of ordinary FKM is above -30°C, which greatly limits its use in extreme working conditions. Facing the new requirements of application scenarios, it is urgent to improve the low temperature performance of FKM. In the process of developing and researching FKM, accurately evaluating its low temperature resistance is crucial. Traditionally, the evaluation of such properties relies on multiple test parameters, including brittle temperature, compression cold resistance coefficient, low temperature retraction temperature, and Tg. Especially in the aerospace field, in the initial research, the brittle temperature and compression cold resistance coefficient are mostly used to characterize the low temperature resistance of materials. However, the determination of these parameters involves mixing FKM raw rubber with materials such as additives and carbon black, and performing vulcanization treatment to prepare vulcanized rubber. This process is not only complex, but also has high labor and time costs. Although Tr10 can provide information about the performance of materials in low temperature environments, the test cycle of this method is long and the energy consumption is high, which is particularly inconvenient in research with limited resources and time. In contrast, the measurement of Tg provides a more efficient option. Tg can be directly measured from FKM raw rubber (a white to light yellow solid) after the polymerization process, greatly reducing energy consumption and reducing operation complexity. Since Tg measurement can save valuable time and resources, and is simple to operate and cost-effective, it has become the preferred index for evaluating the low temperature resistance of FKM. In addition, Tg is not only an important index for measuring the application performance of rubber materials in low temperature environments, but also directly determines the practical temperature range of rubber. Therefore, in the development process, accurately measuring and controlling Tg is crucial to ensure that FKM meets the actual application requirements.

[0003] Tg is a key parameter of rubber materials, which marks the critical point at which rubber transforms from a hard and brittle glassy state to a soft and elastic rubbery state. When the temperature is below Tg, the rubber loses its elasticity, becomes hardened and cannot be used effectively. This transition temperature is not only the critical point at which the rubber chain segments start or stop moving, but is also significantly affected by environmental conditions. Commonly used methods for measuring Tg include differential scanning calorimetry (DSC), dynamic mechanical analysis (DMA) and static thermomechanical analysis (TMA). These techniques can characterize the thermal, mechanical and viscoelastic properties of materials. However, these techniques rely on expensive equipment and a large number of samples, with relatively high research and development and time costs. It is relatively difficult to use these methods in a laboratory with only ordinary equipment and a small amount of sample.

[0004] At the same time, extensive research has been carried out on combining machine learning models to predict Tg. Existing research has demonstrated the effectiveness of machine learning in predicting the Tg of materials and has achieved remarkable results in analyzing the chemical composition, molecular topological structure and chemical descriptors of formulation monomers. However, in the process of material development, truly optimizing material properties requires precise control of formulation design and process parameters. Currently, there is still relatively little research on directly predicting the Tg of materials using basic formulations and process parameters, and the prediction accuracy is relatively low. Summary of the Invention

[0005] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a prediction method based on the glass transition temperature of fluororubber. The present invention solves the problems of the excessively high cost of measuring Tg in the prior art and the limited methods for directly predicting the Tg of materials using basic formulations and process parameters, with low prediction accuracy.

[0006] To achieve the above purpose, the present invention provides the following solution:

[0007] A prediction method based on the glass transition temperature of fluororubber, comprising:

[0008] Obtaining data of the fluororubber to be measured;

[0009] Inputting the data of the fluororubber to be measured into the final glass transition temperature prediction model to obtain a predicted transition temperature result;

[0010] The construction method of the final glass transition temperature prediction model is as follows:

[0011] Collecting fluororubber experimental data, where the fluororubber experimental data includes: process parameters and process formulations;

[0012] Cleaning the fluororubber experimental data to obtain the cleaned data;

[0013] Based on the comprehensive analysis method, align the cleaned data, determine the defining indicators of each process parameter, and determine the characteristic sequence of the cleaned data according to the defining indicators;

[0014] Screen the characteristic sequence to obtain the screened characteristic sequence;

[0015] Construct a first glass transition temperature prediction sub-model, a second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model respectively based on a regressor, a deep learning algorithm, and a clustering analysis algorithm and according to the screened characteristic sequence;

[0016] Determine the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model.

[0017] Preferably, the process formula includes:

[0018] Monomers and additives, wherein the monomers include: tetrafluoroethylene, perfluoromethyl vinyl ether, hexafluoropropylene, and vinylidene fluoride.

[0019] Preferably, the process parameters include:

[0020] Temperature, pressure, feed rate, and reaction medium.

[0021] Preferably, perform data cleaning on the fluororubber experimental data to obtain the cleaned data, including:

[0022] Determine the monomer content value not included in the process formula and fill the monomer content value with zero to obtain the first sub-cleaned data;

[0023] Determine the missing values in the process parameters that cannot be aligned with the original formula or the final performance and delete the missing values to obtain the second sub-cleaned data;

[0024] Determine the cleaned data according to the first sub-cleaned data and the second sub-cleaned data.

[0025] Preferably, the based on the comprehensive analysis method, align the cleaned data, determine the defining indicators of each process parameter, and determine the characteristic sequence of the cleaned data according to the defining indicators, includes:

[0026] Determine the matrix sequence of the cleaned data;

[0027] Obtain a process sequence by performing process parameter estimators on the matrix sequence;

[0028] Obtain a recombinant sequence according to the process sequence;

[0029] Determine the characteristic sequence of the data after cleaning according to the recombinant sequence and the defined index.

[0030] Preferably, screening the characteristic sequence to obtain a screened characteristic sequence, including:

[0031] Determine the data correlation according to the Pearson correlation coefficient calculation formula, where the data correlation includes: the correlation coefficient value between the feature and the performance and the correlation coefficient value between different features;

[0032] Characterize the data correlation according to the heat map to obtain a correlation characterization heat map;

[0033] Obtain a screened characteristic sequence according to the correlation characterization heat map and the recursive feature elimination algorithm.

[0034] Preferably, determining the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model and the third glass transition temperature prediction sub-model includes:

[0035] Determine the first transition temperature prediction sub-result, the second transition temperature prediction sub-result and the third transition temperature prediction sub-result according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model and the third glass transition temperature prediction sub-model respectively;

[0036] Evaluate the first transition temperature prediction sub-result, the second transition temperature prediction sub-result and the third transition temperature prediction sub-result respectively to obtain a first evaluation result, a second evaluation result and a third evaluation result;

[0037] Determine the weights of each prediction sub-model according to the first evaluation result, the second evaluation result and the third evaluation result and perform weighting to determine the final glass transition temperature prediction model.

[0038] The present invention discloses the following technical effects:

[0039] The present invention provides a prediction method based on the glass transition temperature of fluororubber, including: obtaining the data of the fluororubber to be measured; inputting the data of the fluororubber to be measured into the final glass transition temperature prediction model to obtain the prediction result of the transition temperature; the construction method of the final glass transition temperature prediction model is: collecting the experimental data of fluororubber, and the experimental data of fluororubber includes: process parameters and process formulations; cleaning the experimental data of fluororubber to obtain the cleaned data; based on the comprehensive analysis method, aligning the cleaned data to determine the definition indexes of each process parameter and determining the characteristic sequence of the cleaned data according to the definition indexes; screening the characteristic sequence to obtain the screened characteristic sequence; respectively constructing the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model and the third glass transition temperature prediction sub-model based on the regressor, the deep learning algorithm and the clustering analysis algorithm and according to the screened characteristic sequence; determining the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model and the third glass transition temperature prediction sub-model. The present invention considers the formulation design and process parameters to directly predict the material Tg and uses multiple models to predict the results, which can more widely target various types of input data to ensure the best prediction accuracy. Description of the Drawings

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0041] Figure 1 It is a flowchart of a prediction method based on the glass transition temperature of fluororubber provided by an embodiment of the present invention;

[0042] Figure 2 It is a schematic diagram of the experimental data categories provided by an embodiment of the present invention. Detailed Embodiments

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0044] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] As Figure 1 shown, the present invention provides a prediction method based on the glass transition temperature of fluororubber, including:

[0046] Step 100: Obtain the data of the fluororubber to be measured;

[0047] Step 200: Input the data of the fluororubber to be measured into the final glass transition temperature prediction model to obtain the prediction result of the transition temperature;

[0048] The construction method of the final glass transition temperature prediction model is as follows:

[0049] Step 201: Collect the experimental data of fluororubber, and the experimental data of fluororubber includes: process parameters and process recipes;

[0050] Step 202: Clean the experimental data of the fluororubber to obtain the cleaned data;

[0051] Step 203: Based on the comprehensive analysis method, align the cleaned data, determine the definition indexes of each process parameter, and determine the characteristic sequence of the cleaned data according to the definition indexes;

[0052] Step 204: Screen the characteristic sequence to obtain the screened characteristic sequence;

[0053] Step 205: Construct the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model respectively based on the regressor, deep learning algorithm, and clustering analysis algorithm and according to the screened characteristic sequence;

[0054] Step 206: Determine the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model.

[0055] Further, the process recipe includes:

[0056] Monomers and additives, wherein the monomers include: tetrafluoroethylene, perfluoromethyl vinyl ether, hexafluoropropylene, and vinylidene fluoride. The process parameters include: temperature, pressure, feeding amount, and reaction medium.

[0057] Specifically, the data in the actual production experiments of the present invention were systematically sorted out, including 613 experimental record data, which detailedly reflected the experimental values of the monomer components of the formulation, the process, and the performance, etc. According to the confidentiality agreement, the data were encoded with alphanumeric characters. There were a total of 4 types of formulation monomers and 14 process parameters. Each record represented the material type used for producing FKM, the process record parameters, and the physical and chemical properties of the product produced under the formulation monomer components and process control conditions. For example, Figure 2 were the FKM data collected in the experiments. The formulation monomers mainly included four fluorinated olefins: tetrafluoroethylene (TFE), perfluoromethyl vinyl ether (PMVE), hexafluoropropylene (HFP), and vinylidene fluoride (VDF). The process parameters included temperature (gas phase temperature of the reaction kettle, liquid phase temperature of the reaction kettle, reactant feed temperature), pressure (reaction kettle pressure, reactant feed pressure), feed amount, reaction medium, and additives, etc. The performance data included Mooney viscosity and Tg. The original proportion ranges of the four monomers fluctuated little, and HFP was not used in the basic formulation. The maximum value of Tg was -17 °C, and the minimum value was -40 °C.

[0058] Furthermore, the experimental data of the fluororubber were cleaned to obtain the cleaned data, including:

[0059] Determine the monomer content values not included in the process formulation and fill the monomer content values with zero to obtain the first sub-cleaned data;

[0060] Determine the missing values in the process parameters that cannot be aligned with the original formulation or the final performance and delete the missing values to obtain the second sub-cleaned data;

[0061] Determine the cleaned data based on the first sub-cleaned data and the second sub-cleaned data.

[0062] Specifically, the original dataset was collected through manual records. To ensure the quality of the data and the final prediction effect, this study adopted the strategy of deleting or filling the missing values in the formulation, process, and performance data. For the formulation data, to maintain the consistency of the combined formulation and the integrity of the feature dimensions, we chose to fill the monomer content values not included in the original formulation with zero. This can ensure that the data of various formulations have the same feature structure. As for the process data, we noticed that only a small amount of data could not be aligned with the original formulation or the final performance results. On the premise of considering the model prediction effect, these missing values were deleted to avoid inaccurate data caused by filling. We evaluated the overall distribution and fluctuations of the data through parallel coordinates and accurately identified outliers under the guidance of domain experts. For these incorrect values, replacement and correction measures were taken. For those outliers beyond the reasonable fluctuation range, deletion was selected to ensure that it would not affect the training and prediction accuracy of the model. In the development of FKM, in addition to the key monomer ratio, process parameters such as the reactor pressure and temperature are also crucial. Due to the confidentiality of the experimental data, the collected process parameters lack continuous time-series information, and the types of process parameters collected in each process stage are different, resulting in data imbalance and the inability to determine the key definition indicators of the parameters. The data imbalance problem here refers to the inconsistent number of records and the inconsistent types of process parameters recorded, and the inability to determine the parameter definition indicators means that not every record data affects the final result, so the specific impact of each data cannot be determined. Both of the above problems may lead to inaccurate prediction of the final value. Therefore, this study adopted a comprehensive analysis method, considered the dynamic changes of temperature and pressure in the whole preparation process, and after cleaning the data, data alignment was performed to solve the above problems. Through in-depth communication with the FKM research and development team and based on their professional experience, the fluctuations of the liquid phase temperature in the reactor, the reactant feeding temperature, the reactor pressure, and the reactant feeding pressure during the experimental process were analyzed in detail. The same type of process parameters in each test stage and the key parameters recommended by the researchers were extracted, and the statistics of each parameter were calculated as the definition indicators of each batch of data. This process is called alignment.

[0063] Furthermore, based on the comprehensive analysis method, data alignment is performed on the cleaned data, the definition indicators of each process parameter are determined, and the feature sequence of the cleaned data is determined according to the definition indicators, including:

[0064] Determine the matrix sequence of the cleaned data;

[0065] Obtain the process sequence by estimating the process parameters of the matrix sequence;

[0066] Obtain the recombinant sequence according to the process sequence;

[0067] Determine the characteristic sequence of the data after cleaning according to the recombination sequence and the defined index.

[0068] Specifically, taking the original process data as the matrix sequence S = {C1, C2, C3, …, Cn}, where the matrix C can be expressed as the formula;

[0069]

[0070] Among them, Ti-Pi represents various process parameters recorded in each experiment. Ti represents the liquid phase temperature of the reactor, and Pi represents the pressure inside the reactor. Extract the process parameter estimators (minimum value, maximum value, quartile, mean, variance, mode) of this matrix to obtain the process sequence E = {Ti_min, Ti_max, Ti_q1, Ti_q2, …, Pi_q3, Pi_mean, Pi_std, Pi_href}. After alignment, there are 32 process estimators. After encoding the estimator names, the recombination sequence F = {D1, D2, D3, …, Dn} is obtained, where D can be expressed as the formula D = {f7, f8, f9, …, f38};

[0071] The recombination sequence F is initially determined as the process characteristic parameters for prediction. Corresponding the process data with the formula ratio according to the product batch, the characteristic sequence X = {f1, f2, f3, …, f38} is obtained, which includes the product type (f1), product batch number (f2), 4 monomers (f3-f6), 8 liquid phase temperatures of the reactor (f7-f14), 8 reactor pressures (f15-f22), 8 reactant feed temperatures (f23-f30), and 8 reactant feed pressures (f31-f38).

[0072] Furthermore, screen the characteristic sequence to obtain the screened characteristic sequence, including:

[0073] Determine the data correlation according to the Pearson correlation coefficient calculation formula. Among them, the data correlation includes: the correlation coefficient value between the characteristic and the performance and the correlation coefficient value between different characteristics;

[0074] Characterize the data correlation according to the heat map to obtain the correlation characterization heat map;

[0075] Obtain the screened characteristic sequence according to the correlation characterization heat map and the recursive feature elimination algorithm.

[0076] Specifically, the multiplicity of process parameters results in the selection of multi-combination features. A reasonable selection depends to a large extent on the domain knowledge, experience, and prudence of researchers. Not all process parameters are important for Tg prediction in the experiment. Unrelated parameters often lead to an increase in the dimensionality of input features, thereby reducing the prediction ability of the model. Therefore, it is crucial to screen the optimal subset of features. To increase the dimensionality of features and mitigate overfitting, the statistical features of process parameters are combined. Alignment has been done during the data preprocessing process. Subsequently, the correlation coefficient method is used to further refine the feature selection process. Since the original process parameter data is non-continuous data, we use the Pearson correlation coefficient calculation formula:

[0077]

[0078] Calculate the correlation between each sequence feature and the final detection result Tg. Here, X is any process parameter, i.e., the input feature variable, Y is any prediction performance, i.e., the target variable, and E(X) is the mean of any process parameter feature column. The calculation formula is as follows. Similarly, E(Y) is the mean of the target variable;

[0079]

[0080] And σ X is the variance of any attribute feature column. The calculation formula is as follows. Similarly, σ Y is the variance of the target variable:

[0081]

[0082] Finally, calculate the covariance of any attribute feature and the target variable. The formula:

[0083]

[0084] Finally, substitute all the formulas into the original formula to get the following formula:

[0085]

[0086] Eliminate the feature elements with strong correlations through the correlation coefficient values, and retain the key features that affect Tg. To make the data more intuitive, the correlation coefficients are characterized by combining heatmaps. We initially screen out the features for modeling based on the correlation coefficient values between features and the coefficient values between features and properties. Thus, some estimators of f1 (product type), monomers (f3, f4, f5, f6), the liquid phase temperature of the reactor, the reactant feed temperature, the reactor pressure, and the reactant feed pressure are selected as the input features of the model. At the same time, during the model training process, to achieve automatic feature selection, we introduce feature importance analysis, so as to adopt the recursive feature elimination algorithm (RFE). By iteratively selecting the attributes with strong feature importance, recursively delete the features with low contribution in each stage until the required number of features is obtained.

[0087] Furthermore, in the production test, an engineering technology system usually divided into three stages is generally adopted, including small-scale tests, medium-scale tests, and the production test stage at the engineering scale. Although the same type of FKM products are produced in all three stages, there are significant differences in the monomer ratios in the formula and the records of process parameters. To fairly consider all variables and avoid model overfitting, using the data partitioning technique in the field of machine learning, 80% of the data is used as the training set to train and optimize the model, and the remaining 20% is used as the test set to prepare for the further training and verification of the model;

[0088]

[0089] Further, determining the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model includes:

[0090] Determine the first transition temperature prediction sub-result, the second transition temperature prediction sub-result, and the third transition temperature prediction sub-result according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model respectively;

[0091] Evaluate the first transition temperature prediction sub-result, the second transition temperature prediction sub-result, and the third transition temperature prediction sub-result respectively to obtain the first evaluation result, the second evaluation result, and the third evaluation result;

[0092] Determine the weights of each prediction sub-model according to the first evaluation result, the second evaluation result, and the third evaluation result and perform weighting to determine the final glass transition temperature prediction model.

[0093] Specifically, calculate the errors between each predicted sub-result and the actual observed value to obtain the first evaluation result, the second evaluation result, and the third evaluation result. Using the first, second, and third evaluation results, various methods can be employed to determine the weights of each predictive sub-model. For example, the weights can be determined by weighting according to factors such as the magnitude of the error and accuracy. A common method is to use regression analysis or machine learning algorithms to determine the weights of each predictive sub-model, so as to minimize the error between the prediction result of the final model and the actual observed value. Weight the results of each predictive sub-model according to the corresponding weights to obtain the result of the final glass transition temperature prediction model. Final prediction result = weight 1 * first transition temperature prediction sub-result + weight 2 * second transition temperature prediction sub-result + weight 3 * third transition temperature prediction sub-result. For individual special data, the weight can be 0, and the final glass transition temperature prediction model may be a combination of single or multiple sub-models.

[0094] Furthermore, after feature screening of the entire data set, there are 4 monomers, 15 process parameters, and product types as input feature variables. With Tg as the output variable.

[0095] Furthermore, in addition to using machine learning models to achieve accurate prediction of Tg, whether in experimental research or production processes, what is ultimately expected is to discover the key features that affect the Tg value. Therefore, the present invention designs a visualization framework for interpreting model results and exploring production processes. Focusing on feature importance and attribute feature space analysis, and transparent analysis of the model prediction process, this meticulous visualization strategy not only reveals the internal mechanisms of different prediction models, but also greatly enhances the transparency and credibility of the models, providing intuitive visual support for the correlation between the formulation-process parameters-performance of FKM materials.

[0096] To improve the transparency and understandability of the FKM performance prediction analysis and strengthen the basis for decision-making, feature importance analysis is particularly crucial. In the process of constructing a prediction model, each feature is regarded as a potential "driving factor" of the target variable. The feature importance value is a quantitative score of the influence of each feature in the sample, which directly reflects the contribution of the feature to the prediction of the target variable. To deeply explain the prediction ability of the model, we not only consider the direct effects between features, but also analyze the interactions within and between feature groups. For this purpose, we use a network heatmap tool (such as the one provided by https: / / www, omicshare, com) to explore these complex relationships in a visual way. On the one hand, calculate the Pearson correlation coefficient to reveal the correlation between features, and on the other hand, quantify the contribution of features to the model to comprehensively evaluate the importance of features. In addition, this tool can also reveal the internal connections and cross-effects between the prediction results and each feature;

[0097] Combining the heatmap and the network graph to deeply analyze the interactions between features and their impact on the prediction results, the within-group correlation heatmap in the upper right corner shows the Pearson correlation coefficients calculated during the feature selection process. In this heatmap, the horizontal and vertical axes correspond to different features respectively, and the color blocks in each cell are color-coded for the positive and negative of the correlation coefficient, while the size of the color block is proportional to the absolute value of the correlation coefficient.

[0098] In addition, the network connection graph further reveals the correlations between different features and the prediction results. Among them, the connecting lines link the prediction performance with each feature. The thickness of the lines represents the strength of the correlation, and the lines of different colors are used to distinguish the significance levels of feature importance.

[0099] The correlations of each feature in the ENR model when predicting Tg and their contribution degrees to the model prediction. In this network graph, the decisive roles played by the feature product type (f1), monomers VDF and PMVE (f3 and f5) in the prediction model are clearly identified by three purple lines. The remaining light blue lines also indicate the lowest correlation and contribution degree between the Tg value and the monomer TFE (f4). At the same time, the heatmap part highlights the sets with relatively strong correlations between features, several estimators of the reactor liquid phase temperature (f7, f9, f10, f11) and some estimators of the reactant feed temperature (f26, f27, f28, f29). This means that some of them will be given priority consideration when performing feature gradient elimination, and individual estimators will be retained. During the development process, the attention to the estimators of this process parameter can be appropriately reduced.

[0100] To better reveal the detailed changes during the experiment, we conducted an in-depth visual analysis of the two best-performing prediction models. We first analyzed the RFR model, which performs regression analysis by integrating the prediction results of multiple decision trees. To deeply understand the decision-making mechanism of the RFR model, we assigned weights to the best-performing decision trees and selected the subtree with the smallest error through the mean squared error for further analysis. We visualized the decision paths of the subtrees, clearly showing the splitting of input features at each node, the distribution of samples at the leaf nodes, and the decision logic of the prediction process.

[0101] Under the condition of the maximum tree depth limit, how does a single subtree perform regression prediction on the Tg of the FKM material? The path from the root node to the leaf node of the tree is determined by the characteristic test values of each node, and the value of each leaf node represents the average value of the Tg of all training instances under that node, representing the prediction result for these instances. In the figure, the decision paths of three instances on the optimal subtree are marked with different colors. Although the final prediction result of each instance is determined by the mean of the prediction results of all subtrees, through these paths, the key features of the model can be identified, which is very helpful for subsequent visual analysis and discrimination.

[0102] For models without a tree structure, we rely on SHAP values. SHAP (Shapley Additive exPlanations) is a method for explaining additive feature attribution models based on cooperative game theory. It can not only measure the importance of features but also provide local and global explanations of the model. Using a color-coded heatmap, combined with line charts and bar charts, it comprehensively shows the positive and negative impacts of each feature attribute on the model and its association with the model output. The heatmap matrix shows the prediction results of the model, while the bar chart intuitively shows the global importance of each input feature. The horizontal axis in the figure represents the number of instances, and the vertical axis lists different features sorted by overall importance. This method determines that product type, VDF, TFE, PMVE (f1, f4, f5, f3) are the key influencing features in the random forest regression model, highlighting the importance of these features in the prediction process. Compared with single features, the influence of the liquid phase temperature in the reactor (f10, f13, f7) is relatively low. We observe that monomers (f5, f3) and the liquid phase temperature f10 in the reactor have the highest contribution to the ENR model, while the significance of other features is not obvious.

[0103] To enhance the local interpretability of the model, the SHAP values of the RFR and ENR models were used to analyze individual instances, highlighting the specific impacts of different features. Taking three instances as examples, we observed that in Instance 1, when the Tg value fluctuated around -23°C, the characteristic product type f1, formulation monomers (f4, f5, f6), and the liquid phase temperature in the reactor (f9, f12) had a positive impact on the prediction results. Moreover, as can be seen from the width of the red bars, the impacts of the product type (f1) and formulation monomers (f5, f4) were relatively significant. Among them, monomer f3 had a negative impact. In the ENR model, monomers (f3, f5), the liquid phase temperature f10 in the reactor, and the reactant feed pressure (f31, f32) had an impact on the final result. In Instance 3, when the Tg value fluctuated around -31°C, the product type f1, monomers (f4, f5, f6), and the liquid phase temperature in the reactor (f9, f12) showed a negative impact, which played a decisive role in the predicted value. In the ENR model, the characteristic monomers (f3, f5), the liquid phase temperature f10 in the reactor, and the reactant feed pressure (f30, f31) also had a negative impact.

[0104] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other.

[0105] Specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention. At the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation on the present invention.

Claims

1. A prediction method based on the glass transition temperature of fluororubber, characterized in that Including: Obtain the fluoroelastomer data to be measured; Input the fluoroelastomer data to be measured into the final glass transition temperature prediction model to obtain the transition temperature prediction result; The construction method of the final glass transition temperature prediction model is: Collect fluoroelastomer experimental data, and the fluoroelastomer experimental data includes: process parameters and process formulations; Clean the fluoroelastomer experimental data to obtain the cleaned data; Based on the comprehensive analysis method, align the cleaned data, determine the defining indicators of each process parameter, and determine the characteristic sequence of the cleaned data according to the defining indicators; Screen the characteristic sequence to obtain the screened characteristic sequence; Construct the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model respectively based on the regressor, the deep learning algorithm, and the clustering analysis algorithm and according to the screened characteristic sequence; Determine the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model; The process formulation includes: Monomers and additives, wherein the monomers include: tetrafluoroethylene, perfluoromethyl vinyl ether, hexafluoropropylene, and vinylidene fluoride; The process parameters include: Temperature, pressure, feed amount, and reaction medium.

2. The prediction method based on the glass transition temperature of fluororubber according to claim 1, wherein Clean the fluoroelastomer experimental data to obtain the cleaned data, including: Determine the monomer content value not included in the process formulation and fill the monomer content value with zero to obtain the first sub-cleaned data; Determine the missing values in the process parameters that cannot be aligned with the original formulation or the final performance and delete the missing values to obtain the second sub-cleaned data; Determine the cleaned data according to the first sub-cleaned data and the second sub-cleaned data.

3. A prediction method based on the glass transition temperature of fluororubber according to claim 1, characterized in that, The above-mentioned based on the comprehensive analysis method, align the cleaned data, determine the defining indicators of each process parameter, and determine the characteristic sequence of the cleaned data according to the defining indicators, including: Determine the matrix sequence of the cleaned data; Obtain the process sequence by estimating the process parameters of the matrix sequence; Obtain the recombinant sequence according to the process sequence; Determine the characteristic sequence of the cleaned data according to the recombinant sequence and the defining indicators.

4. A prediction method based on the glass transition temperature of fluororubber according to claim 1, characterized in that, The above-mentioned screening of the characteristic sequence to obtain the screened characteristic sequence includes: Determine the data correlation according to the Pearson correlation coefficient calculation formula, wherein the data correlation includes: the correlation coefficient value between the feature and the performance and the correlation coefficient value between different features; Characterize the data correlation according to the heat map to obtain the correlation characterization heat map; Obtain the screened characteristic sequence according to the correlation characterization heat map and the recursive feature elimination algorithm.

5. A prediction method based on the glass transition temperature of fluororubber according to claim 1, characterized in that, The above-mentioned determining the final glass transition temperature prediction model according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model includes: Determine a first transition temperature prediction sub-result, a second transition temperature prediction sub-result, and a third transition temperature prediction sub-result according to the first glass transition temperature prediction sub-model, the second glass transition temperature prediction sub-model, and the third glass transition temperature prediction sub-model, respectively; Evaluate the first transition temperature prediction sub-result, the second transition temperature prediction sub-result, and the third transition temperature prediction sub-result, respectively, to obtain a first evaluation result, a second evaluation result, and a third evaluation result; Determine the weights of the respective prediction sub-models according to the first evaluation result, the second evaluation result, and the third evaluation result and perform weighting to determine a final glass transition temperature prediction model.

Images