Method for predicting polymer performance parameters based on computational simulation and machine learning
Through full atomic simulation and natural language processing, and combined with the CNN-LSTM model, the accuracy and resource consumption problems of polymer performance parameter prediction in the prior art are solved, and efficient and accurate prediction of polymer performance parameter is achieved.
Patent Information
- Application Number
- CN202410966603.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2025-08-08
AI Technical Summary
In the prediction of polymer performance parameters, the prior art has problems such as poor accuracy, high cost, large resource consumption and insufficient generalization ability, especially the impact of polymer sequence structure on performance is not fully considered.
Small sample data sets were constructed through full atomic simulation, feature representation was performed using natural language processing methods, and data sets were expanded using WGAN-GP model, performance parameter prediction was performed in combination with CNN-LSTM model, taking into account the influence of polymer sequence structure.
It realizes efficient and accurate prediction of polymer performance parameters, reduces cost and time consumption, improves the generalization ability of the model, and can be applied to the prediction of different polymer materials and performance indicators.
Smart Images
Figure CN120452583A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of polymer materials, and more particularly, relates to a method for predicting polymer performance parameters based on computational simulation and machine learning. Background Art
[0002] In the field of polymer materials, the relationship between polymer structure and properties has been extensively studied. Currently, methods for obtaining polymer performance parameters mainly include: 1. Experimental determination, 2. Molecular dynamics simulation, and 3. Machine learning and data-driven methods.
[0003] Among them, the limitations of experimental determination methods include: the influence of experimental conditions (such as sample preparation, temperature control, cooling rate, etc.) on the consistency of results; the difficulty in preparing homogeneous samples; the high cost and complex operation of experimental equipment and processes; and the easy introduction of human errors in data analysis.
[0004] Molecular dynamics simulations are a method for calculating the performance parameters of polymers under different conditions. The limitations of this method lie primarily in its demand for extensive computing resources and the impact of the accuracy of the force field model on the results.
[0005] Machine learning and data-driven approaches use experimental or simulated data to build machine learning models to fit and predict performance parameters. Polymer properties are generally influenced by their molecular weight, composition, sequence, and structure. Previous machine learning property prediction models were mostly based on polymer molecular weight and composition data (Journal of Rubber Industry, 2022, 69, 826–829), without considering the impact of sequence structure on performance. Summary of the Invention
[0006] In response to the technical deficiencies of poor accuracy in predicting polymer performance parameters based on existing machine learning and data-driven methods, the present invention provides a method for predicting polymer performance parameters based on computational simulation and machine learning.
[0007] The present invention deeply considers the impact of the sequence structure of a polymer on its performance and uses all-atom simulation to establish the polymer. To save costs, only a small sample data set is obtained through all-atom simulation. Then, a feature representation method is proposed using natural language processing, considering the relationship between each monomer in the polymer sequence structure. The small sample data set is expanded using the WGAN-GP model to increase the sample data set. Based on the expanded data set, a prediction method for polymer performance parameters is proposed with higher accuracy, higher efficiency, and stronger generalization ability.
[0008] The object of the present invention is to provide a method for predicting polymer performance parameters.
[0009] The method for predicting polymer performance parameters comprises the following steps:
[0010] s1. Construct different polymers, each of which has a different sequence structure, and extract the sequence structure of each polymer as the first set of feature data; calculate the performance parameter of each polymer as the first set of label data;
[0011] s2. Perform word vector representation on the first set of feature data;
[0012] s3. Expand the first set of feature data represented by the word vector to obtain the second set of feature data represented by the word vector;
[0013] s4. Calculate the calculated values of the performance parameters corresponding to the second set of feature data represented by the word vector as the second set of label data;
[0014] s5. Combine the first set of feature data represented by the word vector, the first set of label data, the second set of feature data represented by the word vector, and the second set of label data into a data set;
[0015] s6. Constructing a performance parameter prediction model and training the performance parameter prediction model using the data set;
[0016] s7. Input the word vector representation of the polymer to be tested into the performance parameter prediction model trained in step s6 to obtain the predicted value of the performance parameter of the polymer to be tested.
[0017] Compared with the previous application of machine learning methods in the field of polymer materials, this invention starts from the perspective of the relationship between polymer sequence structure and performance, based on a small amount of simulated data (the number of polymers constructed in simulation is small, which is a small sample), uses natural language processing to process the data, and establishes data enhancement models and performance prediction models, achieving low cost and high efficiency, saving a lot of manpower and material resources.
[0018] The polymer can be any existing polymer, such as butadiene rubber, chloroprene rubber, etc. As long as the polymer can be established and performance parameters can be calculated through molecular dynamics simulation, the performance parameters can also be arbitrary, such as wear resistance, fatigue life, glass transition temperature, etc.
[0019] Furthermore, in step s1, the polymer is established based on molecular dynamics simulation; specifically, the polymer can be constructed using all-atom simulation based on molecular dynamics.
[0020] Furthermore, step s2 refers to performing word vector representation on the first set of feature data using natural language processing; preferably, it includes performing word vector representation on the sequence structure of each polymer obtained in step s1 based on the k-mer method and using a word embedding (GloVe) model with a global vector; specifically, it includes the following steps:
[0021] S21, constructing a digital representation of the sequence structure of the polymer in step s1;
[0022] S22. Decompose the digital representation into subsequences of length k based on the k-mer method, where each subsequence represents a word, and map the subsequences into unique integer IDs to obtain word sequences represented by the integer IDs; k ≥ 2, preferably, k = 3-6;
[0023] S23. Use the word embedding model with global vectors (GloVe) to generate word vector representations of word sequences.
[0024] Furthermore, in step s3, the first set of feature data is augmented using the WGAN-GP model. The WGAN-GP model consists of a generator and a discriminator. The generator's network structure consists of five fully connected layers, while the discriminator's network structure consists of four fully connected layers. The generator uses LeakyReLu and Tanh activation functions, while the discriminator uses LeakyReLu as its activation function.
[0025] Furthermore, in step s3, the WGAN-GP model is used to process the first set of feature data represented by the word vectors. The last layer of the discriminator of the WGAN-GP model outputs an approximately fitted Wasserstein distance. The Wasserstein distance is used to construct the loss function of WGAN-GP, and the WGAN-GP model is trained according to the WGAN-GP loss function. The trained WGAN-GP model is used to generate new feature data (the new feature data is represented by the word vectors). The new feature data is the second set of feature data represented by the word vectors. These are all existing technologies and will not be described in detail here.
[0026] Furthermore, in step s4, a KNN algorithm is used to obtain calculated values of performance parameters corresponding to the second set of feature data represented by the word vectors; the calculated values of the performance parameters corresponding to the second set of feature data represented by the word vectors are used as the second set of label data. When using the KNN algorithm, an optimal K value of the KNN algorithm can be obtained through a random grid search algorithm.
[0027] Furthermore, in step s5, the first set of feature data represented by the word vectors, the first set of label data, the second set of feature data represented by the word vectors, and the second set of label data are simply merged together to obtain a dataset. The first set of feature data represented by the word vectors and the second set of feature data represented by the word vectors constitute the feature data in the dataset, and the first set of label data and the second set of label data constitute the label data in the dataset.
[0028] Furthermore, in step s6, a performance parameter prediction model is constructed using Python software. Specifically, this includes constructing a CNN-LSTM prediction model using Python software; the network structure of the CNN-LSTM prediction model sequentially includes: an input layer, a convolutional layer, a pooling layer, an LSTM layer, a dropout layer, an LSTM layer, a dropout layer, an LSTM layer, a dropout layer, a fully connected layer, and an output layer, with each layer connected in sequence; the number of units in each layer is a hyperparameter, and the hyperparameters are all obtained using the "Hyperband" algorithm of the "Keras Tuner" package; thereby obtaining a performance parameter prediction model. That is, the network structure of the performance parameter prediction model includes one convolutional layer, one pooling layer, three LSTM layers, three dropout layers, and one fully connected layer. Data enters the input layer of the performance parameter prediction model, extracts features through the convolutional and pooling layers, is processed by multiple LSTM layers and dropout layers, and finally generates the final output through the fully connected layer.
[0029] Furthermore, in step s7, the characteristic data of the polymer to be tested is extracted from the sequence structure of the polymer to be tested, and its word vector representation is performed with reference to step s2.
[0030] Compared with the prior art, the present invention has the following beneficial effects:
[0031] 1. The present invention obtains the original data set based on molecular dynamics simulation, which effectively reduces the high manpower and material consumption brought by the experiment, saving cost and time.
[0032] 2. Molecular dynamics simulation usually requires a lot of computing resources and time. The present invention only obtains a small sample data set (a small amount of data) based on molecular dynamics simulation, and then uses machine learning methods to expand the small sample data set. On the one hand, it saves computing resources and time costs, and on the other hand, it increases the diversity of data.
[0033] 3. The method of the present invention uses a natural language processing method to perform feature representation on simulation data, fully considers the influence of sequence structure on polymer performance, and realizes the prediction of performance parameters of different sequence structures.
[0034] 4. The model was externally validated using new data, demonstrating that the prediction method of the present invention has good universality and generalization capabilities and can be applied to the prediction of different polymer materials and different performance indicators.
[0035] 5. The method of the present invention has the advantages of simple operation, rapidity, high efficiency, and high accuracy. It can effectively reduce time cost and experimental workload, and can be used to guide the design of high-performance materials. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1This is a flow chart of the polymer performance parameter prediction method constructed based on computational simulation and machine learning in Example 1.
[0037] Figure 2 This is a UMAP visualization comparison of the original data and the generated data in Example 1. UMAP stands for Uniform Manifold Approximation and Projection, a dimensionality reduction method used to reduce word vectors to a two-dimensional space. In the figure, the titles of the horizontal and vertical axes represent the two dimensions of UMAP. In the figure, the left figure shows the UMAP visualization of the original data in Example 1, and the right figure shows the UMAP visualization of the generated data in Example 1.
[0038] The original data corresponds to the first set of feature data in Example 1, and the generated data corresponds to the second set of feature data in Example 1;
[0039] Figure 3 The distribution diagram of the glass transition temperature values (calculated values) of the polymer in Example 1 and the predicted values of the glass transition temperature obtained by the prediction model; the horizontal axis is the calculated value and the vertical axis is the predicted value; in the figure, the yellow dots represent the predicted values of the glass transition temperature of the feature data in the training set, and the red dots represent the predicted values of the glass transition temperature of the feature data in the test set; the upper left corner is the R 2 and R of the test set 2 .
[0040] Figure 4 This is a network structure diagram of the performance parameter prediction model constructed using Python software in Example 1. DETAILED DESCRIPTION
[0041] The present invention will be described in detail below with reference to specific drawings and embodiments. It is necessary to point out that the following embodiments are only used to further illustrate the present invention and cannot be understood as limiting the scope of protection of the present invention. Some non-essential improvements and adjustments made by those skilled in the art to the present invention based on the contents of the present invention still fall within the scope of protection of the present invention.
[0042] A method for predicting polymer performance parameters, comprising steps s1-s7.
[0043] S1. Construct different polymers, each with a different sequence structure. Extract the sequence structure of each polymer as the first set of feature data; and calculate the performance parameter values of each polymer as the first set of label data. Specifically, simulate and construct the sequence structures of multiple polymers, extract the sequence structure of each polymer as the first set of feature data, calculate the performance parameters of each constructed polymer, and obtain the calculated values of the performance parameters; these calculated values of the performance parameters of the constructed different polymers serve as the first set of label data.
[0044] In step s1, the amount of the polymers and the length of the sequence structure are selected according to actual needs and are within a reasonable range.
[0045] In step s1, the sequence structure of the polymer may be constructed based on molecular dynamics simulation; specifically, the sequence structure of the polymer may be constructed based on molecular dynamics and using all-atom simulation.
[0046] In step s1, any existing method for calculating polymer performance parameters can be used to calculate the calculated value of the polymer performance parameter. For example, the glass transition temperature of the polymer can be calculated using a temperature-volume method.
[0047] s2. Perform word vector representation on the first set of feature data; that is, perform word vector representation on the sequence structure of the polymer in step s1.
[0048] Step s2 can use the k-mer method in natural language processing and the word embedding with global vector (GloVe) model to represent the sequence structure of the polymer in step s1 by word vector; specifically, it includes the following steps:
[0049] s21, constructing a digital representation of the sequence structure of the polymer in step s1 to adapt it to the input of the machine learning model;
[0050] s22. Considering the relationship between adjacent monomers in the sequence structure of a polymer, a dictionary is constructed based on the k-mer method. The digital representation of the polymer sequence structure is decomposed into subsequences of length k, each subsequence representing a word, and each subsequence is mapped to a unique integer ID. At this point, the digital representation of the polymer sequence structure is converted into a word sequence represented by an integer ID.
[0051] The value of k is arbitrary and represents the number of numbers contained in the subsequence. Generally, k is greater than or equal to 2, for example, 2-6.
[0052] s23. Further considering the relationship between words, the word embedding (GloVe) model with global vector is used to generate word vector representation of the word sequence represented by the integer ID, and the first set of feature data represented by the word vector is obtained.
[0053] s3. Expand the first set of feature data represented by the word vector to obtain a second set of feature data represented by the word vector; step s3 can specifically use the WGAN-GP model to expand the first set of feature data represented by the word vector.
[0054] In step s3, the WGAN-GP model consists of a generator and a discriminator. The generator's network structure consists of five fully connected layers, while the discriminator's network structure consists of four fully connected layers. The generator's activation function can use LeakyReLu or Tanh, while the discriminator's activation function can use LeakyReLu.
[0055] In step s3, the WGAN-GP model is used to process the first set of feature data represented by the word vectors. The last layer of the discriminator of the WGAN-GP model outputs the approximate fitted Wasserstein distance. The Wasserstein distance is used to construct the WGAN-GP loss function, and the WGAN-GP model is trained based on the WGAN-GP loss function. The trained WGAN-GP model is used to generate new feature data (the new feature data is represented by the word vectors). The new feature data is the second set of feature data represented by the word vectors.
[0056] s4. Calculate the performance parameter corresponding to the second set of feature data represented by the word vectors and use it as the second set of label data. Specifically, the performance parameter corresponding to the second set of feature data can be calculated using a KNN algorithm. When using the KNN algorithm, the optimal K value of the KNN algorithm can be obtained using a random grid search algorithm.
[0057] S5. The first set of feature data represented by the word vectors, the first set of label data, the second set of feature data represented by the word vectors, and the second set of label data are combined to form a dataset. The first set of feature data, the first set of label data, the second set of feature data, and the second set of label data are simply merged together to obtain the dataset. The first set of feature data and the second set of feature data constitute the feature data in the dataset, and the first set of label data and the second set of label data constitute the label data in the dataset.
[0058] s6. Construct a performance parameter prediction model, use the feature data in the dataset as input parameters and the corresponding label data as output parameters, and train the constructed performance parameter prediction model.
[0059] In step s6, a performance parameter prediction model can be constructed using Python software. This includes constructing a CNN-LSTM prediction model using Python software. The network structure of the constructed CNN-LSTM prediction model includes one convolutional layer, one pooling layer, three LSTM layers, three dropout layers, and one fully connected layer. The number of units in each layer is conventionally set. Specifically, the number of units in each layer is a hyperparameter, which is obtained using the "Hyperband" algorithm in the "Keras Tuner" package. Thus, a performance parameter prediction model is obtained.
[0060] The network structure of the performance parameter prediction model constructed includes: input layer, convolution layer, pooling layer, LSTM layer, Dropout layer, LSTM layer, Dropout layer, LSTM layer, Dropout layer, fully connected layer and output layer; each layer is connected in sequence, data enters from the input layer, extracts features through the convolution layer and pooling layer, then passes through multiple LSTM layers and Dropout layers for processing, and finally generates the final output through the fully connected layer. Figure 4 shown.
[0061] Step s7: Input the word vector representation of the characteristic data of the polymer to be tested into the performance parameter prediction model obtained in step s6 to obtain the predicted value of the performance parameter of the polymer to be tested. The characteristic data of the polymer to be tested is extracted from the sequence structure of the polymer to be tested, and its word vector representation is performed as described in step s2.
[0062] Example 1
[0063] This embodiment provides a method for predicting the glass transition temperature of a solution styrene butadiene rubber (SSBR) system based on molecular dynamics simulation and machine learning. The method flow chart is as follows: Figure 1 The specific steps are as follows:
[0064] Step 1: Build the dataset
[0065] Based on molecular dynamics and all-atom simulation, 87 SSBR single chains were established. Each SSBR single chain was repeated 10 times to obtain an SSBR polymer, that is, each SSBR polymer was composed of 10 identical SSBR single chains; 87 polymers were obtained using 87 SSBR single chains. Each SSBR polymer is a solution-polymerized styrene-butadiene rubber system (hereinafter referred to as "SSBR system"). Since each SSBR system is composed of SSBR single chains with the same sequence structure, SSBR single chains are used below to represent SSBR polymers or SSBR systems. The glass transition temperature (Tg) of each polymer was calculated using the temperature-volume method. Its glass transition temperature (Tg) is shown in Tables 1, 2, and 3.
[0066] Step 2: Feature Extraction
[0067] The sequence structures of SSBR single chains from 87 SSBR systems were extracted as the first set of feature data, and their corresponding Tg values were calculated as the second set of tag data. Both the first set of feature data and the second set of tag data contained 87 data points.
[0068] Step 3: Feature Representation
[0069] Each SSBR single chain of the SSBR system is composed of four monomer structures: styrene, 1,2-butadiene, cis-1,4-butadiene, and trans-1,4-butadiene. The four monomer structures are represented by numbers: styrene is represented by "1", 1,2-butadiene is represented by "2", cis-1,4-butadiene is represented by "3", and trans-1,4-butadiene is represented by "4", thereby obtaining the digital representation of the first set of characteristic data (87 SSBR single chain sequence structures). The digital representation of the first set of characteristic data established is shown in Tables 1, 2, and 3. A dictionary was constructed based on the k-mer method, and the digital representation of the first set of characteristic data was converted into a word sequence representation. In this implementation, the value of k is 4.
[0070] For example,
[0071] "113141312231111311111111133111333141131111111212133323434244" is a digital representation of a 60-length SSBR single-chain sequence structure containing 55% styrene, 10% 1,2-butadiene, 25% cis-1,4-butadiene, and 10% trans-1,4-butadiene; in other words, it is a digital representation of a 60-length SSBR single-chain sequence structure containing 55% styrene, 10% 1,2-butadiene, 25% cis-1,4-butadiene, and 10% trans-1,4-butadiene.
[0072] Based on the k-mer feature,
[0073] "1131413122311113111111111331113331411311111111212133323434244" is decomposed into subsequences of length 4, that is, four numbers are iteratively selected from the above digital sequence as subsequences to obtain 57 subsequences. Each subsequence represents a word (for example, from left to right, 1131 represents a word, 1314 represents a word, 3141 represents a word, and so on), and each subsequence is mapped to a unique integer ID (for example, 1131 is mapped to 228, 1314 is mapped to 19, and 3141 is mapped to 61); that is, one integer represents one word; thus, the word sequence represented by integers is obtained [228, 19, 61, 136, 186, 94, 179, 192, 164, 43, 15 At this time, the sequence structure of the SSBR single chain represented by numbers is converted into a word sequence represented by integer ID.
[0074] The word sequence represented in the 57-dimensional representation is converted into a 57*57-dimensional word vector representation using the word embedding model with global vectors. This converts the digital representation of the first set of feature data into a word vector representation, obtaining the first set of feature data represented by the word vector.
[0075] Step 4: Data Augmentation
[0076] The WGAN-GP model is used to expand the first set of feature data represented by the word vectors, and the new feature data represented by the expanded word vectors is used as the second set of feature data, thus obtaining the second set of feature data represented by the word vectors.
[0077] The WGAN-GP model consists of a generator and a discriminator. The generator's network structure consists of five fully connected layers, while the discriminator's network structure consists of four fully connected layers. The generator's activation function can use LeakyReLu or Tanh, while the discriminator's activation function can use LeakyReLu.
[0078] The WGAN-GP model is used to process the first set of feature data represented by word vectors. The last layer of the discriminator of the WGAN-GP model outputs an approximately fitted Wasserstein distance. The Wasserstein distance is used to construct the WGAN-GP loss function, and the WGAN-GP model is trained based on the WGAN-GP loss function. The trained WGAN-GP model is used to generate new feature data represented by 57*57-dimensional word vectors. This new feature data is the second set of feature data.
[0079] The UMAP visualization of the first set of feature data represented by the word vector and the UMAP visualization comparison of the second set of feature data represented by the word vector are as follows: Figure 2 As shown. Figure 2 It can be concluded that the UMAP visualization of the first set of feature data represented by the word vectors is similar to that of the second set of feature data represented by the word vectors. This shows that WGAN-GP captures the feature space distribution of the first set of feature data represented by the word vectors, and the feature space distribution of the second set of feature data represented by the generated word vectors is similar to that of the first set of feature data represented by the word vectors, which can be used for subsequent research.
[0080] Step 5: Calculate labels
[0081] For the second set of feature data represented by the word vectors obtained by expanding the WGAN-GP model, a KNN algorithm is used to assign a Tg label to each feature data in the second set of feature data represented by the word vectors. The obtained Tg labels constitute the second set of label data corresponding to the second set of feature data. In this embodiment, the optimal K value of the KNN algorithm is obtained using a random grid search algorithm, K=14.
[0082] Step 6: Develop a prediction model for the glass transition temperature of SSBR
[0083] A prediction model for glass transition temperature was constructed using Python software (machine learning algorithm). Specifically, the CNN-LSTM prediction model was constructed using Python software. The network structure of the constructed CNN-LSTM prediction model included the following: input layer, convolutional layer, pooling layer, LSTM layer, dropout layer, LSTM layer, dropout layer, LSTM layer, dropout layer, fully connected layer, and output layer. Each layer was connected sequentially. The number of units in each layer was a hyperparameter obtained using the "Hyperband" algorithm in the "Keras Tuner" package. Thus, a prediction model for glass transition temperature was obtained.
[0084] Data enters the input layer of the glass transition temperature prediction model, extracts features through the convolution layer and pooling layer, is processed through multiple LSTM layers and Dropout layers, and finally generates the final output through the fully connected layer. Figure 4 shown.
[0085] The first set of feature data represented by the word vectors, the first set of label data, the second set of feature data represented by the word vectors, and the second set of label data constitute a dataset. The first set of feature data represented by the word vectors, the first set of label data, the second set of feature data represented by the word vectors, and the second set of label data are simply merged together to obtain a dataset. The first set of feature data represented by the word vectors and the second set of feature data represented by the word vectors constitute the feature data in the dataset, and the first set of label data and the second set of label data constitute the label data in the dataset.
[0086] The dataset is randomly divided into two groups: one for training and the other for testing. The ratio of the number of samples in the training set to the number of samples in the testing set is 8:2. A feature data and its corresponding label data are considered a sample.
[0087] The feature data in the training set is used as the input parameter of the glass transition temperature prediction model, and the corresponding label data is used as the output parameter training model to obtain a trained glass transition temperature prediction model.
[0088] The feature data in the training set is used as the input parameter of the trained glass transition temperature prediction model to obtain the predicted value of the glass transition temperature; the square of the correlation coefficient R between the predicted value of the glass transition temperature and the label data is calculated. 2 ;like Figure 3 shown.
[0089] The feature data in the test set is used as the input parameter of the trained glass transition temperature prediction model to obtain the predicted value of the glass transition temperature; the square of the correlation coefficient R between the predicted value of the glass transition temperature and the label data is calculated. 2 ;like Figure 3 shown.
[0090] Figure 3 It shows that the R 2 The R of the test set is 0.94. 2 is 0.92, which is greater than 0.9, indicating that the prediction results of the glass transition temperature prediction model are credible.
[0091] Step 7: External Validation
[0092] Molecular dynamics simulations were used to generate new SSBR single chains, and their corresponding characteristic data were obtained, as shown in Table 4. The characteristic data were converted into word vector representations using the method in step 3. The characteristic data represented by these word vectors were input into the glass transition temperature prediction model in step s6 for prediction, resulting in predicted Tg values, as shown in Table 5. The characteristic data were then input into existing prediction models for prediction, resulting in predicted Tg values, as shown in Table 5. The maximum absolute error, |Δ|, between the predicted Tg values obtained by different prediction models and the simulated values was calculated, as shown in Table 5.
[0093] Table 1
[0094]
[0095] Table 2
[0096]
[0097] Table 3
[0098]
[0099] Table 4
[0100]
[0101] In Tables 1-4, the first column is the sample number, the single-chain sequence structure in the second column represents a polymer, and Tg in the third column refers to the glass transition temperature.
[0102] Table 5
[0103]
[0104] In Table 5, the first column indicates the sample number. CNN, LIST, and CNN-LSTM represent existing prediction models, while OURS represents the entire prediction model established in steps s1-s6 of Example 1. CNN, LSTM, and CNN-LSTM indicate that predictions were made directly using these models without using natural language processing or WGAN-GP methods. The results show that the maximum absolute error of the four new samples using the proposed method is the smallest, at only 3.67, demonstrating the effectiveness and accuracy of the proposed prediction method.
Claims
1. A method for predicting polymer performance parameters, characterized in that: The prediction method comprises the following steps: s1. Construct different polymers, each of which has a different sequence structure, and extract the sequence structure of each polymer as the first set of feature data; calculate the performance parameter of each polymer as the first set of label data; s2. Perform word vector representation on the first set of feature data; s3. Expand the first set of feature data represented by the word vector to obtain the second set of feature data represented by the word vector; s4. Calculate the calculated values of the performance parameters corresponding to the second set of feature data represented by the word vector as the second set of label data; s5. Combine the first set of feature data represented by the word vector, the first set of label data, the second set of feature data represented by the word vector, and the second set of label data into a data set; s6. Constructing a performance parameter prediction model and training the performance parameter prediction model using the data set; s7. Input the word vector representation of the polymer to be tested into the performance parameter prediction model trained in step s6 to obtain the predicted value of the performance parameter of the polymer to be tested.
2. The prediction method according to claim 1, wherein: In step s1, the polymer is constructed based on molecular dynamics simulation; preferably, the polymer is constructed using all-atom simulation.
3. The prediction method according to claim 1, wherein: The step s2 comprises the following steps: S21, constructing a digital representation of the sequence structure of the polymer obtained in step s1; S22. Decompose the digital representation into subsequences of length k based on the k-mer method, where each subsequence represents a word, and obtain a word sequence; S23. Use the word embedding model with the global vector to generate the word vector representation of the word sequence.
4. The prediction method according to claim 3, wherein: k≥2, preferably, k=3-6.
5. The prediction method according to claim 1, wherein: In step s3, the first set of feature data is expanded using the WGAN-GP model; the WGAN-GP model is composed of a generator and a discriminator; the network structure of the generator is composed of five fully connected layers; and the network structure of the discriminator is composed of four fully connected layers.
6. The prediction method according to claim 5, wherein: The activation function of the generator uses LeakyReLu and Tanh; the activation function of the discriminator uses LeakyReLu.
7. The prediction method according to claim 1, wherein: Step s4 is performed using the KNN algorithm.
8. The prediction method according to claim 1, wherein: The step s6 utilizes Python software to construct a performance parameter prediction model.
9. The prediction method according to claim 8, wherein: In step s6, the network structure of the performance parameter prediction model includes 1 convolution layer, 1 pooling layer, 3 LSTM layers, 3 Dropout layers and 1 fully connected layer.
10. The prediction method according to claim 1, wherein: The performance parameter is the glass transition temperature.