Design method of high-performance polyimide material based on deep learning
By extracting key characteristic information of polyimide molecules through the deep learning model COMFO, the problem of high dielectric constant of polyimide materials was solved, and the design of high-performance polyimide materials was realized, which is suitable for the repair of heating pipe linings.
Patent Information
- Application Number
- CN202510901170.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-01
- Publication Date
- 2025-10-17
AI Technical Summary
The dielectric constant of existing commercial polyimide materials is relatively high, which limits their application in the field of heating pipe lining repair. Traditional methods also make it difficult to balance the contradictory relationship between dielectric properties, thermal expansion coefficient and mechanical strength.
Using the deep learning model COMFO, key feature information is extracted from polyimide molecules through multi-angle feature extraction tasks, and combined with high-throughput screening technology, high-performance polyimide materials are predicted and screened.
It has achieved more complete and accurate extraction of key feature information in the polyimide structure, built a high-performance deep learning model, and guided the synthesis of new high-performance polyimide materials to meet the needs of heating pipeline lining repair.
Smart Images
Figure CN120809003A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of high-performance polyimide material structure design, and particularly relates to a design method of a polyimide material with high thermal-mechanical properties and low dielectric constant based on deep learning. BACKGROUND
[0002] To adapt to the environmental characteristics of high temperature and high external load of the heating pipeline, an excellent inner lining repair material should have some key properties, such as a high glass transition temperature (Tg), tensile strength (T M ) and low linear thermal expansion coefficient (CTE). In addition, since the main disease type of the heating pipeline is corrosion, the key factor affecting corrosion is the extraneous charge of the pipeline, and therefore, to ensure that the repaired inner lining pipe has good corrosion resistance, the material should also have a low dielectric constant (ε). As a high-performance engineering material, polyimide exhibits good mechanical properties, heat resistance (Tg> 300℃) and excellent solvent resistance due to its rigid imine group and strong intermolecular interaction force, and has been widely used in flexible display, aerospace and gas separation fields. However, the dielectric constant (3-4) of the existing commercial polyimide is still high, which significantly restricts its direct application in the CIPP repair field of the heating pipeline.
[0003] Polyimide is usually synthesized by polycondensation of diamine and dianhydride or diisocyanate and dianhydride. The variability of monomer structure makes it have flexible molecular design. Through the design of monomer structure and the combination of monomers, people have designed various polyimides. However, the performance of polyimide materials often presents a "trade-off" relationship, for example, the introduction of aliphatic groups improves the optical transparency of PI, but reduces its mechanical properties. The current development of polyimide still highly depends on traditional trial-and-error research, which has the following limitations:
[0004] (1) Low efficiency of functional group introduction: the functional modification of the molecular chain needs to go through a complex synthesis path, and the single experiment cycle is long;
[0005] (2) Difficulty in multi-objective optimization; it is difficult for artificial synthesis to balance the contradictory relationship among dielectric properties, thermal expansion coefficient and mechanical strength.
[0006] In recent years, machine learning (ML) and deep learning (DL) techniques have received increasing attention in the field of material design. This is an efficient method driven by data from traditional experimental tests or simulation calculations. Researchers can establish a quantitative structure-property relationship for polymer materials based on this technology. Machine learning has a much faster calculation speed than experiments and computational simulations. Because machine learning is based on a large and rich database of material structure-property, the results obtained have a wider applicability than computational simulations. ML and DL have been successfully used to predict the glass transition temperature (Tg), cutoff wavelength, gas separation, and dielectric constant of polymers. Feature extraction is a key step in high polymer machine learning, and the effectiveness of extraction is one of the determining factors of model performance and generalization ability. There are currently a variety of feature representation methods, such as molecular fingerprints, molecular graphs, and descriptors. These features have different focuses when extracting polymer molecular information, but it is still unknown whether these feature extraction methods can completely extract hidden information in polymer molecular structures. SUMMARY
[0007] The present application establishes a deep learning model-COMFO, which can extract key feature information in polyimide molecules from three angle learning tasks. Specifically, the three angle feature extraction learning tasks include: 1) extracting feature information in SMILES sequence; 2) extracting information of atoms and bonds of polymer molecules; 3) extracting substructure information of polymer molecules. The COMFO model can integrate, process and establish a regression model for the information extracted in the three types of learning tasks.
[0008] According to the high-performance polyimide material design method based on deep learning provided by the application, the following steps are included:
[0009] (1) Data collection; collect structure-property data related to polyimide heat resistance, mechanical properties and dielectric properties, including glass transition temperature (Tg), linear thermal expansion coefficient (CTE), tensile modulus (TM) and dielectric constant (ε);
[0010] (2) Data preprocessing; data cleaning and normalization processing are performed on the collected structure-property data;
[0011] (3) Feature extraction; according to the processed polyimide structure-property data, a feature extraction task is constructed to extract key feature information contained in the polyimide molecular structure, the feature extraction task includes a SMILES-based feature extraction task, a molecular graph-based feature extraction task and a molecular fingerprint-based feature extraction task;
[0012] (4) Constructing deep learning model; embedding features extracted from three kinds of feature learning tasks are input into integrated module, pooling layer and full connection layer in sequence to construct deep learning model (COMFO);
[0013] (5) Model training and optimization; adjusting key parameters of COMFO model to make performance of the model relatively optimal;
[0014] (6) Model verification; extracting structure-property data of PI from the latest research articles to test the performance of COMFO model;
[0015] (7) Establishment of virtual screening database; collecting existing diamine and dianhydride molecular structures from PubChem database, combining two types of synthesis methods of polyimide to construct hypothetical polyimide, and establishing polyimide screening database;
[0016] (8) Screening of high-performance polyimide; using the trained COMFO model to predict Tg, CTE, TM and ε of polyimide in the virtual screening database, and screening out polyimide with excellent comprehensive performance.
[0017] Further, the specific steps of data cleaning in step (2) are:
[0018] 1) Collecting structure-property data of polyimide from literature;
[0019] 2) Eliminating low molecular weight PI structures;
[0020] Further, the feature extraction process of BET model in the SMILES-based feature extraction task in step (3) is:
[0021] 1) Constructing bidirectional encoder based on Transformer architecture (BET), which is composed of multiple stacked Transformer encoding units, each encoding unit containing a full connection feedforward layer and a self-attention layer with multiple attention heads;
[0022] 2) Using Adam optimizer to pre-train the BET model in a self-supervised manner, setting the weight decay coefficient to 1×10-6 to 1×10-3, and dynamically adjusting the learning rate to 1×10 -4 to 1×10 -6 ;
[0023] 3) Extracting semantic features of SMILES sequence through BET model to generate deep learning fingerprint vectors with 256-1024 dimensions.
[0024] Further, the step (3) of the molecular graph-based feature extraction task aims to introduce the information of atoms and bonds contained in the molecular graph by using the attention FP network, and the feature extraction process is as follows:
[0025] 1) Molecular graph construction: input the SMILES sequence of the polyimide monomer into the RDKit chemical information processing tool to generate molecular graph data containing atomic feature tensors and chemical bond adjacency tensors;
[0026] 2) Feature encoding: the molecular graph data is processed by using the graph attention FP network (Attentive FP), and the network architecture sequentially includes:
[0027] a) Input layer: map the atomic features to low-dimensional initial embedding vectors;
[0028] b) Hidden layer: introduce a graph attention FP convolution layer to extract the atomic features of the input molecule;
[0029] c) Readout layer: generate a low-dimensional molecular graph fingerprint vector by global average pooling;
[0030] Further, the specific steps of the step (3) of the molecular fingerprint-based feature extraction task are as follows:
[0031] 1) Substructure enumeration: traverse the polymer molecular structure by using an improved MorGan fingerprint algorithm, and set the radius parameter r for iterative expansion;
[0032] 2) Frequency statistics: in each iteration process, map the substructure types by using a hash function and record their occurrence frequencies;
[0033] 3) Fingerprint generation: concatenate the frequency statistics results with the binary existence identifier in dimension to output a high-dimensional enhanced fingerprint vector;
[0034] Further, the specific steps of the step (4) of building a deep learning model are as follows:
[0035] 1) Feature integration: integrate the embedding information extracted by the three-dimensional feature learning task.
[0036] 2) Feature dimension reduction: after feature integration, continue to input to the pooling layer, and further pool the extracted features to reduce the feature dimension to low dimension,
[0037] 3) Construction of regression model: input the features read out by the pooling layer into the fully connected layer network for training to construct a four-layer fully connected layer regression model (COMFO), and each fully connected layer is respectively connected with a normalization layer and an activation function;
[0038] Further, the parameter adjustment process in the feature extraction task in the step (5) is as follows:
[0039] 1) Adjust the iteration radius and output feature dimension of MorGan fingerprint and other parameters;
[0040] 2) Set the number of layers of the pooling layer of attention FP within 2-5, adjust the node number of its neurons and the feature embedding dimension of the readout layer; in order to simplify the parameter adjustment process, the output feature embedding dimensions of the two types of learning tasks are set to the same value and are adjusted synchronously.
[0041] Further, the parameter adjustment process of the COMFO model training stage in the step (5) is specifically:
[0042] 1) The training set and the test set of the model are divided in the proportion of 7:3 or 8:2;
[0043] 2) The Adam optimizer is adopted, the initial learning rate is set to 1×10 -4 ~ 1×10 -6 , the weight decay is set to 1×10 -6 ~ 1×10 -3 ;
[0044] 3) The initial iteration number is adjusted;
[0045] 4) The learning rate, the weight decay and the iteration number of the model training are adjusted according to the predicted properties.
[0046] Further, the verification method of the COMFO model in the step (6) is:
[0047] 1) The structure-property number of polyimide is extracted from the latest research article;
[0048] 2) The performance is predicted by using the COMFO model, then the predicted value of the model is compared with the reported value, and the prediction error of the model is calculated.
[0049] The beneficial effects of the present application are:
[0050] This invention provides a deep learning-based design method for high-performance polyimide materials. Collected polyimide structure-property data is cleaned and preprocessed to construct a model training dataset. Three feature extraction tasks are constructed, including: 1) extracting feature information from SMILES sequences using a large language model-bidirectional encoder Transformer; 2) extracting atomic and bond information of polymer molecules from molecular graphs using an attention-based FP network; and 3) extracting substructure information of polymer molecules using molecular fingerprints. The COMFO model integrates and processes the information extracted from these three learning tasks to create a regression model. This method collects existing polyimide synthesis monomers and constructs a polyimide screening database based on standard polyimide synthesis methods. The established COMFO model, combined with high-throughput screening technology, predicts high-performance polyimides. By constructing multi-faceted feature extraction tasks, the invention can more completely and accurately extract key feature information contained in polyimide structures, thereby constructing a high-performance deep learning model. Furthermore, this deep learning model can be combined with high-throughput screening technology to discover potential high-performance polyimide structures, thereby guiding the synthesis of new high-performance polyimide materials. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 It is a flow chart of an implementation case of the present invention.
[0052] Figure 2 This is a diagram of the three major feature extraction tasks in the implementation case of the present invention.
[0053] Figure 3 This is a framework diagram of the COMFO model constructed in the implementation case of the present invention.
[0054] Figure 4 4 is a graph comparing the performance of the COMFO performance prediction model constructed in the embodiment of the present invention and other models / feature representations.
[0055] Figure 5 This is a performance verification diagram of the COMFO performance prediction model constructed in the implementation case of the present invention. DETAILED DESCRIPTION
[0056] The present invention will be described in more detail and more completely below with reference to the accompanying drawings.
[0057] The application aims to provide a high-performance polyimide material design method based on deep learning, which more completely and accurately extracts key feature information contained in the polyimide molecular structure by constructing multi-angle feature extraction tasks, and integrates, pools and inputs the key feature information into a full connection layer for learning by a deep learning model. The constructed deep learning model, combined with high-throughput screening technology, predicts and discovers potential polyimide molecular structures with excellent comprehensive performance, thereby guiding the synthesis of high-performance polyimide materials.
[0058] The method flow of the embodiment of the application is shown in Figure 1 The design method of the high-performance polyimide material based on deep learning comprises the following steps:
[0059] (1) Data collection; collect structure-property data related to the heat resistance, mechanical properties and dielectric properties of polyimide, including glass transition temperature (Tg), linear thermal expansion coefficient (CTE), tensile modulus (TM) and dielectric constant (ε);
[0060] (2) Data preprocessing; data cleaning and normalization processing are performed on the collected structure-property data;
[0061] (3) Feature extraction; according to the processed polyimide structure-property data, a feature extraction task is constructed to extract key feature information contained in the polyimide molecular structure, and the feature extraction task includes a SMILES-based feature extraction task, a molecular graph-based feature extraction task and a molecular fingerprint-based feature extraction task;
[0062] (4) Construction of deep learning model; the embedded features extracted from the three feature learning tasks are sequentially input into an integration module, a pooling layer and a full connection layer to construct a deep learning model (COMFO);
[0063] (5) Model training and optimization; the key parameters of the COMFO model are adjusted to make the performance of the model reach a relatively optimal state, and the parameter adjustment includes hyperparameter adjustment in the three feature learning tasks and model training parameter adjustment;
[0064] (6) Model verification; the structure-property data of PI is extracted from the latest research articles to test the performance of the COMFO model.
[0065] (7) Establishment of virtual screening database; existing diamine, dianhydride and diisocyanate molecular structures are collected from the PubChem database, and polyimides are constructed in combination with two types of polyimide synthesis methods to establish a polyimide screening database;
[0066] (8) High-performance polyimide screening; using the trained COMFO model to predict the Tg, CTE, TM and ε of polyimides in the virtual screening database, and screening out polyimides with excellent comprehensive performance;
[0067] Preferably, the SMILES-based feature extraction task in step (3) aims to extract feature information contained in the SMILES sequence using a large language model.
[0068] Further, the feature extraction process of the BET model in the SMILES-based feature extraction task in step (3) is:
[0069] 1) Construct a bidirectional encoder (BET) based on the Transformer architecture, which consists of 8 stacked Transformer encoding units, each containing a fully connected feedforward layer with an embedding dimension of 1024 and a self-attention layer with 8 attention heads;
[0070] 2) Use the Adam optimizer to pre-train the BET model in a self-supervised manner, set the weight decay coefficient to 0.1, and dynamically adjust the learning rate in the range of 0.0001;
[0071] 3) Extract the deep semantic features of the SMILES sequence through the BET model to generate a deep learning fingerprint vector, and the BET model pre-trained based on the ChEMBL database is used to extract the feature information based on the SMILES language, and the extracted feature embedding dimension is 512. The SMILES-based feature extraction task module is as shown in Figure 2 b.
[0072] Preferably, the molecular graph-based feature extraction task in step (3) aims to introduce an attention FP network to extract the information of atoms and bonds contained in the molecular graph, and the feature extraction process is:
[0073] 1) Molecular graph construction: input the SMILES sequence of the polyimide monomer into the RDKit chemical information processing tool to generate molecular graph data containing atomic feature tensors and chemical bond adjacency tensors;
[0074] 2) Feature encoding: use the graph attention FP network (Attentive FP) to process the molecular graph data, and the network architecture sequentially contains:
[0075] a) Input layer: map the atomic features to 128-dimensional initial embedding vectors;
[0076] b) Hidden layer: introduce 2-5 layers of graph attention FP convolutional layers to extract the atomic features of the input molecule, and set the initial number of neuron nodes to 300;
[0077] c) readout layer: generate 512-dimensional molecular graph fingerprint vector by global average pooling, based on the feature extraction task module of the molecular graph as shown in Figure 2 c.
[0078] Preferably, the definition rule of the atomic feature matrix of the molecular graph in step (3) is shown in Table 1. Table 1
[0079] Preferably, the definition rule of the bond feature matrix of the molecular graph in step (3) is shown in Table 2. Table 2
[0080] Further, the specific steps of the feature extraction task of the molecular fingerprint in step (3) are:
[0081] 1) Substructure enumeration: iterate the polymer substructure by improved MorGan fingerprint algorithm, set the radius parameter r = 3 for iterative expansion;
[0082] 2) Frequency statistics: in each iteration process, map the substructure type by a hash function, and record its frequency;
[0083] 3) Fingerprint generation: concatenate the frequency statistics result with the binary existence identifier in dimension, and output a 1024-dimensional enhanced fingerprint vector;
[0084] Further, the specific parameters of the improved MorGan fingerprint in the feature extraction task of the molecular fingerprint in step (3) are:
[0085] 1) Number of iterations n = 4 times;
[0086] 2) The hash function uses MurmurHash3 algorithm;
[0087] 3) The fingerprint bit number is set to 4096 bits, and is reduced to 512 dimensions;
[0088] 4) The initial iteration radius is set to 2, and is gradually optimized. The molecular fingerprint itself can be input into the regression model for training, but the present application also sets two control groups to determine whether to further extract IMF fingerprint information using the model. Control group 1: extract IMF fingerprint using LSTM model; control group 2: extract IMF fingerprint using FFNN model. The readout feature dimension of the FFNN and LSTM model is still consistent with the original IMF fingerprint. The feature extraction task module based on the molecular fingerprint is as shown in Figure 2 a.
[0089] Preferably, the specific steps of building a deep learning model in step (4) are:
[0090] 1) Feature integration; the embedding information extracted from the three-dimensional feature learning task is integrated by vector splicing.
[0091] 2) Feature dimension reduction; after feature integration, the extracted features are further averaged and pooled in the pooling layer, and the feature dimension is reduced to 256,
[0092] 3) Constructing a regression model; the features read out by the pooling layer are input into the fully connected layer network for training to construct a four-layer fully connected layer regression model (COMFO), and each fully connected layer is connected with a normalization layer and an activation function (ReLU) respectively. According to the different ways of extracting molecular fingerprints, three types of COMFO models are constructed, namely COMFO_ORI, COMFO_FFNN and COMFO_LSTM. The architecture of the COMFO model is shown in Figure 3
[0093] Preferably, the parameter adjustment process in step (5) comprises:
[0094] 1) Feature extraction module of molecular fingerprint, adjusting the iteration radius and output feature dimension of MorGan fingerprint and other parameters;
[0095] 2) Feature extraction module of molecular graph, fixing the number of layers of the attention FP pooling layer to 2, adjusting the number of neurons and the feature embedding dimension of the readout layer. To simplify the parameter adjustment process, the output feature embedding dimensions of the two types of learning tasks are set to the same value and adjusted simultaneously.
[0096] Preferably, the parameter adjustment process in the COMFO model training stage of step (5) is specifically:
[0097] 1) In the training process, the training set and the test set of the model are divided in the ratio of 7:3;
[0098] 2) The global random seed is set to 42, batch_size is set to 30, shuffle is set to True, and dropout is set to 0.3;
[0099] 3) Adam optimizer is used, the initial learning rate is set to 0.0001, and the weight decay is set to 0.000001;
[0100] 4) The initial number of iterations is set to 501;
[0101] 5) The learning rate, weight decay and number of iterations of model training are adjusted according to the predicted properties. The hyperparameters of the optimal model for each property prediction are shown in Table 3. Table 3
[0102] Preferably, the present application adopts coefficient of determination square (R 2 ), root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) to comprehensively evaluate the performance and generalization ability of the model. Figure 4 a-d and Table 4 show the performance of the four property prediction models on the training set and the test set. The results show that the R 2 of all property prediction models are higher than 0.90. Table 4
[0103] Preferably, in order to test the performance of the model, the present application also builds three comparative models such as DNN, RF and XGBoost based on the scikit-learn library. Taking Tg prediction as a representative, the present application compares the performance of the COMFO model with other machine learning models, and the comparison results are shown in Figure 4 e. The results show that the COMFO model built by the present application has the highest R 2 Based on the COMFO model, the integrated fingerprint can still be used for the training of other models. The present application inputs the COMFO fingerprint and the traditional feature representation method (MACCS fingerprint, ECFP fingerprint, IMF fingerprint and RDKit descriptor) into the DNN and XGBoost models for performance testing, and the results show that the model built based on the integrated fingerprint has the lowest RMSE, and the results are shown in Figure 4 f.
[0104] Preferably, the present application selects 8 polyimides reported in previous studies to test the performance of the COMFO model. The four properties of the 8 polyimides have been completely tested. Figure 5 a-d respectively show the comparison results between the experimental values and the predicted values of the COMFO model of the glass transition temperature, tensile modulus, linear thermal expansion coefficient and dielectric constant of the 8 polyimides for verification. In summary, the COMFO deep learning model with multi-feature learning built by the present application has excellent performance and good generalization ability.
[0105] Further, according to two types of synthesis methods of polyimides, including the dehydration synthesis of diamine and dianhydride to synthesize polyimide and the synthesis of polyimide by the decarburization of diamine and diisocyanate, the present application builds a virtual polyimide screening database. According to the COMFO model and high-throughput screening technology, the present application can predict and screen high-performance polyimides.
[0106] The beneficial effects of the present application are:
[0107] The application provides a high-performance polyimide material design method based on deep learning. Collected polyimide structure-property data is cleaned and pretreated to construct a model training data set; three types of feature extraction tasks are constructed, including: 1) using a large language model-bidirectional encoder Transformer to extract feature information in the SMILES sequence; 2) using an attention FP network to extract atom and bond information of the polymer molecule from the molecular graph; and 3) extracting substructure information of the polymer molecule through molecular fingerprint. The COMFO model is used to integrate and process the key feature information extracted in the three types of learning tasks and establish a regression model. The application collects existing polyimide synthesis monomers, constructs a virtual polyimide screening database according to the standard polyimide synthesis method, and can predict high-performance polyimides according to the established COMFO model and high-throughput screening technology.
[0108] Obviously, the above embodiments are only examples for clearly illustrating but not limiting the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. All the embodiments do not need to be exhausted here. The obvious changes or variations derived therefrom are still within the protection scope of the application.
Claims
1. A design method for high-performance polyimide materials based on deep learning, characterized in that: The following steps are involved: (1) Data collection: collecting structure-property data related to the heat resistance, mechanical properties, and dielectric properties of polyimide, including glass transition temperature, linear thermal expansion coefficient, tensile modulus, and dielectric constant; (2) Data preprocessing: cleaning and normalizing the collected structure-property data; (3) Feature extraction: constructing a feature extraction task based on the processed PI structure-property data to extract key feature information contained in the polyimide molecular structure. The feature extraction task includes a SMILES sequence-based feature extraction task, a molecular graph-based feature extraction task, and a molecular fingerprint-based feature extraction task; (4) Constructing a deep learning model; the embedded features extracted from the three feature learning tasks are sequentially input into the integration module, pooling layer, and fully connected layer to construct the COMFO deep learning model; (5) Model training and optimization: adjust the key parameters of the COMFO model to achieve relatively optimal model performance; (6) Model validation: extracting PI structure-performance data from the latest research literature to test the performance of the COMFO model; (7) Establishment of a virtual screening database: Collect existing diamine and dianhydride molecular structures from the PubChem database and construct a PI screening database based on the two types of PI synthesis methods; (8) Screening of high-performance polyimides: The trained COMFO model is used to predict the glass transition temperature, linear thermal expansion coefficient, tensile modulus and dielectric constant of polyimides in the screening database to screen out polyimides with excellent comprehensive performance.
2. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The specific steps of data cleaning in step (2) are: 1) Collect structure-property data of polyimides from the literature; 2) Eliminate low molecular weight polyimide structures; 3. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The SMILES-based feature extraction task in step (3) is achieved by the following steps: 1) Build a bidirectional encoder based on the Transformer architecture, consisting of multiple stacked Transformer encoding units, each of which contains a fully connected feedforward layer and a self-attention layer with multiple attention heads; 2) The BET model was pre-trained using the Adam optimizer, with the weight decay coefficient set to 1×10-6 to 1×10-3 and the learning rate dynamically adjusted to 1×10 -4 to 1×10 -6 ; 3) Extract the semantic features of the SMILES sequence through the BET model and generate a deep learning fingerprint vector.
4. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The molecular graph feature extraction task of step (3) is achieved by the following steps: 1) Molecular graph construction: The SMILES sequence of the polyimide molecular repeating unit is input into the RDKit chemical information processing tool to generate molecular graph data including atomic feature tensors and chemical bond adjacency tensors; 2) Feature Encoding: A graph attention FP network is used to process the molecular graph data. The network architecture includes: a) Input layer: maps atomic features into low-dimensional initial embedding vectors; b) Hidden layer: Introducing graph attention FP convolution layer to extract atomic features of input molecules; c) Readout layer: Generate low-dimensional molecular graph fingerprint vector through pooling.
5. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The molecular fingerprint feature extraction task of step (3) is achieved by the following steps: 1) Substructure enumeration: The polymer molecular structure is traversed by the improved MorGan fingerprint algorithm, and the radius parameter r is set for iterative expansion; 2) Frequency statistics: In each iteration, substructure types are mapped using a hash function and their occurrence frequencies are recorded; 3) Fingerprint generation: The frequency statistics results and the binary presence flags are dimensionally concatenated to output a high-dimensional enhanced fingerprint.
6. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The construction of the deep learning model in step (4) is achieved by the following steps: 1) Feature integration: integrating the embedded information extracted from the three-dimensional feature learning task; 2) Feature dimensionality reduction: After feature integration, the features are input into the pooling layer, where the extracted features are further pooled to reduce the feature dimension to a low dimension. 3) Construct a regression model; input the features read out by the pooling layer into the fully connected layer network for training to construct a fully connected layer regression model, where each fully connected layer is connected to the normalization layer and the activation function.
7. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The parameter adjustment process in step (5) includes: 1) Adjust and improve parameters such as the iteration radius and output feature dimension of MorGan fingerprint; 2) Set the number of pooling layers of the attention FP to 2-5, and adjust the number of neuron nodes and the feature embedding dimension of the readout layer.
8. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The parameter adjustment process of the COMFO model training phase in step (5) is achieved by the following steps: 1) Use a 7:3 or 8:2 ratio to divide the model's training set and test set; 2) Using Adam optimizer, the initial learning rate is set to 1×10 -4 to 1×10 -6 , weight decay is set to 1×10-6 to 1×10-3; 3) Adjust the number of iterations; 4) The learning rate, weight decay, and number of iterations of model training are adjusted based on the predicted properties.
9. The high-performance polyimide material design method based on deep learning according to claim 1, characterized in that: The verification method of the COMFO model in step (6) is: 1) Extract structure-property data of polyimides from the latest literature; 2) The COMFO model was used to predict its performance, and the predicted values of the model were compared with the values reported in the literature to calculate the prediction error.