Methods and applications for predicting the performance of epoxy resins
By combining a graph convolutional neural network with attention and gating mechanisms with a multilayer perceptron machine learning model, along with a crosslinking density descriptor, the problem of low efficiency in epoxy resin structure design is solved, achieving efficient and accurate performance prediction, and applicable to a wide range of epoxy resin systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- EAST CHINA UNIV OF SCI & TECH
- Filing Date
- 2022-11-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies make it difficult to quickly design epoxy resin structures with contradictory properties. Traditional experimental trial-and-error methods are inefficient and costly, and machine learning models are only applicable to specific epoxy resin systems and lack broad applicability.
A machine learning model combining a graph convolutional neural network enhanced with attention and gating mechanisms and a multilayer perceptron, along with a crosslinking density descriptor, is used to construct a Gaussian process regression model. By collecting polyimide data to expand the database, the performance of epoxy resin is predicted.
It enables the rapid design of new materials that are adaptable to a wide range of epoxy resin systems, reduces R&D costs, and improves the accuracy of performance prediction, especially significantly reducing prediction errors in modulus, strength, and elongation at break.
Smart Images
Figure CN116110517B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for predicting the performance of epoxy resins and its application. Background Technology
[0002] Advanced resin-based composite materials refer to composite materials that use high-performance fibers as reinforcement and high-performance resins as the matrix to meet the service conditions of aerospace and weaponry. They possess advantages such as high specific strength and high specific modulus, achieving purposes such as lightweighting and energy saving. While the tensile strength and modulus of carbon fibers have continuously improved, the tensile strength of carbon fiber composite materials has also increased, but the compressive modulus has not seen a significant improvement. This imbalance between tensile and compressive strength greatly limits the overall performance of carbon fiber composite materials, hindering breakthroughs in their structural design.
[0003] Developing high-strength, high-modulus, and high-toughness epoxy matrix resins is an effective way to achieve tensile-compressive balance in third-generation advanced carbon fiber composites. However, the introduction of rigid groups often leads to a decrease in toughness while increasing strength and modulus. The traditional trial-and-error development model of "classical, superficial theory + extensive experimental comparison" not only results in long development cycles and consumes significant human and material resources, but also makes it difficult to solve the contradictory performance problem of simultaneously improving the modulus, strength, and toughness of the epoxy resin matrix.
[0004] Currently, machine learning models for polymers mainly target small molecule systems and linear polymers. Epoxy resins, as cross-linked polymers, have seen limited experimental data and complex cross-linking structures, making them difficult to represent, resulting in few reports in the field of machine learning. Jin et al. calculated the glass transition temperature and stress-strain curves of several specific epoxy resin systems using molecular dynamics simulations and trained a backpropagation neural network using the simulation data, improving the glass transition temperature, Young's modulus, ultimate tensile strength, and elongation of specific epoxy resin systems. However, this model is only suitable for formulation optimization of specific epoxy resin systems and cannot be used for screening epoxy resin systems.
[0005] Therefore, it is essential to develop a method for predicting the performance of epoxy resins that can be adapted to a wide range of epoxy resin systems and can quickly design epoxy resin structures with contradictory properties. Summary of the Invention
[0006] To overcome the shortcomings of existing traditional experimental trial-and-error methods in the research and development process, such as high cost and low efficiency in the design of epoxy resins with conflicting performance characteristics, this invention provides a method for predicting the performance of epoxy resins. This method has high development efficiency, low economic cost, and can realize the rational design of advanced epoxy resins.
[0007] Machine learning models are often built on large datasets. In their research on methods for predicting epoxy resin properties, the inventors addressed the issue of limited experimental data for epoxy resins. Considering that the performance data of polyimide, such as modulus, strength, and toughness, are similar to those of epoxy resin, and that polyimide is a two-component reactant similar to the epoxy resin used in this study, the preferred design specifically included collecting polyimide data to expand the database and build the predictive model. Furthermore, to address the difficulty in representing the crosslinking structure of epoxy resin, the inventors introduced a crosslinking density descriptor to describe the crosslinking density of the crosslinking structure, effectively reducing model error. The introduction of the crosslinking density descriptor also allows for structural differentiation between epoxy resin and polyimide.
[0008] The inventors' machine learning model, which combines a graph convolutional neural network enhanced with specific attention and gating mechanisms with a multilayer perceptron, exhibits the strongest dimensionality reduction capability compared to corresponding models constructed through principal component analysis, minimum absolute value convergence and selection operator algorithms, unified manifold approximation and projection. When used to describe small molecule structures, the resulting prediction model achieves higher accuracy.
[0009] The present invention solves the above-mentioned technical problems through the following technical solutions.
[0010] This invention provides a method for predicting the performance of epoxy resins, comprising the following steps:
[0011] S1. Collect a dataset, which is performance data of epoxy resin;
[0012] A machine learning model combining a graph convolutional neural network with enhanced attention and gating mechanisms and a multilayer perceptron is constructed to automatically learn the latent vectors of small molecular structures of epoxy resin, thereby obtaining a structural representation model of epoxy resin.
[0013] Calculate polymer property descriptors, including crosslinking density descriptors for epoxy resin systems;
[0014] S2. Using the epoxy resin structure representation model, transform the dataset into a potential vector-performance dataset;
[0015] S3. Based on the latent vector-performance dataset, the relative molecular mass of the epoxy resin, and the relationship between the latent vector and computational performance mapped by the polymer feature descriptor, construct a Gaussian process regression model;
[0016] S4. Input the structural information of the epoxy resin to be analyzed into the Gaussian process regression model to obtain the performance data of the epoxy resin to be analyzed.
[0017] In step S1, the dataset typically contains more than 50 data entries. Preferably, the dataset also includes performance data of the polyimide. The introduction of the polyimide performance data can enhance the generalization ability of the model. Those skilled in the art will understand that obtaining the performance data of the epoxy resin and / or polyimide generally includes collecting data from literature and / or during the synthesis and preparation of the epoxy resin and the polyimide.
[0018] In step S1, the performance data preferably includes one or more of modulus data, strength data, and elongation at break data. Specifically, the modulus data preferably includes 20-60 epoxy resin data points and 120-180 polyimide data points; for example, 43 epoxy resin data points and 160 polyimide data points. The strength data preferably includes 15-40 epoxy resin data points and 120-150 polyimide data points; for example, 29 epoxy resin data points and 143 polyimide data points. The elongation at break preferably includes 20-60 epoxy resin data points and 60-100 polyimide data points; for example, 35 epoxy resin data points and 80 polyimide data points. Those skilled in the art will understand that the more performance data points, the better.
[0019] In step S1, the preferred method for constructing the machine learning model combining an attention- and gating-enhanced graph convolutional neural network with a multilayer perceptron is to encode the epoxy resin structure in the form of a molecular graph and use it as the frame input, and to use a computable molecular descriptor as an auxiliary property of the epoxy resin structure and use it as the frame output; the preferred method for obtaining the latent vector of the small molecular structure of the epoxy resin is to reduce the dimensionality of the molecular descriptor through the machine learning model to obtain the latent vector of the small molecular structure of the epoxy resin; the latent vector is the dimensionality-reduced molecular descriptor.
[0020] The method for obtaining the epoxy resin structure is conventional in the art, generally involving searching a database for small molecule structures suitable for transfer learning. These small molecule structures are epoxy and amino groups. The database is conventional in the art, preferably a publicly available database, such as SciFinder, PubChem, or ChemSpider. The number of small molecule structures is likely to be over 200,000, preferably 230,000-240,000, for example, 238,674. Those skilled in the art will understand that these small molecule structures are suitable for transfer learning.
[0021] The method of encoding the epoxy resin structure in the form of a molecular diagram preferably includes: representing it through a molecular adjacency matrix and the atomic number of each atom, the number of connected atoms, the number of connected hydrogen atoms, the number of implicit hydrogens, and the aromaticity index.
[0022] The dimension reduction can be 8, 16, 32, 64, or 128, with 8 dimensions being preferred. That is, the model has an eight-dimensional hidden layer.
[0023] Preferably, the method for calculating the molecular descriptors is to use the Mordred package to calculate the descriptors of epoxy resins, and the molecular descriptors can be 1081 types.
[0024] In step S1, the method for obtaining the polymer characteristic descriptor is preferably to use the Mordred package to calculate the polymer characteristic descriptor of the epoxy resin. Those skilled in the art will know that the polymer characteristics generally also include the number of epoxy or amino functional groups and the mass ratio of each group in the epoxy resin. The method for obtaining the crosslinking density is preferably to calculate it based on the theoretical derivation of Miller and Macosko according to classical gel theory; wherein, preferably, the calculation of the crosslinking density can be set with the following conditions:
[0025] a. The change in the density of epoxy resin in the experimental samples is negligible; b. The degree of polymerization of all epoxy resins is 0.9.
[0026] In step S1, the steps of collecting the dataset, obtaining the epoxy resin structure representation model, and defining the polymer property descriptor are not in any particular order.
[0027] In step S3, the Gaussian process regression model can be a high-throughput computation, a performance surrogate, or a quantitative structure-performance relationship model, preferably a quantitative structure-performance relationship model. The ratio of the training set to the test set of the Gaussian process regression model is preferably 4:1.
[0028] In step S4, the performance data of the epoxy resin to be analyzed is preferably represented in a material property space. The epoxy resin to be analyzed may include known epoxy resin structures and hypothetical epoxy resin structures.
[0029] The method for obtaining the virtual epoxy resin structure preferably includes the following steps:
[0030] (1) Define the gene types of epoxy resin structures, including epoxy compounds and amines;
[0031] (2) Use databases to search for genes, and split or combine existing chemical structures to obtain diverse gene pools;
[0032] The search for the gene preferably satisfies the following conditions: 1) the gene must be a complete molecule containing one or more epoxy or amino groups; 2) epoxy and amino groups do not exist simultaneously on the same molecule; 3) the number of nodes in the molecular graph does not exceed 50; the method for splitting or combining the existing chemical structure is preferably the Recap algorithm, the BRICS algorithm, or a method implemented by the user based on a programming language, with the BRICS algorithm being preferred; the gene pool may include 700-800 of the epoxy compounds and 300-400 of the amine molecules, for example, 746 of the epoxy compounds and 327 of the amine molecules.
[0033] (3) Based on the gene pool, the virtual epoxy resin structure is obtained by combination; the combination method is preferably implemented by Python programming; the number of virtual epoxy resin structures can be more than 200,000, preferably 240,000 to 250,000, for example 243,972.
[0034] The virtual epoxy resin structure can be a cross-linked structure formed by the reaction of one or more epoxy compounds and one or more amines, preferably an epoxy compound and an amine.
[0035] Preferably, after obtaining the material property space, the method further includes screening the property space of the epoxy resins to be analyzed, thereby further screening the epoxy resins to be analyzed. The screening method preferably employs a scalar function in the form of a weighted product or a Pareto dominance relation, and more preferably a scalar function in the form of a weighted product.
[0036] Following step S4, the method further includes analyzing the relationship between genes in the epoxy resin structure and the properties to identify key genes affecting the performance data. The method for analyzing the relationship between genes in the epoxy resin structure and the properties may involve statistically analyzing the frequency of gene occurrence in the preferred structure or statistically analyzing the changes in performance distribution caused by the introduction of genes; preferably, the frequency of gene occurrence leads to changes in performance distribution.
[0037] Those skilled in the art will understand that the structural information of the epoxy resin to be analyzed is generally obtained by converting the structure of the epoxy resin to be analyzed into the corresponding SMILES input. Furthermore, those skilled in the art will understand that the epoxy resin structures described in this invention are generally represented using SMILES.
[0038] Those skilled in the art will know that SMILES, short for Simplified molecular input lineentry specification, is a specification that explicitly describes molecular structures using ASCII strings.
[0039] In this invention, the machine learning model that combines a graph convolutional neural network with enhanced attention and gating mechanisms with a multilayer perceptron can be simply referred to as the GCN+a+g-MLP model.
[0040] This invention provides the application of the above-mentioned epoxy resin performance prediction method in predicting epoxy resin performance.
[0041] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.
[0042] The reagents and raw materials used in this invention are all commercially available.
[0043] The positive and progressive effects of this invention are as follows:
[0044] 1. The epoxy resin performance prediction method of this invention is adaptable to a wide range of epoxy resin systems, enabling rapid design of novel epoxy resin materials with excellent comprehensive properties, and identifying key factors affecting epoxy resin performance. It boasts high R&D efficiency, low economic cost, and facilitates the rational design of advanced epoxy resins. In predicting the modulus, strength, and elongation at break of epoxy resins, the performance prediction method of this invention achieves the following results: for modulus performance, the RMSE and MAE in the model's test set are 0.91 GPa and 0.69 GPa, respectively. For strength performance, the lowest RMSE value is 2.42 MPa, and the lowest MAE value is 1.76 MPa. For elongation at break performance, the average RMSE value is 1.94%, and the average MAE value is 1.43%.
[0045] 2. The performance prediction method of this invention verifies the structural performance through experimental synthesis and can analyze the key factors affecting the performance of epoxy resin. The obtained structure is verified experimentally, and its performance meets expectations. This structure is expected to be applied in aerospace and other technological fields. Attached Figure Description
[0046] Figure 1 This is a flowchart of the prediction model for tensile modulus, strength, and elongation at break in Example 1.
[0047] Figure 2 This is a comparison of the dimensionality reduction performance of the GCN+a+g-MLP model in Example 1 with other traditional models.
[0048] Figure 3 The performance of the models in Example 1 and Comparative Example 4 is compared.
[0049] Figure 4 This is the composition of the small molecule descriptor and polymer descriptor in Example 1.
[0050] Figure 5This refers to the accuracy of the quantitative structure-property relationship model of epoxy resin in Example 1.
[0051] Figure 6 The performance space of the candidate virtual epoxy resin structure predicted in Example 1.
[0052] Figure 7 The stress-strain curves are from the experimental tests in Example 1.
[0053] Figure 8 The performance of the epoxy resin designed in Example 1 is compared with that of epoxy resins known in the literature.
[0054] Figure 9 This is the result of gene analysis for high modulus, high strength, and high toughness in Example 1. Detailed Implementation
[0055] The present invention is further illustrated below by way of embodiments, but the invention is not limited to the scope of the embodiments described herein. Experimental methods in the following embodiments that do not specify specific conditions were performed according to conventional methods and conditions, or as selected according to the product instructions.
[0056] Example 1
[0057] S1. Collect datasets on modulus, strength, and elongation at break of epoxy resin and polyimide from the literature. The modulus data includes 43 data points for epoxy resin and 160 data points for polyimide, the strength data includes 29 data points for epoxy resin and 143 data points for polyimide, and the elongation at break includes 35 data points for epoxy resin and 80 data points for polyimide.
[0058] S2. Construct a machine learning model combining a graph convolutional neural network enhanced with attention and gating mechanisms with a multilayer perceptron to obtain the latent vector of small molecule structures, specifically:
[0059] (1) Using the public databases SciFinder, PubChem, and ChemSpider, we searched for small molecule structures for transfer learning. A total of 238,674 small molecule structures with epoxy and amino groups were found.
[0060] (2) Use the Mordred package to calculate 1081 molecular descriptors for the small molecules in step (1).
[0061] (3) The small molecule structure in step (1) is represented in the form of a molecular diagram, specifically by the molecular adjacency matrix and the atomic number of each atom, the number of connected atoms, the number of connected hydrogen atoms, the number of implicit hydrogens and the aromaticity index.
[0062] (4) Using the molecular diagram as model input and the 1081 small molecule descriptors calculated by the Mordred package as model output, a structural representation model of GCN+a+g-MLP is constructed. An eight-dimensional hidden layer is set in the model, and the hidden layer result is the dimensionality reduction result (e.g., ...). Figure 1 ).
[0063] S3. The Mordred package was used to calculate polymer property descriptors. Polymer properties include the crosslinking density descriptor of the epoxy resin system, the number of epoxy or amino functional groups, and the mass ratio of each group in the epoxy resin. The crosslinking density descriptor was obtained by calculating the crosslinking density of the epoxy resin system based on classical gel theory and the theoretical derivation of Miller and Macosko. The calculation of the crosslinking density was set under the following conditions: a. The change in epoxy resin density in the experimental sample is negligible; b. The degree of polymerization of all epoxy resins is 0.9. The developed crosslinking density descriptor is as follows: Figure 4 As shown.
[0064] S4. Using a structural representation model, transform the dataset from step S1 into a latent vector-performance dataset;
[0065] S5. Based on the latent vector-performance dataset, relative molecular mass, and polymer feature descriptors mapping the relationship between latent vectors and computational performance, a Gaussian process regression model is constructed; that is, the latent vector, relative molecular mass, and polymer characteristic descriptors are used as input parameters, and the performance data of modulus, strength, and elongation at break of epoxy resin and polyimide are used as output parameters; a quantitative structure-performance relationship model is constructed, with a training set to test set split ratio of 4:1. The accuracy and error of the training set and test set are shown in [reference needed]. Figure 5 .
[0066] S6. Input the structural information of the epoxy resin to be analyzed into the Gaussian process regression model to obtain the performance data of the epoxy resin to be analyzed.
[0067] The epoxy resin to be analyzed includes known epoxy resin structures and virtual epoxy resin structures. The method for obtaining the virtual epoxy resin structure includes the following steps:
[0068] (1) Define the gene types of epoxy resin structures, including epoxy compounds and amines;
[0069] (2) Genes were searched using the public databases SciFinder, PubChem, and ChemSpider. Existing chemical structures were disassembled to obtain a gene pool of 746 epoxides and 327 amine molecules.
[0070] The search for the gene must meet the following conditions: 1) The gene must be a complete molecule containing one or more epoxy or amino groups; 2) Epoxy and amino groups do not exist simultaneously on the same molecule; 3) The number of nodes in the molecular graph does not exceed 50.
[0071] The BRICS algorithm is a method for splitting and combining existing chemical structures.
[0072] (3) Based on the gene pool, 243,972 virtual epoxy resin structures were obtained.
[0073] Comparative Example 1
[0074] The difference between Comparative Example 1 and Example 1 is that a machine learning model is constructed using PCA (principal component analysis).
[0075] Comparative Example 2
[0076] The difference between Comparative Example 2 and Example 1 is that the machine learning model is constructed using LASSO (least absolute shrinkage and selection operator).
[0077] Comparative Example 3
[0078] The difference between Comparative Example 3 and Example 1 is that the machine learning model is constructed using UMAP (Uniform Manifold Approximation and Projection).
[0079] Application Example 1
[0080] The models of Example 1 and Comparative Examples 1-3 were constructed and run randomly 50 times. The RMSE (root mean square error) and MSE (mean square error) of the obtained models are shown in [reference needed]. Figure 2 In the graph, the orange area represents the error of the training set, and the blue area represents the error of the test set. Figure 2 From a1 and a2, we can see that, for modulus performance, the average RMSE and MAE of the GCN+a+g-MLP model on the test set are 0.91 GPa and 0.69 GPa, respectively, which are lower than the values of the models in comparative examples 1-3. From Figure 2 From b1 and b2, we can see that, for strength performance, the GCN+a+g-MLP model has an average RMSE of 2.42 MPa and an average MAE of 1.76 MPa, which are the lowest among the four models. Figure 2As can be seen from c1 and c2, for elongation at break, the average RMSE of the GCN+a+g-MLP model is 1.94% and the average MAE is 1.43%, which are the lowest compared to the models in Comparative Examples 1-3.
[0081] Comparative Example 4
[0082] The difference between Comparative Example 4 and Example 1 is that the polymer property descriptor in step S1 does not include a crosslinking density descriptor. The RMSE (root mean square error) and MSE (mean square error) of the models with and without crosslinking descriptors are shown below. Figure 3 . Figure 3 In this context, a1, b1, and c1 represent the root mean square errors of the modulus, strength, and elongation at break of the models of Example 1 (with crosslinking density symbol) and Comparative Example 4 (without crosslinking density symbol). Figure 3 In the figure, a2, b2, and c2 represent the mean square errors of the modulus, strength, and elongation at break of the models in Example 1 (with crosslinking density symbol) and Comparative Example 4 (without crosslinking density symbol). The orange area in the figure represents the error of the training set, and the blue area represents the error of the test set. Figure 3 It can be seen that introducing cross-linking descriptors can effectively reduce model errors.
[0083] Example 1: Design and verification of epoxy resin based on the materials genome method.
[0084] In this embodiment, based on the framework of the materials genome approach, the Gaussian process regression model established in Example 1 was applied to screen out preferred epoxy resins that exhibit superior tensile modulus, strength, and toughness compared to existing epoxy resins. The research steps include:
[0085] S1. Calculate the small molecule descriptors and molecular diagrams of the structures of the 243,972 virtual epoxy resins in Example 1; use the Gaussian process regression model in Example 1 to obtain the performance data of their virtual epoxy resin structures.
[0086] S2. A scalar function in the form of a general weighted product (Formula 1) is used to represent the three properties with a comprehensive score, and the comprehensive score is used to color the material property space, such as... Figure 6 As shown. The colors from purple to red represent overall performance from low to high. Next, considering the feasibility of synthesis, one of the top 20% of preferred structures with a higher probability of synthesis was selected for experimental verification.
[0087]
[0088] Where M, S, and E represent tensile modulus, tensile strength, and elongation at break, respectively.
[0089] S3. The selected NPEP resin was synthesized experimentally, and its tensile modulus, tensile strength, and elongation at break were tested. Its properties after curing are shown in [see figure]. Figure 7 The prepared NPEP resin had a tensile modulus (room temperature) of 3.85±0.19 GPa, a tensile strength (room temperature) of 89.25±0.88 MPa, and an elongation at break (room temperature) of 6.71±0.94%, which were similar to the results predicted using the machine learning model (tensile modulus (room temperature) 3.57±1.43 GPa, tensile strength (room temperature) 99.0±38.4 MPa, elongation at break (room temperature) 6.08±3.17%). The performance of the screened NPEP was compared with that of existing epoxy resin systems in the literature. Figure 8 The results show that the selected epoxy resin has superior tensile modulus, strength and toughness compared to existing epoxy resins.
[0090] S4. By statistically analyzing the changes in performance distribution under different gene frequencies, key genes affecting tensile modulus, tensile strength, and elongation at break were identified, such as... Figure 9 As shown.
Claims
1. A method for predicting the performance of epoxy resin, characterized in that, It includes the following steps: S1. Collect a dataset, which includes performance data of epoxy resin; construct a machine learning model that combines a graph convolutional neural network with attention and gating mechanisms and a multilayer perceptron to automatically learn the latent vectors of small molecular structures of epoxy resin and obtain an epoxy resin structure representation model. Calculate polymer property descriptors, including crosslinking density descriptors for epoxy resin systems; S2. Using the epoxy resin structure representation model, transform the dataset into a potential vector-performance dataset; S3. Based on the latent vector-performance dataset, the relative molecular mass of the epoxy resin, and the polymer property descriptor mapping the relationship between latent vectors and computational performance, construct a Gaussian process regression model; S4. Input the structural information of the epoxy resin to be analyzed into the Gaussian process regression model to obtain the performance data of the epoxy resin to be analyzed.
2. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The dataset also includes performance data for polyimide; And / or, the types of performance data include one or more of modulus, strength, and elongation at break data.
3. The method for predicting the performance of epoxy resin as described in claim 2, characterized in that, The modulus data includes 20-60 data points for epoxy resins and 120-180 data points for polyimides.
4. The method for predicting the performance of epoxy resin as described in claim 3, characterized in that, The modulus data includes 43 data points for epoxy resin and 160 data points for polyimide.
5. The method for predicting the performance of epoxy resin as described in claim 2, characterized in that, The strength data includes 15-40 data points for epoxy resin and 120-150 data points for polyimide.
6. The method for predicting the performance of epoxy resin as described in claim 5, characterized in that, The strength data includes 29 data points for epoxy resin and 143 data points for polyimide.
7. The method for predicting the performance of epoxy resin as described in claim 2, characterized in that, The elongation at break data includes 20-60 data points for epoxy resins and 60-100 data points for polyimides.
8. The method for predicting the performance of epoxy resin as described in claim 7, characterized in that, The elongation at break data includes 35 data points for epoxy resins and 80 data points for polyimides.
9. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The method for constructing a machine learning model that combines a graph convolutional neural network enhanced with attention and gating mechanisms with a multilayer perceptron is as follows: the epoxy resin structure is encoded in the form of a molecular graph and used as the frame input; a computable molecular descriptor is used as an auxiliary property of the epoxy resin structure and used as the frame output; the method for obtaining the latent vector of the small molecular structure of epoxy resin is to reduce the dimensionality of the molecular descriptor through the machine learning model to obtain the latent vector of the small molecular structure of epoxy resin.
10. The method for predicting the performance of epoxy resin as described in claim 9, characterized in that, The epoxy resin structure was obtained by searching a database for small molecule structures for transfer learning, wherein the small molecule structures are epoxy groups and amino groups.
11. The method for predicting the performance of epoxy resin as described in claim 10, characterized in that, The database in question is a public database.
12. The method for predicting the performance of epoxy resin as described in claim 11, characterized in that, The database is SciFinder, PubChem, or ChemSpider.
13. The method for predicting the performance of epoxy resin as described in claim 9, characterized in that, The number of small molecular structures is over 200,000.
14. The method for predicting the performance of epoxy resin as described in claim 13, characterized in that, The number of small molecular structures is 230,000 to 240,000.
15. The method for predicting the performance of epoxy resin as described in claim 14, characterized in that, The number of small molecular structures is 238,674.
16. The method for predicting the performance of epoxy resin as described in claim 9, characterized in that, Encoding the epoxy resin structure in the form of a molecular diagram includes: representing it through a molecular adjacency matrix and the atomic number of each atom, the number of connected atoms, the number of connected hydrogen atoms, the number of implicit hydrogens, and the aromaticity index.
17. The method for predicting the performance of epoxy resin as described in claim 9, characterized in that, The dimensions of the dimensionality reduction are 8, 16, 32, 64, and 128.
18. The method for predicting the performance of epoxy resin as described in claim 17, characterized in that, The dimension reduction is 8-dimensional.
19. The method for predicting the performance of epoxy resin as described in claim 9, characterized in that, The calculable molecular descriptor is obtained by using the Mordred package to calculate the descriptor of the epoxy resin.
20. The method for predicting the performance of epoxy resin as described in claim 19, characterized in that, There are 1081 molecular descriptors.
21. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The polymer property descriptor is obtained by using the Mordred package to calculate the polymer property descriptor of the epoxy resin.
22. The method for predicting the performance of epoxy resin as described in claim 21, characterized in that, The polymer properties also include the number of epoxy or amino functional groups and the mass ratio of each group in the epoxy resin.
23. The method for predicting the performance of epoxy resin as described in claim 21, characterized in that, The crosslinking density is obtained by setting the following conditions for its calculation: a. The change in the density of epoxy resin in the experimental sample is negligible; b. The degree of polymerization of all epoxy resins is 0.
9.
24. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The Gaussian process regression model is a high-throughput computation, performance surrogate quantity, or quantitative structure-performance relationship model.
25. The method for predicting the performance of epoxy resin as described in claim 24, characterized in that, The Gaussian process regression model is a quantitative structure-performance relationship model.
26. The method for predicting the performance of epoxy resin as described in claim 24, characterized in that, The ratio of the training set to the test set of the Gaussian process regression model is 4:
1.
27. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The performance data of the epoxy resin to be analyzed are expressed in a material property space.
28. The method for predicting the performance of epoxy resin as described in claim 27, characterized in that, After obtaining the material property space of the epoxy resins to be analyzed, the method further includes screening the property space to further screen the epoxy resins to be analyzed.
29. The method for predicting the performance of epoxy resin as described in claim 28, characterized in that, The screening method is to use a scalar function in the form of a weighted product or a Pareto dominance relation.
30. The method for predicting the performance of epoxy resin as described in claim 29, characterized in that, The filtering method is a scalar function in the form of a weighted product.
31. The method for predicting the performance of epoxy resin as described in claim 1, characterized in that, The epoxy resin to be analyzed includes known epoxy resin structures and virtual epoxy resin structures.
32. The method for predicting the performance of epoxy resin as described in claim 31, characterized in that, The method for obtaining a virtual epoxy resin structure includes the following steps: (1) Define the gene types of epoxy resin structures, including epoxy compounds and amines; (2) Use databases to search for genes, disassemble existing chemical structures, and obtain diverse gene pools; (3) Based on the gene pool, virtual epoxy resin candidate structures are combined; the virtual epoxy resin candidate structures obtained by the combination are more than 200,000.
33. The method for predicting the performance of epoxy resin as described in claim 32, characterized in that, The combination yields 240,000 to 250,000 virtual epoxy resin candidate structures.
34. The method for predicting the performance of epoxy resin as described in claim 33, characterized in that, The combination yielded 243,972 virtual epoxy resin candidate structures.
35. The method for predicting the performance of epoxy resin as described in any one of claims 31-34, characterized in that, The search gene must meet the following conditions: 1) The gene must be a complete molecule containing one or more epoxy or amino groups; 2) Epoxy and amino groups do not exist simultaneously on the same molecule; 3) The number of nodes in the molecular graph does not exceed 50. And / or, the method for splitting or combining the existing chemical structures is the Recap algorithm, the BRICS algorithm, or a method implemented by the user based on a programming language; And / or, the gene pool comprises 700-800 of the epoxy compound and 300-400 of the amine molecules; And / or, the virtual epoxy resin structure is a cross-linked structure generated by the reaction of one or more epoxy compounds and one or more amines.
36. The method for predicting the performance of epoxy resin as described in claim 35, characterized in that, The method for splitting or combining the existing chemical structures is the BRICS algorithm.
37. The method for predicting the performance of epoxy resin as described in claim 35, characterized in that, The gene pool comprises 746 of the epoxy compounds and 327 amine molecules.
38. The method for predicting the performance of epoxy resin as described in claim 35, characterized in that, The virtual epoxy resin structure consists of an epoxy compound and an amine.
39. The method for predicting the performance of epoxy resin as described in any one of claims 27-30, characterized in that, After obtaining the material property space of the epoxy resins to be analyzed, the method further includes screening the property space to further screen the epoxy resins to be analyzed; after step S4, the method further includes analyzing the relationship between the genes of the epoxy resin structure and the properties to obtain the key genes that affect the performance data.
40. The method for predicting the performance of epoxy resin as described in claim 39, characterized in that, The method for analyzing the relationship between genes in epoxy resin structure and its properties is to statistically analyze the frequency of gene occurrence in the preferred structure or to statistically analyze the changes in performance distribution caused by the introduction of genes.
41. The method for predicting the performance of epoxy resin as described in claim 40, characterized in that, The method for analyzing the relationship between genes in epoxy resin structure and its properties is based on the fact that the frequency of gene occurrence leads to changes in property distribution.
42. The application of a method for predicting the properties of an epoxy resin as described in any one of claims 1-41 in predicting the properties of an epoxy resin to be analyzed.
43. The application of the epoxy resin performance prediction method as described in claim 42 in predicting the performance of the epoxy resin to be analyzed, characterized in that, The properties mentioned are modulus, strength, and elongation at break.