A method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks

Abstract

This invention discloses a method for predicting the molecular structure-modulus performance relationship of glass fibers based on graph neural networks, comprising the following steps: screening composition-modulus data for multi-component glass fiber systems from a database; constructing graph data based on the screened composition data; and predicting the modulus using a graph neural network based on the graph data. This invention innovatively proposes an improved graph data construction method, introduces joint modeling of "node representation + edge features + composition vectors," introduces a triple information fusion of "local-global-composition," and introduces structural consistency regularization. This allows the model to consider both the direct contribution of local structure to the modulus and the regulatory effect of the overall composition on macroscopic performance, thereby improving prediction accuracy and generalization. This invention maintains high prediction accuracy and physical rationality, demonstrating strong generalization ability and practical application value.

Description

Technical Field

[0001] This invention belongs to the fields of glass fiber and artificial intelligence technology, and in particular, it is a method for predicting the relationship between molecular structure and modulus properties of glass fiber based on graph neural networks. Background Technology

[0002] Glass fiber is an important inorganic non-metallic material widely used in construction, transportation, energy, and electronics. Its mechanical properties, especially its modulus, directly affect its effectiveness in composite reinforcement, structural support, and the manufacture of high-performance products. Increasing the modulus not only improves the load-bearing capacity of materials but also enhances the stability and durability of composite materials. Therefore, accurately predicting the modulus properties of glass fiber during the composition design stage is a key issue in materials design and industrial applications.

[0003] Existing modulus prediction methods mainly rely on two types. One type is experimental determination, which can obtain relatively accurate results, but the experimental process is often time-consuming, costly, and difficult to cover the vast compositional space. The other type is empirical formulas and semi-empirical models, which estimate modulus by establishing an empirical relationship between chemical composition ratios and modulus. These methods are simple, but because glass is a complex amorphous system with a lack of periodicity in its molecular structure and diverse network configurations, empirical formulas cannot accurately characterize the correspondence between composition and properties, resulting in limited prediction accuracy.

[0004] In recent years, machine learning methods have been increasingly applied to predicting glass performance. For example, regression models, support vector machines, and random forests can use component vectors (such as the proportions of SiO2, Al2O3, and CaO) as input to build predictive models, improving prediction efficiency to some extent. However, these methods only use component proportions as features, ignoring the interactions between atoms and the characteristics of the glass network structure. Furthermore, feature engineering relies on manual design, resulting in significant deficiencies in generalization ability and prediction accuracy.

[0005] With the development of artificial intelligence, Graph Neural Networks (GNNs) have demonstrated unique advantages in predicting molecular and material properties. GNNs can represent systems as graph structures, where nodes can represent atoms or oxide units, and edges can represent bonding relationships or interactions. They learn the correlation between local structures and the global network through message passing mechanisms. Compared to traditional methods, GNNs can automatically extract high-order structural features, avoiding the limitations of manual feature construction, and exhibit superior accuracy and scalability in areas such as organic molecular property prediction and crystal performance estimation.

[0006] However, research on amorphous glass systems remains limited. Glass molecules have highly complex structures and diverse coordination environments, making it difficult for traditional composition vectors to effectively reflect their true structure. Existing research largely focuses on crystalline or organic molecular systems, failing to adequately address the structural characterization challenges in predicting the modulus of glass fibers. Therefore, a novel method combining compositional information with amorphous structural characteristics is urgently needed to improve the accuracy and applicability of predictions. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the prior art by providing a method for predicting the relationship between molecular structure and modulus properties of glass fibers based on graph neural networks.

[0008] The technical solution to achieve the objective of this invention is as follows: On the one hand, a method for predicting the relationship between the molecular structure and modulus properties of glass fibers based on graph neural networks is provided, the method comprising the following steps:

[0009] Step 1: Filter composition-modulus data for multi-component glass fiber systems from the database;

[0010] Step 2: Based on the component data selected in Step 1, construct the graph data;

[0011] Step 3: Based on the graph data, perform modulus prediction using a graph neural network.

[0012] Furthermore, the multi-component glass fiber system in step 1 is 16-yuan system.

[0013] Furthermore, step 2, which involves constructing graph data based on the component data selected in step 1, specifically includes:

[0014] Step 2-1, construct node features;

[0015] The m oxides in the multi-component glass fiber system are defined as m nodes in the figure;

[0016] For the i-th node, its feature vector is represented as :

[0017]

[0018] In the formula, Represents the i-th node Mole fraction in sample, i.e., component data; Let j represent the m-th physicochemical property of the i-th node, where j = 1, 2, ..., m; This represents the k-th external descriptor of the i-th node;

[0019] Step 2-2, construct the edge features as follows:

[0020]

[0021] in,

[0022]

[0023]

[0024]

[0025] In the formula, These are the features of the first side, the second side, and the third side, respectively. This represents the mole fraction of the j-th node in the sample, i.e., the component data; , Let be the four-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, represented as . ; Let be the three-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, denoted as . ;

[0026] Steps 2-3, formalize each sample as follows:

[0027]

[0028] In the formula, A set of nodes; Let it be the set of edges; The node feature matrix; The edge feature matrix; It is the modulus.

[0029] Furthermore, step 3 specifically includes:

[0030] Step 3-1: Establish an improved information transmission mechanism for graph neural networks:

[0031]

[0032] in,

[0033]

[0034]

[0035] In the formula, For the message sent from the j-th node in the t-th layer of a graph neural network to the i-th node, For the aggregated message of the i-th node, Let be the importance weight of the j-th node to the i-th node in the t-th layer. Let j be the feature representation of the j-th node in the t-th layer. Let be the feature representation of the i-th node in the (t+1)-th layer. Let be the feature representation of the i-th node in the t-th layer. The layer normalization function is represented. This represents the activation function. This represents a multilayer perceptron that integrates all information. This is a non-linear encoding of edge features. For parameter matrices; These are trainable weight matrices used to linearly map the input node representation, neighbor representation, edge features, and global component vector to a vector space of uniform dimension. Let i be the set of nodes adjacent to the i-th node. It is a component vector;

[0036] Step 3-2: Implement message passing and node state update in the graph neural network based on the mechanism in Step 3-1;

[0037] Step 3-3: Calculate the graph-level representation of the sample;

[0038] Steps 3-4: Calculate the modulus based on graph-level representation.

[0039] Furthermore, in step 3-3, the graph-level representation of the sample is calculated using the following formula:

[0040]

[0041] in,

[0042]

[0043]

[0044] In the formula, For the graph-level representation of the samples, For the readout stage, a multilayer perceptron. For the introduced component vector, This indicates the characteristics of local weighted pooling. Represents global statistical characteristics. This represents the local weight of the i-th node. Let be the representation of the i-th node in the T-th layer of the graph neural network, where T is the depth of the graph neural network. Represents the mean function, represents the standard deviation function; N represents the total number of nodes.

[0045] Furthermore, the calculation formula for the modulus based on graph-level representation described in steps 3-4 is as follows:

[0046]

[0047] In the formula, This is the predicted value of the modulus. These are the weight matrix and bias term of the readout layer for modulus prediction, respectively.

[0048] Furthermore, the loss function L of the graph neural network in step 3 is:

[0049]

[0050] in,

[0051]

[0052]

[0053] In the formula, For modulus prediction error, For consistency loss function, For regularization terms, All are hyperparameters; Let be the predicted value of the structure descriptor of the i-th node. Let i be the structure descriptor of the i-th node; This represents the total number of nodes. To utilize the node's structure descriptor Representation of the final layer nodes The function that performs inverse mapping constraints.

[0054] On the other hand, a graph neural network-based system for predicting the molecular structure-modulus relationship of glass fibers is provided, the system comprising:

[0055] The first module is used to: filter composition-modulus data for multi-component glass fiber systems from the database;

[0056] The second module is used to construct graph data based on the component data selected in the first module.

[0057] The third module is used to perform modulus prediction based on the graph data using a graph neural network.

[0058] On the other hand, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for predicting the molecular structure-modulus relationship of glass fibers based on graph neural networks.

[0059] On the other hand, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method for predicting the molecular structure-modulus relationship of glass fibers based on graph neural networks.

[0060] Compared with the prior art, the significant advantages of this invention are:

[0061] (1) In the process of constructing the graph data, in addition to considering conventional descriptors such as oxide composition and ionic radius, structural descriptors obtained through molecular dynamics calculations are also introduced for node features; for edge features, information on the chemical element combination and structural role combination of oxides is introduced. Through the above improvements, the local structural features and interatomic interaction relationships of the oxide system can be more comprehensively characterized, thereby improving the physical interpretability and discriminability of node features; at the same time, the chemical and structural combination information integrated in the edge features effectively reveals the coupling relationship and structural synergistic effect between different components, making the constructed graph structure more consistent with the real physical properties of the material system. This method can significantly enhance the sensitivity and expressive power of graph data to the microstructure of oxide systems, thereby improving the accuracy and generalization ability of the prediction model.

[0062] (2) Based on the traditional GNN message passing, a joint modeling of “node representation + edge features + component vector” is introduced; the local message passing is modulated by the global component vector c, so that the same local structure gets different weights in different component systems; the model’s ability to capture the chemical-structure-performance coupling relationship in amorphous glass is improved, and the prediction accuracy and generalization performance across component systems are enhanced.

[0063] (3) Based on the traditional readout function (readout layer), the "local-global-component" triple information fusion is introduced, so that the model can take into account both the direct contribution of local structure to the modulus and the adjustment effect of overall ratio on macro performance, thereby improving prediction accuracy and generalization.

[0064] (4) By introducing structural consistency regularization, the node embeddings learned by the model not only serve modulus prediction, but also remain consistent with physically observable structural representations. This design improves the physical reliability of the model and its generalization ability across compositional systems.

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0066] Figure 1 This is a flowchart of a method for predicting the molecular structure-modulus relationship of glass fibers based on graph neural networks in one embodiment.

[0067] Figure 2 This is a schematic diagram of the structural data for a glass formulation of 60% SiO2-20% Al2O3-10% MgO-10% CaO in one embodiment. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0069] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0070] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0071] In one embodiment, combined Figure 1 This paper provides a method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks. The method includes the following steps:

[0072] Step 1: Filter composition-modulus data for multi-component glass fiber systems from the database;

[0073] Step 2: Based on the component data selected in Step 1, construct the graph data;

[0074] Step 3: Based on the graph data, perform modulus prediction using a graph neural network.

[0075] More preferably, the multi-element glass fiber system in step 1 is 16-yuan system.

[0076] Furthermore, in one embodiment, step 2, which involves constructing graph data based on the component data selected in step 1, specifically includes:

[0077] Step 2-1, construct node features;

[0078] The m oxides in the multi-component glass fiber system are defined as m nodes in the figure;

[0079] For the i-th node, its feature vector is represented as :

[0080]

[0081] In the formula, Represents the i-th node Mole fraction in sample, i.e., component data; Let m represent the m-th physicochemical property of the i-th node (such as electronegativity, ionic radius, etc.), j=1,2,...,m; This represents the kth external descriptor of the i-th node (preferably, a structural descriptor obtained through molecular dynamics calculations).

[0082] Step 2-2, construct the edge features as follows:

[0083]

[0084] in,

[0085] The oxide composition of the nodes is as follows:

[0086]

[0087]

[0088]

[0089] In the formula, These are the features of the first side, the second side, and the third side, respectively. This represents the mole fraction of the j-th node in the sample, i.e., the component data; , Let be the four-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, represented as . ; Let be the three-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, denoted as . ; Represents the Kronecker product;

[0090] Here, the 16 oxides are classified according to their chemical elements as follows: non-metallic oxides SiO2 and B2O3; main group metal oxides Li2O, Na2O, K2O, MgO, CaO, SrO, and Al2O3; transition metal oxides TiO2, ZrO2, Fe2O3, and ZnO; and rare earth oxides La2O3, CeO2, and Eu2O3.

[0091] Here, the 16 oxides are categorized into three main types based on their structural roles: network-forming agents, intermediates, and modifiers. First, network-forming agents primarily function to form the framework structure of the glass network, providing the basic three-dimensional silicon-oxygen or boron-oxygen framework. In this system, network-forming agents include SiO2, B2O3, and Al2O3. Second, intermediates, these oxides cannot form a stable glass network independently, but under specific conditions, they can enter the network structure, enhancing, stabilizing, or regulating the network. Intermediates in this system include TiO2, ZrO2, ZnO, Fe2O3, CeO2, La2O3, Eu2O3, and Y2O3. Finally, modifiers primarily function to break strong covalent bonds in the network structure, thereby reducing the degree of polymerization and affecting properties such as viscosity, coefficient of thermal expansion, and modulus. In this system, modifiers include MgO, CaO, Li2O, Na2O, K2O, and SrO.

[0092] Steps 2-3, formalize each sample as follows:

[0093]

[0094] In the formula, A set of nodes; Let it be the set of edges; The node feature matrix; The edge feature matrix; It is the modulus.

[0095] Furthermore, step 3 specifically includes:

[0096] Step 3-1: Establish an improved information transmission mechanism for graph neural networks:

[0097]

[0098] in,

[0099]

[0100]

[0101] In the formula, For the message sent from the j-th node in the t-th layer of a graph neural network to the i-th node, For the aggregated message of the i-th node, Let be the importance weight of the j-th node to the i-th node in the t-th layer. Let j be the feature representation of the j-th node in the t-th layer. Let be the feature representation of the i-th node in the (t+1)-th layer. Let be the feature representation of the i-th node in the t-th layer. The layer normalization function is represented. This represents the activation function. This represents a multilayer perceptron that integrates all information. This is a non-linear encoding of edge features. For parameter matrices; These are trainable weight matrices used to linearly map the input node representations, neighbor representations, edge features, and global component vectors to a vector space of uniform dimension, so that attention scores can be calculated subsequently. Let i be the set of nodes adjacent to the i-th node. It is a component vector;

[0102] It should be noted here that the attention score is first defined as:

[0103]

[0104] To achieve trainable fusion, The result is obtained by concatenating elements one by one and then passing the result through a scalar MLP:

[0105] .

[0106] Step 3-2: Implement message passing and node state update in the graph neural network based on the mechanism in Step 3-1;

[0107] Step 3-3: Calculate the graph-level representation of the sample;

[0108] Steps 3-4: Calculate the modulus based on graph-level representation.

[0109] Preferably, in some embodiments, the graph-level representation of the sample is calculated in step 3-3, and the specific calculation formula is as follows:

[0110]

[0111] in,

[0112]

[0113]

[0114] In the formula, For the graph-level representation of the samples, For the readout stage, a multilayer perceptron. For the introduced component vector, This indicates the characteristics of local weighted pooling. Represents global statistical characteristics. This represents the local weight of the i-th node. Let be the representation of the i-th node in the T-th layer of the graph neural network, where T is the depth of the graph neural network. Represents the mean function, represents the standard deviation function; N represents the total number of nodes.

[0115] Here, the present invention proposes a multi-scale readout method based on component conditions. This method fuses local node information with global statistical features and introduces the component vector c as a condition to enhance the adaptability of the prediction model to different ratio systems.

[0116] Preferably, in some embodiments, the calculation of the modulus based on graph-level representation in steps 3-4 is performed using the following formula:

[0117]

[0118] In the formula, This is the predicted value of the modulus. These are the weight matrix and bias term of the readout layer for modulus prediction, respectively.

[0119] Preferably, in some embodiments, the loss function L of the graph neural network in step 3 is:

[0120]

[0121] in,

[0122]

[0123]

[0124] In the formula, For modulus prediction error, For consistency loss function, For regularization terms, All are hyperparameters; Let be the predicted value of the structure descriptor of the i-th node. Let i be the structure descriptor of the i-th node; This represents the total number of nodes. To utilize the node's structure descriptor Representation of the final layer nodes Functions that perform inverse mapping constraints (which can improve the physical consistency and interpretability of model prediction results).

[0125] In one embodiment, a graph neural network-based system for predicting the molecular structure-modulus relationship of glass fibers is provided, the system comprising:

[0126] The first module is used to: filter composition-modulus data for multi-component glass fiber systems from the database;

[0127] The second module is used to construct graph data based on the component data selected in the first module.

[0128] The third module is used to perform modulus prediction based on the graph data using a graph neural network.

[0129] Specific limitations regarding the graph neural network-based glass fiber molecular structure-modulus property prediction system can be found in the limitations of the graph neural network-based glass fiber molecular structure-modulus property prediction method described above, and will not be repeated here. Each module in the graph neural network-based glass fiber molecular structure-modulus property prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module.

[0130] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the graph neural network-based method for predicting the molecular structure-modulus relationship of glass fibers.

[0131] Step 1: Filter composition-modulus data for multi-component glass fiber systems from the database;

[0132] Step 2: Based on the component data selected in Step 1, construct the graph data;

[0133] Step 3: Based on the graph data, perform modulus prediction using a graph neural network.

[0134] For specific limitations on each step, please refer to the limitations of the method for predicting the relationship between molecular structure and modulus properties of glass fibers based on graph neural networks mentioned above, which will not be repeated here.

[0135] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the graph neural network-based method for predicting the molecular structure-modulus relationship of glass fibers.

[0136] Step 1: Filter composition-modulus data for multi-component glass fiber systems from the database;

[0137] Step 2: Based on the component data selected in Step 1, construct the graph data;

[0138] Step 3: Based on the graph data, perform modulus prediction using a graph neural network.

[0139] For specific limitations on each step, please refer to the limitations of the method for predicting the relationship between molecular structure and modulus properties of glass fibers based on graph neural networks mentioned above, which will not be repeated here.

[0140] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0141] Step 1, collect data from the database. Composition-modulus data for a 16-element system.

[0142] Step 2: Convert the extracted glass composition data into graph structure data. Figure 2 (A schematic diagram of the graph structure data for a glass formulation of 60% SiO2-20% Al2O3-10% MgO-10% CaO) is shown. A graph neural network is used for training to obtain a predictive model for the glass modulus.

[0143] Step 3: The method proposed in this invention is tested using 10 different formulations. The specific molar ratios are shown in Table 1. Table 2 shows the actual modulus and the prediction results and errors of three prediction models (the model of this invention, random forest, and SVM).

[0144] surface Molar ratio composition information of ten different glass formulations in a 16-element oxide system

[0145]

[0146] surface Comparison of the predicted and actual values ​​of glass modulus using this method and other methods.

[0147]

[0148] The comparison shows that the model of this invention is significantly superior to other methods in terms of overall accuracy and stability: its average error is controlled within about 1%, and the correlation between predicted and true values ​​is the highest, indicating that the model has a stronger ability to capture the nonlinear relationship between glass composition and modulus. Random Forest and SVM performed poorly, with average errors of approximately 3.5% and 4.5%, respectively, exhibiting large prediction fluctuations and significant overestimation or underestimation of some samples, indicating that both have limited adaptability to complex multi-component systems.

[0149] Overall, the method of this invention can maintain high prediction accuracy and physical rationality, demonstrating strong generalization ability and practical application value.

[0150] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks, characterized in that, The method includes the following steps: Step 1: Filter composition-modulus data for multi-component glass fiber systems from the database; Step 2: Based on the component data selected in Step 1, construct the graph data; Step 3: Based on the graph data, perform modulus prediction using a graph neural network; Step 2, which involves constructing graph data based on the component data selected in Step 1, specifically includes: Step 2-1, construct node features; The m oxides in the multi-component glass fiber system are defined as m nodes in the figure; For the i-th node, its feature vector is represented as : ； In the formula, Represents the i-th node Mole fraction in sample, i.e., component data; This represents the m-th physicochemical property of the i-th node, where i = 1, 2, ..., m; This represents the k-th external descriptor of the i-th node; Step 2-2, construct the edge features as follows: ； in, ； ； ； In the formula, These are the features of the first side, the second side, and the third side, respectively. This represents the mole fraction of the j-th node in the sample, i.e., the component data; , Let be the four-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, represented as . ; Let be the three-dimensional one-hot vectors corresponding to the i-th node and the j-th node, respectively, denoted as . ; Steps 2-3, formalize each sample as follows: ； In the formula, A set of nodes; Let it be the set of edges; The node feature matrix; The edge feature matrix; Modulus; Step 3 specifically includes: Step 3-1: Establish an improved information transmission mechanism for graph neural networks: ； in, ； ； In the formula, For the message sent from the j-th node in the t-th layer of a graph neural network to the i-th node, For the aggregated message of the i-th node, Let be the importance weight of the j-th node to the i-th node in the t-th layer. Let j be the feature representation of the j-th node in the t-th layer. Let be the feature representation of the i-th node in the (t+1)-th layer. Let be the feature representation of the i-th node in the t-th layer. The layer normalization function is represented. This represents the activation function. This represents a multilayer perceptron that integrates all information. This is a non-linear encoding of edge features. For parameter matrices; These are trainable weight matrices used to linearly map the input node representation, neighbor representation, edge features, and global component vector to a vector space of uniform dimension. Let i be the set of nodes adjacent to the i-th node. It is a component vector; Step 3-2: Implement message passing and node state update in the graph neural network based on the mechanism in Step 3-1; Step 3-3: Calculate the graph-level representation of the sample; Steps 3-4: Calculate the modulus based on graph-level representation; Step 3-3 calculates the graph-level representation of the sample. The specific calculation formula is as follows: ； in, ； ； In the formula, For the graph-level representation of the samples, For the readout stage, a multilayer perceptron. For the introduced component vector, This indicates the characteristics of local weighted pooling. Indicates global statistical characteristics, This represents the local weight of the i-th node. Let be the representation of the i-th node in the T-th layer of the graph neural network, where T is the depth of the graph neural network. Represents the mean function, represents the standard deviation function; N represents the total number of nodes.

2. The method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks according to claim 1, characterized in that, The multi-component glass fiber system in step 1 is 16-yuan system.

3. The method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks according to claim 1, characterized in that, Steps 3-4 describe the calculation of the modulus based on graph-level representation. The calculation formula is as follows: ； In the formula, This is the predicted value of the modulus. These are the weight matrix and bias term of the readout layer for modulus prediction, respectively.

4. The method for predicting the molecular structure-modulus property relationship of glass fibers based on graph neural networks according to claim 1, characterized in that, In step 3, the loss function L of the graph neural network is: ； in, ； ； In the formula, For modulus prediction error, For consistency loss function, For regularization terms, All are hyperparameters; Let be the predicted value of the structure descriptor of the i-th node. Let i be the structure descriptor of the i-th node; This represents the total number of nodes. To utilize the node's structure descriptor Representation of the final layer nodes The function that performs inverse mapping constraints.

5. A graph neural network-based system for predicting the molecular structure-modulus relationship of glass fibers based on the method of any one of claims 1 to 4, characterized in that, The system includes: The first module is used to: filter composition-modulus data for multi-component glass fiber systems from the database; The second module is used to construct graph data based on the component data selected in the first module. The third module is used to perform modulus prediction based on the graph data using a graph neural network.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

Images