A graph neural network method for predicting the transmittance of silicone coatings

Abstract

This invention relates to the field of materials characterization technology, specifically disclosing a graph neural network prediction method for the transmittance of organosilicon coatings. The method includes: initializing the features of atomic nodes and chemical bond edges in the molecular structure based on the organic atom types and bond orders of the organosilicon coating molecular structure data, obtaining node feature vectors and edge feature vectors; constructing a heterogeneous graph structure of the organosilicon coating based on these two types of vectors; inputting the heterogeneous graph into a pre-trained graph neural network model, performing context encoding on the target nodes to obtain updated feature vectors; performing mode decomposition on the updated feature vectors according to the node types of the heterogeneous graph to obtain multiple types of feature vectors; multi-source aggregation of these updated feature vectors with the overall molecular feature vectors to obtain a full graph representation vector; and finally performing a linear transformation on the full graph representation vector to output the predicted transmittance value as the final result. This invention can improve the accuracy of transmittance prediction for organosilicon coatings.

Description

Technical Field

[0001] This invention relates to the field of materials characterization technology, and in particular to a graph neural network method for predicting the transmittance of organosilicon coatings. Background Technology

[0002] Currently, the determination of the transmittance of silicone coatings mostly relies on physical samples and specialized testing equipment. The entire process includes multiple steps such as sample preparation, surface preparation, equipment calibration, and on-site testing. Each step requires a fixed amount of time, resulting in a long overall testing cycle and making it impossible to obtain transmittance data instantly. Ambient light, ambient temperature, and minute impurities on the sample surface can all interfere with the equipment's signal acquisition, causing fluctuations in the final transmittance data and poor data stability, making it difficult to handle continuous testing of large batches of coating samples.

[0003] Existing conventional prediction methods rely solely on the basic physicochemical parameters of the coating, failing to uncover the deep connections between atoms, chemical bonds, and light transmittance within the molecular structure of silicone coatings, and thus unable to fully capture the core factors affecting transmittance. The predicted results obtained using these methods differ significantly from the actual transmittance values ​​of the coatings, indicating insufficient reliability. This makes it impossible to provide accurate data support for research and development work related to silicone coating formulation adjustments and performance optimization. Therefore, improving the accuracy of silicone coating transmittance prediction has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a graph neural network method for predicting the transmittance of organosilicon coatings to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a graph neural network method for predicting the transmittance of an organosilicon coating, comprising: C1. Based on the organic atom types and bond orders in the molecular structure data of the organosilicon coating, the atomic nodes and chemical bond edges in the molecular structure data are initialized with features to obtain the node feature vectors of the atomic nodes and the edge feature vectors of the chemical bonds in the organosilicon coating. C2. Based on the node feature vectors and the edge feature vectors, construct the heterogeneous graph structure of the organosilicon coating; C3. Input the heterogeneous graph structure into a pre-trained graph neural network model, perform context encoding on the target nodes in the heterogeneous graph structure, and obtain the updated feature vector of the target nodes; C4. Based on the node types of the heterogeneous graph structure, perform mode decomposition on the updated feature vector to obtain multiple type feature vectors of the target node; C5. Perform multi-source aggregation of the multiple types of feature vectors with the molecular overall feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating; C6. Input the full-image representation vector into the graph neural network model, perform a linear transformation on the full-image representation vector to obtain the transmittance prediction value of the full-image representation vector, and use the transmittance prediction value as the transmittance prediction result of the organosilicon coating.

[0006] In a preferred embodiment, the step of initializing the atomic nodes and chemical bond edges in the molecular structure data based on the organic atom types and bond orders in the organosilicon coating to obtain the node feature vectors of the atomic nodes and the edge feature vectors of the chemical bonds in the organosilicon coating includes: The molecular structure data of the organosilicon coating is obtained, and the molecular structure data is used for feature recognition to obtain the organic atom types of the molecular structure data; Based on the organic atom type, the atomic attribute set of the atomic nodes in the molecular structure data is extracted, and the atomic attribute set is vectorized to obtain the node feature vector of the atomic node. The chemical bond edges in the molecular structure data are subjected to attribute extraction to obtain the bond order of the chemical bond edges; The bond order is numerically encoded to obtain the initial edge vector of the bond order, and the initial edge vector is stacked sequentially based on the original arrangement order of the chemical bond edges to obtain the edge feature vector of the chemical bond.

[0007] In a preferred embodiment, constructing the heterogeneous graph structure of the organosilicon coating based on the node feature vectors and the edge feature vectors includes: Using the atomic nodes in the molecular structure data as graph nodes and the chemical bond edges in the molecular structure data as graph edges, an initial heterogeneous graph structure for the organosilicon coating is constructed. Based on the original arrangement order of the atomic nodes and the chemical bond edges, the graph nodes and graph edges are identified and assigned, resulting in unique node identifiers for the graph nodes and unique edge identifiers for the graph edges. Based on the atom type of the atomic node and the bond order of the chemical bond edge, the graph node and the graph edge are labeled with type to obtain the node type label of the graph node and the edge type label of the graph edge; Each row vector in the node feature vector is assigned to the graph node to obtain the node feature row vector, and the node feature row vector is used as the initial node feature of the graph node; Each row vector in the edge feature vector is assigned to the graph edge to obtain the edge feature row vector, and the edge feature row vector is used as the initial edge feature of the graph edge; In the initial heterogeneous graph structure, the unique node identifier, the unique edge identifier, the node type label, the edge type label, the initial node feature, and the initial edge feature are fused to obtain the heterogeneous graph structure of the organosilicon coating.

[0008] In a preferred embodiment, the step of inputting the heterogeneous graph structure into a pre-trained graph neural network model and performing context encoding on the target nodes in the heterogeneous graph structure to obtain the updated feature vectors of the target nodes includes: Call the pre-trained graph neural network model and read the aggregate weight matrix, edge weight matrix and update weight matrix of the target node in the heterogeneous graph structure from the parameter set of the graph neural network model; The target node in the heterogeneous graph structure is expanded to obtain the neighboring nodes that are directly connected to the chemical bond edge. The neighboring nodes are then feature-encoded to obtain the current neighboring node feature vector and the chemical bond edge feature vector. Based on the edge weight matrix, the chemical bond edge feature vector is linearly mapped to obtain the edge contribution vector of the adjacent node; Based on the aggregation weight matrix, the current node feature vector of the target node, the current neighboring node feature vector, and the edge contribution vector are weighted and combined to obtain the neighborhood aggregation vector of the target node. Based on the updated weight matrix, the neighborhood aggregation vector and the current node feature vector are fused and updated to obtain the updated feature vector of the target node. The original feature vector of the target node is then replaced with the updated feature vector to obtain the updated feature vector of the target node.

[0009] In a preferred embodiment, the step of weighting and combining the current node feature vector, the current neighbor node feature vector, and the edge contribution vector of the target node based on the aggregation weight matrix to obtain the neighborhood aggregation vector of the target node includes: The aggregated weight matrix is ​​decoupled to obtain the self-weight sub-vector of the current node feature vector, the neighbor weight sub-vector of the current neighbor node feature vector, and the edge contribution weight sub-vector of the edge contribution vector. The product of the weight coefficients in the self-weighted sub-vector and the dimension components in the current node feature vector is taken as the self-weighted component sequence of the current node feature vector; The weight coefficients in the neighbor weight sub-vectors are superimposed on the dimension components in the feature vectors of the current neighbor nodes to obtain the neighbor weighted component sequence of the feature vectors of the current neighbor nodes. The weight coefficients in the edge contribution weight sub-vector are multiplied by the dimension components in the edge contribution vector to obtain the edge contribution weighted component sequence of the edge contribution vector. The components at the same position in the self-weighted component sequence, the neighbor-weighted component sequence, and the edge contribution weighted component sequence are superimposed to obtain the fused components at different positions in the target node. The fused components are then concatenated to obtain the neighborhood aggregation vector of the target node.

[0010] In a preferred embodiment, the modal decomposition of the updated feature vector based on the node type of the heterogeneous graph structure to obtain multiple type feature vectors of the target node includes: Based on the node types of the heterogeneous graph structure, the updated feature vector is categorized to obtain multiple type groups of the updated feature vector; Group pooling is performed on the updated feature vectors within the multiple type groups to obtain the type feature vectors of the multiple type groups; Based on the inherent fixed order of the node type, the type feature vectors are sequentially merged to obtain multiple type feature vectors of the target node.

[0011] In a preferred embodiment, the step of performing group pooling on the updated feature vectors within the plurality of type groups to obtain the type feature vectors of the plurality of type groups includes: Traverse the vector dimensions of the updated feature vectors, extract the element values ​​of all updated feature vectors in the multiple type groups on the same vector dimension, and construct the set of elements to be aggregated in the vector dimension; The aggregated element values ​​of the vector dimension are obtained by summing and aggregating all element values ​​in the set of elements to be aggregated. Based on the original dimensional order of the updated feature vector, the aggregated element values ​​are sequentially concatenated to obtain a preliminary aggregated vector with the same dimensions as the updated feature vector; The aggregated element values ​​in the initial aggregated vector are averaged and reduced to the total number of updated feature vectors within the multiple type groups to obtain the mean feature vector of the multiple type groups, and the mean feature vector is used as the type feature vector of the multiple type groups.

[0012] In a preferred embodiment, the step of multi-source aggregation of the multiple types of feature vectors with the molecular overall feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating includes: Feature decoding is performed on the multiple types of feature vectors to obtain the node type labels of the multiple types of feature vectors; The molecular structure data is globally aggregated to obtain the overall molecular feature vector of the molecular structure data. Based on the fixed order of the node type labels, the multiple type feature vectors are serialized and combined to obtain a combined feature vector of the multiple type feature vectors; The combined feature vector is jointly encoded with the overall molecular feature vector to obtain the full-image representation vector of the organosilicon coating.

[0013] In a preferred embodiment, the step of inputting the full-image representation vector into the graph neural network model, performing a linear transformation on the full-image representation vector to obtain a transmittance prediction value of the full-image representation vector, and using the transmittance prediction value as the transmittance prediction result of the silicone coating includes: The parameters of the regression output layer in the graph neural network model are analyzed to obtain the weight parameter vector and bias parameters of the regression output layer. Based on the weight parameter vector, the dimensional components in the full graph representation vector are multiplied dimension by dimension to obtain multiple product results of the full graph representation vector; By summing the multiple product results, the intermediate accumulated value of the full-image representation vector is obtained; Based on the bias parameter, the intermediate accumulated value is corrected for deviation to obtain the transmittance prediction value of the full-image representation vector; The predicted transmittance value is used as the predicted transmittance result of the silicone coating.

[0014] In a preferred embodiment, the formula for calculating the predicted transmittance value is as follows: ; In the formula, The predicted transmittance value is... The total number of dimensions of the full graph representation vectors. The dimension index variable is used to represent the vector of the entire graph. The weight parameter vector is the first... Each weighting coefficient The first element in the full graph representation vector is... Each dimension component The bias parameter is denoted as .

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention, through feature initialization of atomic nodes and chemical bond edges, fully preserves the core structural information of organosilicon coating molecules, such as atomic types and bond orders. The constructed heterogeneous graph structure can accurately represent the topological relationships between atoms and chemical bonds within the molecule. By using context encoding of target nodes and modal decomposition by node type, it achieves the classification, extraction, and orderly integration of different types of atomic features, enhancing the specificity and comprehensiveness of molecular structural feature expression.

[0016] 2. This method generates a comprehensive characterization vector that fully represents the molecular properties of organosilicon coatings by fusing node type features and overall molecular features through multi-source polymerization. Linear transformation based on this comprehensive characterization vector directly outputs predicted transmittance values, improving the accuracy and stability of transmittance prediction results. This provides precise data for the formulation design and performance control of organosilicon coatings, accelerating the research and development of related products. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a graph neural network prediction method for the transmittance of an organosilicon coating provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a graph neural network method for predicting the transmittance of an organosilicon coating. The execution subject of this graph neural network method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the graph neural network method for predicting the transmittance of an organosilicon coating can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a schematic flowchart of a graph neural network method for predicting the transmittance of an organosilicon coating according to an embodiment of the present invention. In this embodiment, the graph neural network method for predicting the transmittance of an organosilicon coating includes: C1. Based on the organic atom types and bond orders in the molecular structure data of the organosilicon coating, the atomic nodes and chemical bond edges in the molecular structure data are initialized with features to obtain the node feature vectors of the atomic nodes and the edge feature vectors of the chemical bonds in the organosilicon coating. In this embodiment of the invention, the step of initializing the atomic nodes and chemical bond edges in the molecular structure data based on the organic atom types and bond orders in the organosilicon coating to obtain the node feature vectors of the atomic nodes and the edge feature vectors of the chemical bonds in the organosilicon coating includes: The molecular structure data of the organosilicon coating is obtained, and the molecular structure data is used for feature recognition to obtain the organic atom types of the molecular structure data; Based on the organic atom type, the atomic attribute set of the atomic nodes in the molecular structure data is extracted, and the atomic attribute set is vectorized to obtain the node feature vector of the atomic node. The chemical bond edges in the molecular structure data are subjected to attribute extraction to obtain the bond order of the chemical bond edges; The bond order is numerically encoded to obtain the initial edge vector of the bond order, and the initial edge vector is stacked sequentially based on the original arrangement order of the chemical bond edges to obtain the edge feature vector of the chemical bond.

[0021] The molecular structure data of the organosilicon coating is obtained from publicly available standard molecular structure databases, experimental records of organosilicon material synthesis, or molecular three-dimensional structure information obtained by instrumental characterization. The data fully includes the elemental identifier of each atom, the position number of the atom in the molecule, and the connection relationship between adjacent atoms.

[0022] The process iterates through all atomic units in the molecular structure data one by one, matching the element identifier corresponding to each atomic unit with a pre-established organic element lookup table. This lookup table is based on the general element classification rules of organic chemistry and includes all possible element types that may appear in organosilicon coating molecules. After the matching is completed, the organic atom type corresponding to each atom is obtained.

[0023] Based on the obtained organic atom types, the inherent physicochemical properties corresponding to each atomic node are extracted. These properties include the number of outer electron shells, the number of valence electrons, the atomic radius, electronegativity, and the number of bonds formed by the atom in the molecule. All properties are inherent to the atom itself and can be found in standard chemical handbooks.

[0024] All the extracted attributes are arranged in a fixed order according to a pre-defined order. This order is determined based on the weight of each attribute in relation to the transmittance in the study of the optical properties of organic materials. After the arrangement is completed, an atomic attribute set corresponding to each atomic node is formed.

[0025] For discrete attributes in the atomic attribute set, they are converted into a numerical sequence that matches the number of attribute categories. The sequence contains only two values, 0 and 1, which correspond to different category states of the attribute.

[0026] For continuous attributes in the atomic attribute set, their values ​​are converted to standard values ​​between 0 and 1. The conversion process is to subtract the minimum value of the attribute among all atoms from the original value of the attribute, and then divide by the difference between the maximum and minimum values ​​of the attribute among all atoms.

[0027] Arrange the converted values ​​of all attributes in each atomic attribute set in the original order of the atomic attribute set to form a one-dimensional numerical sequence that corresponds one-to-one with the atomic node. This sequence is the node feature vector of the atomic node.

[0028] The process iterates through all chemical bond units connecting two adjacent atomic nodes in the molecular structure data, extracting the bonding type identifier corresponding to each chemical bond unit. This identifier is directly derived from the definition of interatomic bonds in the molecular structure data.

[0029] The extracted bonding type identifiers are matched with a pre-established bond order correspondence table, which is based on valence bond theory and clearly defines the bond order values ​​corresponding to single bonds, double bonds, triple bonds, and coordinate bonds. After the matching is completed, the bond order of each chemical bond edge is obtained.

[0030] The bond order value corresponding to each chemical bond, as well as the bond length and bond angle attributes corresponding to that chemical bond, are converted into standard values ​​according to the above-mentioned conversion method for continuous attributes.

[0031] The values ​​after bond order conversion are arranged in a fixed order with the values ​​after bond length and bond angle conversion, forming a one-dimensional numerical sequence of uniform length. This sequence is the initial vector of the edge of the chemical bond.

[0032] According to the original order of appearance of chemical bond edges in the molecular structure data, the initial vectors of all chemical bonds are arranged as independent rows to form a two-dimensional numerical set with the number of rows equal to the total number of chemical bonds and the number of columns equal to the length of the initial vectors of the edges. This set is the edge feature vector of the chemical bond.

[0033] The beneficial effects are that it completely and systematically extracts all the core properties of atoms and chemical bonds in the molecular structure of organosilicon coatings, ensuring that all feature expression dimensions are consistent and have clear meanings, avoiding the loss or confusion of feature information, providing accurate and reproducible basic data support for the subsequent construction of heterogeneous map structures, and ensuring the stability and consistency of the transmittance prediction process.

[0034] C2. Based on the node feature vectors and the edge feature vectors, construct the heterogeneous graph structure of the organosilicon coating; In this embodiment of the invention, constructing the heterogeneous graph structure of the organosilicon coating based on the node feature vector and the edge feature vector includes: Using the atomic nodes in the molecular structure data as graph nodes and the chemical bond edges in the molecular structure data as graph edges, an initial heterogeneous graph structure for the organosilicon coating is constructed. Based on the original arrangement order of the atomic nodes and the chemical bond edges, the graph nodes and graph edges are identified and assigned, resulting in unique node identifiers for the graph nodes and unique edge identifiers for the graph edges. Based on the atom type of the atomic node and the bond order of the chemical bond edge, the graph node and the graph edge are labeled with type to obtain the node type label of the graph node and the edge type label of the graph edge; Each row vector in the node feature vector is assigned to the graph node to obtain the node feature row vector, and the node feature row vector is used as the initial node feature of the graph node; Each row vector in the edge feature vector is assigned to the graph edge to obtain the edge feature row vector, and the edge feature row vector is used as the initial edge feature of the graph edge; In the initial heterogeneous graph structure, the unique node identifier, the unique edge identifier, the node type label, the edge type label, the initial node feature, and the initial edge feature are fused to obtain the heterogeneous graph structure of the organosilicon coating.

[0035] Extract the set of all atomic nodes and the set of all chemical bond edges from the molecular structure data. Map each atomic node in the set to an independent graph node in the initial heterogeneous graph structure. The number of graph nodes is exactly equal to the number of atomic nodes. Map each chemical bond edge in the set to an independent graph edge in the initial heterogeneous graph structure. The two endpoints of each graph edge correspond to the graph nodes mapped to the two atomic nodes connected by the chemical bond, respectively. After the mapping is completed, the initial heterogeneous graph structure of the organosilicon coating containing only the topological connection relationship between nodes and edges is obtained.

[0036] According to the original arrangement order of atomic nodes in the molecular structure data, starting from the first atomic node, its corresponding graph nodes are assigned consecutive and non-repeating numerical identifiers. Each graph node is assigned only one numerical identifier, and the numerical identifiers of all graph nodes are different from each other. After the assignment is completed, a unique node identifier of the graph node is obtained.

[0037] Based on the original arrangement order of chemical bond edges in the molecular structure data, starting from the first chemical bond edge, its corresponding graph edges are assigned consecutive and non-repeating numerical identifiers. Each graph edge is assigned only one numerical identifier, and all graph edges have different numerical identifiers. After the assignment is completed, a unique edge identifier is obtained for each graph edge.

[0038] Based on the atomic type of the atomic node corresponding to each graph node, a corresponding node type label is assigned to each graph node. Graph nodes with the same atomic type are assigned the same node type label, and graph nodes with different atomic types are assigned different node type labels. The value of the label is determined based on the pre-established correspondence rules between atomic types and labels. After the assignment is completed, the node type label of the graph node is obtained.

[0039] Based on the bond order of the chemical bond corresponding to each graph edge, a corresponding edge type label is assigned to each graph edge. Graph edges with the same bond order are assigned the same edge type label, and graph edges with different bond orders are assigned different edge type labels. The value of the label is determined based on the pre-established correspondence rules between bond order and label. After the assignment is completed, the edge type label of the graph edge is obtained.

[0040] Extract each row vector from the node feature vectors, and assign the row vector to the graph node mapped by the atomic node corresponding to the row vector. After the assignment, the row vector becomes the node feature row vector of the graph node. After all graph nodes have been assigned, the node feature row vector is used as the initial node feature of the corresponding graph node.

[0041] Extract each row vector from the edge feature vectors, and assign the row vector to the graph edge mapped by the chemical bond edge corresponding to the row vector. After assignment, the row vector becomes the edge feature row vector of the graph edge. After all graph edges have been assigned, the edge feature row vector is used as the initial edge feature of the corresponding graph edge.

[0042] Each graph node's unique node identifier, node type label, and initial node features are associated and bound together, so that each graph node simultaneously contains its own identifier information, type information, and feature information. Similarly, each graph edge's unique edge identifier, edge type label, and initial edge features are associated and bound together with the unique node identifiers of the two graph nodes connected to that edge, so that each graph edge simultaneously contains its own identifier information, type information, feature information, and connection relationship information. All the bound graph nodes and edges are then integrated into the initial heterogeneous graph structure to obtain the heterogeneous graph structure of the silicone coating.

[0043] The beneficial effects include accurately restoring the topological connections of atoms and chemical bonds within the organosilicon coating molecule, and unifying and integrating the identifiers, types, and feature information of nodes and edges. The constructed heterogeneous graph structure can completely and orderly carry all the core information of the molecular structure, providing a standardized and reproducible structural basis for subsequent node context encoding and feature extraction, and ensuring the consistency and accuracy of the transmittance prediction process.

[0044] C3. Input the heterogeneous graph structure into a pre-trained graph neural network model, perform context encoding on the target nodes in the heterogeneous graph structure, and obtain the updated feature vector of the target nodes; In this embodiment of the invention, the step of inputting the heterogeneous graph structure into a pre-trained graph neural network model and performing context encoding on the target nodes in the heterogeneous graph structure to obtain the updated feature vectors of the target nodes includes: Call the pre-trained graph neural network model and read the aggregate weight matrix, edge weight matrix and update weight matrix of the target node in the heterogeneous graph structure from the parameter set of the graph neural network model; The target node in the heterogeneous graph structure is expanded to obtain the neighboring nodes that are directly connected to the chemical bond edge. The neighboring nodes are then feature-encoded to obtain the current neighboring node feature vector and the chemical bond edge feature vector. Based on the edge weight matrix, the chemical bond edge feature vector is linearly mapped to obtain the edge contribution vector of the adjacent node; Based on the aggregation weight matrix, the current node feature vector of the target node, the current neighboring node feature vector, and the edge contribution vector are weighted and combined to obtain the neighborhood aggregation vector of the target node. Based on the updated weight matrix, the neighborhood aggregation vector and the current node feature vector are fused and updated to obtain the updated feature vector of the target node. The original feature vector of the target node is then replaced with the updated feature vector to obtain the updated feature vector of the target node.

[0045] The step of weighting and combining the current node feature vector, the current neighbor node feature vector, and the edge contribution vector of the target node based on the aggregation weight matrix to obtain the neighborhood aggregation vector of the target node includes: The aggregated weight matrix is ​​decoupled to obtain the self-weight sub-vector of the current node feature vector, the neighbor weight sub-vector of the current neighbor node feature vector, and the edge contribution weight sub-vector of the edge contribution vector. The product of the weight coefficients in the self-weighted sub-vector and the dimension components in the current node feature vector is taken as the self-weighted component sequence of the current node feature vector; The weight coefficients in the neighbor weight sub-vectors are superimposed on the dimension components in the feature vectors of the current neighbor nodes to obtain the neighbor weighted component sequence of the feature vectors of the current neighbor nodes. The weight coefficients in the edge contribution weight sub-vector are multiplied by the dimension components in the edge contribution vector to obtain the edge contribution weighted component sequence of the edge contribution vector. The components at the same position in the self-weighted component sequence, the neighbor-weighted component sequence, and the edge contribution weighted component sequence are superimposed to obtain the fused components at different positions in the target node. The fused components are then concatenated to obtain the neighborhood aggregation vector of the target node.

[0046] The pre-trained graph neural network model is invoked. This model learns and solidifies parameters using existing silicone coating molecular structure samples and corresponding measured transmittance data. From the parameter set stored in the model's internal structure, the aggregate weight matrix, edge weight matrix, and update weight matrix, which correspond one-to-one with the target node type in the current heterogeneous graph structure, are extracted. These three types of matrices provide fixed transformation rules for node features, edge features, and feature fusion processes, respectively.

[0047] Locate the target node to be processed in the heterogeneous graph structure, traverse all graph edges in the heterogeneous graph structure, filter out all graph edges that connect one end to the target node and the other end to other graph nodes, and determine all graph nodes connected to the other end of these graph edges as adjacent nodes of the target node. Adjacent nodes only include graph nodes that are directly connected to the target node through a single graph edge, and do not include any graph nodes that are indirectly connected.

[0048] Extract the initial features of each neighboring node currently stored, and use them as the current neighboring node feature vector of that neighboring node. At the same time, extract the initial features of the edges currently stored in the connection graph between the target node and each neighboring node, and use them as the chemical bond edge feature vector of the corresponding connection relationship.

[0049] For each chemical bond edge feature vector, based on its corresponding edge type label, the feature transformation rule corresponding to the edge type is matched from the edge weight matrix. Each dimension component of the chemical bond edge feature vector is transformed one by one according to the transformation rule. The one-dimensional numerical sequence with the same length as the original chemical bond edge feature vector after transformation is the edge contribution vector corresponding to the adjacent node.

[0050] According to the pre-defined structural boundaries during model construction, the aggregated weight matrix is ​​split into three independent sub-parts. The lengths of the three sub-parts are exactly the same as the lengths of the current node feature vector, the current neighbor node feature vector, and the edge contribution vector, respectively. After splitting, we obtain the self-weight sub-vector of the current node feature vector, the neighbor weight sub-vector of the current neighbor node feature vector, and the edge contribution weight sub-vector of the edge contribution vector.

[0051] Multiply the weight coefficient of each position in the weighted subvector of the current node by the corresponding dimensional component of the same position in the feature vector of the current node. This results in a set of numerical sequences with the same dimensions as the feature vector of the current node, arranged in the original dimensional order. This sequence is the self-weighted component sequence of the feature vector of the current node.

[0052] The weight coefficient of each position in the neighbor weight sub-vector is added to the corresponding dimension component in the feature vector of the current neighbor node, resulting in a set of numerical sequences with the same dimension as the feature vector of the current neighbor node and arranged in the original dimensional order. This sequence is the neighbor weighted component sequence of the feature vector of the current neighbor node.

[0053] Multiply the weight coefficient at each position in the edge contribution weight sub-vector by the corresponding dimension component at the same position in the edge contribution vector to obtain a set of numerical sequences with the same dimension as the edge contribution vector and arranged in the original dimensional order. This sequence is the edge contribution weighted component sequence of the edge contribution vector.

[0054] The values ​​at the same position in the self-weighted component sequence, the neighbor-weighted component sequence, and the edge contribution weighted component sequence are added together to obtain the fusion component corresponding to each dimension position. Then, all the fusion components are arranged in the original dimensional order to form a one-dimensional numerical sequence, which is the neighborhood aggregation vector of the target node.

[0055] The neighborhood aggregation vector and the current node feature vector of the target node are combined dimension by dimension according to the feature fusion rule corresponding to the updated weight matrix. The dimensional order remains unchanged during the combination process. The resulting one-dimensional numerical sequence is the updated feature vector of the target node.

[0056] The obtained updated feature vector is used to completely replace the original initial features of the target node, so that the stored features of the target node are updated to new features containing neighborhood context information, and finally the updated feature vector of the target node is obtained.

[0057] The beneficial effects are that the features of the target node fully integrate its own inherent properties, the properties of adjacent nodes, and the properties of connecting chemical bonds, which can accurately capture the influence of the local molecular topology on atomic features, eliminate the limitations of single node feature expression, improve the expressive power and discriminative power of node features, provide a high-quality feature foundation containing complete contextual information for subsequent mode decomposition and full-map representation construction, and ensure the accuracy and stability of the transmittance prediction process.

[0058] C4. Based on the node types of the heterogeneous graph structure, perform mode decomposition on the updated feature vector to obtain multiple type feature vectors of the target node; In this embodiment of the invention, the modal decomposition of the updated feature vector based on the node type of the heterogeneous graph structure to obtain multiple type feature vectors of the target node includes: Based on the node types of the heterogeneous graph structure, the updated feature vector is categorized to obtain multiple type groups of the updated feature vector; Group pooling is performed on the updated feature vectors within the multiple type groups to obtain the type feature vectors of the multiple type groups; Based on the inherent fixed order of the node type, the type feature vectors are sequentially merged to obtain multiple type feature vectors of the target node.

[0059] The step of performing group pooling on the updated feature vectors within the multiple type groups to obtain the type feature vectors of the multiple type groups includes: Traverse the vector dimensions of the updated feature vectors, extract the element values ​​of all updated feature vectors in the multiple type groups on the same vector dimension, and construct the set of elements to be aggregated in the vector dimension; The aggregated element values ​​of the vector dimension are obtained by summing and aggregating all element values ​​in the set of elements to be aggregated. Based on the original dimensional order of the updated feature vector, the aggregated element values ​​are sequentially concatenated to obtain a preliminary aggregated vector with the same dimensions as the updated feature vector; The aggregated element values ​​in the initial aggregated vector are averaged and reduced to the total number of updated feature vectors within the multiple type groups to obtain the mean feature vector of the multiple type groups, and the mean feature vector is used as the type feature vector of the multiple type groups.

[0060] Extract the pre-bound and stored node type labels of each graph node in the heterogeneous graph structure, traverse all the node type labels of all graph nodes, remove duplicate label content, retain all unique node types, and gather these unique node types into a node type set, where each element in the set corresponds to an independent atomic node type.

[0061] Iterate through all the updated feature vectors obtained through context encoding, read the node type label corresponding to the graph node to which each updated feature vector belongs, and assign the updated feature vector to the corresponding exclusive group in the node type set. Each updated feature vector is assigned to only one group, and there is no case of repeated assignment across groups. The total number of groups is exactly equal to the number of elements in the node type set. After all updated feature vectors are assigned, multiple type groups of updated feature vectors are obtained.

[0062] Starting from the first dimension of the updated feature vector, each vector dimension is processed sequentially according to the pre-defined dimensional arrangement of the updated feature vectors. For the currently processed vector dimension, the element values ​​corresponding to all updated feature vectors in the current type group in that dimension are extracted, and these element values ​​are aggregated into an independent set of values, which is the set of elements to be aggregated for that vector dimension.

[0063] The process involves sequentially summing all element values ​​in the set to be aggregated, adding all the element values ​​contained in the set one by one, and finally obtaining a single numerical result. This numerical result is the aggregate element value corresponding to the current vector dimension being processed, and a unique aggregate element value is generated for each vector dimension.

[0064] Following the original dimensional order of the updated feature vectors, the aggregated element values ​​corresponding to all vector dimensions are arranged sequentially. During the arrangement process, the dimensional order is strictly kept to be completely consistent with the dimensional order of the original updated feature vectors. After the arrangement is completed, a one-dimensional numerical sequence is formed. The length of this sequence is exactly equal to the length of a single updated feature vector. This sequence is the preliminary aggregated vector for the current type grouping.

[0065] The total number of updated feature vectors contained in the currently processed type group is counted. This number is equal to the number of updated feature vectors assigned to that group. The aggregate element value corresponding to each position in the initial aggregate vector is divided by the total number of updated feature vectors in that type group to obtain the mean element value corresponding to each dimension position.

[0066] Arrange the mean element values ​​corresponding to all dimensional positions in strict accordance with the original dimensional arrangement order of the updated feature vector. After the arrangement is completed, a one-dimensional numerical sequence is formed, which is the mean feature vector of the current type group.

[0067] The mean feature vector obtained by processing each type group as described above is directly used as the type feature vector corresponding to that type group. After all the above processing steps are completed for all type groups, multiple type feature vectors with the same number of elements as the node type set are obtained.

[0068] All feature vectors are arranged according to a pre-defined fixed order inherent to the node type. This fixed order is based on the atomic type classification rules commonly used in the field of organic chemistry and the atomic number order of the periodic table. This order remains constant throughout all processing. After the arrangement is completed, an ordered set of feature vectors is formed, which is the multiple feature vectors of the target node.

[0069] The beneficial effects include the accurate classification, separation, and independent aggregation of node features of different atomic types, the complete preservation of the differential influence of different atomic types on molecular light transmittance, the effective elimination of individual feature fluctuations between nodes of the same type through group pooling operations, the improvement of the stability and representativeness of type features, and the consistency of feature arrangement based on fixed-order sequence merging, which provides clear, semantically clear, and formatted feature inputs for subsequent multi-source aggregation, ensuring the construction quality of the full-image representation vector and the accuracy of light transmittance prediction results.

[0070] C5. Perform multi-source aggregation of the multiple types of feature vectors with the molecular overall feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating; In this embodiment of the invention, the step of performing multi-source aggregation of the multiple types of feature vectors with the molecular overall feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating includes: Feature decoding is performed on the multiple types of feature vectors to obtain the node type labels of the multiple types of feature vectors; The molecular structure data is globally aggregated to obtain the overall molecular feature vector of the molecular structure data. Based on the fixed order of the node type labels, the multiple type feature vectors are serialized and combined to obtain a combined feature vector of the multiple type feature vectors; The combined feature vector is jointly encoded with the overall molecular feature vector to obtain the full-image representation vector of the organosilicon coating.

[0071] Read the pre-bound and stored association identifier for each type feature vector. This association identifier is generated synchronously when the type feature vector is generated by modal decomposition and is bound to the corresponding type feature vector one by one. There is a unique correspondence between the association identifier and the node type label. The node type label corresponding to each type feature vector is obtained by matching through this correspondence.

[0072] Global attribute information of molecules is extracted from the molecular structure data of organosilicon coatings. This information includes the total number of atoms, the total number of chemical bonds, the number of atoms corresponding to each node type, the total mass of the molecule, and the spatial size parameters of the molecule. All attribute information is directly derived from the inherent descriptive content of the molecular structure data.

[0073] All extracted global molecular attribute information is arranged in a pre-defined fixed order, which is based on the importance of the overall molecular attributes in the characterization of organic materials. After the arrangement is completed, all attribute information is converted into a one-dimensional numerical sequence, which is the overall molecular feature vector of the molecular structure data.

[0074] Extract the node type labels corresponding to all types of feature vectors, match the node type labels with the pre-defined fixed order of node types, which is completely consistent with the node type order used in the modality decomposition step, and arrange all types of feature vectors according to the order obtained from the matching.

[0075] All the arranged feature vectors of each type are connected end to end in sequence. The last dimension element of the previous feature vector is directly connected to the first dimension element of the next feature vector. During the connection process, the dimensional order within each feature vector remains unchanged. After the connection is completed, a unified one-dimensional numerical sequence is formed. This sequence is the combined feature vector of multiple feature vectors.

[0076] The combined feature vector and the molecular overall feature vector are concatenated in a pre-defined fixed order, while maintaining the dimensional order within each feature vector and the molecular overall feature vector. After concatenation, a unified one-dimensional numerical sequence containing all types of feature information and molecular overall feature information is formed. This sequence is the full-image representation vector of the organosilicon coating. The multi-source data involved in this multi-source aggregation specifically includes multiple types of feature vectors obtained by classifying and aggregating nodes of different atomic types, as well as the molecular overall feature vector representing the global properties of the molecule.

[0077] The beneficial effect is that by multi-source polymerization, the local structural features of different atomic types within the molecule and the global overall property features of the molecule are integrated simultaneously, comprehensively covering various structural factors that affect the light transmittance of organosilicon coatings. This eliminates the one-sidedness of single-source feature expression, and the constructed full-image characterization vector can completely and accurately reflect the overall structural and performance correlation features of organosilicon coating molecules, providing a comprehensive and reliable feature basis for subsequent light transmittance prediction, and ensuring the accuracy and stability of the prediction results.

[0078] C6. Input the full-image representation vector into the graph neural network model, perform a linear transformation on the full-image representation vector to obtain the transmittance prediction value of the full-image representation vector, and use the transmittance prediction value as the transmittance prediction result of the organosilicon coating.

[0079] In this embodiment of the invention, the step of inputting the full-image representation vector into the graph neural network model, performing a linear transformation on the full-image representation vector to obtain a transmittance prediction value of the full-image representation vector, and using the transmittance prediction value as the transmittance prediction result of the organosilicon coating includes: The parameters of the regression output layer in the graph neural network model are analyzed to obtain the weight parameter vector and bias parameters of the regression output layer. Based on the weight parameter vector, the dimensional components in the full graph representation vector are multiplied dimension by dimension to obtain multiple product results of the full graph representation vector; By summing the multiple product results, the intermediate accumulated value of the full-image representation vector is obtained; Based on the bias parameter, the intermediate accumulated value is corrected for deviation to obtain the transmittance prediction value of the full-image representation vector; The predicted transmittance value is used as the predicted transmittance result of the silicone coating.

[0080] The formula for calculating the predicted transmittance value is as follows: ; In the formula, The predicted transmittance value is... The total number of dimensions of the full graph representation vectors. The dimension index variable is used to represent the vector of the entire graph. The weight parameter vector is the first... Each weighting coefficient The first element in the full graph representation vector is... Each dimension component The bias parameter is denoted as .

[0081] The final layer of the pre-trained graph neural network model is located. This structure is a regression output layer specifically designed to convert high-dimensional molecular structure features into continuous transmittance values. All parameters of the regression output layer have been learned and stored during the model training phase using a large number of silicone-coated molecular structure samples and corresponding measured transmittance data.

[0082] Extract all fixed parameters from the structured parameter storage area of ​​the regression output layer, and split them into two parts according to their function and arrangement: one part is a weight parameter vector arranged in a fixed order, and the other part is a single bias parameter.

[0083] The length of the weight parameter vector and the total dimension of the full graph representation vector They are completely equal. The value of is equal to the sum of the dimensions of all types of feature vectors in the multi-source aggregation step and the sum of the dimensions of the overall molecular feature vector. It is determined by the process of serializing and combining multiple types of feature vectors in a fixed order according to node type and then splicing them with the overall molecular feature vector. During the splicing process, the dimensions of each type of feature vector and the overall molecular feature vector remain unchanged, so the total number of dimensions is a fixed value.

[0084] The first weight parameter vector Each weight coefficient is , This is a dimension index variable for the full graph representation vector, used to identify the position of each dimension component in the full graph representation vector. During the model training phase, the model is obtained through iterative learning using a large number of silicone coating molecular structure samples and corresponding measured transmittance data. During the training process, each weight coefficient is adjusted to minimize the error between the predicted and measured values. Finally, the model is solidified and stored in the structured parameter region of the regression output layer. Each weight coefficient corresponds to the degree of influence of the same position dimension component in the full-image representation vector on the transmittance.

[0085] The bias parameter is It is obtained through synchronous iterative learning with the weight parameter vector during the model training phase. It is used to correct the systematic benchmark offset of the model prediction results, eliminate the overall bias caused by the overall distribution of the training data, and make the benchmark value of the prediction results consistent with the benchmark value of the actual transmittance of the silicone coating.

[0086] Starting from the first dimension component of the full graph representation vector, each dimension component is multiplied with the corresponding weight parameter in the weight parameter vector in a corresponding order, strictly following the pre-defined dimensional arrangement of the full graph representation vector.

[0087] The first vector in the full graph representation Each dimension component is The full-image representation vector originates from the organosilicon coating generated by the multi-source polymerization step. The full-image representation vector is obtained by serializing and combining multiple type feature vectors according to the inherent fixed order of node type to obtain a combined feature vector, and then jointly encoding it with the overall molecular feature vector. Each dimension component contains the molecular local structural features or global attribute feature information at the corresponding position.

[0088] Each dimension component is multiplied by its corresponding weight parameter to generate an independent numerical result. After all dimension components have undergone the corresponding multiplication operation, multiple independent numerical values ​​are obtained that are exactly equal to the number of dimensions of the full graph representation vector. These numerical values ​​are the multiple product results of the full graph representation vector.

[0089] Strictly following the original dimensional arrangement order of the full-image representation vector, all product results are continuously accumulated in the order of generation. Each accumulation operation adds the current accumulated result to the next product result that has not yet participated in the operation, until all product results have been accumulated, and finally a single numerical result is obtained. This value is the intermediate accumulated value of the full-image representation vector, which reflects the comprehensive weighted contribution of all dimensional components in the full-image representation vector to the transmittance.

[0090] The obtained intermediate accumulated values ​​are compared with the bias parameters extracted from the regression output layer. The addition operation is performed, maintaining the same numerical precision as the full-image representation vector throughout the calculation. The resulting single numerical value is the predicted transmittance value corresponding to the full-image representation vector. .

[0091] This formula is the mathematical expression of the linear transformation process of the regression output layer of the graph neural network model. It uses weight parameters to quantize each dimension component of the full graph representation vector, clarifying the contribution of different molecular structural features to transmittance. It completes the overall benchmark correction through bias parameters, and finally converts the high-dimensional full graph representation vector containing the comprehensive features of local and global molecules into a continuous numerical result that can directly reflect the transmittance performance of the organosilicon coating. This is completely consistent with the previous technical steps of inputting the full graph representation vector into the regression output layer for linear transformation to obtain the transmittance prediction value.

[0092] The obtained transmittance prediction value The transmittance prediction result of the silicone coating is directly used as a continuous value that can directly reflect the visible light transmission capability of the silicone coating. It can be directly used for subsequent work such as performance evaluation, formulation screening and optimization design of silicone coating.

[0093] The beneficial effects are that a direct and reproducible mapping from the full-image representation vector to the transmittance value is achieved through linear transformation based on the model curing parameters. The conversion process is logically clear and has no black-box operation, which can quickly generate accurate transmittance prediction results without the need for cumbersome processes such as physical sample preparation, equipment calibration and instrument testing. This greatly improves the efficiency of evaluating the transmittance performance of silicone coatings, while ensuring the stability and consistency of the prediction results, providing reliable data support for the research and development and production of silicone coatings.

[0094] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0095] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A graph neural network method for predicting the transmittance of an organosilicon coating, characterized in that, The method includes: C1. Based on the organic atom types and bond orders in the molecular structure data of the organosilicon coating, the atomic nodes and chemical bond edges in the molecular structure data are initialized with features to obtain the node feature vectors of the atomic nodes and the edge feature vectors of the chemical bonds in the organosilicon coating. C2. Based on the node feature vectors and the edge feature vectors, construct the heterogeneous graph structure of the organosilicon coating; C3. Input the heterogeneous graph structure into a pre-trained graph neural network model, perform context encoding on the target nodes in the heterogeneous graph structure, and obtain the updated feature vector of the target nodes; C4. Based on the node types of the heterogeneous graph structure, perform mode decomposition on the updated feature vector to obtain multiple type feature vectors of the target node; C5. Perform multi-source aggregation of the multiple types of feature vectors with the molecular overall feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating; C6. Input the full-image representation vector into the graph neural network model, perform a linear transformation on the full-image representation vector to obtain the transmittance prediction value of the full-image representation vector, and use the transmittance prediction value as the transmittance prediction result of the organosilicon coating.

2. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, The process involves initializing the atomic nodes and chemical bond edges in the molecular structure data based on the organic atom types and bond orders in the organosilicon coating, thereby obtaining the node feature vectors and edge feature vectors of the atomic nodes and chemical bonds in the organosilicon coating. This includes: The molecular structure data of the organosilicon coating is obtained, and the molecular structure data is used for feature recognition to obtain the organic atom types of the molecular structure data; Based on the organic atom type, the atomic attribute set of the atomic nodes in the molecular structure data is extracted, and the atomic attribute set is vectorized to obtain the node feature vector of the atomic node. The chemical bond edges in the molecular structure data are subjected to attribute extraction to obtain the bond order of the chemical bond edges; The bond order is numerically encoded to obtain the initial edge vector of the bond order, and the initial edge vector is stacked sequentially based on the original arrangement order of the chemical bond edges to obtain the edge feature vector of the chemical bond.

3. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, The process of constructing the heterogeneous graph structure of the organosilicon coating based on the node feature vectors and the edge feature vectors includes: Using the atomic nodes in the molecular structure data as graph nodes and the chemical bond edges in the molecular structure data as graph edges, an initial heterogeneous graph structure for the organosilicon coating is constructed. Based on the original arrangement order of the atomic nodes and the chemical bond edges, the graph nodes and graph edges are identified and assigned, resulting in unique node identifiers for the graph nodes and unique edge identifiers for the graph edges. Based on the atom type of the atomic node and the bond order of the chemical bond edge, the graph node and the graph edge are labeled with type to obtain the node type label of the graph node and the edge type label of the graph edge; Each row vector in the node feature vector is assigned to the graph node to obtain the node feature row vector, and the node feature row vector is used as the initial node feature of the graph node; Each row vector in the edge feature vector is assigned to the graph edge to obtain the edge feature row vector, and the edge feature row vector is used as the initial edge feature of the graph edge; In the initial heterogeneous graph structure, the unique node identifier, the unique edge identifier, the node type label, the edge type label, the initial node feature, and the initial edge feature are fused to obtain the heterogeneous graph structure of the organosilicon coating.

4. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, The step of inputting the heterogeneous graph structure into a pre-trained graph neural network model and performing context encoding on the target nodes in the heterogeneous graph structure to obtain the updated feature vectors of the target nodes includes: Call the pre-trained graph neural network model and read the aggregate weight matrix, edge weight matrix and update weight matrix of the target node in the heterogeneous graph structure from the parameter set of the graph neural network model; The target node in the heterogeneous graph structure is expanded to obtain the neighboring nodes that are directly connected to the chemical bond edge. The neighboring nodes are then feature-encoded to obtain the current neighboring node feature vector and the chemical bond edge feature vector. Based on the edge weight matrix, the chemical bond edge feature vector is linearly mapped to obtain the edge contribution vector of the adjacent node; Based on the aggregation weight matrix, the current node feature vector of the target node, the current neighboring node feature vector, and the edge contribution vector are weighted and combined to obtain the neighborhood aggregation vector of the target node. Based on the updated weight matrix, the neighborhood aggregation vector and the current node feature vector are fused and updated to obtain the updated feature vector of the target node. The original feature vector of the target node is then replaced with the updated feature vector to obtain the updated feature vector of the target node.

5. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 4, characterized in that, The step of weighting and combining the current node feature vector, the current neighbor node feature vector, and the edge contribution vector of the target node based on the aggregation weight matrix to obtain the neighborhood aggregation vector of the target node includes: The aggregated weight matrix is ​​decoupled to obtain the self-weight sub-vector of the current node feature vector, the neighbor weight sub-vector of the current neighbor node feature vector, and the edge contribution weight sub-vector of the edge contribution vector. The product of the weight coefficients in the self-weighted sub-vector and the dimension components in the current node feature vector is taken as the self-weighted component sequence of the current node feature vector; The weight coefficients in the neighbor weight sub-vectors are superimposed on the dimension components in the feature vectors of the current neighbor nodes to obtain the neighbor weighted component sequence of the feature vectors of the current neighbor nodes. The weight coefficients in the edge contribution weight sub-vector are multiplied by the dimension components in the edge contribution vector to obtain the edge contribution weighted component sequence of the edge contribution vector. The components at the same position in the self-weighted component sequence, the neighbor-weighted component sequence, and the edge contribution weighted component sequence are superimposed to obtain the fused components at different positions in the target node. The fused components are then concatenated to obtain the neighborhood aggregation vector of the target node.

6. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, Based on the node types of the heterogeneous graph structure, mode decomposition is performed on the updated feature vector to obtain multiple type feature vectors of the target node, including: Based on the node types of the heterogeneous graph structure, the updated feature vector is categorized to obtain multiple type groups of the updated feature vector; Group pooling is performed on the updated feature vectors within the multiple type groups to obtain the type feature vectors of the multiple type groups; Based on the inherent fixed order of the node type, the type feature vectors are sequentially merged to obtain multiple type feature vectors of the target node.

7. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 6, characterized in that, The step of performing group pooling on the updated feature vectors within the multiple type groups to obtain the type feature vectors of the multiple type groups includes: Traverse the vector dimensions of the updated feature vectors, extract the element values ​​of all updated feature vectors in the multiple type groups on the same vector dimension, and construct the set of elements to be aggregated in the vector dimension; The aggregated element values ​​of the vector dimension are obtained by summing and aggregating all element values ​​in the set of elements to be aggregated. Based on the original dimensional order of the updated feature vector, the aggregated element values ​​are sequentially concatenated to obtain a preliminary aggregated vector with the same dimensions as the updated feature vector; The aggregated element values ​​in the initial aggregated vector are averaged and reduced to the total number of updated feature vectors within the multiple type groups to obtain the mean feature vector of the multiple type groups, and the mean feature vector is used as the type feature vector of the multiple type groups.

8. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, The step of performing multi-source aggregation of the multiple types of feature vectors with the overall molecular feature vector in the molecular structure data to obtain the full-image representation vector of the organosilicon coating includes: Feature decoding is performed on the multiple types of feature vectors to obtain the node type labels of the multiple types of feature vectors; The molecular structure data is globally aggregated to obtain the overall molecular feature vector of the molecular structure data. Based on the fixed order of the node type labels, the multiple type feature vectors are serialized and combined to obtain a combined feature vector of the multiple type feature vectors; The combined feature vector is jointly encoded with the overall molecular feature vector to obtain the full-image representation vector of the organosilicon coating.

9. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 1, characterized in that, The step of inputting the full-image representation vector into the graph neural network model, performing a linear transformation on the full-image representation vector to obtain the transmittance prediction value of the full-image representation vector, and using the transmittance prediction value as the transmittance prediction result of the silicone coating includes: The parameters of the regression output layer in the graph neural network model are analyzed to obtain the weight parameter vector and bias parameters of the regression output layer. Based on the weight parameter vector, the dimensional components in the full graph representation vector are multiplied dimension by dimension to obtain multiple product results of the full graph representation vector; By summing the multiple product results, the intermediate accumulated value of the full-image representation vector is obtained; Based on the bias parameter, the intermediate accumulated value is corrected for deviation to obtain the transmittance prediction value of the full-image representation vector; The predicted transmittance value is used as the predicted transmittance result of the silicone coating.

10. The graph neural network prediction method for the transmittance of an organosilicon coating as described in claim 9, characterized in that, The formula for calculating the predicted transmittance value is as follows: ； In the formula, The predicted transmittance value is... The total number of dimensions of the full graph representation vectors. The dimension index variable is used to represent the vector of the entire graph. The weight parameter vector is the first... Each weighting coefficient The first element in the full graph representation vector is... Each dimension component The bias parameter is denoted as .

Images