OLED material stability prediction method based on heterogeneous graph neural network

Abstract

The invention discloses an OLED material stability prediction method based on a heterogeneous graph neural network, and the method comprises the steps: constructing a molecular graph based on the molecular structure of an OLED material, carrying out the data enhancement of the molecular graph, fusing the features of the data-enhanced molecular graph with multi-modal data, and obtaining the feature data of the molecular structure of the OLED material; oLED material molecular structure feature data are transmitted to a heterogeneous graph neural network, deep feature information in a molecular graph is learned, and OLED material molecular structure fusion features are obtained; and constructing an OLED material thermal stability prediction model based on the depth operator network, and inputting the OLED material molecular structure feature data and the OLED material molecular structure fusion features into the OLED material thermal stability prediction model to obtain a prediction result of the OLED material stability. According to the method, the stability prediction performance of the OLED material can be remarkably improved, and a new path is provided for material design and optimization.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and materials science, and particularly to a method for predicting the stability of OLED materials based on a heterogeneous graph neural network. Background Art

[0002] With the continuous development of display technology, organic light-emitting diodes (OLEDs) have been widely used in the fields of displays and lighting due to their advantages such as high contrast ratio, self-luminescence, thinness, lightness, and flexibility. However, the stability problem of OLED materials has always been one of the key bottlenecks restricting their large-scale application and performance improvement. Factors such as the molecular structure, chemical bond strength, and thermal stability of materials directly affect the service life and display effect of OLEDs. Therefore, how to efficiently predict and optimize the stability of OLED materials has become an important research direction in the field of materials science. Traditional research on OLED materials mainly relies on experimental and theoretical calculation methods. Experimental tests evaluate the stability of materials through accelerated aging tests. Although the results are reliable, the test cycle is long and the cost is high, making it difficult to meet the need for quickly screening a large number of candidate materials. Theoretical calculation methods can simulate the relationship between the structure and properties of materials at the atomic scale, but their computational complexity increases exponentially with the increase in molecular scale, making it difficult to efficiently process complex OLED material systems. Therefore, there is an urgent need to develop efficient and accurate prediction methods to make up for the deficiencies of traditional methods.

[0003] Currently, the research on the stability of OLED materials mainly relies on experimental tests and theoretical calculations. Experimental methods usually evaluate the stability of materials through accelerated aging tests. Although this method is reliable, it is costly and time-consuming, making it difficult to quickly screen a large number of candidate materials. Theoretical calculation methods (such as density functional theory, DFT) can simulate the stability of materials at the atomic scale, but their computational complexity increases exponentially with the increase in molecular scale, making it difficult to meet the requirements of high-throughput material screening.

[0004] To solve the efficiency problem of traditional methods, machine learning technologies have been widely applied to material property prediction in recent years. For example, traditional machine learning models such as Random Forest and Support Vector Machine (SVM), or deep learning models based on Convolutional Neural Network (CNN) and Graph Neural Network (GNN) are used. These methods achieve relatively high prediction efficiency by learning the relationship between the features and properties of materials from experimental or computational data. In the prior art, an ensemble learning model LightGBM is proposed to predict the thermal stability of OLED materials, and a high-quality dataset is constructed and active learning-guided density functional theory is adopted to enable traditional machine learning models to learn the molecular state of OLED materials.

[0005] However, these technologies still have the following disadvantages:

[0006] (1) Feature extraction relies on expert knowledge: Traditional machine learning models require manual design of material features (such as molecular descriptors), making it difficult to capture the complex structural characteristics of materials.

[0007] (2) Homogenization assumption limits the modeling ability: Most GNNs assume that the attributes of nodes and edges in the graph are homogeneous, and cannot fully describe the diverse relationships between different chemical bonds and atoms in OLED materials.

[0008] (3) Insufficient generalization ability: Existing models perform well on small-scale datasets, but the prediction accuracy drops significantly when dealing with new materials.

[0009] At the same time, the Heterogeneous Graph Neural Network (HGNN), as an extended graph neural network model, can handle data of multiple types of nodes and edges and has shown potential in materials science. However, the existing HGNN methods still have the following problems when applied to OLED materials: 1) Lack of global mapping ability: HGNN mainly focuses on feature aggregation in the local neighborhood and is difficult to effectively capture the global structural characteristics of materials; 2) Limited ability to process high-dimensional complex data: The stability of OLED materials is usually affected by multiple physical and chemical factors, and existing HGNNs are difficult to handle these complex high-dimensional relationships.

[0010] Therefore, the existing technologies have problems such as low efficiency, insufficient accuracy, and limited generalization ability in the prediction of OLED material stability.

[0011] Therefore, the existing technologies still have defects. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to provide a method for predicting the stability of OLED materials based on a heterogeneous graph neural network in view of the above defects of the existing technologies. The technical solution adopted by the present invention is as follows:

[0013] In a first aspect, the present invention provides a method for predicting the stability of OLED materials based on a heterogeneous graph neural network, wherein the method includes:

[0014] Construct a molecular graph based on the molecular structure of the OLED material, perform data augmentation on the molecular graph, and fuse the features of the data-augmented molecular graph with multi-modal data to obtain the molecular structure feature data of the OLED material;

[0015] Transmit the molecular structure feature data of the OLED material into the heterogeneous graph neural network to learn the deep feature information in the molecular graph and obtain the molecular structure fusion feature of the OLED material;

[0016] Construct an OLED material thermal stability prediction model based on a deep operator network, and input the OLED material molecular structure feature data and the OLED material molecular structure fusion feature into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability.

[0017] In one implementation, the node set in the molecular graph represents the atoms in the molecule, and different types of atoms have their own unique physical and chemical properties, and the physical and chemical properties include electronegativity, atomic radius, and electron arrangement;

[0018] The edges in the molecular graph represent the chemical bonds between atoms, and the attributes of the edges include bond type and bond energy.

[0019] In one implementation, the features of the data-augmented molecular graph are fused with multimodal data to obtain the OLED material molecular structure feature data, including:

[0020] Extract the global features and local features of the OLED material molecular structure from the data-augmented molecular graph to obtain molecular descriptors and molecular fingerprint features, and the molecular descriptors and the molecular fingerprint features are the key inputs for multimodal feature fusion;

[0021] Fuse the molecular descriptors and the molecular fingerprint features with the node feature matrix and the adjacency matrix in the molecular graph to obtain the OLED material molecular structure feature data.

[0022] In one implementation, the OLED material molecular structure feature data is passed to a heterogeneous graph neural network to learn the deep feature information in the molecular graph to obtain the OLED material molecular structure fusion feature, including:

[0023] In the message passing process of the heterogeneous graph neural network, the neighbor nodes are dynamically adjusted through an attention mechanism, and in edge type aggregation, weights are assigned to different types of edges by combining the attention mechanism;

[0024] Pool the node features in the molecular graph, and fuse the pooled global features with the molecular descriptors to obtain the OLED material molecular structure fusion feature.

[0025] In one implementation, the OLED material thermal stability prediction model based on the deep operator network includes a backbone network and a branch network, where the backbone network is used to learn the global characteristic mapping of the OLED material molecular thermal stability index, and the branch network is used to learn the local structure information of the OLED material molecular structure.

[0026] In one implementation, the backbone network consists of 7 fully-connected layers, and each fully-connected layer applies the ReLU activation function to perform layer-by-layer non-linear feature transformation on the fused features of the OLED material molecular structure and reduce the dimension of the features;

[0027] The branch network consists of a single-layer RNN layer and 6 fully-connected layers. The RNN layer is responsible for extracting the spatial position relationship between molecules in the OLED material molecular structure feature data to capture the molecular structure features. Each fully-connected layer applies the ReLU activation function to perform layer-by-layer non-linear feature transformation on the OLED material molecular structure feature data and reduce the dimension of the features.

[0028] In one implementation, the OLED material molecular structure feature data and the OLED material molecular structure fused features are input into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability, including:

[0029] Input the OLED material molecular structure fused features into the backbone network to output low-dimensional fused features;

[0030] Input the OLED material molecular structure feature data into the branch network to output low-dimensional molecular structure features;

[0031] Based on the low-dimensional fused features and the low-dimensional molecular structure features, obtain the prediction result of the OLED material stability, and the prediction result includes: the predicted value of the glass transition temperature and the predicted value of the decomposition temperature.

[0032] In a second aspect, an embodiment of the present invention further provides an OLED material stability prediction system based on a heterogeneous graph neural network. Among them, the OLED material stability prediction system based on the heterogeneous graph neural network is used to implement the steps of the OLED material stability prediction method based on the heterogeneous graph neural network. The system includes:

[0033] A molecular graph construction and multi-modal fusion module, which is used to construct a molecular graph based on the OLED material molecular structure, perform data augmentation on the molecular graph, and fuse the features of the data-augmented molecular graph with multi-modal data to obtain OLED material molecular structure feature data;

[0034] A heterogeneous graph neural network construction and feature extraction module, which is used to transfer the OLED material molecular structure feature data to the heterogeneous graph neural network, learn the deep feature information in the molecular graph, and obtain the OLED material molecular structure fused features;

[0035] The OLED material thermal stability prediction module is used to construct an OLED material thermal stability prediction model based on a deep operator network, and input the OLED material molecular structure feature data and the OLED material molecular structure fusion feature into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability.

[0036] In a third aspect, an embodiment of the present invention further provides a terminal. The terminal includes a memory, a processor, and an OLED material stability prediction program based on a heterogeneous graph neural network stored in the memory and executable on the processor. When the processor executes the OLED material stability prediction program based on the heterogeneous graph neural network, the steps of the OLED material stability prediction method based on the heterogeneous graph neural network in any one of the above solutions are implemented.

[0037] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium. A computer-readable storage medium stores an OLED material stability prediction program based on a heterogeneous graph neural network. When the OLED material stability prediction program based on the heterogeneous graph neural network is executed by a processor, the steps of the OLED material stability prediction method based on the heterogeneous graph neural network in any one of the above solutions are implemented.

[0038] Beneficial effects: Compared with the prior art, the present invention provides an OLED material stability prediction method based on a heterogeneous graph neural network. First, the present invention constructs a molecular graph based on the OLED material molecular structure, performs data augmentation on the molecular graph, and fuses the features of the data-augmented molecular graph with multi-modal data to obtain the OLED material molecular structure feature data. Then, the OLED material molecular structure feature data is transmitted to the heterogeneous graph neural network to learn the deep feature information in the molecular graph to obtain the OLED material molecular structure fusion feature. Finally, an OLED material thermal stability prediction model based on a deep operator network is constructed, and the OLED material molecular structure feature data and the OLED material molecular structure fusion feature are input into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability. By combining the fine modeling ability of the heterogeneous graph neural network for the molecular structure and the efficient learning ability of the heterogeneous graph neural network architecture for complex non-linear mappings, the present invention realizes high-precision prediction of the OLED material stability, and solves the problems of low efficiency, insufficient generalization ability, and inability to comprehensively capture the multi-dimensional characteristics of materials in the prior art. Description of the Drawings

[0039] Figure 1 It is a flowchart of a preferred embodiment of the OLED material stability prediction method based on a heterogeneous graph neural network provided by an embodiment of the present invention.

[0040] Figure 2 This is a schematic structural diagram of the OLED material thermal stability prediction model based on the deep operator network provided by an embodiment of the present invention.

[0041] Figure 3 This is a schematic architecture diagram of the OLED material stability prediction system based on the heterogeneous graph neural network provided by an embodiment of the present invention.

[0042] Figure 4 This is a schematic block diagram of the principle of the terminal provided by an embodiment of the present invention. Detailed implementation manners

[0043] To make the objectives, technical solutions and effects of the present invention clearer and more definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all the contents and operations or steps, nor do they necessarily need to be executed in the described order. For example, some operations or steps can also be decomposed, combined or partially merged, so the actual execution order may be changed according to the actual situation.

[0045] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0046] It should be understood that, in order to facilitate a clear description of the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, terms such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and effects. For example, the first control information and the second control information are only used to distinguish different control information, and do not limit their sequence.

[0047] Those skilled in the art can understand that the terms "first", "second", etc. do not limit the quantity and execution order, and the terms "first", "second", etc. do not necessarily mean different.

[0048] It should also be understood that the term " / and" used in the specification of the present invention and the appended claims refers to any combination and all possible combinations of one or more of the related listed items, and includes these combinations.

[0049] In recent years, with the rapid development of artificial intelligence technology, the application of machine learning and deep learning methods in materials science has gradually emerged. By learning the complex relationships between material structures and properties from experimental data or simulation data, these technologies have significantly accelerated the process of material design and optimization. Among them, the Graph Neural Network (GNN), as a deep learning model capable of processing graph-structured data, has shown great potential in predicting material properties because it can directly model the graph structures of molecules or crystals. For example, GNN can capture molecular structure features through the relationships between nodes (atoms) and edges (chemical bonds) to achieve accurate prediction of material properties. However, traditional graph neural networks usually assume that the attributes of nodes and edges in the graph are homogeneous, which has limitations in dealing with complex OLED materials. OLED materials have a high degree of heterogeneity, and their molecular structures may contain various types of atoms and chemical bonds, and these relationships have different effects on the stability of the materials. Therefore, the Heterogeneous Graph Neural Network (HGNN), as an extended graph neural network model, can more precisely describe the complex relationships in OLED materials by introducing the modeling ability of multiple types of nodes and edges. In addition, HGNN can also combine multi-modal data of materials (such as molecular structures, crystal properties, experimental parameters, etc.) to achieve a comprehensive modeling of material properties. This method not only improves the prediction accuracy of the model but also provides a new perspective for understanding the relationship between material properties and structures, promoting the efficient screening and performance optimization of OLED materials. By applying HGNN technology to the prediction of OLED material stability, it can provide important technical support for the development of display technology.

[0050] The method for predicting the stability of OLED materials based on the heterogeneous graph neural network in this embodiment can be applied to terminals, and the terminals include intelligent product terminals such as computers and mobile phones. Specifically, as Figure 1 shown, the method for predicting the stability of OLED materials based on the heterogeneous graph neural network includes the following steps:

[0051] Step S100: Construct a molecular graph based on the molecular structure of the OLED material, perform data augmentation on the molecular graph, and fuse the features of the data-augmented molecular graph with multi-modal data to obtain the molecular structure feature data of the OLED material.

[0052] The thermal stability and service life of OLED (organic light-emitting diode) materials are key issues for their commercialization. The thermal stability is directly affected by the molecular structure and its characteristics of the materials, but traditional experimental methods such as thermogravimetric analysis (TGA) and differential scanning calorimetry (DSC) are time-consuming and costly. To quickly and efficiently predict the thermal stability of OLED materials, data-driven methods based on machine learning have become a research hotspot. Combining the heterogeneous graph neural network (HGNN) and the deep learning framework, in this embodiment, the molecular structure of the OLED material is represented as a molecular graph G=(V, E), where V represents the node set, that is, atoms, and the edge E represents chemical bonds. The aim is to comprehensively capture the complex relationship between the molecular structure and the thermal stability.

[0053] In this molecular graph, the node set V in the molecular graph represents the atoms (such as carbon, hydrogen, oxygen) in the molecule, and each node and edge contain attribute information, where the node attribute is X v , and the attributes of each node v∈V include electronegativity, atomic radius, and electron arrangement; the edge E in the molecular graph represents the chemical bond between atoms, and the edge attribute is X e , and the attributes of each edge e∈E include the type of chemical bond (such as single bond, double bond, aromatic bond) and bond energy. In this embodiment, through the construction of the molecular graph, it supports the modeling of multiple types of nodes and edges and is suitable for describing the complex molecular structure characteristics. Then, the node set V is converted into a node feature matrix F, and the attribute X vi of each node is extracted and stacked into the node feature matrix F, representing the characteristics of the nodes. The edge set E is converted into an adjacency matrix A, and the connection relationship between nodes is represented by the adjacency matrix A.

[0054] Next, this embodiment performs data preprocessing. The purpose of data preprocessing is to remove noise, fill in missing values, and standardize the samples to ensure the consistency and reliability of the data. This embodiment can perform denoising and missing value processing on the molecular structure (that is, the node feature matrix F and the adjacency matrix A) to ensure data consistency.

[0055] Specifically, for the molecular structure set G=(F, A), statistical methods are used to detect outliers:

[0056]

[0057] If the sample F ij satisfies |F ii -μ j |>kσ j (k = 3), it is determined as an outlier and removed. Among them, μ j represents the sample mean, σ j represents the sample standard deviation, k is the multiple for outlier determination, and F ij is the node v iThe j-th eigenvalue. N is the number of nodes, and j represents the feature dimension.

[0058] Furthermore, for the positions M with missing values in the node feature matrix F, the mean filling method is used for supplementation:

[0059]

[0060] where N j is the number of samples with non-missing values in the j-th column, and F ij is an element in the node feature matrix, representing the value of the j-th feature of the i-th sample.

[0061] Furthermore, to eliminate the influence of feature dimensions, each feature is normalized. The normalized eigenvalue F' ij is expressed as:

[0062]

[0063] where μ j and σ j are the mean and standard deviation of the j-th column feature respectively. After data preprocessing, the preprocessed molecular structure set is G = (F1, A1).

[0064] Furthermore, in this embodiment, data augmentation is performed on the preprocessed molecular structure. According to the preprocessed molecular structure set G = (F1, A1), molecular simulation is used to generate an extended sample set to increase sample diversity. In this embodiment, molecular dynamics (MD) simulation can be used to generate new molecular structure data. Specifically, according to the molecular structure G = (F1, A1), new samples G' are generated through simulation:

[0065] G' = MD(G, Δt, T) (4)

[0066] where Δt is the time step and T represents the simulation temperature.

[0067] Furthermore, equivalent samples are generated using molecular symmetry. The node feature matrix F2 after rotating by the angle θ is:

[0068] F2 = R(θ) · F1 (5)

[0069] where R(θ) represents the rotation matrix. Random Gaussian noise ∈ ∼ N(0, σ 2 ) is added to the original node feature matrix F1 to generate augmented samples:

[0070] F2 = F1 + ∈ (6)

[0071] The molecular structure set after data augmentation is G=(F2, A2).

[0072] Furthermore, in this embodiment, feature selection is performed on the molecular structure set G=(F2, A2) after data augmentation. Key feature vectors are selected through feature importance analysis to reduce redundant features and improve the model efficiency. Specifically, for each feature dimension of the node feature matrix F2, the Pearson correlation coefficient r d with the target values (thermal decomposition temperature T g ) y i is calculated. j .

[0073]

[0074] Among them, is the value of the j-th feature of the i-th sample, represents the mean of the j-th feature, μ y is the mean of the target values. Features are screened according to the absolute value of the correlation coefficient, and a threshold γ is set:

[0075] R = {j||r j | > γ} (8)

[0076] Among them, R is the set of indexes of the key features screened out. γ is the correlation threshold.

[0077] Then, principal component analysis is performed to calculate the covariance matrix τ:

[0078]

[0079] The covariance matrix is decomposed into eigenvalues α1, α2, …, α d and the corresponding eigenvectors v1, v2, …, v d . The screened node feature matrix is projected onto the principal component space to generate the reduced-dimensional node feature matrix F f , and the adjacency matrix remains unchanged. The finally obtained molecular structure set is G=(F f , A f ).

[0080] To further enhance the feature expression ability of the model, this embodiment also introduces multimodal data fusion, jointly modeling the experimentally measured material properties (such as thermal decomposition temperature, energy gap) and computational simulation data (such as molecular trajectories, electron density distributions) with the features of the molecular graph. This fusion method can comprehensively extract multi-dimensional information related to material stability, providing a solid feature basis for subsequent prediction of important indicators (glass transition temperature and decomposition temperature) for measuring the thermal stability of OLED materials. Specifically, this embodiment extracts the global features and local features of the molecular structure of OLED materials from the data-augmented molecular graph to obtain molecular descriptors and molecular fingerprint features. The molecular descriptors and the molecular fingerprint features are the key inputs for multimodal feature fusion. Then, the molecular descriptors and the molecular fingerprint features are subjected to multimodal data fusion with the node feature matrix and the adjacency matrix in the molecular graph to obtain the molecular structure feature data of OLED materials.

[0081] In practical applications, the molecular descriptors and the molecular fingerprint features are the global features and local features extracted from the molecular graph G=(F f , A f ), respectively used to quantify the global properties and local structural information of the molecule. These features are calculated from the selected node feature matrix F f and the adjacency matrix A f through specific algorithms, forming the key inputs for multimodal feature fusion.

[0082] This embodiment can calculate the global descriptors of the molecule using the statistical information of the nodes and edges. Molecular mass: m i represents the atomic mass of the node, which is obtained from the selected node feature matrix F f ; Polar surface area: polar(v i ) represents the polar property of the node; Total bond energy: e ij is the bond energy property of the edge, which is obtained from the selected adjacency matrix A f .

[0083] Furthermore, calculate topological properties, including molecular diameter D mol and molecular complexity C mol . Similarly, it is obtained from the selected adjacency matrix A f .

[0084] Finally, the molecular descriptor D can be expressed as: D = [M, PSA, E bond , D mol , C mol .

[0085] Further, molecular fingerprint feature extraction is performed. The molecular fingerprint features are used to capture the local structure of the molecule, and the obtained molecular fingerprint features F fp .

[0086] Finally, the node feature matrix and the adjacency matrix are generated. In this embodiment, after the molecular descriptors and the molecular fingerprint features are extracted, the molecular descriptors D, the molecular fingerprint features F fp , the original node feature matrix F, and the adjacency matrix A are fused to obtain the OLED material molecular structure feature data Z:

[0087] Z = con(D, F fp , F, A) (10).

[0088] Step S200: Transmit the OLED material molecular structure feature data to the heterogeneous graph neural network to learn the deep - level feature information in the molecular graph, and obtain the OLED material molecular structure fusion feature.

[0089] The heterogeneous graph neural network (HGNN) is one of the core modules of this embodiment, which is used to extract deep - level feature information from the heterogeneous graph of the OLED material. By designing a message - passing mechanism for multiple types of nodes and edges, HGNN can effectively capture the complex relationships between different atoms and chemical bonds. To expand from local features to global characteristics, the HGNN of this embodiment adopts a multi - layer structure. Each layer gradually expands the neighborhood range of the nodes, and extracts the global features of the molecular structure through multiple iterative aggregations. In addition, to improve the computational efficiency and reduce redundant information, this embodiment designs a feature dimension reduction module to reduce redundant information, screen the key features related to the stability of the OLED material, and remove the noise features. This modular design not only improves the modeling ability of HGNN, but also significantly improves the computational efficiency and generalization performance of the model.

[0090] Using the OLED material molecular structure feature data Z obtained in the above - mentioned step S100 as the input, it is transmitted to the heterogeneous graph neural network (HGNN) to learn the deep - level features of the molecular graph. Due to the heterogeneity of nodes (atoms) and edges (chemical bonds) in the molecular graph, the contributions of different nodes and edges to the target characteristics are not equal. Some key atoms or functional groups may play a dominant role in molecular stability, while other parts contribute less. The traditional uniform message - passing mechanism is difficult to distinguish this difference, which may lead to information redundancy or weakening of key features. Therefore, this embodiment designs to combine the attention mechanism into the message - passing part, enabling the model to dynamically allocate weights to nodes and edges, thereby highlighting the structural features with high influence on the task.

[0091] Specifically, the core of the heterogeneous graph neural network is the message passing mechanism, which is used to spread information between nodes in the graph. For heterogeneous graphs, the types of nodes and edges may be different, so message passing needs to consider node types and edge types.

[0092] First, initialize the node features. The initial feature vector of each node v i is:

[0093]

[0094] where F i is the feature of node v i .

[0095] Furthermore, message passing and node feature update. Information is passed from neighbor node v j and edge e ij to node v i . At each layer l, the node features are updated through neighbor node and edge features:

[0096]

[0097] where, is the feature representation of node v i at the l-th layer. is the self-loop weight matrix. represents the weight matrix for edge type r. For heterogeneous graphs, the information of different types of nodes and edges is aggregated separately:

[0098] Node type aggregation:

[0099]

[0100] where, t1, t2, …, t T represent node types.

[0101] Edge type aggregation:

[0102]

[0103] where, R is the set of edge types.

[0104] To enhance the model's attention to important nodes and edges and improve the effectiveness of message passing, this embodiment introduces an attention mechanism. During the message passing process, the neighbor nodes are dynamically adjusted through the attention mechanism, and according to formula (12), it can be changed to:

[0105]

[0106] where, the attention weight represents the neighbor node v j to node vi The importance, and its calculation formula is:

[0107]

[0108] Among them, W att represents the attention weight matrix, which maps the node features to the attention space. a is the attention vector, which is used to calculate the correlation of node pairs.

[0109] Furthermore, in edge type aggregation, the attention mechanism is combined to assign weights to different types of edges, and according to formula (14), it can be changed to:

[0110]

[0111] Among them, the edge type weight is calculated through edge type embedding:

[0112]

[0113] Among them, represents the edge type embedding vector. is the normalized edge type weight.

[0114] Secondly, graph-level feature extraction and aggregation are performed. To extract the global features of the graph, pooling is performed on the node features, and average pooling is used:

[0115]

[0116] The pooled global features are fused with the molecular descriptor D to obtain the fused feature h of the OLED material molecular structure f :

[0117] h f = concat(h g , D).

[0118] Step S300, construct an OLED material thermal stability prediction model based on a deep operator network, and input the OLED material molecular structure feature data and the OLED material molecular structure fused feature into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability.

[0119] The problem of predicting the glass transition temperature and decomposition temperature in OLED materials can be expressed as a non - linear operator learning task. The Deep Operator Network (DeepONet architecture) is good at dealing with complex non - linear mapping problems. Therefore, in this embodiment, an OLED material thermal stability prediction model is designed based on the DeepONet architecture to achieve high - precision prediction of the glass transition temperature and decomposition temperature. In this embodiment, the OLED material thermal stability prediction model based on the deep operator network includes a backbone network and a branch network. Among them, the backbone network is used to learn the global characteristic mapping of the OLED material molecular thermal stability indicators (glass transition temperature and decomposition temperature). Therefore, the input of the backbone network is the fused feature h of the OLED material molecular structure obtained in step S200 f . The branch network is used to learn the local structural information of the OLED material molecular structure, which helps to capture the relationship between the molecular structure and stability. Therefore, the input of the branch network is the OLED material molecular structure feature data Z obtained in step S100. The structure of the OLED material thermal stability prediction model based on the deep operator network in this embodiment is as Figure 2 shown. The OLED material thermal stability prediction model includes a backbone network, a branch network, a dot - product layer, and a Sigmoid activation function layer. Among them, the backbone network consists of 7 fully - connected layers, and each fully - connected layer applies the ReLU activation function to perform layer - by - layer non - linear feature transformation on the fused feature of the OLED material molecular structure and reduce the dimension of the features. The branch network consists of a single - layer RNN layer and 6 fully - connected layers. The RNN layer is responsible for extracting the spatial position relationship between molecules in the OLED material molecular structure feature data to capture the molecular structure features. Each fully - connected layer applies the ReLU activation function to perform layer - by - layer non - linear feature transformation on the OLED material molecular structure feature data and reduce the dimension of the features. Therefore, by inputting the fused feature of the OLED material molecular structure into the backbone network, a low - dimensional fused feature Lh f can be output. By inputting the OLED material molecular structure feature data into the branch network, a low - dimensional molecular structure feature LZ can be output. Finally, in this embodiment, based on the low - dimensional fused feature Lh f and the low - dimensional molecular structure feature LZ, the prediction result of the OLED material stability is obtained. The prediction result includes: the predicted value of the glass transition temperature and the predicted value of the decomposition temperature. The specific formula is as follows:

[0120] The specific calculation formula is as follows:

[0121]

[0122]

[0123] Where is the predicted glass transition temperature, is the predicted decomposition temperature, is the p-th element in the low-dimensional fusion feature and the p-th element in the LZ low-dimensional molecular structure feature.

[0124] It can be seen that in this embodiment, DeepONet (Deep Operator Network) is introduced into the modeling framework. DeepONet further improves the prediction accuracy by learning the complex non-linear mapping from the OLED material feature space to the thermal stability indicators (glass transition temperature and decomposition temperature). DeepONet is good at dealing with complex non-linear mapping problems. Its branch network is used to learn the input features of the material molecular structure, and the backbone network learns the global feature mapping of the stability indicators. The global features extracted by HGNN are used as the input of DeepONet. The local patterns in the molecular structure are captured through the branch network, while the backbone network learns the complex relationship between the material properties and stability. The branch network of DeepONet is used to learn the specific input features of the molecular structure, and the backbone network learns the global feature mapping of the stability indicators. Combining the two generates the final prediction result. Through this dual-network structure, DeepONet can effectively make up for the deficiencies of HGNN in dealing with high-dimensional non-linear mapping. In addition, DeepONet has strong generalization ability and can adapt to different types of material data, providing a reliable technical support for the rapid screening and optimization of OLED materials.

[0125] In summary, a method for predicting the stability of OLED materials based on heterogeneous graph neural network provided in this embodiment combines HGNN (heterogeneous graph neural network) and DeepONet (deep operator network) for the first time to predict the stability of OLED materials. HGNN can finely model the molecular structure and capture the complex relationships between different atoms and chemical bonds, while DeepONet is good at dealing with complex non-linear mapping problems. The combination of the two significantly improves the prediction accuracy. Secondly, by introducing a multi-modal data fusion strategy, the experimentally measured material property data and computational simulation data (such as molecular trajectories, electron density distributions) are jointly modeled with the heterogeneous graph features to comprehensively capture the multi-dimensional information related to the stability of OLED materials and enhance the feature expression ability of the model. Thirdly, by introducing an attention mechanism in HGNN, the model can dynamically allocate the weights of nodes and edges, highlight the structural features with high influence on the task, reduce information redundancy, and improve the effectiveness of message passing. In addition, the global features of HGNN are extracted through the pooling layer and fused with the molecular descriptors to obtain the fusion features of the OLED material molecular structure, providing a more comprehensive feature basis for subsequent stability prediction.

[0126] Based on the above embodiments, the present invention further provides an OLED material stability prediction system based on a heterogeneous graph neural network. The system of this embodiment implements the steps of the OLED material stability prediction method based on a heterogeneous graph neural network in the above embodiments. Specifically, as Figure 3 shown, the system includes: a molecular graph construction and multimodal fusion module 10, a heterogeneous graph neural network construction and feature extraction module 20, and an OLED material thermal stability prediction module 30. Specifically, the molecular graph construction and multimodal fusion module 10 is used to construct a molecular graph based on the molecular structure of the OLED material, perform data augmentation on the molecular graph, and fuse the features of the data-augmented molecular graph with multimodal data to obtain OLED material molecular structure feature data. The heterogeneous graph neural network construction and feature extraction module 20 is used to transfer the OLED material molecular structure feature data into a heterogeneous graph neural network, learn the deep feature information in the molecular graph, and obtain the OLED material molecular structure fusion feature. The OLED material thermal stability prediction module 30 is used to construct an OLED material thermal stability prediction model based on a deep operator network, and input the OLED material molecular structure feature data and the OLED material molecular structure fusion feature into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability.

[0127] The working principles of the various modules in the OLED material stability prediction system based on a heterogeneous graph neural network in this embodiment are the same as those of the various steps in the above method embodiment, and will not be elaborated here.

[0128] Each of the above modules in the OLED material stability prediction device based on a heterogeneous graph neural network can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in or independent of the processor in the terminal in the form of hardware, or stored in the memory in the terminal in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above modules.

[0129] Based on the above embodiments, the present invention further provides a terminal, and the principle block diagram of the terminal can be as Figure 4 shown. The terminal may include one or more processors 100 ( Figure 4 only one is shown in

[0130] In one embodiment, the so-called processor 100 may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0131] In one embodiment, the memory 101 may be an internal storage unit of the electronic device, such as the hard disk or memory of the electronic device. The memory 101 may also be an external storage device of the electronic device, such as a plug-in hard disk equipped on the electronic device, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 101 may also include both the internal storage unit and the external storage device of the electronic device. The memory 101 is used to store computer programs and other programs and data required by the terminal. The memory 101 may also be used to temporarily store data that has been output or is to be output.

[0132] Those skilled in the art can understand that Figure 4 the principle block diagram shown in

[0133] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, operational database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0134] 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 them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or equivalently replace some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting OLED material stability based on heterogeneous graph neural network, characterized in that: The method comprises: Building a molecular graph based on the molecular structure of the OLED material, performing data enhancement on the molecular graph, and fusing the features of the data-enhanced molecular graph with the multimodal data to obtain molecular structure feature data of the OLED material; The molecular structure characteristic data of the OLED material is transmitted to a heterogeneous graph neural network to learn the deep characteristic information of the molecular graph and obtain the molecular structure fusion characteristics of the OLED material; A thermal stability prediction model for OLED materials based on a deep operator network is constructed, and the molecular structure characteristic data of the OLED material and the molecular structure fusion characteristics of the OLED material are input into the thermal stability prediction model for OLED materials based on a deep operator network to obtain a prediction result of the stability of the OLED material.

2. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 1, characterized in that: The node set in the molecular graph represents the atoms in the molecule, and different types of atoms have their own unique physical and chemical properties, including electronegativity, atomic radius and electronic configuration; The edges in the molecular graph represent chemical bonds between atoms, and the attributes of the edges include bond type and bond energy.

3. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 1, characterized in that: The features of the enhanced molecular graph are fused with the multimodal data to obtain the molecular structure feature data of OLED materials, including: Extracting global and local features of the molecular structure of the OLED material from the data-enhanced molecular graph to obtain molecular descriptors and molecular fingerprint features, wherein the molecular descriptors and molecular fingerprint features are key inputs for multimodal feature fusion; The molecular descriptors and the molecular fingerprint features are subjected to multimodal data fusion with the node feature matrix and the adjacency matrix in the molecular graph to obtain the molecular structure feature data of the OLED material.

4. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 3, characterized in that: The molecular structure feature data of the OLED material is transferred to the heterogeneous graph neural network to learn the deep feature information in the molecular graph and obtain the molecular structure fusion features of the OLED material, including: In the message passing process of heterogeneous graph neural networks, the neighbor nodes are dynamically adjusted through the attention mechanism, and in the edge type aggregation, the attention mechanism is combined to assign weights to different types of edges; The node features in the molecular graph are pooled, and the pooled global features are fused with the molecular descriptors to obtain the fused features of the molecular structure of OLED materials.

5. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 1, characterized in that: The OLED material thermal stability prediction model based on deep operator network includes a trunk network and a branch network, wherein the trunk network is used to learn the global characteristic mapping of the thermal stability index of OLED material molecules, and the branch network is used to learn the local structural information of the OLED material molecular structure.

6. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 5, characterized in that: The backbone network consists of 7 fully connected layers, and each fully connected layer applies a ReLU activation function to perform layer-by-layer nonlinear feature transformation on the molecular structure fusion features of the OLED material and reduce the dimension of the features; The branch network consists of a single RNN layer and 6 fully connected layers. The RNN layer is responsible for extracting the spatial position relationship between molecules in the molecular structure feature data of OLED materials to capture the molecular structure characteristics. Each fully connected layer applies the ReLU activation function to perform layer-by-layer nonlinear feature transformation on the molecular structure feature data of OLED materials and reduce the dimension of the features.

7. The method for predicting OLED material stability based on heterogeneous graph neural network according to claim 6, characterized in that: Inputting the molecular structure characteristic data of the OLED material and the molecular structure fusion characteristics of the OLED material into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability, including: Inputting the fusion features of the molecular structure of the OLED material into the backbone network and outputting low-dimensional fusion features; Inputting the molecular structure characteristic data of the OLED material into the branch network, and outputting low-dimensional molecular structure characteristics; Based on the low-dimensional fusion characteristics and the low-dimensional molecular structure characteristics, a prediction result of the stability of the OLED material is obtained, and the prediction result includes: a predicted value of the glass transition temperature and a predicted value of the decomposition temperature.

8. An OLED material stability prediction system based on heterogeneous graph neural network, characterized in that: The OLED material stability prediction system based on heterogeneous graph neural network is used to implement the steps of the OLED material stability prediction method based on heterogeneous graph neural network according to any one of claims 1 to 7, and the system comprises: A molecular graph construction and multimodal fusion module is used to construct a molecular graph based on the molecular structure of the OLED material, perform data enhancement on the molecular graph, and fuse the features of the data-enhanced molecular graph with the multimodal data to obtain the molecular structure feature data of the OLED material; A heterogeneous graph neural network construction and feature extraction module is used to transfer the molecular structure feature data of the OLED material to the heterogeneous graph neural network, learn the deep feature information of the molecular graph, and obtain the molecular structure fusion feature of the OLED material; The OLED material thermal stability prediction module is used to construct an OLED material thermal stability prediction model based on a deep operator network, and input the OLED material molecular structure feature data and the OLED material molecular structure fusion features into the OLED material thermal stability prediction model based on the deep operator network to obtain the prediction result of the OLED material stability.

9. A terminal, characterized in that: The terminal includes a memory, a processor, and an OLED material stability prediction program based on a heterogeneous graph neural network stored in the memory and executable on the processor. When the processor executes the OLED material stability prediction program based on a heterogeneous graph neural network, the steps of the OLED material stability prediction method based on a heterogeneous graph neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an OLED material stability prediction program based on a heterogeneous graph neural network. When the OLED material stability prediction program based on a heterogeneous graph neural network is executed by a processor, the steps of the OLED material stability prediction method based on a heterogeneous graph neural network as described in any one of claims 1 to 7 are implemented.

Images