Polymer informatics method based on graph neural network and attention mechanism
By combining graph neural networks and self-attention mechanisms, the problem of neglecting connectivity features in polymer material processing is solved, enabling efficient polymer property prediction and reverse design, and generating novel materials that conform to chemical rules.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNIV
- Filing Date
- 2024-06-04
- Publication Date
- 2026-07-31
AI Technical Summary
Existing polymer material processing methods ignore the inherent connectivity characteristics of polymers, and ordinary graph neural networks (GNNs) cannot meet the complex connectivity of polymers. Experts manually improving the accuracy of GNNs is time-consuming and labor-intensive.
A graph neural network and attention mechanism-based approach is adopted, which combines graph convolutional neural networks (GCN) and the self-attention mechanism of Transformer to capture the relationships between polymer atoms, and combines genetic algorithms to perform reverse design and property prediction of polymers.
It improves the accuracy of polymer property prediction and data population update rate, the generated polymer materials are more in line with chemical rules, shorten the material research and development cycle, and increase the possibility of discovering new materials.
Smart Images
Figure CN118571365B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence algorithms, and more specifically to a polymer informatics method based on graph neural networks and attention mechanisms. Background Technology
[0002] Polymer materials, due to their superior chemical, physical, biological, and mechanical properties, have permeated into human daily life. Polymers are typically composed of long-chain molecules, which are formed by numerous organic molecules linked by covalent bonds. The vast variety of available monomers and polymerization methods results in an enormous search space for polymers. Consequently, traditional empirical and trial-and-error methods are often constrained by computational and experimental limitations.
[0003] For the digital processing of polymers, finding suitable polymer fingerprints for complex computations is crucial. Traditional methods include structured feature extraction and self-supervised language model training. Structured feature extraction includes chemical tool extraction and manual extraction. Chemical tools such as RDKit and Chem use traditional extraction methods like ECFP to directly extract important chemical features from polymer monomers. Manual extraction requires expert experience to manually encode characteristics such as polymer chain length, morphology, and polymerization mode. Self-supervised language model training uses chemical languages like SMILES for self-supervised training, treating the word embeddings of polymer monomer SMILES as features. While these two methods can represent polymers to some extent, they neglect the inherent connectivity features of polymers. Currently, the most advanced method is to use graph neural networks (GNNs) to train and model polymer monomers. Top universities such as MIT and Georgia Tech are working on applying GNNs to polymer informatics. However, ordinary GNNs cannot handle the complex connectivity relationships of polymers, and manually improving GNN accuracy using expert knowledge is time-consuming and labor-intensive. Therefore, a purely data-driven approach combined with GNNs is needed to extract and represent polymer information. Currently, the most advanced method for capturing the relationships between various elements is the self-attention mechanism in Transformers. This mechanism has led to the rapid growth of large language models in recent years, such as ChatGPT and Wenxin Yiyan. Methods combining GNNs with self-attention mechanisms have also been extensively studied, but none are currently applicable to the polymer science field. Summary of the Invention
[0004] This invention provides a polymer informatics method based on graph neural networks and attention mechanisms. Its purpose is to accelerate the screening and discovery of polymer materials by using artificial intelligence (AI) and machine learning (ML) technologies through a data-driven approach.
[0005] This invention is achieved through the following technical solution:
[0006] A polymer informatics method based on graph neural networks and attention mechanisms uses a graph convolutional neural network (GCN) to model and train polymers, and incorporates a self-attention mechanism module into the GCN to efficiently capture the relationships between polymer atoms. The specific steps include:
[0007] Step (1) Use RDKit to extract features from polymer data. The extracted features include atomic features and bond features.
[0008] Step (2) Represent the polymer as graph data based on the connection relationship between atoms, model the atomic features and bond features in step (1) into node matrix and edge matrix, model the connection relationship between atoms into adjacency matrix, and model other positional relationships such as the shortest distance between nodes into positional codes;
[0009] Step (3) sends the polymer graph data obtained in step (2) into GCN for training, and the node matrix of the graph data is dynamically updated by the message passing steps of GCN.
[0010] Step (4) After the message passing operation of GCN is completed, the node matrix is updated using the self-attention mechanism module of Transformer. The node features are regarded as word embeddings and are sent to the self-attention mechanism module after being encoded at the fusion position.
[0011] Step (5) The graph updated using the self-attention mechanism has information about the relationships between atoms. After compressing the node matrix, the graph embedding can be obtained, and then the graph embedding is used to train the downstream task.
[0012] The self-attention mechanism module described in step (6) guides polymer reverse design. After obtaining efficient graph data, the graph data and the atomic self-attention scores obtained from the attention layer are used to perform downstream reverse design tasks.
[0013] The reverse design described in step (7) is a genetic algorithm that treats the atomic structure of the polymer as individuals in a population;
[0014] Step (8) uses the atom self-attention score obtained in step (5) to assign a certain weight to each atom. Atoms with higher weights are more likely to be updated by the genetic algorithm, effectively using the information from the training process of GNN and attention mechanism;
[0015] Step (9) Under the framework of this genetic algorithm, the fitness of the population is dynamically designed and scored according to the trained multilayer perceptron agent model. After the population evolution is completed, the polymer material with the required properties can be obtained. Materials with the required properties are selected from the population to update the database.
[0016] Furthermore, the atomic characteristics in step (1) include atomic number, degree, number of extranuclear electrons, chiral label, number of hydrogen atoms, whether it is a tetrahedral aromatic atom, hybridization mode, and atomic mass.
[0017] Furthermore, the bond features in step (1) include bond type, conjugate bond, and whether it forms a ring.
[0018] Furthermore, in step (7), several polymers constitute a population, and the population will be dynamically updated in the genetic algorithm to allow population individuals that are more in line with the target properties to survive.
[0019] The technical principle of this invention is as follows: By using a Graph Convolutional Network (GCN) and the self-attention mechanism module of the Transformer, the graph data of the polymer is dynamically updated for node features and relationships are modeled. The GCN is responsible for initial message passing and feature extraction, while the self-attention mechanism further captures the complex interrelationships between nodes and generates graph embeddings for training downstream tasks.
[0020] In the polymer reverse engineering process, the advantages of genetic algorithms and deep learning are combined. The genetic algorithm treats the atomic structure of the polymer as individuals in a population and uses weights obtained through a self-attention mechanism to guide the population's evolution. A multilayer perceptron surrogate model evaluates the population's fitness, optimizes the design process, and ultimately generates polymer materials with the target properties.
[0021] Beneficial effects: By using the graph embedding representation of this invention, the property prediction ability on datasets of homopolymers and copolymers is improved compared to existing pure GNN methods. By using the graph embedding representation of this invention and using atomic self-attention scores to guide inverse design, the data population update rate and population fitness are improved compared to ordinary algorithm processes, and the generated polymer materials are more chemically consistent. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the framework structure of the polymer informatics method based on graph neural networks and attention mechanisms of the present invention. Detailed Implementation
[0023] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.
[0024] Example:
[0025] like Figure 1This paper presents a polymer informatics framework based on graph neural networks and attention mechanisms. First, a dataset is selected from a polymer database, transformed into graph data, and fed into a graph neural network for training. Then, a self-attention mechanism layer is used to obtain updated graph embeddings and attention scores between nodes in the graph. Here, the graph embeddings can use machine learning methods such as multilayer perceptrons for property prediction, resulting in a surrogate model. In the reverse design phase, a genetic algorithm is used, iteratively evolving the graph embeddings and attention scores as input. The surrogate model is used to score individuals in the population. The final output of the algorithm is the polymer data that conforms to the specified properties. The specific implementation includes the following three aspects:
[0026] 1. Establishment of a surrogate model for predicting polymer properties:
[0027] The physicochemical properties of polymers are predicted using a Graph Convolutional Neural Network (GCN) model. First, the polymer is represented as graph data, including features such as element types, masses, and connectivity of atoms. This graph data is then fed into the GCN model for training. During training, the model extracts polymer structural features through a message-passing mechanism and captures the complex relationships between atoms using a self-attention mechanism.
[0028] This method can efficiently and accurately predict various properties of polymers, such as thermal conductivity, mechanical strength, and solubility. Through this prediction, researchers can quickly screen polymers with potentially superior properties, significantly shortening the materials development cycle.
[0029] 2. Reverse engineering of polymers:
[0030] Inverse design of polymers is achieved using GCN and self-attention mechanisms. First, a graph representation and self-attention scores of the polymer are obtained through a GCN model. Then, a genetic algorithm is used to treat the polymer's atomic structure as a population, assigning weights to each atom based on its self-attention score. Atoms with higher weights are more likely to be selected and updated in the genetic algorithm.
[0031] This reverse design method dynamically adjusts population fitness, and through continuous evolution, ultimately generates polymer materials with specific properties. This method can significantly improve polymer design efficiency and greatly increase the likelihood of discovering novel materials.
[0032] 3. Construction of a polymer materials database:
[0033] Graph neural networks and self-attention mechanisms are used to process large-scale polymer datasets, including structural and property information of different types of polymers. The data is then fed into a materials database after inference by the trained model.
[0034] Establish a large database covering the properties of a wide range of polymer materials. This database will not only be able to store and query the properties of existing polymers, but also predict the properties of new polymers through machine learning models, providing important data support and decision-making basis for materials science research.
[0035] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A polymer informatics method based on graph neural networks and attention mechanisms, characterized in that, The polymer was modeled and trained using a Graph Convolutional Neural Network (GCN), and a self-attention mechanism was added to the GCN to efficiently capture the relationships between polymer atoms. The specific steps included: Step (1) Use RDKit to extract features from polymer data. The extracted features include atomic features and bond features. Step (2) Represent the polymer as graph data based on the connection relationship between atoms, model the atomic features and bond features in step (1) into node matrix and edge matrix, model the connection relationship between atoms into adjacency matrix, and model other positional relationships such as the shortest distance between nodes into positional encoding; Step (3) sends the polymer graph data obtained in step (2) into GCN for training, and the node matrix of the graph data is dynamically updated by the message passing steps of GCN. Step (4) After the message passing operation of GCN is completed, the node matrix is updated using the self-attention mechanism module of Transformer. The node features are regarded as word embeddings and are sent to the self-attention mechanism module after being encoded at the fusion position. Step (5) The graph updated using the self-attention mechanism has information about the relationships between atoms. After compressing the node matrix, the graph embedding can be obtained, and then the graph embedding is used to train the downstream task. The self-attention mechanism module described in step (6) guides polymer reverse design. After obtaining efficient graph data, the graph data and the atomic self-attention scores obtained from the attention layer are used to perform downstream reverse design tasks. The reverse design described in step (7) is a genetic algorithm that treats the atomic structure of the polymer as individuals in a population; Step (8) uses the atom self-attention score obtained in step (5) to assign a certain weight to each atom. Atoms with higher weights are more likely to be updated by the genetic algorithm, effectively using the information from the training process of GNN and attention mechanism; Step (9) Under the framework of this genetic algorithm, the fitness of the population is dynamically designed and scored according to the trained multilayer perceptron agent model. After the population evolution is completed, the polymer material with the required properties can be obtained. Materials with the required properties are selected from the population to update the database.
2. The graph neural network and attention mechanism based polymer informatics method of claim 1, wherein, The atomic characteristics in step (1) include atomic number, degree, number of extranuclear electrons, chiral label, number of hydrogen atoms, whether it is a tetrahedral atom, hybridization mode, and atomic mass.
3. The graph neural network and attention mechanism based polymer informatics method of claim 1, wherein, The bond features in step (1) include bond type, conjugate bond, and whether it forms a ring.
4. The graph neural network and attention mechanism based polymer informatics method of claim 1, wherein, In step (7), several polymers constitute a population. The population will be dynamically updated in the genetic algorithm so that population individuals that are more in line with the target properties can survive.