Polymer graph convolutional neural network method based on uncertainty guidance

Abstract

The invention belongs to the field of material science and chemical informatics, and particularly discloses a polymer graph convolutional neural network method based on uncertainty guidance, and the method comprises the steps: carrying out the graph structure coding of a polymer monomer; performing property prediction and uncertainty evaluation on the data set; training an uncertainly guided active learning acceleration model; a global attention mechanism identifies key functional groups; and high-throughput screening of polymers with specific target properties is carried out. According to the method, through the representation capability of the complex topological relation of the graph neural network and uncertainty quantification of evidence deep learning, the polymer features are effectively extracted, efficient sample training is achieved, and the method is particularly suitable for the data scarcity condition; the relation between the polymer structure and the performance can be deeply excavated, and recognition of high-quality polymers is accelerated.

Description

Technical Field

[0001] The present invention relates to the technical field of material science and chemical informatics, and in particular to a method of a polymer graph convolutional neural network based on uncertainty guidance. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The applications of polymer materials in modern society range from daily consumer products to high-performance electronic devices, medical devices, and aerospace. Machine learning technology has gradually become a hot tool for the development of new polymers due to its advantages in processing complex data structures and predicting properties.

[0004] In the process of polymer structure-property prediction, researchers have adopted a variety of machine learning methods, including random forest (RF), Gaussian process regression (GPR), K nearest neighbor algorithm (KNN), and graph neural network (GNN). Although these methods can predict the properties of polymers to a certain extent, they have some significant shortcomings: (1) they often rely on a large number of manual descriptors as input features, and the imperfect design of these descriptors may lead to insufficient prediction accuracy of the model; (2) they have limited ability to deal with complex polymer structural relationships, such as: they are not effective in dealing with direct environmental effects beyond specific atoms or bonds, and it is difficult to capture the deep-level features inside the polymer; (3) the computational complexity is high when dealing with large data sets, resulting in slow model training and prediction; (4) the poor interpretability of existing complex models is also a problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method based on uncertainty-guided polymer graph convolutional neural network to improve prediction accuracy, data utilization efficiency and model interpretability, and screen potential candidate polymers with specific properties to accelerate polymer discovery.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A method for a polymer graph convolutional neural network based on uncertainty guidance comprises the following steps: S1. Extract the polymer monomer information, convert the SMILES code representing the polymer monomer molecule into Morgan fingerprint, and transmit it to the polymer graph convolutional neural network to capture the structural characteristics of the molecule; S2. The node and edge feature vectors of the molecule are aggregated through convolution and embedded with different weights by the global soft attention mechanism to form the feature vector of the polymer monomer molecular graph structure to capture the topological relationship within the polymer monomer molecule; S3. Input the feature vector of the graph structure into the evidence layer (uncertainty estimation layer) to predict the properties of the polymer and provide uncertainty estimates; S4. Select data samples through uncertainty-guided active learning methods to improve data utilization and model training efficiency; S5. Use the trained polymer graph convolutional neural network model to perform high-throughput screening of a large number of polymers to identify potential candidate polymers with specific properties; S6. Identify important functional groups within polymer monomer molecules by focusing on the global mechanism and associate them with specific material properties, providing physical mechanistic insights.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) The uncertainty-guided polymer graph convolutional neural network method proposed in the present invention directly uses SMILES strings to represent polymer structures, avoiding tedious manual descriptor design and improving the ease of use and generalization ability of the model; (2) The present invention can effectively capture the complex atomic interactions within the polymer, thereby providing a deeper understanding and characterization of the structural characteristics of the polymer and improving the accuracy of the prediction; (3) The uncertainty estimation introduced in this invention enhances the interpretability of the model and provides an explanation at the evidence level, making the prediction results of the model more transparent and credible, especially more reliable when data is scarce; (4) The present invention uses an active learning strategy to prioritize high-uncertainty data for training, thereby optimizing the utilization efficiency of training data and increasing the training speed of the model. At the same time, it also ensures the realization of a lightweight model, and can efficiently perform high-throughput screening of polymer properties. (5) The present invention recognizes functional groups through an attention mechanism, which can automatically identify and emphasize functional groups that play an important role in polymer properties, helping to explain the intrinsic relationship between polymer structure and properties; BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 A schematic diagram of an uncertainty-guided polymer graph convolutional neural network prediction process involved in the present invention. Figure 2 This is a graph of ten-fold cross-validation results of an uncertainty-guided polymer graph convolutional neural network on a RadonPy dataset involved in the present invention. Figure 3 This is a comparison result of the prediction accuracy of the uncertainty-guided polymer graph convolutional neural network involved in the present invention and other methods. Figure 4This is a comparison chart of training error results of active learning and random learning of an uncertainty-guided polymer graph convolutional neural network involved in the present invention for data set uncertainty. Figure 5 This is a result graph of high-throughput screening of polymer properties on RadonPy dataset and PI1M dataset based on an uncertainty-guided polymer graph convolutional neural network involved in the present invention. Figure 6 The present invention provides a method for the uncertainty-guided polymer graph convolutional neural network to screen the frequently occurring functional groups in polymers in the RadonPy dataset, which is used to classify and summarize the results. Figure 7 The present invention relates to a method of uncertainty-guided polymer graph convolutional neural network based on the attention mechanism to identify key functional group results. DETAILED DESCRIPTION

[0010] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0011] In this example, a method based on uncertainty-guided polymer graph convolutional neural network is disclosed, which encodes the graph structure features of polymer monomers, explores the intrinsic mechanism relationship between polymer structure and properties, performs polymer property prediction and uncertainty assessment, and accelerates high-throughput screening to explore polymers with specific target properties.

[0012] The specific implementation of this method is as follows: (1) Data preprocessing: Given a batch of SMILES data representing the molecular structure of polymers, the rdkit tool is used to extract the structural information of monomers and polymers. The extracted atomic type, atomic degree, hybridization mode, bond type, stereostructure, and conjugation mode information are encoded and converted into Morgan fingerprints, that is, binary vectors with nodes fixed to 163 lengths and edges fixed to 13 lengths. (2) Model architecture: The feature vector is embedded in the polymer graph convolutional neural network model. The model performs convolution and pooling steps on the complex topological relationships within the molecule and assigns different weights to generate the graph feature vector of the molecule. There are four convolutional layers in total. In the graph convolution layer: z (u,j) =v i ||v j ||u (i,j) Represents all nodes v i , v j and adjacent edge u(i,j) The concatenation of feature vectors, which are dynamically updated at each iteration t. f ,W s ,b f ,b s Represents different weights and biases, σ and g are sigmoid and softplus functions respectively. (3) Global soft attention mechanism: After graph convolution, a global soft attention pool mechanism is used to assign different weights to the embedding of each node: h gate and h are two fully connected neural networks, is a graph embedding for predicting polymer properties. (4) Uncertainty assessment: In the final stage, the evidence layer (uncertainty estimation layer) is used to quantify the prediction uncertainty and extract the uncertainty value from the distribution parameters. The prediction mean μ represents the model's prediction of the target attribute and reflects the expected value of the prediction. The prediction variance σ 2 Quantifies the uncertainty around the predicted mean expressed as a confidence interval. The target value y i Following the unknown mean and variance (μ,σ 2 )’s Gaussian distribution: (5) High-throughput screening: Using the trained polymer graph convolutional neural network model, the PI1M discovery dataset containing 1 million hypothetical polymers is subjected to high-throughput screening. The model quickly evaluates the predicted properties of polymers and screens out potential polymer candidates with target properties (such as low thermal conductivity, high thermal conductivity, etc.). This process can be completed in a short time, accelerating the discovery and development of polymer materials. (6) Identification of key functional groups: Through a global attention mechanism, the important functional groups related to specific properties in the polymer molecular structure are analyzed. Specifically, the model can identify which atoms or atomic groups have a significant impact on the target properties of the polymer, thereby providing a deep understanding of the structure-property relationship. This key functional group information helps to modify the molecular structure in a targeted manner when designing new materials to optimize their performance.

[0013] The above are preferred implementation cases of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method based on uncertainty-guided polymer graph convolutional neural network, characterized in that: The steps include: S1. Extract the polymer monomer information, convert the SMILES code representing the polymer monomer molecule into Morgan fingerprint, and transmit it to the polymer graph convolutional neural network to capture the structural characteristics of the molecule; S2. The node and edge feature vectors of the molecule are aggregated through convolution and embedded with different weights by the global soft attention mechanism to form the feature vector of the polymer monomer molecular graph structure to capture the topological relationship within the polymer monomer molecule; S3. Input the feature vector of the graph structure into the evidence layer to predict the properties of the polymer and provide uncertainty estimates; S4. Select data samples through uncertainty-guided active learning methods to improve data utilization and model training efficiency; S5. Use the trained polymer graph convolutional neural network model to perform high-throughput screening of a large number of polymers to identify potential candidate polymers with specific properties; S6. Identify important functional groups within polymer monomer molecules by focusing on the global mechanism and associate them with specific material properties to provide physical mechanistic insights.

2. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S1, the Morgan fingerprint is a fixed-length binary vector that is input into a neural network for unified input processing and batch calculation to capture the local and global structural features of the molecule.

3. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S2, the graph representation is achieved by describing each atom and its interconnected chemical bonds as nodes and edges respectively, and dynamically embedding the weights of each node and edge through a global soft attention mechanism.

4. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S3, in addition to property prediction, the evidence layer is also used to quantify the uncertainty in the model prediction and show the confidence of the model prediction under different circumstances.

5. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S4, the uncertainty-guided active learning method comprises the following steps: a. Use the initial model to estimate the uncertainty of the data sample; b. Select data samples for training based on the level of uncertainty; c. Update the model weights and repeat the selection and training steps until the predetermined amount of training data is reached.

6. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S5, the high-throughput screening is performed on a single GPU using a lightweight (the model occupies 60Kb of storage space) polymer graph convolutional neural network model to screen candidate polymers with target properties.

7. The method of uncertainty-guided polymer graph convolutional neural network according to claim 1, characterized in that: In S6, the contribution of each functional group to the material properties is dynamically evaluated through the global attention mechanism of the neural network, revealing the intrinsic correlation between molecular structure and properties.

Images