CN114818948B - Data-mechanism driven material attribute prediction method of graph neural network - Google Patents

Abstract

The invention discloses a data-mechanism driven material attribute prediction method of a graph neural network, which comprises the following steps: s1, obtaining descriptor characteristics and graph structures of molecules of a material to be predicted; s2, screening out a final feature descriptor by utilizing a feature engineering; s3, extracting molecular diagram features of different layers by using a diagram volume and a diagram attention network; s4, fusing the molecular diagram features and the descriptor features by using a feature fusion layer; s5, a correction module is used for better fusing the calculated value and the experimental value; the calculated value is a numerical value obtained by first-nature principle simulation calculation, and the experimental value is an actual material attribute measured by an experiment; and S6, fusing the calculated value of the mechanism driving model and the deep learning data driving model for model reasoning, and outputting the numerical value of the prediction attribute. The invention integrates the descriptor characteristics and the graph structure characteristics of the molecules and overcomes the problems that the graph structure data information is incomplete and the descriptor characteristics ignore the molecular attributes.

Description

A Data-Mechanism-Driven Material Property Prediction Method Based on Graph Neural Network

technical field

The invention relates to the technical fields of material discovery and graph neural network, in particular to a data-mechanism-driven material attribute prediction method of graph neural network.

Background technique

Molecular materials are widely used in medical and health, food, daily chemical and other fields. Therefore, accelerating the discovery of new molecular materials is of great significance to promote the development of science and society. Currently, the study of molecular materials is very time-consuming and requires a lot of effort to determine certain target properties and optimize the synthesis conditions of molecules. Theoretical high-throughput computational methods are often used to predict the properties of molecules. Such mechanism-driven computational models with plausible explanations can effectively accelerate the discovery of new materials. However, the mechanism-driven computational model is a theoretical model with parameter simplifications. It ignores the effects of material imperfections, real environment, facilities, researcher skills, etc. These factors may lead to inaccuracies in forecasts.

In recent years, big data-driven AI methods have been widely used in fields such as computer vision, natural language processing, medicine, and transportation. Due to the powerful nonlinear capability and feasibility of molecular big data, the prediction of material properties based on machine learning and deep learning has attracted extensive attention from researchers. Currently, there are two main aspects of AI approaches in materials. One is descriptor-based machine learning prediction, which needs to find descriptors with a strong correlation with the target attribute; the other is an end-to-end deep learning model based on graph neural network, which is an As input, the neural network can extract abstract information from the molecular graph structure and map to the target attributes. However, graph neural networks suffer from the same problems as other machine learning methods, namely lack of generalization and easy to reach the limit of training data. Especially for the discovery of new materials, the predictions of deep learning methods may not be accurate. When a real molecule is abstracted into a graph structure, it will lose part of its three-dimensional structure information and extranuclear electronic information. And it can lead to inaccurate predictions of outcomes. Therefore, fusing descriptor features with graph-structured features can not only solve the problem of descriptors lacking graph structure information, but also solve the problem of poor generalization of graph structures.

Contents of the invention

The invention provides a data-mechanism-driven material attribute prediction method of a graph neural network to solve the problems that the mechanism-driven calculation model ignores material attributes and the deep learning network has low generalization performance. By inputting descriptor features and graph structure features into a deep learning network for training, and using a mechanism-driven model to adjust the output of deep learning, the accuracy of molecular property prediction is improved.

In order to solve the problems of the technologies described above, the present invention provides the following technical solutions:

The present invention provides a data-mechanism-driven material property prediction method of a graph neural network, including:

S1, obtain the descriptor features and graph structure of the material molecules to be predicted;

S2, using feature engineering to screen out the final feature descriptor;

S3, using graph convolution and graph attention networks to extract molecular graph features at different levels;

S4, using the feature fusion layer to fuse molecular graph features and descriptor features;

S5, using the correction module to better integrate the calculated value and the experimental value; wherein, the calculated value is a value obtained by first-principle simulation calculation, and the experimental value is an actual material property measured experimentally;

In S6, the calculated value of the mechanism-driven model is fused with the deep learning data-driven model for model reasoning, and the value of the predicted attribute is output.

Further, in the S1, the descriptor feature and graph structure of the material molecule to be predicted are obtained, including:

From the PubChem website, the json file with graph structure information and experimental data is obtained through the Restful API interface provided by the website, where the graph structure information includes the properties of atoms and keys. The corresponding descriptor features of the molecules were collected online using the open source cheminformatics software rdkit. Molecular descriptors cover the basic properties, electronic properties, and topological properties of molecules.

Further, in said S2, feature engineering is used to screen out the final feature descriptors, including:

For the initial feature description, polynomial features are constructed to perform feature screening to reduce the difficulty of data fitting. The correlation of polynomial features is sorted using the Pearson correlation coefficient and the maximum information coefficient to filter out the final feature descriptors.

Further, the polynomial features include:

The feature descriptor can be expressed as X={x ₁ , x ₂ , x ₃ , . . . , x _k }, where k represents the dimension of the feature. Use x ² , x ^1/2 , ln(1+x) to perform feature polynomial transformation, and the constructed polynomial feature is

Further, in the S3, graph convolution and graph attention networks are used to extract molecular graph features at different levels, including:

Construct feature extraction layers with different numbers of convolutions to extract molecular features at different levels, shallow graph convolutions to obtain local feature information, and deep graph convolutions to obtain global long-range feature information. Graph convolution can gather the neighbor information of atoms in the molecule. The more layers of graph convolution, the wider the feature information that can be obtained. Graph attention can adjust the weights of neighbor nodes around each atom. The input of the graph network is a triplet {V,E,A}, where V represents the characteristic matrix of atoms constituting the molecule, E represents the characteristic matrix of edges, and A represents the adjacency matrix of the graph. The specific feature pyramid construction method is as follows:

X ₀ =GCN(V,E,A) (1)

X ₁ =GAT(X ₀ ) (2)

X ₂ =GAT(GCN(GCN(X ₁ ))) (3)

X ₃ =GAT(GCN(GCN(GCN(GCN(X ₂ ))))) (4)

X _Out ＝Set2Set(mean(X ₁ +X ₂ +X ₃ )) (5)

Among them, GCN represents the graph convolutional network layer, GAT represents the graph attention network layer; X ₀ represents the initial features obtained through graph convolution, X ₁ , X ₂ , and X ₃ represent the features obtained through different layers of graph convolutional networks. Features; mean means to calculate the average value of features, and Set2Set means to read out the global pool.

Further, in said S4, the feature fusion layer is used to fuse the molecular graph feature and the descriptor feature, including:

The feature fusion layer fuses the descriptor features obtained through feature engineering screening with the molecular graph features obtained through graph neural network feature extraction. Descriptor features are modeled based on electron and peripheral structures, and these calculation methods contain a large amount of three-dimensional information and interaction information of extranuclear electrons, which can be used to supplement the information loss problem of graph structures. The specific feature fusion method is as follows:

X _Cat ＝Concat((X _Out ,X _f )) (6)

X _fusion = ReLU(FC(X _Cat )) (7)

X _Out represents graph structure feature information, X _Cat represents feature descriptor information; Concat represents splicing features according to the dimension of features; FC represents fully connected layer; ReLU represents nonlinear activation function; X _fusion represents fused features.

Further, in the S5, the correction module is used to better integrate the calculated value and the experimental value; wherein, the calculated value is a value obtained by first-principles simulation calculation, and the experimental value is measured experimentally Actual material properties, including:

In order to better integrate the calculated value and the experimental value, the experimental value is used to construct the predicted label L of deep learning. The deep learning prediction label L not only considers the prediction results of the deep learning model, but also considers the real experimental data. The specific prediction label creation method is as follows:

L=F(L _c ,L _e )=αL _c +βL _e (α+β=1) (8)

Among them, L _c is the output of the deep learning network, which is trained by the calculation results of the theoretical model, L _e represents the real experimental data, α and β are the weighting coefficients of the calculation results and experimental data; bz represents the batch size during the model training process Size, Loss is the amount of loss.

Further, in said S6, the calculated value of the mechanism-driven model is fused with the deep learning data-driven model for model reasoning, and the value of the predicted attribute is output, including:

The mechanism-driven model calculates the attribute values of molecules through formulas, and the results produced by the mechanism-driven model are credible. Modulate the deep learning model through the mechanism-driven model to enhance the accuracy of the deep learning model output. The specific model output calculation method is as follows:

_OD represents the prediction of the deep learning network model, _OM represents the result of the mechanism model, and represents the final output; ε ₁ and ε ₂ represent the modulation coefficients of the mechanism model and the deep learning model, respectively, where ε ₁ +ε ₂ =1.

The beneficial effects brought by the technical solution provided by the present invention at least include:

The above technical scheme of the present invention provides an initial molecular descriptor feature and a graph structure acquisition method; using feature engineering to select the most relevant descriptor feature from the initial molecular descriptor feature; using graph convolution and graph attention network Extract molecular graph features at different levels; use the feature fusion layer to fuse molecular graph features and descriptor features; use the correction module to better integrate calculated values and experimental values; combine the calculated values of the mechanism-driven model with deep learning data The driving model fusion is used for model reasoning and outputs the value of predicted properties; the data calculated by the mechanism model and the features predicted by the deep learning model are combined and complement each other, improving the accuracy of molecular property prediction.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Figure 1 is a schematic diagram of the execution flow of the data-mechanism-driven material property prediction method of the graph neural network provided by the embodiment of the present invention;

2 is a schematic diagram of the overall processing flow of the data-mechanism-driven material property prediction method of the graph neural network provided by the embodiment of the present invention;

Fig. 3 is a schematic diagram of obtaining the descriptor features and graph structure of the material molecule to be predicted provided by the embodiment of the present invention;

4 is a schematic diagram of a deep learning network structure provided by an embodiment of the present invention;

Fig. 5 is a schematic diagram of a correction module provided by an embodiment of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the implementation manner of the present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, an embodiment of the present invention provides a data-mechanism-driven material property prediction method of a graph neural network, which includes:

S1, obtain the descriptor features and graph structure of the material molecules to be predicted;

It should be noted that since molecular datasets usually only contain structural features of molecules without descriptor features, these feature descriptors can be a good supplement to the three-dimensional structure information of molecules. Therefore, in this example, in order to solve the problem that the graph structure features are not rich enough, the descriptor features of molecules are obtained for information complementation.

S2, using feature engineering to screen out the final feature descriptor;

It should be noted that this embodiment uses a feature engineering method to filter out the most relevant features. Specifically, x ² , x ^1/2 , ln(1+x) are used to perform feature polynomial transformation to construct polynomial features. Use the Pearson correlation coefficient and the maximum information coefficient to sort the correlation, and select the final feature descriptor.

S3, using graph convolution and graph attention networks to extract molecular graph features at different levels;

Specifically, in this embodiment, the method for extracting molecular graph features at different levels is as follows: constructing feature extraction layers with different numbers of convolutions to extract molecular features at different levels, shallow graph convolutions to obtain local feature information, and deep layers Graph convolution obtains global long-range feature information. The feature information of different levels is fused by splicing and input to the global pooling layer to obtain the final graph features. Graph convolution can gather the neighbor information of atoms in the molecule. The more layers of graph convolution, the wider the feature information that can be obtained. Graph attention can adjust the weights of neighbor nodes around each atom. The input of the graph network is a triplet {V,E,A}, where V represents the characteristic matrix of atoms constituting the molecule, E represents the characteristic matrix of edges, and A represents the adjacency matrix of the graph. The specific feature pyramid construction method is as follows:

X ₀ =GCN(V,E,A) (1)

X ₁ =GAT(X ₀ ) (2)

X ₂ =GAT(GCN(GCN(X ₁ ))) (3)

X ₃ =GAT(GCN(GCN(GCN(GCN(X ₂ ))))) (4)

X _Out ＝Set2Set(mean(X ₁ +X ₂ +X ₃ )) (5)

Among them, GCN represents the graph convolutional network layer, GAT represents the graph attention network layer; X ₀ represents the initial features obtained through graph convolution, X ₁ , X ₂ , and X ₃ represent the features obtained through different layers of graph convolutional networks. Features; mean means to calculate the average value of features, and Set2Set means to read out the global pool.

S4, using the feature fusion layer to fuse molecular graph features and descriptor features;

Specifically, in this embodiment, the feature fusion layer fuses the descriptor features obtained through feature engineering screening with the molecular graph features extracted through graph neural forgetting features. The two splice features according to the same dimension, and use fully connected layers and nonlinear activation functions to better fuse features. Descriptor features are modeled based on electron and peripheral structures, and these calculation methods contain a large amount of three-dimensional information and interaction information of extranuclear electrons, which can be used to supplement the information loss problem of graph structures. The specific feature fusion method is as follows:

X _Cat ＝Concat((X _Out ,X _f )) (6)

X _fusion = ReLU(FC(X _Cat )) (7)

X _Out represents graph structure feature information, X _Cat represents feature descriptor information; Concat represents splicing features according to the dimension of features; FC represents fully connected layer; ReLU represents nonlinear activation function; X _fusion represents fused features.

S5, using the correction module to better integrate the calculated value and the experimental value; wherein, the calculated value is a value obtained by first-principle simulation calculation, and the experimental value is an actual material property measured experimentally;

Specifically, in this embodiment, in order to better integrate the calculated value and the experimental value, the experimental value is used to construct the predicted label L of deep learning. The deep learning prediction label L not only considers the prediction results of the deep learning model, but also considers the real experimental data. The specific prediction label creation method is as follows:

L=F(L _c ,L _e )=αL _c +βL _e (α+β=1) (8)

Among them, L _c is the output of the deep learning network, which is trained by the calculation results of the theoretical model, L _e represents the real experimental data, α and β are the weighting coefficients of the calculation results and experimental data; bz represents the batch size during the model training process Size, Loss is the amount of loss.

S6, merging the calculated value of the mechanism-driven model with the deep learning data-driven model for model reasoning, and outputting the value of the predicted attribute;

Specifically, in this embodiment, the mechanism-driven model calculates the property values of molecules through formulas, and the results generated by the mechanism-driven model are credible. Modulate the deep learning model through the mechanism-driven model to enhance the accuracy of the deep learning model output. The specific model output calculation method is as follows:

_OD represents the prediction of the deep learning network model, _OM represents the result of the mechanism model, and represents the final output; ε ₁ and ε ₂ represent the modulation coefficients of the mechanism model and the deep learning model, respectively, where ε ₁ +ε ₂ =1.

Further, the network structure of the network model adopted by the data-mechanism-driven material property prediction method of the graph neural network in this embodiment is shown in FIG. 2 .

Embodiment one

In this embodiment, the collected molecular data set is used to verify the effect of the data-mechanism-driven material property prediction method of the graph neural network. The molecular data set includes 4208 pieces of data, and each molecule includes: graph structure, descriptor, experimental value and calculated value. The graph structure includes the characteristics of the atoms in the graph and the characteristics of the bonds. Descriptors include: relative molecular mass, relative mass of heavy atoms, number of amino and hydroxyl groups, number of nitro groups, number of hydrogen acceptors, number of hydrogen donors, number of rotatable bonds, number of valence electrons, relative molecular mass, relative weight of heavy atoms mass, number of amino and hydroxyl groups. The collection method of the molecular dataset is shown in Fig. 3.

The batch size of this model training is 32, and the Adam optimizer is used for parameter optimization and its initial learning rate is set to 0.001, which is then reduced to 0.0001 according to the results of the verification set. The entire training set is trained for a total of 120 epochs, each with a batch size of 32. The data set is divided into three parts: training set, verification set and test set, with a ratio of 8:1:1. The evaluation index is mean absolute error.

Table 1 Experimental data (graph structure and descriptors) and their interpretation

The specific implementation steps are:

(1) Obtain the descriptor features and graph structure of the material molecules to be predicted. The molecular feature collection method is shown in the figure, and the graph structure information and experimental data are obtained from the PubChem website. Molecular descriptor features were collected using the open source cheminformatics software rdkit.

(2) Use the method of feature engineering to screen molecular descriptor features.

(3) Graph convolution and graph attention networks with different layers are used to extract structural feature information of molecular graphs.

(4) Fusing the descriptor features with the output of the graph neural network. Use the correction module to use the experimental values to construct more accurate labels to guide the training of the graph neural network.

(5) The calculated value of the mechanism-driven model is fused with the deep learning data-driven model for model reasoning, and the value of the predicted attribute is output.

Further, the data-mechanism-driven material property prediction method of the graph neural network in this embodiment is denoted as MD-GNN, and the detailed structure of the network is shown in FIG. 4 . In order to prove the effectiveness of the feature fusion method proposed in this example in improving the accuracy of molecular attribute prediction, three sets of experiments were set up: only use graph structure for attribute prediction, and the models used include ene-s2s, GAT, GraphSage and SchNet , using only molecular descriptors for attribute prediction, the traditional machine learning models used include random forests, gradient boosting decision trees and multi-layer perceptrons, the MD-GNN model that combines the two, when verifying the accuracy of feature fusion, for Control the influencing factors, do not use the correction module, and directly use the calculated value as the label for training.

Table 2 Prediction accuracy of roughness category from different angles (%)

Table 2 shows the prediction accuracy of roughness categories from different angles. Using graph structure + descriptor features is better than using graph structure alone or using descriptor features. The fusion model reduces the loss by 0.151 compared to the ene-s2s model alone, reduces the loss by 0.156 compared to the GAT model, reduces the loss by 0.175 compared to the GraphSages model, and reduces the loss by 0.158 compared to the SchNet model. At the same time, by comparing the use of only graph structure features and only descriptor features, it can be found that the loss of graph structure features is lower than that of description features, indicating that graph structure features are more important for molecular structures.

To illustrate the effectiveness of the correction module, comparative experiments are carried out on the proposed MD-GNN model with ene-s2s, GAT, GraphSage and SchNet. In the calibration block, the labels consist of experimental and calculated data. Then, the mechanism-based model is fused into the AI model via the M-D fusion block to condition model predictions. Compared with ene-s2s, GAT, GraphSage and SchNet, Table 3 lists the comparison of the experimental results with and without the correction module.

Table 3 Comparison of experimental results between adding correction module and not adding correction module

In order to observe the performance of the correction block, in Table 3, ene-s2s, GAT, GraphSage and SchNet are compared with the uncorrected correction block, and in the uncorrected case, the calculated and experimental values are used as labels, respectively , and calculated the error between the model output and the experimental value on the test set. The model is only marked by the calculated value, and there is a large error between the actual experimental value. The error is also high when using experimental values as labels, because the data in the experiment is not completely clean, making it difficult for the model to fit the distribution.

The MAEs in Table 4 show that combining mechanism-driven data is crucial for predicting real experimental properties, and the errors are all reduced after combining the proposed correction blocks.

To sum up, the method of this embodiment fuses molecular descriptor features and graph structure features into the network model to predict the properties of molecules. At the same time, in order to better improve the feature learning ability of the graph neural network, a correction module is introduced to narrow the gap between the calculated value and the experimental value.

It should also be noted that in this article, relational terms such as first and second etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations Any such actual relationship or order exists between. The term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or end-equipment comprising a set of elements includes not only those elements but also items not expressly listed other elements, or also include elements inherent in such a process, method, article, or end-equipment. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising said element.

Finally, it should be noted that the above description is a preferred embodiment of the present invention, and it should be pointed out that although the preferred embodiment of the present invention has been described, for those skilled in the art, once the basic creative concepts of the present invention are understood , under the premise of not departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. Therefore, the appended claims are intended to be construed to cover the preferred embodiment and all changes and modifications which fall within the scope of the embodiments of the present invention.

Claims (4)

1. A data-mechanism driven material property prediction method for a graph neural network, comprising:

s1, obtaining descriptor characteristics and graph structure information of a material molecule to be predicted, wherein json files with graph structure information and experimental data are obtained through a website, the graph structure information comprises atomic attributes and bond attributes, the descriptor characteristics corresponding to the molecule are collected on line by using open source chemical information software rdkit, and the molecule descriptor covers basic properties, electronic properties and topological properties of the molecule;

s2, screening out a final feature descriptor by utilizing feature engineering, wherein for initial feature description, polynomial features are constructed for feature screening, correlation degrees of the polynomial features are ranked by using Pearson correlation coefficients and maximum information coefficients, and the final feature descriptor is screened out;

the polynomial feature comprising:

the feature descriptor is represented by X = { X = ₁ ,x ₂ ,x ₃ ,...,x _k H, k denotes the dimension of the feature, x denotes the descriptor feature, using x ² ,x ^1/2 Ln (1 + x) is subjected to characteristic polynomial transformation, and the constructed polynomial is characterized by

S3, extracting molecular diagram features of different layers by using a diagram volume and a diagram attention network;

s4, fusing the molecular diagram features with the feature descriptors by using a feature fusion layer;

s5, fusing the calculated value and the experimental value by using a correction module; the calculated value is a numerical value obtained by first-nature principle simulation calculation, and the experimental value is an actual material attribute measured by an experiment;

s6, fusing a calculated value of the mechanism driving model and the deep learning data driving model for model reasoning, and outputting a numerical value of the prediction attribute;

the mechanism driving model calculates the attribute value of the molecule through a formula, the result generated by the mechanism driving model is interpretable and credible, the deep learning model is modulated through the mechanism driving model, the accuracy of the output of the deep learning model is enhanced, and the model output calculation mode is as follows:

O _D representing predictions of a deep-learning network model, O _M The results of the mechanism model are shown, representing the final output;ε ₁ ，ε ₂ represents the modulation coefficients of the mechanism model and the deep learning model, respectively, where ₁ +ε ₂ ＝1。

2. The data-mechanism driven material property prediction method of the graph neural network as claimed in claim 1, wherein in S3, extracting molecular graph features of different levels by using graph convolution and graph attention network comprises:

the method comprises the steps that feature extraction layers with different convolution quantities are constructed to extract molecular features of different levels, shallow graph convolution obtains local feature information, deep graph convolution obtains global feature information in a long range, the graph convolution is used for gathering neighbor information of atoms in a molecule, the more the number of graph convolution layers is, the wider the obtained feature information is, the more the graph convolution layers are, the graph attention is used for adjusting the weight of neighbor nodes around each atom, the input of a graph network is a triple group { V, E, A }, wherein V represents a feature matrix of atoms forming the molecule, E represents a feature matrix of an edge, and A represents an adjacent matrix of the graph.

3. The data-mechanism driven material property prediction method of graph neural network of claim 1, wherein in S4, fusing molecular graph features with feature descriptors using a feature fusion layer, comprises:

the feature fusion layer fuses descriptor features obtained through feature engineering screening and molecular diagram features obtained through diagram neural network feature extraction, and the descriptor features are modeled based on electrons and peripheral structures.

4. The data-mechanism driven material property prediction method of graph neural network of claim 1, wherein in S5, a correction module is used to better fuse the calculated value and the experimental value; the calculation value is a numerical value obtained by first-nature principle simulation calculation, and the experimental value is an actual material attribute measured by an experiment, and comprises the following steps:

the experimental values are used to construct a prediction label L for deep learning.

Images