Dynamic prediction method for the structure of amorphous materials

The dynamic prediction model is constructed through graph neural network and interatomic potential model, which solves the problem of change prediction of amorphous material structure under environmental disturbance, and realizes high-precision dynamic prediction of amorphous material structure, which promotes amorphous silicon, alloys and Application of fiber optic materials.

CN114067929BActive Publication Date: 2024-11-26CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111349530.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-15
Publication Date
2024-11-26
Estimated Expiration
2041-11-15

AI Technical Summary

Technical Problem

The existing amorphous material structure prediction methods cannot effectively consider environmental disturbances, resulting in low prediction accuracy and efficiency, and cannot be directly applied to the study of dynamic structural changes of amorphous material.

Method used

Graph neural network is used to combine Lammps software and interatomic potential model, and high-dimensional vector embedding is carried out by constructing graph models, dynamic prediction models are trained, and the impact of environmental changes on the amorphous material structure is considered, and the aggregated information quality is optimized to improve prediction accuracy.

Benefits of technology

It realizes dynamic prediction of amorphous material structure under environmental changing conditions, improves the reliability and practicality of prediction, and can be applied to the research and development of amorphous silicon, amorphous alloys and optical fiber structures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114067929B_ABST
    Figure CN114067929B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for dynamic prediction of amorphous material structure, including obtaining crystal structure information of target amorphous material; obtaining structure information of target amorphous material at several consecutive moments and modeling; embedding high-dimensional vectors into graph models to obtain input vectors; training graph neural networks with input vectors to obtain trained dynamic prediction models; and using dynamic prediction models to dynamically predict amorphous material structures. The method of the present invention introduces weights based on distance information to aggregate information of edges and nodes, optimizes the aggregated information quality, and improves prediction accuracy; establishes a coordinate system to describe the position of each particle, and the final output of the model is also the atomic coordinates obtained on this basis, so that the representation of the material structure is more accurate and the model can better predict; therefore, the method of the present invention can not only predict the material structure based on changes in external conditions, but also has higher reliability and better practicality.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of computational chemistry, and in particular relates to a method for dynamically predicting the structure of an amorphous material. Background Art

[0002] As the basic framework for building matter, the structure of matter is important information for in-depth exploration of the various physical and chemical properties of matter. To this end, in recent years, many researchers have begun to try to use various methods to determine the structure of matter, and then understand some special chemical properties or physical characteristics of matter.

[0003] When computers were not yet developed, the prediction of material structure was often achieved through a large number of statistical experimental results. However, this method is often affected by various external environments. If the environment is slightly disturbed, the material structure will undergo unpredictable changes, which leads to the fact that the accuracy and efficiency of the material structure predicted by experiments cannot be effectively guaranteed. With the development of computers and various search algorithms, many methods have gradually emerged to predict material structure through computer simulation and simulation of real environment. These methods will not be affected by environmental disturbances, which greatly improves the prediction accuracy.

[0004] Recently, some scholars have proposed and developed a software CALYPSO for material structure search: the software determines the structure of the material based on the chemical components contained in the material and the external environment (pressure, temperature and other influencing factors) in which the material is located. The purpose is to make theoretical structure predictions based on the particle swarm optimization algorithm and find the structure with the lowest global energy in a large group of structures of a certain material. Based on this software, researchers have successfully determined many stable material structures with excellent performance, and these material structures have also been confirmed by corresponding experiments. For example, a large number of metastable phases of Si and Ge were synthesized in high-pressure experiments, but many phase structures could not be determined experimentally. Using CALYPSO, the metastable tetragonal structures of Si and Ge discovered in the experiment were successfully determined. However, these material structures determined by computer simulation are often stable structures obtained by mining under a certain specific environment, and the changes in material structure caused by environmental disturbances have not been effectively solved. Even if this method is used to predict the material structure of the material in the changed environment, it still requires a certain amount of calculation time, which is relatively cumbersome.

[0005] With the continuous development of artificial intelligence in recent years, deep learning algorithms have gradually been applied to various fields in life. For example, T. Keck et al. used neural network algorithms to successfully predict the long-term dynamic evolution of glass systems from the positions of initial particles. R. Mandal et al. trained neural networks under supervised conditions and predicted the positions of force chains generated when a particle system deforms given an undeformed static structure.

[0006] The dynamic structure prediction of amorphous materials plays an important role in the fields of chemistry and chemical engineering. However, none of the above studies involve how the structure of amorphous materials will change when the environment is disturbed. Therefore, the current structural prediction research of amorphous materials cannot be directly applied to the dynamic structure prediction of amorphous materials, which seriously restricts the prediction research of amorphous material structure. Summary of the invention

[0007] The object of the present invention is to provide a method for dynamic prediction of amorphous material structure which can take into account the environment and dynamic disturbances and has high reliability and good practicality.

[0008] The method for dynamically predicting the structure of an amorphous material provided by the present invention comprises the following steps:

[0009] S1. Obtaining crystal structure information of the target amorphous material;

[0010] S2. According to the crystal structure information obtained in step S1, the structure information of the target amorphous material at several consecutive moments is obtained and modeled;

[0011] S3. Perform high-dimensional vector embedding on the graph model constructed in step S2 to obtain an input vector;

[0012] S4. Input the input vector obtained in step S3 into the graph neural network for training, thereby obtaining a trained dynamic prediction model;

[0013] S5. Using the dynamic prediction model obtained in step S4, dynamically predict the structure of the amorphous material.

[0014] The step S2, based on the crystal structure information obtained in step S1, obtains the structure information of the target amorphous material at a number of consecutive moments and performs modeling, which specifically includes the following steps:

[0015] A. According to the crystal structure information of the target amorphous material, Lammps software is used to obtain the material structure of the amorphous material at multiple times;

[0016] B. Select the interatomic potential model to determine the chemical bonds between atoms;

[0017] C. Consider the principles as nodes of the graph neural network, the chemical bonds between atoms as edges, and construct a coordinate system based on the crystal structure information of the target amorphous material;

[0018] D. Generate embeddings for node and edge features; for nodes, the features considered include the radius of each atom at the previous moment, the coordinates of the center of the circle of each atom at the previous moment, and the radius of each atom at the next moment, and the coordinates of the center of the circle of each atom at the next moment; for edges, the features considered include the chemical bonds, bond lengths, and bond angles between atoms at the previous moment, and the chemical bonds, bond lengths, and bond angles between atoms at the next moment;

[0019] E. Divide the data set obtained in step D into a training set, a validation set, and a test set according to the set ratio;

[0020] F. The input of the graph neural network is the graph structure G = (V, E, A), the initial feature matrix of the node and the initial feature matrix of the edge; where V is the set of nodes, E is the set of edges, and A is the adjacency matrix; the initial feature matrix of the node is in is the initial feature matrix of the i-th node, is the jth initial feature vector of the ith node, N is the number of nodes, num_node_features is the number of node features considered; the initial feature matrix of the edge is in is the initial feature matrix of the edge between node i and node j and is the kth eigenvector of the edge between node i and node j, and num_edge_features is the number of edge features considered.

[0021] The selected interatomic potential model described in step B, thereby determining the chemical bonds between atoms, specifically includes the following steps:

[0022] The interatomic potential model includes the repulsive potential model and the LJ potential model:

[0023] Repulsive potential model: If the distance between any two particles of a substance is less than the sum of the radii of the two particles, there is a chemical bond; if the distance between any two particles of a substance is greater than or equal to the sum of the radii of the two particles, there is no chemical bond;

[0024] LJ potential model: Using the formula Calculate the potential energy between any two particles that satisfies There is a repulsive force between the two particles, satisfying When r>2.5σ, there is attraction between the two particles; when r>2.5σ, there is no chemical bond between the two particles; ε is the potential energy parameter; σ is the particle radius; r is the distance between the centers of the two particles.

[0025] Step S3 performs high-dimensional vector embedding on the graph model constructed in step S2 to obtain an input vector, which specifically includes the following steps:

[0026] The following linear transformation is used to embed the initial feature matrix X of the input node and the initial feature matrix Y of the edge into a high-dimensional vector:

[0027]

[0028]

[0029] Where W x is the weight matrix corresponding to the node embedding and its dimension is h*col(X), b x is the bias of node feature embedding, W y is the weight matrix corresponding to the edge embedding and its dimension is h*col(Y), b y The offset of edge feature embedding, h is a hyperparameter, and col() is an operation to obtain the number of columns of the matrix.

[0030] Step S4, in which the input vector obtained in step S3 is input into the graph neural network for training, thereby obtaining a trained dynamic prediction model, specifically includes the following steps:

[0031] Use the following expression to update each node:

[0032]

[0033] In the formula is the embedded feature vector of node i to be updated in layer l; is the embedded feature vector of node i to be updated at the l-1 layer; the set of neighboring nodes of node i is S(i); nodes i and j are neighboring nodes, is the embedding feature vector of the edge between node i and node j at the l-1 layer; σ[x] = 1 / (1+e -x ) is a nonlinear activation function; b l-1 is the offset; W s and W o are all weight matrices obtained through training; f l-1 () is the aggregation function; is the aggregation function f l-1 () Aggregate the information obtained from the neighboring nodes and the edges between them, and w j is the weight parameter for establishing the distance between each node in the neighborhood and the node to be updated i and k is the number of neighboring nodes of node i, d ij is the distance between node i and node j;

[0034] After all nodes are updated, the environment change information C is sent to the graph neural network to obtain the node coordinates of each node at the next moment after the environment changes from the previous moment;

[0035] The distance between the real coordinates and the predicted coordinates is used as the loss function and optimized;

[0036] The weights of each layer are trained using stochastic gradient descent;

[0037] The embedding dimension h and the update number n l Carry out different combinations and training to obtain the optimal training configuration.

[0038] The dynamic prediction method for the structure of amorphous materials provided by the present invention introduces weights based on distance information to aggregate information of edges and nodes, optimizes the quality of aggregated information, and improves prediction accuracy; a coordinate system is established to describe the position of each particle, and the final output of the model is the atomic coordinates obtained on this basis, so that the representation of the material structure is more accurate and the model can perform better predictions; therefore, the method of the present invention can not only predict the material structure based on changes in external conditions, but also has higher reliability and better practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The figure is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0040] like Figure 1 The method flow chart of the method of the present invention is shown as follows: The method for dynamically predicting the structure of an amorphous material provided by the present invention comprises the following steps:

[0041] S1. Obtaining crystal structure information of the target amorphous material;

[0042] S2. According to the crystal structure information obtained in step S1, the structure information of the target amorphous material at several consecutive moments is obtained and modeled; specifically, the steps include:

[0043] A. According to the crystal structure information of the target amorphous material, Lammps software is used to obtain the material structure of the amorphous material at multiple times;

[0044] B. Select the interatomic potential model to determine the chemical bonds between atoms; specifically, the following steps are included:

[0045] The interatomic potential model includes the repulsive potential model and the LJ potential model:

[0046] Repulsive potential model: If the distance between any two particles of a substance is less than the sum of the radii of the two particles, there is a chemical bond; if the distance between any two particles of a substance is greater than or equal to the sum of the radii of the two particles, there is no chemical bond;

[0047] LJ potential model: Using the formula Calculate the potential energy between any two particles that satisfies There is a repulsive force between the two particles, satisfying There is attraction between the two particles when r>2.5σ, and there is no chemical bond between the two particles when r>2.5σ; ε is the potential energy parameter; σ is the particle radius; r is the distance between the centers of the two particles;

[0048] C. Consider the principles as nodes of the graph neural network, the chemical bonds between atoms as edges, and construct a coordinate system based on the crystal structure information of the target amorphous material;

[0049] D. Generate embeddings for node and edge features; for nodes, the features considered include the radius of each atom at the previous moment, the coordinates of the center of the circle of each atom at the previous moment, and the radius of each atom at the next moment, and the coordinates of the center of the circle of each atom at the next moment; for edges, the features considered include the chemical bonds, bond lengths, and bond angles between atoms at the previous moment, and the chemical bonds, bond lengths, and bond angles between atoms at the next moment; the bond angle is the distance between atoms and the direction of the unit vector on the relative distance;

[0050] E. Divide the data set obtained in step D into a training set, a validation set, and a test set according to a set ratio (e.g., 8:1:1);

[0051] F. The input of the graph neural network is the graph structure G = (V, E, A), the initial feature matrix of the node and the initial feature matrix of the edge; where V is the set of nodes, E is the set of edges, A is the adjacency matrix, if there is a chemical bond connection between node i and node j, then A(i, j) = A(j, i) = 1, otherwise A(i, j) = A(j, i) = 0; the initial feature matrix of the node is in is the initial feature matrix of the i-th node, is the jth initial feature vector of the ith node, N is the number of nodes, num_node_features is the number of node features considered; the initial feature matrix of the edge is in is the initial feature matrix of the edge between node i and node j and is the kth feature vector of the edge between node i and node j, and num_edge_features is the number of edge features considered;

[0052] S3. Perform high-dimensional vector embedding on the graph model constructed in step S2 to obtain an input vector; specifically, the steps include:

[0053] The following linear transformation is used to embed the initial feature matrix X of the input node and the initial feature matrix Y of the edge into a high-dimensional vector:

[0054]

[0055]

[0056] Where W x is the weight matrix corresponding to the node embedding and its dimension is h*col(X), b x is the bias of node feature embedding, W y is the weight matrix corresponding to the edge embedding and its dimension is h*col(Y), b y The offset of edge feature embedding, h is a hyperparameter, and col() is an operation to obtain the number of columns of the matrix;

[0057] S4. Input the input vector obtained in step S3 into the graph neural network for training, thereby obtaining a trained dynamic prediction model; specifically comprising the following steps:

[0058] Use the following expression to update each node:

[0059]

[0060] In the formula is the embedded feature vector of node i to be updated in layer l; is the embedded feature vector of node i to be updated at the l-1 layer; the set of neighboring nodes of node i is S(i); nodes i and j are neighboring nodes, is the embedding feature vector of the edge between node i and node j at the l-1 layer; σ[x] = 1 / (1+e -x ) is a nonlinear activation function; b l-1 is the offset; W s and W o are all weight matrices obtained through training; f l-1 () is the aggregation function; is the aggregation function f l-1 () Aggregate the information obtained from the neighboring nodes and the edges between them, and w j is the weight parameter for establishing the distance between each node in the neighborhood and the node to be updated i and k is the number of neighboring nodes of node i, d ij is the distance between node i and node j;

[0061] After all nodes are updated, the environment change information C is sent to the graph neural network to obtain the node coordinates of each node at the next moment after the environment changes from the previous moment;

[0062] The distance between the real coordinates and the predicted coordinates is used as the loss function and optimized; in specific implementation, if it is two-dimensional data, the loss function is If it is three-dimensional data, the loss function is Where (a i ,b i ) or (a i ,b i ,c i ) is the predicted coordinate of node i at the next moment, or is the real coordinate of node i at the next moment, and N is the number of input nodes;

[0063] The weights of each layer are trained using stochastic gradient descent;

[0064] The embedding dimension h and the update number n l Carry out different combinations and training to obtain the optimal training configuration;

[0065] S5. Using the dynamic prediction model obtained in step S4, dynamically predict the structure of the amorphous material.

[0066] The dynamic prediction method of the structure of amorphous materials provided by the present invention can better predict the structure of amorphous materials; after predicting the structure of amorphous materials by the method of the present invention, ① a certain amorphous silicon material that can replace high-purity silicon can be found, and by predicting, studying and researching its structural characteristics, it can be applied to the solar cell industry. Since the main part of the battery is amorphous silicon, its light absorption efficiency is higher than that of crystalline silicon, and a 0.5μm amorphous silicon film is sufficient to completely absorb the available sunlight. ② The method of the present invention can also further predict the material structure of amorphous alloys, and studies have shown that amorphous alloys can be used as catalysts in the chemical industry. Through the method of the present invention, a certain chemical structure suitable for amorphous alloys as catalysts can be found, so that the structure can greatly increase the contact area with the reactants, thereby facilitating the adsorption of reactants, and having more surface active centers to improve the catalytic efficiency. At the same time, since the amorphous alloy maintains a nearly ideal uniformity, there will be no structural factors that are not conducive to catalysis, such as grain boundaries, dislocations, etc. This makes the amorphous alloy catalyst have higher catalytic activity and selectivity. In today's increasingly severe energy crisis and environmental pollution, amorphous alloys as new catalytic materials have broad application prospects. ③ The method of the present invention can also predict the optical fiber structure made of amorphous materials. By studying the characteristics of existing optical fiber structures, it is possible to obtain an amorphous optical fiber structure that can expand the optical fiber communication capacity, improve the anti-interference ability of optical fiber, reduce the weight and hardness of optical fiber, facilitate optical fiber laying, and have good temperature stability, thereby improving the quality and speed of information transmission.

Claims

1. A method for dynamically predicting the structure of an amorphous material, comprising the following steps: S1. Obtaining crystal structure information of the target amorphous material; S2. According to the crystal structure information obtained in step S1, the structure information of the target amorphous material at several consecutive moments is obtained and modeled; specifically, the steps include: A. According to the crystal structure information of the target amorphous material, Lammps software is used to obtain the material structure of the amorphous material at multiple times; B. Select the interatomic potential model to determine the chemical bonds between atoms; C. Consider the principles as nodes of the graph neural network, the chemical bonds between atoms as edges, and construct a coordinate system based on the crystal structure information of the target amorphous material; D. Generate embeddings for node and edge features; for nodes, the features considered include the radius of each atom at the previous moment, the coordinates of the center of the circle of each atom at the previous moment, and the radius of each atom at the next moment, and the coordinates of the center of the circle of each atom at the next moment; for edges, the features considered include the chemical bonds, bond lengths, and bond angles between atoms at the previous moment, and the chemical bonds, bond lengths, and bond angles between atoms at the next moment; E. Divide the data set obtained in step D into a training set, a validation set, and a test set according to the set ratio; F. The input of the graph neural network is the graph structure G = (V, E, A), the initial feature matrix of the node and the initial feature matrix of the edge; where V is the set of nodes, E is the set of edges, and A is the adjacency matrix; the initial feature matrix of the node is in is the initial feature matrix of the i-th node, is the jth initial feature vector of the ith node, N is the number of nodes, num_node_features is the number of node features considered; the initial feature matrix of the edge is in is the initial feature matrix of the edge between node i and node j and is the kth feature vector of the edge between node i and node j, and num_edge_features is the number of edge features considered; S3. Perform high-dimensional vector embedding on the graph model constructed in step S2 to obtain an input vector; S4. Input the input vector obtained in step S3 into the graph neural network for training, thereby obtaining a trained dynamic prediction model; S5. Using the dynamic prediction model obtained in step S4, dynamically predict the structure of the amorphous material.

2. The method for dynamic prediction of amorphous material structure according to claim 1, characterized in that The selected interatomic potential model described in step B, thereby determining the chemical bonds between atoms, specifically includes the following steps: The interatomic potential model includes the repulsive potential model and the LJ potential model: Repulsive potential model: If the distance between any two particles of a substance is less than the sum of the radii of the two particles, there is a chemical bond; if the distance between any two particles of a substance is greater than or equal to the sum of the radii of the two particles, there is no chemical bond; LJ potential model: Using the formula Calculate the potential energy between any two particles that satisfies There is a repulsive force between the two particles, satisfying When r>2.5σ, there is attraction between the two particles; when r>2.5σ, there is no chemical bond between the two particles; ε is the potential energy parameter; σ is the particle radius; r is the distance between the centers of the two particles.

3. The method for dynamic prediction of amorphous material structure according to claim 1, characterized in that Step S3 performs high-dimensional vector embedding on the graph model constructed in step S2 to obtain an input vector, which specifically includes the following steps: The following linear transformation is used to embed the initial feature matrix X of the input node and the initial feature matrix Y of the edge into a high-dimensional vector: Where W x is the weight matrix corresponding to the node embedding and its dimension is h*col(X), b x is the bias of node feature embedding, W y is the weight matrix corresponding to the edge embedding and its dimension is h*col(Y), b y The offset of edge feature embedding, h is a hyperparameter, and col() is an operation to obtain the number of columns of the matrix.

4. The method for dynamic prediction of amorphous material structure according to claim 3, characterized in that Step S4, in which the input vector obtained in step S3 is input into the graph neural network for training, thereby obtaining a trained dynamic prediction model, specifically includes the following steps: Use the following expression to update each node: In the formula is the embedded feature vector of node i to be updated in layer l; is the embedded feature vector of node i to be updated at the l-1 layer; the set of neighboring nodes of node i is S(i); nodes i and j are neighboring nodes, is the embedded feature vector of the edge between node i and node j at the l-1 layer; σ[x]=1 / (1+e -x ) is a nonlinear activation function; b l-1 is the offset; W s and W o are all weight matrices obtained through training; f l-1 () is the aggregation function; is the aggregation function f l-1 () Aggregate the information obtained from the neighboring nodes and the edges between them, and w j is the weight parameter for establishing the distance between each node in the neighborhood and the node i to be updated and k is the number of neighboring nodes of node i, d ij is the distance between node i and node j; After all nodes are updated, the environment change information C is sent to the graph neural network to obtain the node coordinates of each node at the next moment after the environment changes from the previous moment; The distance between the real coordinates and the predicted coordinates is used as the loss function and optimized; The weights of each layer are trained using stochastic gradient descent; The embedding dimension h and the update number n l Carry out different combinations and training to obtain the optimal training configuration.