Method, system and storage medium for generating indefinite 3D structure based on graph neural network

Abstract

The present invention discloses a method and system for generating an indeterminate 3D structure based on a graph neural network. The method includes an indeterminate 3D crystal structure generation model training phase and a 3D structure crystal generation phase, wherein the indeterminate 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregression module, wherein the conditional variational autoencoder module is used to obtain a guide vector of a crystal sample, the asymmetric reconstruction module is used to obtain a key information vector of a crystal sample, and the conditional autoregression module is used to obtain a type prediction and position prediction of each atom of the crystal sample; the type prediction and position prediction are used to calculate errors with expected results, and the error is reversely conducted to train the model; the material data is input into the trained model, and the model obtains a prediction result of the indeterminate 3D crystal structure. The present invention is used for crystal material discovery, and the real data distribution can be learned through the model, and the generated results can prove the effectiveness of the model.

Description

Technical Field

[0001] The present invention relates to the field of crystal material discovery, and in particular to a method, system and storage medium for generating an indefinite 3D crystal structure based on a graph neural network. Background Art

[0002] Materials are the foundation for the development of all industries. From infrastructure to aerospace, they are inseparable from the use of high-performance materials. However, the exploration of new materials is often time-consuming and unstable. Rapidly developing new materials based on the different requirements for materials in various fields is of great significance to scientific research and social construction. In recent years, machine learning has been widely used in materials science, greatly accelerating the discovery and application of new materials. However, the mining and modeling of complex crystalline material data is not enough. Existing machine learning algorithms have not fully extracted the characteristics of existing material data, and it is difficult to meet the various needs of academia and industry for material exploration.

[0003] Crystalline materials are solid materials composed of crystalline substances. The atoms, ions, molecules or particle groups they contain have periodic arrangement rules. The discovery of new crystalline materials refers to finding stable material molecular structures that may meet the requirements under the requirements of specific material properties. Since crystalline materials can be regarded as 3D point sets arranged periodically in space, we can use graph structures and certain boundary conditions to represent crystalline materials, and then use graph neural network methods to model materials. Therefore, the generation of crystalline materials can be regarded as an indefinite 3D periodic structure generation problem based on graph neural networks.

[0004] The number of possible crystalline materials is so huge, in fact, that it has been estimated that the theoretical total number of materials could be as high as 10 100 There are 10 possible quaternary compounds alone. 13 However, traditional material discovery methods, such as material preparation experiments and first-principles calculations, are very inefficient and cannot meet the requirements of today's society for the speed of exploring new materials. Based on the existing large number of public databases, machine learning methods can greatly accelerate the material discovery process. Using machine learning models as the first screening threshold for material discovery can reduce a lot of experimental time, allowing valuable experimental resources to be concentrated on more promising materials. Therefore, machine learning has far-reaching significance for accelerating material exploration.

[0005] For existing crystal material discovery research, high-throughput screening frameworks based on traditional machine learning or deep learning are the most common. Their idea is to use the high-speed prediction characteristics of machine learning to input a large number of possible crystal material structures into the model, and conduct preliminary screening of the guessed materials through the prediction results. In particular, the widespread use of graph neural networks in this field since 2018 has enabled deep learning models to achieve or even exceed the accuracy of first-principles calculations in predicting material properties. Although the use of graph neural networks has achieved great success, most graph neural networks currently do not fully exploit the spatial geometric information of crystal material data. Therefore, in the application of crystal materials, graph neural networks still have great development potential. Secondly, the currently widely used high-throughput screening methods are subject to the conjecture of material structure. The materials to be screened are usually structures that are considered reasonable in people's common sense, which brings great limitations and makes it difficult to explore the entire material space. Therefore, it is quite necessary and challenging to explore a more reasonable generation framework. Summary of the invention

[0006] In response to the above problems, the present invention provides a method, system and storage medium for generating an indefinite 3D crystal structure based on a graph neural network, which are used to generate a 3D crystal structure of any crystal system and solve the limitations in the field of crystal material discovery.

[0007] In a first aspect of the present invention, a method for generating an indefinite 3D crystal structure based on a graph neural network is provided, the method comprising the following steps:

[0008] Indeterminate 3D crystal structure generation model training stage: The indeterminate 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module, wherein the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, the asymmetric reconstruction module is used to obtain the key information vector of the crystal sample, and the conditional autoregressive module is used to obtain the type prediction and position prediction of each atom of the crystal sample according to the input guidance vector and key information vector; the error is calculated by using the type prediction and position prediction and the expected result, and the error is reversely transmitted to train the indeterminate 3D crystal structure generation model;

[0009] 3D structure crystal generation stage: the material data is input into the trained indefinite 3D crystal structure generation model, and the conditional autoregressive model obtains the prediction result of the indefinite 3D crystal structure. If the prediction result of the indefinite 3D crystal structure meets the conditions for stopping generation, the generation is stopped. If the prediction result of the indefinite 3D crystal structure does not meet the conditions for stopping generation, the input material data is updated.

[0010] Furthermore, the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, and the specific steps include:

[0011] Let the conditional generation probability be p θ (x|y), x represents the real crystal data, y represents the input conditional variable, in the conditional variational autoencoder module, using p θ The variational lower bound of (x|y) is used as the optimization objective of the conditional variational autoencoder module:

[0012]

[0013] KL div represents the KL divergence, Represents the low-dimensional space vector z in the corresponding distribution The mathematical expectation of represents the conditional probability distribution of the low-dimensional space vector z, which is used to encode the crystal graph, p θ (z|y) represents the prior conditional distribution of z, and the conditional distribution p θ The realization of (x|y,z) is the key to the indefinite 3D crystal structure generation model, introducing the probability distribution of L with fixed dimensions The p θ (x|y,z) decomposes into:

[0014]

[0015] Among them, the real crystal data x is expressed as (R ≤n ,Z ≤n ,L),R ≤n ,Z ≤n Represent the spatial coordinates R of the first n atoms in the crystal data x ≤n and atomic number Z ≤n , r i ,z i Respectively represent the spatial coordinates and atomic number of the i-th atom to be generated, and the conditional distribution p θ (x|y,z) are all constrained by the hidden vector z and the input variable y, so that the guide vector Conditional distribution p θ (x|y,z) can be written as:

[0016]

[0017] Furthermore, the asymmetric reconstruction module is used to obtain the key information vector of the crystal sample. The specific method includes using the low-dimensional feature vector of the crystal sample to perform asymmetric reconstruction to obtain the key information vector of the crystal sample, wherein the low-dimensional feature vector of the crystal sample is part of the guidance vector.

[0018] Furthermore, the conditional autoregressive module is used to obtain the prediction of each atom type and atom position of the crystal sample according to the input guidance vector and key information vector. The specific steps include:

[0019] The implementation formula of the conditional autoregressive module is:

[0020]

[0021] Where R represents the spatial coordinate of the atom, Z represents the atomic number, L is the introduced fixed dimensional variable, and p(stop|R ≤n ,Z ≤n ,x',z g ) means that when R is satisfied ≤n ,Z ≤n ,x',z g In the case of , the probability that the conditional autoregressive module stops generating, x'={L,T n ,P d} represents the key information vector of the real crystal data x, represents the number of atoms in the crystal to be generated, k represents the total number of atomic types, P d represents the atomic position prediction p pos The weak constraint condition in p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x',z g ) is expressed as:

[0022] p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x′,z g )=p type *p pos

[0023] p type =p type (z i |R ≤i-1 ,Z ≤i-1 ,z g ,L,f(T n ,Z ≤i-1 ))

[0024] p pos =p pos (r i |R ≤i-1 ,Z ≤i-1 ,z i ,z g ,P d )

[0025] Among them, the atomic type prediction p type A strong constraint f(T n,Z ≤i-1 )=max(T n -Z ≤i-1 ,0), used to represent the remaining atoms that need to be generated, and predict p at the atomic position pos A weak constraint P is added d , b d represents the atomic position prediction p in the conditional autoregressive module pos The relative distance distribution between atoms in the .

[0026] A second aspect of the present invention provides a system for generating an indefinite 3D crystal structure based on a graph neural network, the system comprising:

[0027] An indeterminate 3D crystal structure generation model training unit is used to train an indeterminate 3D crystal structure generation model, wherein the indeterminate 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module, wherein the conditional variational autoencoder module is used to obtain a guide vector of a crystal sample, the asymmetric reconstruction module is used to obtain a key information vector of the crystal sample, and the conditional autoregressive module is used to obtain a type prediction and a position prediction of each atom of the crystal sample according to the input guide vector and the key information vector; the error is calculated by using the type prediction and the position prediction and the expected result, and the error is reversely transmitted to train the indeterminate 3D crystal structure generation model;

[0028] The 3D structure crystal generation unit is used to generate 3D structure crystals using material data. The material data is input into a trained indefinite 3D crystal structure generation model. The conditional autoregressive model obtains the prediction result of the indefinite 3D crystal structure. If the prediction result of the indefinite 3D crystal structure meets the conditions for stopping generation, the generation is stopped. If the prediction result of the indefinite 3D crystal structure does not meet the conditions for stopping generation, the input material data is updated.

[0029] The third aspect of the present invention provides an indefinite 3D crystal structure generation system based on a graph neural network, comprising: a processor; and a memory, wherein a computer executable program is stored in the memory, and when the computer executable program is executed by the processor, the above-mentioned indefinite 3D crystal structure generation method based on a graph neural network is executed.

[0030] In a fourth aspect of the present invention, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed by a processor, the processor executes the above-mentioned method for generating an indefinite 3D crystal structure based on a graph neural network.

[0031] In order to solve the problem of indefinite 3D structure generation based on graph neural network, the present invention proposes an indefinite 3D crystal structure generation model, which is suitable for crystal material generation and can model and search for crystal materials of any crystal system and any number of atoms. The entire model is realized by decomposing the complex generation model into a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module. By training the indefinite 3D crystal structure generation model of the present invention on the public database OQMD, it is proved that the model of the present invention can learn the real data distribution, and the conditional generation results show the effectiveness of the indefinite 3D crystal structure generation model of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 3D crystal structure generation model training process diagram in an embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of the structure of an indefinite 3D crystal structure generation system based on a graph neural network in an embodiment of the present invention;

[0034] Figure 3 is the architecture of the computer device in the embodiment of the present invention;

[0035] Figure 4 It is a crystal material structure diagram generated by three data sets in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to further explain the technical solution of the present invention in detail, this embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation method and specific steps.

[0037] The embodiments of the present invention provide the following embodiments for a method, system and storage medium for generating an indefinite 3D crystal structure based on a graph neural network:

[0038] Embodiment 1 based on the present invention

[0039] This embodiment is used to illustrate an indefinite 3D crystal structure generation method based on a graph neural network. The indefinite 3D crystal structure generation method based on a graph neural network includes two stages: an indefinite 3D crystal structure generation model training stage and a 3D structure crystal generation stage:

[0040] like Figure 1 FIG. 1 is a flow chart of the training phase of the indefinite 3D crystal structure generation model of the generation method according to an embodiment of the present invention.

[0041] The indefinite 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module, wherein the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, the asymmetric reconstruction module is used to obtain the key information vector of the crystal sample, and the conditional autoregressive module is used to obtain the type prediction and position prediction of each atom of the crystal sample according to the input guidance vector and key information vector; the error is calculated using the type prediction and position prediction and the expected result, and the error is reversely transmitted to train the indefinite 3D crystal structure generation model.

[0042] In the specific implementation process, the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample. The specific steps include:

[0043] Let the conditional generation probability be p θ (x|y), x represents the real crystal data, y represents the input conditional variable, in the conditional variational autoencoder module, using p θ The variational lower bound of (x|y) is used as the optimization objective of the conditional variational autoencoder module:

[0044]

[0045] KL div stands for KL divergence, which is used to describe the distance between two probability distributions. Represents the low-dimensional space vector z in the corresponding distribution The mathematical expectation of represents the conditional probability distribution of the low-dimensional space vector z, which is used to encode the crystal graph and can be implemented by any graph network. θ (z|y) represents the prior conditional distribution of z, which is defined as a normal distribution in the embodiment, and the conditional distribution p θ The realization of (x|y,z) is the key to generating models of indefinite 3D crystal structures, but the model is too complicated due to the nearly infinite combinations of atoms and various lattices in crystalline materials. θ (x|y,z) is implemented without any additional constraints, introducing a probability distribution of L with fixed dimension The p θ (x|y,z) decomposes into:

[0046]

[0047] Among them, the real crystal data x is expressed as (R ≤n ,Z ≤n ,L),R ≤n ,Z ≤n Represent the spatial coordinates R of the first n atoms in the crystal data x ≤n and atomic number Z ≤n , ri ,z i Represent the spatial coordinates and atomic number of the i-th atom to be generated, respectively. θ (x|y,z) decomposition, complex conditional distribution p θ (x|y,z) is decomposed into two realizable probability distributions, one is the probability distribution of L with fixed dimension Another is a probability distribution with variable dimensions represents the joint conditional distribution of atom types and positions in the lattice. Since atoms exist within the periodic boundaries defined by the lattice, After defining the lattice, we can start to define the components of the crystal. Note that both of the above conditional distributions are constrained by the hidden vector z and the input variable y. Because they work in a similar way to the guided generation process, the present invention defines the guided vector The above formula can be written as:

[0048]

[0049]

[0050] Furthermore, the asymmetric reconstruction module is used to obtain the key information vector of the crystal sample. The specific method includes using the low-dimensional feature vector of the crystal sample to perform asymmetric reconstruction to obtain the key information vector of the crystal sample, wherein the low-dimensional feature vector z of the crystal sample is a part of the guide vector and satisfies

[0051] In the specific implementation process, in order to make the probability of separation The meaning of is more generalized. as a separate reconstruction module to reconstruct a part of x to help the self-conditional autoregressive module p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,L,z g ) regression process. Therefore, the present invention modifies the variational lower bound of the conditional autoregressive model to:

[0052]

[0053] Where x' represents the key information vector of the real crystal data x, including the lattice. The reconstruction module of the conditional variational autoencoder module is finally divided into two parts and Plus two other conditional probability models and p θ (z|y), conditional generative model p θ(x|y) becomes realizable.

[0054] Furthermore, the conditional autoregressive module is used to obtain the prediction of each atom type and atom position of the crystal sample according to the input guidance vector and key information vector. The specific steps include: after introducing the guidance vector z g Under the premise of partially reconstructing the crystal x', the formula for implementing the conditional autoregressive module is:

[0055]

[0056] Where R represents the spatial coordinate of the atom, Z represents the atomic number, L is the introduced fixed dimensional variable, and p(stop|R ≤n ,Z ≤n ,x',z g ) means that when R is satisfied ≤n ,Z ≤n ,x',z g In the case of , the probability that the conditional autoregressive module stops generating, x'={L,T n ,P d} represents the key information vector of the real crystal data x, represents the number of atoms in the crystal to be generated, k represents the total number of atomic types, P d represents the atomic position prediction p pos The weak constraints in

[0057] p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x',z g ) is expressed as:

[0058] p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x′,z g )=p type *p pos

[0059] p type =p type (z i |R ≤i-1 ,Z ≤i-1 ,z g ,L,f(T n ,Z ≤i-1 ))

[0060] p pos =p pos (r i |R≤i-1 ,Z ≤i-1 ,z i ,z g ,P d )

[0061] Among them, the atomic type prediction p type A strong constraint f(T n ,Z ≤i-1 )=max(T n -Z ≤i-1 ,0), used to represent the remaining atoms that need to be generated, and predict p at the atomic position pos A weak constraint P is added d , b d The number of bins representing the discrete distances, i.e., the atomic position prediction p in the conditional autoregressive module pos The relative distance distribution between atoms in the .

[0062] During the implementation, we found that the conditional autoregressive module is very good at predicting the position of atoms, but it is not good enough in predicting the type of atoms. On the contrary, the variational autoencoder module is more suitable for reconstructing the type of atoms. Function f(T n ,Z ≤i-1 )=max(T n -Z ≤i-1 ,0) represents the remaining atoms that need to be generated, which gives the atom type prediction p type Generate a direct hint for a given atom type. The inventors have experimentally shown that the strong constraint f(T n ,Z ≤i-1 )=max(T n -Z ≤i-1 ,0) will improve the accuracy of atom type prediction during both training and generation. There are many ways to introduce these additional conditions into the network p type and p pos In the embodiment of the present invention, no limitation is imposed on the specific practice of the graph neural network, and only two output layers of the graph network are defined, namely output type Layers and output pos layer, introducing these parameters by aggregating them together in the two output layers.

[0063] p type =output type (V graph ,z g ,f(T n ,Z ≤i-1 ))=Softmax(relu(MLPf(T n ,Z ≤i-1))+relu(MLP(V graph ))+relu(MLP(z g )))

[0064] In the above formula, Softmax, relu, and MLP represent the corresponding neural network layers, V graph A graph structure representing the input material data.

[0065] output pos Have the same structure.

[0066] Furthermore, the coordinates of each atom in the material reconstructed in the conditional autoregressive module are expressed as discrete relative distances.

[0067] For the atomic position prediction p decomposed in the conditional autoregressive module pos , where the next atom r i The probability of the position is calculated by multiplying the distance likelihood between the possible position and all generated points, as shown in the following formula:

[0068]

[0069] Where α is the normalization constant, d ji Represents the distance between atom i and the already placed atom j.

[0070] Furthermore, the conditional variational autoencoder module, asymmetric reconstruction module and conditional autoregressive module of the indefinite 3D crystal structure generation model are generated separately, but trained at the same time, and the expected result is the crystal in the database. For the loss function of the indefinite 3D crystal structure generation model training, assuming that there are already n atoms in the input part of the crystal, the total loss function of sample i in the model can be written as:

[0071] L i =KL div (p type ||Z n+1 )+δL pos +μL rec +σL reg +ηL val

[0072]

[0073]

[0074]

[0075] L val =MSE(y||y')

[0076] in k represents the total number of atom types in the entire dataset, where b d Indicates the number of bins of discrete distances. During parallel training, the number of atoms in each crystal in each batch is not equal. The present invention uses the maximum number of atoms in the batch to construct the batch processing. And fill the blanks with 0. During the training process, use a mask to filter out the filled 0s. represents the distribution of the distance between the jth (j≤n) atom and the n+1th atom, smoothed by a Gaussian process. Due to the periodic boundary conditions, the distance d j,n+1 is the shortest distance between atom j and atom n+1 in the same or different unit cells. Reconstructing the lattice where b abc The number of bins representing the three discrete lengths of the crystal axes a, b, and c of the discrete lattice. The last bin covers the maximum length of the crystal axis in the database. αβγ The number of bins representing the three discrete crystal axes α, β, and γ. The last bin contains the largest possible angle, 180°. List of target atom types in A one-hot vector representing the type of the i-th atom. The number of all atoms in the crystal to be generated is represented as a vector In addition, δ, μ, σ, and η in the above formula are all set hyperparameters, which are used to adjust the weights of each part of the loss function.

[0077] 3D structure crystal generation stage: input the material data into the trained indeterminate 3D crystal structure generation model, and the conditional autoregressive model obtains the prediction result of the indeterminate 3D crystal structure. If the prediction result of the indeterminate 3D crystal structure meets the condition p (stop|R ≤n ,Z ≤n ,x',z g ), the generation is stopped. If the prediction result of the indefinite 3D crystal structure does not meet the conditions for stopping the generation, the input material data is updated.

[0078] Embodiment 2 of the present invention

[0079] Below, refer to Figure 2The system corresponding to the method according to the first embodiment of the present disclosure is described, and a system for generating an indefinite 3D crystal structure based on a graph neural network is described. The system 100 includes: an indefinite 3D crystal structure generation model training unit 101, which is used to train an indefinite 3D crystal structure generation model. The indefinite 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module, and a conditional autoregression module, wherein the conditional variational autoencoder module is used to obtain a guide vector of a crystal sample, the asymmetric reconstruction module is used to obtain a key information vector of a crystal sample, and the conditional autoregression module is used to obtain a crystal sample according to an input guide vector and a key information vector. Type prediction and position prediction of each atom in the volume sample; using the type prediction and position prediction to calculate the error with the expected result, and reversely conduct the error to train the indefinite 3D crystal structure generation model; 3D structure crystal generation unit 102, used to generate 3D structure crystals using material data, input material data into the trained indefinite 3D crystal structure generation model, and obtain the prediction result of the indefinite 3D crystal structure by the conditional autoregressive model. If the prediction result of the indefinite 3D crystal structure meets the condition for stopping generation, then stop generating, if the prediction result of the indefinite 3D crystal structure does not meet the condition for stopping generation, update the input material data. In addition to the above two units, the system 100 may also include other components, however, since these components are not related to the content of the embodiments of the present disclosure, their illustration and description are omitted here.

[0080] The specific working process of a graph neural network-based indefinite 3D crystal structure generation system 100 refers to the description of the above-mentioned graph neural network-based indefinite 3D crystal structure generation method embodiment 1, which will not be repeated here.

[0081] Embodiment 3 based on the present invention

[0082] The device according to the embodiment of the present invention can also be used by Figure 3 The architecture of the computing device shown is implemented. Figure 3 The architecture of the computing device is shown. Figure 3 As shown, computer system 201, system bus 203, one or more CPUs 204, input / output 202, memory 205, etc. Memory 205 can store various data or files used for computer processing and / or communication and program instructions including the methods of Embodiments 1-2 executed by the CPU. Figure 3 The architecture shown is only exemplary and can be adjusted according to actual needs when implementing different devices. Figure 3 One or more components in .

[0083] Embodiment 4 based on the present invention

[0084] The embodiment of the present invention can also be implemented as a computer-readable storage medium. According to the computer-readable storage medium of embodiment 4, computer-readable instructions are stored thereon. When the computer-readable instructions are executed by a processor, the indefinite 3D crystal structure generation method based on a graph neural network according to embodiments 1 to 3 of the present invention described with reference to the above drawings can be executed.

[0085] The embodiments of the present invention are directed to the above-mentioned indefinite 3D crystal structure generation method, system and storage medium based on graph neural network. In order to construct the indefinite 3D crystal structure generation model in the above four embodiments, a large amount of crystal material data is collected, each of which contains a crystal structure description file and the corresponding physical properties of the crystal. These data mainly come from the following three sources.

[0086] (1) Open Quantum Material Database (OQMD): It records data on more than 561K crystalline materials, most of which are calculated from first-principles calculations, and a small number are experimentally measured results.

[0087] (2) Material Project (MP): It records data on more than 144K crystalline materials, most of which are calculated results obtained by first-principles calculations, and a small number are experimentally measured results.

[0088] (3) Inorganic Crystal Structures Database (ICSD): It contains data on more than 250K crystal materials, and all materials have been experimentally confirmed to exist in reality. Since the data obtained by first-principles (DFT) calculations and experiments are recognized by the physics community, they are all valid experimental data.

[0089] Using Examples 1-4 and the above three data sets, three generated crystal material structure diagrams are obtained as follows: Figure 4 As shown, based on the above-mentioned indefinite 3D crystal structure generation method, system and storage medium based on graph neural network, the present invention proposes an indefinite 3D crystal structure generation model to solve the problem of indefinite 3D structure generation based on graph neural network. The model is suitable for crystal material generation and can model and search for crystal materials of any crystal system and any number of atoms. The entire model is realized by decomposing the complex generation model into a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module. By training the indefinite 3D crystal structure generation model of the present invention on the public database OQMD, it is proved that the model of the present invention can learn the real data distribution, and the conditional generation results show the effectiveness of the indefinite 3D crystal structure generation model of the present invention.

[0090] In this document, the terms "comprises," "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that includes a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such step or method.

[0091] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for generating an indefinite 3D crystal structure based on a graph neural network, characterized in that: The method comprises the following steps: Indeterminate 3D crystal structure generation model training stage: The indeterminate 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module, wherein the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, the asymmetric reconstruction module is used to obtain the key information vector of the crystal sample, and the conditional autoregressive module is used to obtain the type prediction and position prediction of each atom of the crystal sample according to the input guidance vector and key information vector; the error is calculated by using the type prediction and position prediction and the expected result, and the error is reversely transmitted to train the indeterminate 3D crystal structure generation model; 3D structure crystal generation stage: input material data into the trained indefinite 3D crystal structure generation model, and the conditional autoregressive model obtains the prediction result of the indefinite 3D crystal structure. If the prediction result of the indefinite 3D crystal structure meets the conditions for stopping generation, the generation is stopped. If the prediction result of the indefinite 3D crystal structure does not meet the conditions for stopping generation, the input material data is updated. Among them, the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, and the specific steps include: Let the conditional generation probability be p θ (x|y), x represents the real crystal data, y represents the input conditional variable, in the conditional variational autoencoder module, using p θ The variational lower bound of (x|y) is used as the optimization objective of the conditional variational autoencoder module: KL div represents the KL divergence, Represents the low-dimensional space vector z in the corresponding distribution The mathematical expectation of represents the conditional probability distribution of the low-dimensional space vector z, which is used to encode the crystal graph, p θ (z|y) represents the prior conditional distribution of z, and the conditional distribution p θ The realization of (x|y,z) is the key to the indefinite 3D crystal structure generation model, introducing the probability distribution of L with fixed dimensions The p θ (x|y,z) decomposes into: Among them, the real crystal data x is expressed as (R ≤n ,Z ≤n ,L),R ≤n ,Z ≤n Represent the spatial coordinates R of the first n atoms in the crystal data x ≤n and atomic number X ≤n , r i ,z i Respectively represent the spatial coordinates and atomic number of the i-th atom to be generated, and the conditional distribution p θ (x|y,z) are all constrained by the hidden vector z and the input variable y, so that the guide vector Conditional distribution p θ (x|y,z) can be written as: The asymmetric reconstruction module is used to obtain the key information vector of the crystal sample. The specific method includes using the low-dimensional feature vector of the crystal sample to perform asymmetric reconstruction to obtain the key information vector of the crystal sample, wherein the low-dimensional feature vector of the crystal sample is a part of the guide vector and satisfies The conditional autoregressive module is used to obtain the prediction of each atom type and atom position of the crystal sample according to the input guidance vector and key information vector. The specific steps include: The implementation formula of the conditional autoregressive module is: Where R represents the spatial coordinate of the atom, Z represents the atomic number, L is the introduced fixed dimensional variable, and p(stop|R ≤n ,X ≤n ,x',z g ) means that when R is satisfied ≤n ,Z ≤n ,x',z g In the case of , the probability that the conditional autoregressive module stops generating, x'={L,T n ,P d } represents the key information vector of the real crystal data x, represents the number of atoms in the crystal to be generated, k represents the total number of atomic types, P d represents the atomic position prediction p pos The weak constraints in p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x',z g ) is expressed as: p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x,z g )=p type *p pos p type =p type (z i |R ≤i-1 ,Z ≤i-1 ,z g ,L,f(T n ,Z ≤i-1 )) p pos =p pos (r i |R ≤i-1 ,Z ≤i-1 ,z i ,z g ,P d ) Among them, the atomic type prediction p type A strong constraint f(T n ,Z ≤i-1 )=max(T n -Z ≤i-1 ,0), used to represent the remaining atoms that need to be generated, and predict p at the atomic position pos A weak constraint P is added d , b d represents the atomic position prediction p in the conditional autoregressive module pos The relative distance distribution between atoms in the .

2. A system for generating indefinite 3D crystal structures based on graph neural networks, characterized in that: The system includes: An indeterminate 3D crystal structure generation model training unit is used to train an indeterminate 3D crystal structure generation model, wherein the indeterminate 3D crystal structure generation model includes a conditional variational autoencoder module, an asymmetric reconstruction module and a conditional autoregressive module, wherein the conditional variational autoencoder module is used to obtain a guide vector of a crystal sample, the asymmetric reconstruction module is used to obtain a key information vector of the crystal sample, and the conditional autoregressive module is used to obtain a type prediction and a position prediction of each atom of the crystal sample according to the input guide vector and the key information vector; the error is calculated by using the type prediction and the position prediction and the expected result, and the error is reversely transmitted to train the indeterminate 3D crystal structure generation model; A 3D structure crystal generation unit is used to generate a 3D structure crystal using material data, input the material data into a trained indefinite 3D crystal structure generation model, and obtain a prediction result of the indefinite 3D crystal structure using a conditional autoregressive model. If the prediction result of the indefinite 3D crystal structure meets a condition for stopping generation, generation is stopped; if the prediction result of the indefinite 3D crystal structure does not meet the condition for stopping generation, the input material data is updated; Among them, the conditional variational autoencoder module is used to obtain the guidance vector of the crystal sample, and the specific steps include: Let the conditional generation probability be p θ (x|y), x represents the real crystal data, y represents the input conditional variable, in the conditional variational autoencoder module, using p θ The variational lower bound of (x|y) is used as the optimization objective of the conditional variational autoencoder module: KL div represents the KL divergence, Represents the low-dimensional space vector z in the corresponding distribution The mathematical expectation of represents the conditional probability distribution of the low-dimensional space vector z, which is used to encode the crystal graph, p θ (z|y) represents the prior conditional distribution of z, and the conditional distribution p θ The realization of (x|y,z) is the key to the indefinite 3D crystal structure generation model, introducing the probability distribution of L with fixed dimensions The p θ (x|y,z) decomposes into: Among them, the real crystal data x is expressed as (R ≤n ,Z ≤n ,L),R ≤n ,Z ≤n Represent the spatial coordinates R of the first n atoms in the crystal data x ≤n and atomic number Z ≤n , r i ,z i Respectively represent the spatial coordinates and atomic number of the i-th atom to be generated, and the conditional distribution p θ (x|y,z) are all constrained by the hidden vector z and the input variable y, so that the guide vector Conditional distribution p θ (x|y,z) can be written as: The asymmetric reconstruction module is used to obtain the key information vector of the crystal sample. The specific method includes using the low-dimensional feature vector of the crystal sample to perform asymmetric reconstruction to obtain the key information vector of the crystal sample, wherein the low-dimensional feature vector of the crystal sample is a part of the guide vector and satisfies The conditional autoregressive module is used to obtain the prediction of each atom type and atom position of the crystal sample according to the input guidance vector and key information vector. The specific steps include: The implementation formula of the conditional autoregressive module is: Where R represents the spatial coordinate of the atom, Z represents the atomic number, L is the introduced fixed dimensional variable, and p(stop|R ≤n ,Z ≤n ,x',z g ) means that when R is satisfied ≤n ,Z ≤n ,x',z g In the case of , the probability that the conditional autoregressive module stops generating, x'={L,T n ,P d } represents the key information vector of the real crystal data x, represents the number of atoms in the crystal to be generated, k represents the total number of atomic types, P d represents the atomic position prediction p pos The weak constraints in p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x',z g ) is expressed as: p θ (r i ,z i |R ≤i-1 ,Z ≤i-1 ,x,z g )=p type *p pos p type =p type (z i |R ≤i-1 ,Z ≤i-1 ,z g ,L,f(T n ,Z ≤i-1 )) p pos =p pos (r i |R ≤i-1 ,Z ≤i-1 ,z i ,z g ,P d ) Among them, the atomic type prediction p type A strong constraint f(T n ,Z ≤i-1 )=max(T n -Z ≤i-1 ,0), used to represent the remaining atoms that need to be generated, and predict p at the atomic position pos A weak constraint P is added d , b d represents the atomic position prediction p in the conditional autoregressive module pos The relative distance distribution between atoms in the .

3. A system for generating indefinite 3D crystal structures based on graph neural networks, characterized in that: include: processor; and a memory, wherein a computer executable program is stored in the memory, and when the computer executable program is executed by the processor, the indefinite 3D crystal structure generation method based on graph neural network according to claim 1 is executed.

4. A computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for generating an indefinite 3D crystal structure based on a graph neural network as claimed in claim 1 is implemented.

Images