Application method for crystal structure optimization based on graph convolutional neural network generation model

Abstract

The present invention discloses an application method for optimizing crystal structures based on a graph convolutional neural network generation model, which relates to the technical field of crystal structures and includes: obtaining initial crystal structure data, and performing data augmentation on the initial crystal structure data to obtain respective crystal structure feature data to be screened; training a graph convolutional neural network, and performing a first screening on the respective crystal structure feature data to be screened according to the graph convolutional neural network to obtain respective target crystal structures, and obtaining respective target crystal structure eigenvalue; performing stability verification on the respective target crystal structures to obtain respective target crystal structure performance evaluation values, and performing a second screening according to the respective target crystal structure eigenvalue and the respective target crystal structure performance evaluation values to obtain respective qualified crystal structures and perform feedback output. The present invention greatly improves the efficiency and accuracy of crystal structure optimization and reduces the consumption of manpower and material resources through an intelligent mode driven by data.

Description

Technical Field

[0001] The present invention relates to the technical field of crystal structure prediction, and specifically to an application method for optimizing crystal structures based on a graph convolutional neural network generation model. Background Art

[0002] With the continuous progress of modern technology and the increasingly stringent requirements for the performance of new materials in various industries, optimizing crystal structures has become a key link in enhancing the comprehensive performance of materials. The method for optimizing crystal structures based on a graph convolutional neural network generation model is an innovative and promising technical approach. However, when dealing with complex crystal structure data, this method faces many problems and challenges, such as high model complexity, large consumption of computing resources, and strong dependence on large-scale data.

[0003] Existing means for optimizing crystal structures mainly rely on traditional experimental tests and numerical calculation methods based on physical models to proceed. By repeating experiments multiple times, finely adjusting parameters, and using classical calculation theories for simulation, etc., efforts are made to optimize and improve crystal structures, so as to meet the material performance requirements in different application scenarios.

[0004] For example, the invention patent with the publication number CN108268750B discloses a method for predicting the hypothetical inorganic crystal structure based on the enumeration of Wyckoff position combinations. In the first step, the target space group for structure prediction is selected, and the ranges of unit cell parameters and framework density are determined; in the second step, Wyckoff position combinations are enumerated; in the third step, a structure model is built. Under a specific space group, through the enumerated Wyckoff position combinations, atoms are fixed at Wyckoff positions to build a hypothetical inorganic crystal structure model, and finally, the prediction of the hypothetical inorganic crystal structure is completed.

[0005] For example, the invention patent with the publication number CN116343962A discloses a method for optimizing the design of self-supporting crystal structures based on topological optimization. The technology of the present invention mainly includes two stages: subdivision and simplification. In the subdivision stage, this method subdivides the initial coarse crystal structure composed of several self-supporting crystal units based on a topological optimization framework. By assigning a state variable to each crystal unit, it is determined whether the structural unit is subdivided. The state variables of the units in each subdivision level are converted into the virtual density of the rods, so that continuous calculation can be incorporated. The self-support of the rods is ensured by a carefully designed unit topology, while the self-support at the nodes is realized by a self-support filter introduced on the rod density under each newly introduced node. By repeatedly applying subdivision, an adaptive and optimized self-supporting crystal structure can be obtained.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of the present application, it is found that the above technologies have at least the following technical problems: In practical applications, since the method based on the enumeration of Wyckoff position combinations relies on fixed rules, there may be a poor adaptability to complex systems, resulting in a decrease in prediction accuracy and affecting the subsequent crystal R & D process. Summary of the Invention

[0007] In view of the deficiencies of the prior art, the present invention provides an application method for crystal structure optimization based on a graph convolutional neural network generation model, which can effectively solve the problems involved in the above background technology.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: The first aspect of the present invention provides an application method for crystal structure optimization based on a graph convolutional neural network generation model, including: obtaining initial crystal structure data, and performing data augmentation on the initial crystal structure data to obtain respective crystal structure feature data to be screened.

[0009] Training a graph convolutional neural network, and performing a first screening on the respective crystal structure feature data to be screened according to the graph convolutional neural network to obtain respective target crystal structures, and obtaining respective target crystal structure feature values.

[0010] Performing stability verification on the respective target crystal structures to obtain respective target crystal structure performance evaluation values, and performing a second screening according to the respective target crystal structure feature values and the respective target crystal structure performance evaluation values to obtain respective qualified crystal structures and perform feedback output.

[0011] As a further method, the process of performing data augmentation on the initial crystal structure data to obtain respective crystal structure feature data to be screened is specifically analyzed as follows: Extracting molecular coordinate data from the initial crystal structure data, obtaining a random number variable by random extraction, and transforming the molecular coordinate data according to the random number variable to obtain respective crystal structure feature data to be screened, and the respective crystal structure feature data to be screened includes the respective molecular coordinates of the respective crystal structures to be screened.

[0012] As a further method, the process of training the graph convolutional neural network is specifically analyzed as follows: Setting up a graph convolutional neural network and initializing the graph structure; setting up a generative adversarial network, including initializing the model weights of the generator and the discriminator, initializing the loss function and optimization parameters, and setting the initial training parameters of the generative adversarial network; importing the respective crystal structure feature data to be screened into the graph convolutional neural network and training the graph convolutional neural network.

[0013] As a further method, the crystal structure feature data to be screened are screened once according to the graph convolutional neural network. The specific analysis process is as follows: Obtain the number of molecules with critical position transformation and the critical molecular transformation displacement from the materials science database; Extract the corresponding molecular coordinates from the initial crystal structure data and the crystal structure feature data to be screened respectively, and further obtain the molecular transformation displacement in each crystal structure to be screened. Through comprehensive analysis, the crystal structure feature values of each crystal structure to be screened are obtained; Compare the crystal structure feature values of each crystal structure to be screened with the preset crystal structure feature threshold in the materials science database to obtain each target crystal structure.

[0014] As a further method, the process of comparing the crystal structure feature values of each crystal structure to be screened with the preset crystal structure feature threshold in the materials science database to obtain each target crystal structure is as follows: If the crystal structure feature value of a crystal structure to be screened is greater than or equal to the preset crystal structure feature threshold, mark the crystal structure to be screened as a target crystal structure; If the crystal structure feature value of a crystal structure to be screened is less than the preset crystal structure feature threshold, mark the crystal structure to be screened as a nonqualified target crystal structure; Count to obtain each target crystal structure.

[0015] As a further method, the process of verifying the stability of each target crystal structure to obtain the performance evaluation value of each target crystal structure is as follows: The stability verification includes mechanical stability verification, thermodynamic stability verification and kinetic stability verification; Through the stability verification, the stability data of each target crystal structure are obtained. The stability data of each target crystal structure include the number of negative elements in the elastic constant matrix of each target crystal structure, the crystal system energy at each monitoring time point and the number of imaginary frequencies of the phonon spectrum; According to the stability data of each target crystal structure, through comprehensive analysis, the performance evaluation value of each target crystal structure is obtained.

[0016] As a further method, the process of comprehensively analyzing according to the stability data of each target crystal structure to obtain the performance evaluation value of each target crystal structure is as follows: Obtain the critical number of negative elements in the elastic constant matrix, the critical crystal system energy at the monitoring time point and the critical number of imaginary frequencies of the phonon spectrum from the materials science database; According to the processing of the stability data of each target crystal structure, the performance evaluation value of each target crystal structure is obtained. The performance evaluation value of the target crystal structure is used to quantitatively evaluate the stability of each target crystal structure.

[0017] As a further method, the process of secondary screening according to the crystal structure feature values of each target crystal structure and the performance evaluation values of each target crystal structure is as follows: According to the crystal structure feature values of each crystal structure to be screened and the performance evaluation values of each target crystal structure, through comprehensive analysis, the quality evaluation index of each target crystal structure is obtained;

[0018] If the quality evaluation index of a target crystal structure is greater than or equal to the preset threshold of the quality evaluation index of the target crystal structure, the target crystal structure is marked as a qualified crystal structure; if the quality evaluation index of a target crystal structure is less than the preset threshold of the quality evaluation index of the target crystal structure, the target crystal structure is marked as an unqualified crystal structure; count all the qualified crystal structures and perform feedback output.

[0019] As a further method, the process of counting all the qualified crystal structures and performing feedback output is as follows: according to the quality evaluation indexes of all the qualified crystal structures, sort all the qualified crystal structures in descending order to obtain the output order of the qualified crystal structures; perform display output on all the qualified crystal structures according to the output order of the qualified crystal structures.

[0020] Compared with the prior art, the embodiments of the present invention have at least the following beneficial effects:

[0021] (1) By providing a crystal structure optimization application method based on a graph convolutional neural network generation model, the present invention deeply mines the complex topological relationships between crystal atoms, transforms the crystal structure optimization from a traditional rule- and experience-based mode to a data-driven intelligent mode, greatly improves the efficiency and accuracy of crystal structure optimization, and reduces the consumption of human and material resources.

[0022] (2) By using a graph convolutional neural network to model and analyze the interactions between atoms and lattice arrangements in a crystal structure, the present invention comprehensively considers the complexity and diversity of the crystal structure, makes the optimization of the crystal structure more comprehensive and in-depth, can accurately predict the stability and performance of the crystal structure, provides a scientific basis for the selection and modification of materials, and promotes the improvement of the performance and the reduction of the cost of new materials.

[0023] (3) By establishing a dynamic feedback mechanism for crystal structure optimization, the present invention real-time monitors and evaluates the performance changes of the crystal structure during the optimization process, realizes the intelligent control and adaptive adjustment of the crystal structure optimization process, reduces the interference of human factors on the optimization results, improves the stability and reliability of the optimization process, and provides an efficient and stable solution for the industrial large-scale production of high-performance crystal materials.

[0024] Of course, it is not necessary for any product implementing the present invention to simultaneously achieve all the above-mentioned advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic flowchart of the method of the present invention.

[0026] Figure 2 It is a schematic diagram of the functional relationship between the characteristic values of the target crystal structure and the quality evaluation index of the target crystal structure of the present invention. Detailed implementation manners

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0029] Referring to Figure 1 As shown, the first aspect of the present invention provides an application method for optimizing crystal structures based on a graph convolutional neural network generation model, including: obtaining initial crystal structure data, and performing data augmentation on the initial crystal structure data to obtain each crystal structure feature data to be screened.

[0030] Specifically, performing data augmentation on the initial crystal structure data to obtain each crystal structure feature data to be screened, the specific analysis process is: extracting molecular coordinate data from the initial crystal structure data, obtaining a random number variable by random extraction, and transforming the molecular coordinate data according to the random number variable to obtain each crystal structure feature data to be screened, and the each crystal structure feature data to be screened includes the molecular coordinates of each crystal structure to be screened.

[0031] In this embodiment, known crystallographic structure data is obtained from a materials science database. The crystallographic structure is an approximate structure similar to the predicted crystallographic structure data. Data preprocessing is performed. Before predicting the crystallographic structure, all input variables are made independent of each other to improve the effectiveness of data prediction. The data preprocessing also includes converting the three-dimensional crystallographic vectors in the original data into six independent crystallographic constants. Random number variables randomly selected are added to the original molecular coordinate data to achieve the purpose of translating the crystal structure in three-dimensional space. The data augmentation method includes translation transformation and rotation transformation. By using the data augmentation method, the problem of severely insufficient materials science data is solved. By directly using crystal vectors and molecular coordinates for crystallographic structure data operation, the problem of information loss caused by data dimensionality reduction in the data representation link is avoided. For example, when performing rotation transformation, the positions of the three molecular coordinates are mutually converted. Through data augmentation, the diversity and complexity of the data set can be increased, thereby improving the generalization ability and robustness of the machine learning model.

[0032] It should be understood that in this embodiment, the approximate structure refers to a crystallographic structure that is structurally similar or close, and may have certain similarities in atomic arrangement, chemical bond formation, or unit cell parameters, but is not exactly the same. Being independent of each other means that there is no direct dependence relationship between variables, that is, the change of one variable will not affect the value of another variable. The crystal structure is composed of a large number of molecules, and the position of the molecules in the crystal is described by molecular coordinates. The molecular coordinates determine the position of the molecules in three-dimensional space, and the properties of the crystal are often closely related to the relative positions of the molecules. For example, the distance between molecules affects the intermolecular force, which in turn affects the physical and chemical properties of the crystal. By extracting the molecular coordinate data from the initial crystal structure data. Through random selection, a series of random number variables are generated. The random number variables will be used for subsequent coordinate transformation to simulate the small changes of the crystal structure under different conditions. For example, by generating uniformly distributed or normally distributed random numbers, different parameters can be provided for different molecular coordinate transformation operations. The random number variables can be one or more, depending on the transformation method adopted. According to the generated random number variables, the molecular coordinate data is transformed, and the transformation can be translation transformation, rotation transformation, scaling transformation, etc. By transforming the molecular coordinate data, different crystal structure feature data to be screened are obtained, and each crystal structure feature data to be screened contains the molecular coordinates of each crystal structure to be screened.

[0033] Train a graph convolutional neural network. According to the graph convolutional neural network, each crystal structure feature data to be screened is screened once to obtain each target crystal structure, and the feature values of each target crystal structure are obtained.

[0034] Specifically, training a graph convolutional neural network, the specific analysis process is as follows: Set up a graph convolutional neural network and initialize the graph structure; Set up a generative adversarial network, including initializing the model weights of the generator and discriminator, initializing the loss function and optimization parameters, and setting the initial training parameters of the generative adversarial network; Import each crystal structure feature data to be screened into the graph convolutional neural network and train the graph convolutional neural network.

[0035] In this embodiment, during the training of the graph convolutional neural network, the crystal structure feature data to be screened, the initialized graph structure, the initial values of the model weights of the generator and discriminator, the initial values of the loss function and optimization parameters, and the initialization training parameters of the generative adversarial network are input, and the trained graph convolutional neural network and the trained generative adversarial network are output. Among them, the crystal structure feature data to be screened is the basis for training the graph convolutional neural network, which contains various feature information of the crystal structure, such as the coordinates of atoms, the types and strengths of chemical bonds, etc.; the initialized graph structure is used to construct the topological structure of the graph convolutional neural network, describing the connection relationship between atoms in the crystal; the initial values of the model weights of the generator and discriminator need to be initialized when setting up the generative adversarial network; the initial values of the loss function and optimization parameters are used to guide the training process of the generative adversarial network, enabling the network to be optimized according to the loss function; the initialization training parameters of the generative adversarial network include the learning rate, the number of iterations, and the batch size, etc.; the trained graph convolutional neural network is used to screen the crystal structure feature data; the trained generative adversarial network helps to improve the performance of the graph convolutional neural network, such as generating more diverse samples or assisting in feature extraction and discrimination. Set up the graph convolutional neural network, determine the graph structure, according to the characteristics of the crystal structure, take the atoms in the crystal as nodes and the chemical bonds between atoms as edges to construct the graph structure, where the connection relationship of the graph can be represented by an adjacency matrix or an edge list. For node features, various attributes of the atoms (such as atomic number, electronegativity, etc.) can be used as the feature vectors of the nodes. Select graph convolutional layers. Common graph convolutional layers include GCN (Graph Convolutional Network) layers and GAT (Graph Attention Network) layers, etc. Build the network architecture, stack multiple graph convolutional layers together to form a deep network, add a pooling layer (such as GraphPooling) to reduce the scale of the graph and extract higher-level features, and add a fully connected layer to map the features of the graph to the required output dimension, such as for classification or regression tasks. Select activation functions and optimizers. Activation functions can be selected from ReLU (Rectified Linear Unit), Sigmoid (S-shaped curve function), and Tanh (Hyperbolic Tangent) etc., and select the appropriate activation function according to the specific task; optimizers can use Adam (Adaptive Moment Estimation) and SGD (Stochastic Gradient Descent) etc., and adjust parameters such as the learning rate according to the training situation of the network.Set up a generative adversarial network (GAN), initialize the generator and the discriminator. The generator is a neural network whose input is random noise and output is the generated data, and its structure can be a variant of a multi-layer perceptron (MLP) or a convolutional neural network (CNN). The discriminator is also a neural network, with input being real data or data generated by the generator, and output being a probability value representing the probability that the input data is real data, and its structure can be similar to that of the generator. Initialize the model weights, and random initialization methods such as normal distribution or uniform distribution can be used to initialize the weights of the generator and the discriminator. Initialize the loss function. The commonly used loss function is binary cross-entropy loss, which is used to measure the ability of the discriminator to distinguish real data from generated data. For the generator, its goal is to minimize the probability that the generated data is misjudged as fake by the discriminator. For the discriminator, its goal is to correctly distinguish real data from generated data. Set the optimization parameters, select a suitable optimizer such as the Adam optimizer, and set parameters such as the learning rate and weight decay. Different learning rates or optimization strategies are required for the generator and the discriminator to balance their training processes. Set the training parameters, determine the number of training iterations to prevent overfitting or underfitting, and select a suitable batch size, usually adjusted according to the size of the dataset and hardware resources. The trained graph convolutional neural network can be used to screen crystal structure feature data. Input the data to be screened into the trained network, and the network will perform screening based on the learned features and patterns to obtain the target crystal structure. The generative adversarial network can assist in the training of the graph convolutional neural network. For example, the samples generated by the generative adversarial network can be used as additional data to expand the training set, or help the graph convolutional neural network learn more discriminative features, thereby improving the accuracy and performance of screening. Based on the obtained target crystal structure, the characteristic values of the target crystal structure can be further calculated, providing a basis for subsequent research on crystal material property analysis, structure optimization, etc.

[0036] It should be understood that in this embodiment, the graph convolutional neural network is a deep learning model specifically designed to process graph-structured data. The crystal structure adapts to the graph-structured data representation form because the atoms or molecules in the crystal can be regarded as nodes in the graph, and the chemical bonds or interactions between them are the edges. Build a GCNN (Graph Convolutional Neural Network) to process graph data, such as social networks, molecular structures, and knowledge graphs, etc. Determine hyperparameters such as the number of layers of the network, the number and size of convolutional kernels in each layer, and the type of activation function. For example, for a social network dataset, the task is to classify nodes. Build a GCNN, where the number of network layers is 3. The number and size of convolutional kernels: in the first layer, there are 32 convolutional kernels with a size of 3; in the second layer, there are 64 convolutional kernels with a size of 3; in the third layer, there are 128 convolutional kernels with a size of 3 (or choose a global average pooling layer to aggregate node features according to the task requirements). The type of activation function is ReLU. This configuration can be adjusted according to specific tasks and datasets. If the dataset is large and the network structure is complex, the number of network layers or the number of convolutional kernels can be increased; if the dataset is small and the task is relatively simple, the number of network layers or the number of convolutional kernels can be reduced to reduce the risk of overfitting. Initialization is to assign initial values to the nodes and edges in the graph structure. For example, assign an initial feature vector representing information such as atomic species and charge to the nodes, and set initial weights representing chemical bond strength and type for the edges, providing a starting state for subsequent calculations and allowing the network to gradually update these values to fit the data pattern. Initializing the graph structure means defining the initial features of the nodes and edges in the graph, and these features may include the type of atoms, the type of bonds, spatial positions, etc. The generative adversarial network consists of two core components: a generator and a discriminator. The generator is used to try to generate realistic data to make it look similar to the real data; the discriminator is used to try to distinguish between real data and the data generated by the generator. Initializing the model weights is to assign random initial values to the parameters of the neural network layers in the generator and the discriminator. For example, put starting weights on the trays on both sides of the balance, and these weights will be continuously adjusted during subsequent training to make them reach balance. Reasonable initial values can speed up the convergence speed, avoid getting stuck in local optimal solutions, and enable the model to learn data features faster. The loss function usually includes two parts: one part is used to measure the difference between the data generated by the generator and the real data, and the other part is used to measure the ability of the discriminator to distinguish between real and generated data. Common ones include mean squared error, cross-entropy loss, etc. Optimizing parameters involves learning rate, momentum, etc., controlling the step size and direction of each parameter update, and adjusting the learning rhythm of the model. Setting the initialization training parameters of the generative adversarial network includes the number of training epochs and batch size, etc. The number of epochs determines the number of times the entire training process loops, and the batch size refers to the amount of data input into the model each time. Gradually find the optimal parameter combination through experiments, adjusting parameters, and observing training results.For example, in the GAN training parameter settings on the MNIST dataset, for the network structure settings, the generator adopts a multi-layer perceptron (MLP) or a convolutional neural network (CNN) structure. The input is a random noise vector, and the output is a grayscale image of 28x28 pixels. The discriminator adopts an MLP or a CNN structure. The input is a grayscale image of 28x28 pixels, and the output is a scalar representing the probability that the image is a real image; for the learning rate settings, the learning rate of the generator is 0.0002, which is a hyperparameter that controls the update speed of the generator's parameters. A smaller learning rate means slower parameter updates, which helps to stabilize the training process. The learning rate of the discriminator is 0.0002. The same learning rate as the generator can maintain the balance between the two; the batch size is set to 64, which is the number of samples input to the network each time during training. A larger batch size can accelerate the training process, but may also increase memory consumption and computing time; for the number of training epochs settings, the total number of training epochs is 10,000, which is the number of times the entire training process will iterate. More training epochs may help improve the performance of the model, but may also lead to overfitting. The discriminator is updated once every 5 updates of the discriminator, and the generator is updated once. This helps to maintain the advantage of the discriminator, thereby guiding the generator to generate more realistic images; the dimension of the noise vector is set to 100, which is the dimension of the random noise vector input to the generator. A larger dimension may increase the expressive ability of the generator, but may also increase the computational complexity; the optimizer is set to the Adam optimizer, which is a commonly used adaptive learning rate optimization algorithm with a fast convergence speed and good performance. Import the data of each crystal structure feature to be screened into the graph convolutional neural network and train the graph convolutional neural network. Send the data of each crystal structure feature to be screened obtained from the previous processing into the GCNN in sequence according to the set batch size. Inside the network, the data continuously propagates and updates the features of nodes and edges according to the graph convolutional operation rules. At the same time, combined with the GAN (Generative Adversarial Network) mechanism, the discriminator distinguishes between real data and the data generated by the generator, and the generator optimizes its own output based on the feedback of the discriminator, prompting the GCNN to gradually fit the data distribution and achieve the purpose of selecting the target crystal structure from the data to be screened.

[0037] Specifically, perform a primary screening on the data of each crystal structure feature to be screened according to the graph convolutional neural network. The specific analysis process is as follows: Obtain the number of critical position transformation molecules and the critical molecular transformation displacement from the materials science database; extract the corresponding molecular coordinates from the initial crystal structure data and the data of each crystal structure feature to be screened respectively, and further obtain the molecular transformation displacement in each crystal structure to be screened. Through comprehensive analysis, obtain the feature values of each crystal structure to be screened; compare the feature values of each crystal structure to be screened with the preset crystal structure feature threshold in the materials science database to obtain each target crystal structure.

[0038] It should be understood that in this embodiment, the number of molecules with critical position transformation refers to the number of molecules whose positions change in the crystal during the process of crystal structure change when reaching a certain critical state (such as phase change, chemical reaction, etc.). For example, at the phase transition point of the crystal, the positions of some molecules change, and the number of these molecules is the number of molecules with critical position transformation. The critical molecular transformation displacement refers to the distance of molecular position change when the crystal structure change reaches the critical state, that is, the displacement amount of the molecule moving from the original position to the new position. This displacement amount is crucial for studying the change of crystal structure. First, extract the corresponding molecular coordinates from the initial crystal structure data and the characteristic data of each crystal structure to be screened respectively. The initial crystal structure data serves as the original reference, and the characteristic data of the crystal structures to be screened are multiple groups of samples after data augmentation and other processes. Based on the two sets of coordinate data, the transformation displacement of each molecule in each crystal structure to be screened is obtained. The number of molecules with position transformation and the molecular transformation displacement can be obtained through molecular dynamics simulation software such as LAMMPS (Large-scale Atomic / Molecular Massively Parallel Simulator). During the simulation of the crystal structure change process, the position change trajectory of each molecule can be traced. By comparing the initial state and the changed state, the difference in molecular position is calculated to determine the molecular transformation displacement. At the same time, by counting the number of molecules with obvious position changes, the number of molecules with position transformation is obtained. The number of molecules with critical position transformation and the critical molecular transformation displacement can be directly obtained from the materials science database.

[0039] In a specific embodiment, the method for obtaining the characteristic values of each crystal structure to be screened is as follows:

[0040] ;

[0041] In the formula, represents the j-th characteristic value of the crystal structure to be screened, e represents the natural constant, represents the number of molecules with position transformation of the j-th crystal structure, represents the t-th molecular transformation displacement of the j-th crystal structure, represents the preset number of molecules with critical position transformation, represents the preset critical molecular transformation displacement, represents the influence weight of the characteristic values of each crystal structure to be screened corresponding to the preset number of molecules with position transformation, represents the influence weight of the characteristic values of each crystal structure to be screened corresponding to the preset molecular transformation displacement. i represents the number of the crystal structure to be screened, j = 1, 2, 3,..., n, n represents the total number of crystal structures to be screened, t represents the number of each molecule in the crystal structure, t = 1, 2, 3,..., m, and m represents the total number of molecules in the crystal structure.

[0042] When performing the crystal structure features to be screened, and it is possible to directly obtain the influence weights of each crystal structure feature value corresponding to the number of position-transformed molecules and the molecular transformation displacement from the materials science database. These weight values respectively reflect the degree of influence on the crystal structure feature values to be screened, and there is a preset mapping rule between their corresponding relationships. For example, the number of molecules in a crystal forms a mapping set with the influence weights of each crystal structure feature value corresponding to the number of position-transformed molecules and the molecular transformation displacement obtained from the materials science database. By inputting the number of molecules in the crystal into the mapping set, the influence weights of each crystal structure feature value corresponding to the number of position-transformed molecules and the molecular transformation displacement can be obtained. The mapping method can be either one-to-one or many-to-one. In this example, the value range of the weight is limited to between 0 and 1 (excluding 0 and 1).

[0043] In this embodiment, the crystal structure feature values to be screened are used to quantitatively evaluate the stability and potential performance of the crystal structure. The fewer the number of position-transformed molecules, or the smaller the molecular transformation displacement, the smaller the crystal structure feature value to be screened, and the lower the stability and potential performance of the crystal structure.

[0044] The algorithm of this embodiment combines the number of position-transformed molecules and the molecular transformation displacement, and comprehensively analyzes to obtain the crystal structure feature values to be screened. In this formula, the number of position-transformed molecules and the molecular transformation displacement affect each other. When the crystal is affected by external or internal factors, if the number of position-transformed molecules is more, it means that more molecules have changed their positions, resulting in a more obvious overall change in the crystal structure, and the molecular transformation displacement is also larger. By comprehensively analyzing the number of position-transformed molecules and the molecular transformation displacement, the crystal structure feature values to be screened can be accurately obtained, which quantitatively reflects the degree of change and complexity of the crystal structure.

[0045] Specifically, comparing each crystal structure feature value to be screened with the preset crystal structure feature threshold in the materials science database to obtain each target crystal structure. The specific analysis process is as follows: if a crystal structure feature value to be screened is greater than or equal to the preset crystal structure feature threshold, then mark this crystal structure feature to be screened as the target crystal structure; if a crystal structure feature value to be screened is less than the preset crystal structure feature threshold, then mark this crystal structure feature to be screened as a non-qualified target crystal structure; count to obtain each target crystal structure.

[0046] It should be understood that in this embodiment, a preset crystal structure feature threshold is stored in the materials science database, which is used to measure whether the crystal structure meets the standard. Compare the feature value corresponding to each crystal structure to be screened with the crystal structure feature threshold. If the feature value of a crystal structure to be screened is greater than or equal to the crystal structure feature threshold, it means that the crystal structure meets or exceeds the established standard in terms of stability, rationality of molecular arrangement, or other key performance aspects, and then mark the crystal structure to be screened as the target crystal structure. On the contrary, if the feature value of a crystal structure to be screened is less than the crystal structure feature threshold, it indicates that the crystal structure does not meet the standard and there are some problems such as unreasonable molecular spacing and insufficient structural stability, and then mark the crystal structure to be screened as an unqualified target crystal structure. After completing the comparison and marking process for all crystal structures to be screened, count all the individuals marked as the target crystal structure and summarize them to form the final target crystal structure set. Each target crystal structure is a crystal sample that initially meets the requirements and has further analysis or application value after screening. The feature values of each target crystal structure are obtained by screening the feature values of each crystal structure to be screened.

[0047] Perform stability verification on each target crystal structure to obtain the performance evaluation values of each target crystal structure, and conduct secondary screening based on the feature values of each target crystal structure and the performance evaluation values of each target crystal structure to obtain each qualified crystal structure and perform feedback output.

[0048] Specifically, perform stability verification on each target crystal structure to obtain the performance evaluation values of each target crystal structure. The specific analysis process is as follows: The stability verification includes mechanical stability verification, thermodynamic stability verification, and kinetic stability verification; through the stability verification, the stability data of each target crystal structure is obtained, and the stability data of each target crystal structure includes the number of negative elements in the elastic constant matrix of each target crystal structure, the crystal system energy at each monitoring time point, and the number of imaginary phonon frequencies; based on the stability data of each target crystal structure, comprehensively analyze to obtain the performance evaluation values of each target crystal structure.

[0049] In this embodiment, the mechanical stability is verified through the elastic constant matrix. The crystal elastic constant matrix is a symmetric second-order tensor, which contains the relationship between the internal stress and strain in the crystal when subjected to external forces. If all elements of the elastic constant matrix are positive values, the crystal has elastic stability in all directions and will not crack or deform. If there are negative elements, the crystal is unstable in the corresponding direction and is prone to cracking or deformation. The thermodynamic stability is verified through molecular dynamics simulation. By simulating the behavior of the crystal at different temperatures and pressures, the stability of the crystal structure can be evaluated. The stability of the crystal can be determined by observing that there are no large fluctuations in the energy of the crystal system over time, which may indicate that the crystal structure is stable. The dynamic stability is verified through the phonon dispersion spectrum. The phonon spectrum in the crystal describes the relationship between the energy and momentum of phonons in the crystal. If there are no imaginary frequencies in the phonon spectrum, that is, the frequencies of all phonons are real numbers, then it can be considered that the crystal structure is dynamically stable. If there are imaginary frequencies, it indicates that the crystal structure is dynamically unstable and may undergo phase transitions or structural deformations, etc. A phase transition is a process in which matter changes from one structure to another under different conditions (such as temperature, pressure, etc.). In crystals, a phase transition involves the reorganization of atomic arrangements and will lead to significant changes in crystal properties. For example, a metal may undergo a phase transition from an ordered crystal structure to a disordered liquid structure when heated to a sufficiently high temperature.

[0050] It should be understood that in this embodiment, the mechanical stability verification is to examine the stability of the crystal structure under external force. The elastic constant matrix reflects the response of the crystal to strain. If there are negative elements in the elastic constant matrix, it means that the crystal may experience structural instability when stressed. The number of negative elements is a key indicator for measuring mechanical stability. The thermodynamic stability verification focuses on the stability of the crystal structure under different temperature conditions. The energy of the crystal system recorded at each monitoring time point can reflect the energy state of the crystal in the thermal environment. The lower the energy, the more thermodynamically stable the crystal structure is; if there are abnormal fluctuations in the system energy or changes that do not conform to the laws of thermodynamics, it implies that there are thermodynamically unstable factors in the crystal structure. The dynamic stability verification mainly analyzes the motion of atoms and molecules in the crystal. The number of imaginary frequencies in the phonon spectrum is the key basis. The phonon spectrum describes the vibration modes of crystal atoms. Normally, the vibration frequency should be a real number. Once an imaginary frequency appears, it indicates that the atomic vibration mode is unstable, which will cause the dynamic instability of the crystal structure. The number of negative elements in the elastic constant matrix of each target crystal structure can be calculated by first-principles calculation software, such as VASP (Vienna Ab-initio Simulation Package). When calculating the elastic constant matrix, VASP applies a small strain to the crystal and then calculates the corresponding stress change to obtain the elastic constant matrix, and the number of negative elements can be directly counted from the calculation results. The energy of the crystal system at each monitoring time point can be obtained by molecular dynamics simulation software, such as LAMMPS. During the dynamic simulation of the crystal, the energy of the crystal system is calculated and recorded in real time. The energy includes multiple parts such as kinetic energy and potential energy (such as intermolecular potential energy, chemical bond potential energy, etc.). The system energy value is output at different monitoring time points. The number of imaginary frequencies in the phonon spectrum can be calculated by first-principles calculation software, such as Phonopy (Phonon properties calculator). By combining with first-principles calculation software (such as VASP), the force constant matrix of the crystal is calculated, and then the phonon vibration frequency is solved using this matrix. If an imaginary frequency appears, the software will record the number of imaginary frequencies.

[0051] Specifically, according to the stability data of each target crystal structure, the performance evaluation value of each target crystal structure is comprehensively analyzed. The specific analysis process is as follows: Obtain the critical number of negative elements in the elastic constant matrix, the critical energy of the crystal system at the monitoring time point, and the critical number of imaginary frequencies in the phonon spectrum from the materials science database; according to the processing of the stability data of each target crystal structure, the performance evaluation value of each target crystal structure is obtained, and the performance evaluation value of the target crystal structure is used to quantitatively evaluate the stability of each target crystal structure.

[0052] In a specific embodiment, the way to obtain the performance evaluation value of each target crystal structure is as follows:

[0053] ;

[0054] Wherein, represents the performance evaluation value of the i-th target crystal structure, e represents the natural constant, represents the number of negative elements of the elastic constant matrix of the i-th crystal structure, represents the energy of the crystal system at the monitoring time point of the i-th crystal structure, represents the number of imaginary frequencies of the phonon spectrum of the i-th crystal structure, represents the preset critical number of negative elements of the elastic constant matrix, represents the preset critical energy of the crystal system at the monitoring time point, represents the preset critical number of imaginary frequencies of the phonon spectrum, represents the influence weight of the performance evaluation value of the target crystal structure corresponding to the preset number of negative elements of the elastic constant matrix, represents the influence weight of the performance evaluation value of the target crystal structure corresponding to the preset energy of the crystal system at the monitoring time point, represents the influence weight of the performance evaluation value of the target crystal structure corresponding to the preset number of imaginary frequencies of the phonon spectrum. i represents the number of the target crystal structure, i = 1, 2, 3,..., k, and k represents the total number of target crystal structures.

[0055] When performing the performance evaluation of the target crystal structure, , and can directly obtain the influence weights of the performance evaluation values of the target crystal structure corresponding to the number of negative elements of the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies of the phonon spectrum from the materials science database. These weight values respectively reflect their influence degrees on the performance evaluation value of the target crystal structure, and there is a preset mapping rule between their corresponding relationships. For example, the number of molecules of the crystal forms a mapping set with the influence weights of the performance evaluation values of the target crystal structure corresponding to the number of negative elements of the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies of the phonon spectrum obtained from the materials science database. Inputting the number of molecules of the crystal into the mapping set, the influence weights of the performance evaluation values of the target crystal structure corresponding to the number of negative elements of the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies of the phonon spectrum can be obtained. Among them, the mapping method can be either one-to-one or many-to-one. In this example, the value range of the weight is limited between 0 and 1 (excluding 0 and 1).

[0056] In this embodiment, the performance evaluation value of the target crystal structure is used to quantitatively evaluate the stability of each target crystal structure. The fewer the number of negative elements of the elastic constant matrix, or the smaller the energy of the crystal system at the monitoring time point, or the fewer the number of imaginary frequencies of the phonon spectrum, the larger the performance evaluation value of the target crystal structure, and the higher the stability of the target crystal structure.

[0057] In this embodiment, the algorithm combines the number of negative elements in the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies in the phonon spectrum, and comprehensively analyzes to obtain the evaluation value of the target crystal structure performance. In this formula, the number of negative elements in the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies in the phonon spectrum affect each other. The number of negative elements in the elastic constant matrix directly reflects the mechanical stability of the crystal structure. When there are negative elements in the elastic constant matrix, it indicates that the crystal structure is unstable under stress. Mechanical instability will cause changes in the atomic vibration mode of the crystal, thus affecting the number of imaginary frequencies in the phonon spectrum, because the instability of the structure may lead to abnormal atomic vibration modes. The energy of the crystal system at the monitoring time point reflects the thermodynamic stability of the crystal structure. From the energy perspective, the change in the energy of the crystal system at the monitoring time point will affect the structural state of the crystal, and then affect the elastic constant matrix. For example, the higher the energy of the crystal system at the monitoring time point, the more the crystal structure expands or deforms, affecting the elastic constant matrix and the larger the number of negative elements. At the same time, thermodynamic instability may also cause the atomic vibration to intensify, change the phonon spectrum, resulting in an increase in the number of imaginary frequencies, and affecting the dynamic stability. The number of imaginary frequencies in the phonon spectrum mainly reflects the dynamic stability of the crystal structure. The appearance of imaginary frequencies means that the atomic vibration mode is unstable. The fewer the number of imaginary frequencies, the more stable the crystal structure is dynamically. Dynamic stability and mechanical stability affect each other. The stability of the atomic vibration mode will affect the microstructure of the crystal, and then affect the elastic constant matrix. Moreover, dynamic instability leads to the redistribution of the internal energy of the crystal, causing the energy of the crystal system at the monitoring time point to change, affecting the thermodynamic stability. By comprehensively analyzing the number of negative elements in the elastic constant matrix, the energy of the crystal system at the monitoring time point, and the number of imaginary frequencies in the phonon spectrum, the evaluation value of the target crystal structure performance can be accurately obtained, which quantitatively reflects the comprehensive stability degree of the target crystal structure in terms of mechanics, thermodynamics, and dynamics.

[0058] Specifically, according to the characteristic values of each target crystal structure and the evaluation values of each target crystal structure performance, a secondary screening is carried out. The specific analysis process is as follows: According to the characteristic values of each crystal structure to be screened and the evaluation values of each target crystal structure performance, the quality evaluation index of each target crystal structure is comprehensively analyzed and obtained.

[0059] In a specific embodiment, the obtaining method of the quality evaluation index of each target crystal structure is as follows:

[0060] ;

[0061] In the formula, represents the quality evaluation index of the i-th target crystal structure, e represents the natural constant, represents the characteristic value of the i-th target crystal structure, represents the evaluation value of the i-th target crystal structure performance, represents the influence weight of the target crystal structure quality evaluation index corresponding to the preset target crystal structure feature value, represents the influence weight of the target crystal structure quality evaluation index corresponding to the preset target crystal structure performance evaluation value. i represents the number of the target crystal structure, i = 1, 2, 3,..., k, and k represents the total number of the target crystal structures.

[0062] In a specific embodiment, as Figure 2 shown, = = 0.5. When = 0.8, the functional relationship between the target crystal structure feature value and the target crystal structure quality evaluation index is as shown by curve a; when = 1.5, the target crystal structure feature value and the target crystal structure quality evaluation index are as shown by curve b; when = 2.7, the target crystal structure feature value and the target crystal structure quality evaluation index are as shown by curve c.

[0063] When performing the target crystal structure quality evaluation, and can directly obtain the influence weights of the target crystal structure feature value and the target crystal structure performance evaluation value on the

[0064] corresponding target crystal structure quality evaluation index from the materials science database. These weight values respectively reflect their influence degrees on the target crystal structure quality evaluation index, and there is a preset mapping rule between their corresponding relationships. For example, the number of molecules of the crystal forms a mapping set with the influence weights of the target crystal structure feature value and the target crystal structure performance evaluation value obtained from the materials science database corresponding to the target crystal structure quality evaluation index. Inputting the number of molecules of the crystal into the mapping set, the influence weights of the target crystal structure feature value and the target crystal structure performance evaluation value corresponding to the target crystal structure quality evaluation index can be obtained. The mapping method can be either one-to-one or many-to-one. In this example, the value range of the weight is limited between 0 and 1 (excluding 0 and 1).

[0065] In this embodiment, the target crystal structure quality evaluation index is used to quantitatively evaluate the overall quality and applicability of the target crystal structure. The larger the target crystal structure feature value or the target crystal structure performance evaluation value, the larger the target crystal structure quality evaluation index, and the higher the overall quality and applicability of the target crystal structure.

[0066] The algorithm of this embodiment combines the target crystal structure eigenvalue and the target crystal structure performance evaluation value, and comprehensively analyzes to obtain the target crystal structure quality evaluation index. In this formula, the target crystal structure eigenvalue and the target crystal structure performance evaluation value affect each other. The greater the molecular displacement, the greater the target crystal structure eigenvalue, which may lead to more negative elements in the elastic constant matrix, and the smaller the target crystal structure performance evaluation value. By comprehensively analyzing the target crystal structure eigenvalue and the target crystal structure performance evaluation value, the target crystal structure quality evaluation index can be accurately obtained, which quantitatively reflects the balance and synergy between the structure optimization potential and performance stability of the target crystal structure.

[0067] If the quality evaluation index of a certain target crystal structure is greater than or equal to the preset target crystal structure quality evaluation index threshold, then mark this target crystal structure as a qualified crystal structure; if the quality evaluation index of a certain target crystal structure is less than the preset target crystal structure quality evaluation index threshold, then mark this target crystal structure as an unqualified crystal structure; count all the qualified crystal structures and perform feedback output.

[0068] It should be understood that in this embodiment, the quality evaluation indexes of each target crystal structure are compared with the preset target crystal structure quality evaluation index threshold in the materials science database. If the quality evaluation index of a certain target crystal structure is greater than or equal to the target crystal structure quality evaluation index threshold, it means that this crystal structure meets or exceeds the established requirements in key dimensions such as the rationality of structural characteristics and performance stability, then mark this target crystal structure as a qualified crystal structure. On the contrary, if the quality evaluation index of a certain target crystal structure is less than the target crystal structure quality evaluation index threshold, it means that the comprehensive performance of this crystal structure does not meet the requirements, then mark this target crystal structure as an unqualified crystal structure. After completing the comparison and marking operations on all target crystal structures, the crystals marked as qualified crystal structures are unified and summarized, and finally feedback output.

[0069] Specifically, counting all the qualified crystal structures and performing feedback output, the specific process is as follows: According to the quality evaluation indexes of each qualified crystal structure, sort each qualified crystal structure in descending order to obtain the output order of the qualified crystal structures; display and output each qualified crystal structure according to the output order of the qualified crystal structures.

[0070] It should be understood that the quality evaluation index of the qualified crystal structure in this embodiment comprehensively reflects the overall quality and applicability of the crystal structure. The higher the value, the better the crystal performs in terms of the rationality of structural characteristics, performance stability, etc. Therefore, according to the ranking of the quality evaluation indexes of the qualified crystal structures, the superiority and inferiority of different qualified crystals are distinguished. All the qualified crystal structures are arranged in descending order of the quality evaluation index to determine the output order of the qualified crystal structures. Such a ranking method makes the crystal with the best quality ranked at the front, facilitating subsequent rapid positioning and prioritized attention to samples with outstanding quality. According to the determined output order of the qualified crystal structures, these qualified crystals are displayed and output one by one. The form of display output can be a data list, a chart, or a dedicated crystal structure visualization interface, which is convenient for users to intuitively view the screening results, quickly obtain the information of the crystal structure with the best quality, and provide clear and orderly references for further experimental research, material application selection, and other work.

[0071] The materials science database is used to store various key data information in the field of materials science, including the number of molecules with critical position transformation, the critical molecular transformation displacement, the number of negative elements in the critical elastic constant matrix, the energy of the crystal system at the critical monitoring time point, and the number of imaginary frequencies of the critical phonon spectrum, etc. The data in the materials science database can be obtained from channels such as professional materials experimental institutions, the achievement data of academic research teams, the research project data related to materials science in universities, and the R & D test data of large materials enterprises. These data provide an important data basis and reference for materials science research, crystal structure screening and analysis, and new material development.

[0072] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor limit the invention to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments to better explain the principle and practical application of the present invention, so that those skilled in the art can well understand and utilize the present invention. As long as it does not deviate from the structure of the present invention or exceed the scope defined by the present invention, it should fall within the protection scope of the present invention.

Claims

1. A crystal structure optimization application method based on a graph convolutional neural network generation model, characterized in that: include: Acquire initial crystal structure data, and perform data augmentation on the initial crystal structure data to obtain characteristic data of each crystal structure to be screened; Training a graph convolutional neural network, screening the characteristic data of each crystal structure to be screened according to the graph convolutional neural network, obtaining each target crystal structure, and obtaining characteristic values ​​of each target crystal structure; Perform stability verification on each target crystal structure to obtain a performance evaluation value of each target crystal structure, perform secondary screening based on the characteristic value of each target crystal structure and the performance evaluation value of each target crystal structure, obtain each qualified crystal structure and output feedback; The graph convolutional neural network is used to screen the characteristic data of each crystal structure to be screened, and the specific analysis process is as follows: Obtain the number of critical position transformation molecules and critical molecular transformation displacement from the materials science database; Extracting the corresponding molecular coordinates from the initial crystal structure data and the characteristic data of each crystal structure to be screened, further obtaining the transformation displacement of each molecule in each crystal structure to be screened, and comprehensively analyzing to obtain the characteristic values ​​of each crystal structure to be screened; Compare the characteristic values ​​of each crystal structure to be screened with the crystal structure characteristic thresholds preset in the materials science database to obtain each target crystal structure; The stability of each target crystal structure is verified to obtain a performance evaluation value of each target crystal structure. The specific analysis process is: obtaining stability data of each target crystal structure through stability verification, and obtaining a performance evaluation value of each target crystal structure through comprehensive analysis based on the stability data of each target crystal structure; The secondary screening according to the characteristic values ​​of each target crystal structure and the performance evaluation value of each target crystal structure includes: comprehensively analyzing the characteristic values ​​of each target crystal structure and the performance evaluation value of each target crystal structure to obtain the quality evaluation index of each target crystal structure.

2. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 1 is characterized in that: The initial crystal structure data is augmented to obtain characteristic data of each crystal structure to be screened. The specific analysis process is as follows: Molecular coordinate data are extracted from the initial crystal structure data, random number variables are obtained by random extraction, and the molecular coordinate data are transformed according to the random number variables to obtain characteristic data of each crystal structure to be screened, wherein the characteristic data of each crystal structure to be screened includes the molecular coordinates of each crystal structure to be screened.

3. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 1 is characterized in that: The specific analysis process of training the graph convolutional neural network is as follows: Set up the graph convolutional neural network and initialize the graph structure; Set up the Generative Adversarial Network, including initializing the model weights of the generator and discriminator, initializing the loss function and optimization parameters, and setting the initial training parameters of the Generative Adversarial Network; The characteristic data of each crystal structure to be screened are imported into the graph convolutional neural network, and the graph convolutional neural network is trained.

4. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 1 is characterized in that: The characteristic values ​​of each crystal structure to be screened are compared with the crystal structure characteristic thresholds preset in the material science database to obtain each target crystal structure. The specific analysis process is as follows: If a characteristic value of a crystal structure to be screened is greater than or equal to a preset crystal structure characteristic threshold, the crystal structure to be screened is marked as a target crystal structure; If a characteristic value of a crystal structure to be screened is less than a preset crystal structure characteristic threshold, the crystal structure to be screened is marked as an unqualified target crystal structure; The target crystal structures were statistically obtained.

5. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 1 is characterized in that: The stability verification of each target crystal structure to obtain a performance evaluation value of each target crystal structure also includes: The stability verification includes mechanical stability verification, thermodynamic stability verification and kinetic stability verification; The stability data of each target crystal structure include the number of negative elements in the elastic constant matrix of each target crystal structure, the energy of the crystal system at each monitoring time point, and the number of imaginary frequencies of the phonon spectrum.

6. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 5 is characterized in that: According to the stability data of each target crystal structure, a comprehensive analysis is performed to obtain the performance evaluation value of each target crystal structure. The specific analysis process is as follows: Obtain the number of negative elements in the critical elastic constant matrix, the energy of the crystal system at the critical monitoring time point, and the number of imaginary frequencies of the critical phonon spectrum from the materials science database; By processing the stability data of each target crystal structure, a performance evaluation value of each target crystal structure is obtained, and the target crystal structure performance evaluation value is used to quantitatively evaluate the stability of each target crystal structure.

7. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 1 is characterized in that: The secondary screening according to the characteristic values ​​of each target crystal structure and the performance evaluation values ​​of each target crystal structure also includes: If a target crystal structure quality assessment index is greater than or equal to a preset target crystal structure quality assessment index threshold, the target crystal structure is marked as a qualified crystal structure; If a target crystal structure quality assessment index is less than a preset target crystal structure quality assessment index threshold, the target crystal structure is marked as an unqualified crystal structure; The qualified crystal structures are statistically obtained and fed back and output.

8. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 7 is characterized in that: The statistics of each qualified crystal structure are obtained and feedback is output. The specific process is as follows: According to the quality evaluation index of each qualified crystal structure, the qualified crystal structures are sorted in descending order to obtain the output order of the qualified crystal structures; Each qualified crystal structure is displayed and output according to the output order of the qualified crystal structures.

9. The crystal structure optimization application method based on the graph convolutional neural network generation model according to claim 7 is characterized in that: The method for obtaining the quality evaluation index of each target crystal structure is as follows: ； In the formula, represents the i-th target crystal structure quality assessment index, e represents the natural constant, represents the eigenvalue of the i-th target crystal structure, represents the performance evaluation value of the i-th target crystal structure, Indicates the impact weight of the target crystal structure quality assessment index corresponding to the preset target crystal structure characteristic value, represents the target crystal structure quality evaluation index influence weight corresponding to the preset target crystal structure performance evaluation value, i represents the number of the target crystal structure, i=1, 2, 3, ..., k, k represents the total number of target crystal structures.

Images