Crystal Structure Prediction Method and System Based on Hybrid Potential Deep Learning Model

Abstract

The present invention discloses a crystal structure prediction method and system based on a hybrid potential energy deep learning model, which relates to the field of material structure, including inputting basic information of the material to be predicted, obtaining structural characteristic values ​​of each crystal structure after obtaining rationality evaluation results; performing corresponding feature learning on the local environment between atoms of each crystal structure, and outputting the preliminary energy of each crystal structure; obtaining each stable crystal structure of the material to be predicted, calculating the short-range interaction information between atoms, and processing to obtain the empirical potential of each stable crystal structure. The present invention combines the deep learning model MEGNet with Buckingham potential to accurately describe the local structure and interaction between atoms, thereby significantly improving the accuracy of crystal structure energy prediction. The search process of the crystal structure is optimized, avoiding the possible local minimum problem, and ensuring that the global optimal or near-optimal stable structure is found.

Description

Technical Field

[0001] The present invention relates to the field of material structures, and specifically to a crystal structure prediction method and system based on a hybrid potential deep learning model. Background Art

[0002] The accurate prediction of crystal structures is one of the core tasks in material design and the discovery of new materials, especially when searching for new materials with stable and excellent properties. The crystal structure of a material determines its physical, chemical, and mechanical properties, etc. Therefore, accurately predicting the crystal structure can not only accelerate the material screening process but also provide a theoretical basis for the practical application of materials. Traditional computational methods, such as density functional theory (DFT), with its strict quantum mechanical foundation, can provide very high computational accuracy and are thus widely used in energy calculations and structure optimizations in materials science. However, the DFT method has extremely high computational overhead when dealing with large-scale material systems. Especially when conducting large-scale material screening, its computational cost and time cost greatly limit the application scope of this method.

[0003] To solve this problem, in recent years, data-driven machine learning methods, especially graph neural networks, have gradually emerged in the field of materials science. Compared with traditional physics-based models, the GNN method can effectively capture the local environmental characteristics of crystal structures by simulating the complex relationships between atoms, thereby predicting the properties of materials. Especially when dealing with molecules and crystal systems with highly non-Euclidean structures, graph neural networks can directly perform modeling in the graph structure, showing significant advantages in molecular and crystal structure prediction.

[0004] The prior art, such as the invention patent with the publication number: CN117393089B, is a crystal evolution simulation method based on a single-mode Bessel crystal phase field model, which relates to the technical field of phase field models and includes the following steps: constructing a single-mode Bessel crystal phase field model through Bessel functions and a classical density functional model; substituting the density distribution functions of different phases into the single-mode Bessel crystal phase field model respectively to solve the parameters at the minimum free energy of different phases; constructing a crystal phase diagram according to the parameters at the minimum free energy of different phases by the common tangent method; discretizing the crystal phase field evolution equation by the Fourier spectral method to obtain a crystal phase field iteration equation; and simulating the evolution process of the crystal according to the crystal phase diagram and the crystal phase field iteration equation. A new single-mode crystal phase field model is constructed based on the two-point correlation function obtained by fitting with Bessel functions and combined with the classical density functional model.

[0005] The prior art is an invention patent with the publication number of CN112466411B, which is a method for predicting crystal structures based on crystal topology theory and relates to the technical field of material structures. The method includes: systematically and comprehensively searching for and obtaining energy-stable crystal structures using crystal topology methods; obtaining the mechanical stability of the research object; obtaining the mechanical properties of the research object. Analyzing the structural characteristics of the research object and obtaining crystal structures of different topological types. Obtaining the results of the analysis of the crystal structure performance of the carbon system based on the energy stability, mechanical stability, and mechanical properties of the research object.

[0006] Based on the above scheme, it can be seen that when the prior art analyzes and simulates crystal structures, it often uses models or algorithms that consume a large amount of computing power and are only applicable to the phase transition processes of specific materials, lacking a certain degree of generality. When dealing with complex interatomic interactions, traditional machine learning methods usually simplify the modeling of these interactions, which leads to relatively low prediction accuracy of machine learning methods in the process of energy prediction and structure optimization. Especially in the modeling of short-range interactions, it may not accurately reflect the repulsive and attractive forces between atoms, affecting the prediction accuracy of the overall energy. At the same time, the prior art pays more attention to the geometric structure of crystals and ignores the influence of the electronic structure of crystals. Therefore, there are problems of inaccurate prediction caused by insufficient consideration in current crystal structure prediction. Summary of the Invention

[0007] Aiming at the deficiencies of the prior art, the present invention provides a method and system for predicting crystal structures based on a hybrid potential deep learning model. To achieve the above objectives, the present invention is realized through the following technical solutions: The method for predicting crystal structures based on a hybrid potential deep learning model includes:

[0008] Input the basic information of the material to be predicted into the structure prediction platform, generate the characteristic values of the basic information of the material to be predicted and conduct a rationality evaluation. After obtaining the rationality evaluation result, based on the characteristic values of the basic information of the material to be predicted, retrieve the crystal structure data sets that conform to the material to be predicted from the structure prediction platform and obtain the structural characteristic values of each crystal structure.

[0009] Based on the structural characteristic values of each crystal structure, conduct corresponding feature learning on the local environment between atoms of each crystal structure, output the preliminary energy of each crystal structure, denoted as the neural network potential of each crystal structure, and analyze and process to obtain the evaluation result of the neural network potential of each crystal structure.

[0010] Based on the evaluation results of the neural network potential of each crystal structure, screen out the stable crystal structures of the material to be predicted, calculate the short-range interaction information between atoms of each stable crystal structure, and obtain the empirical potential of each stable crystal structure through cumulative processing.

[0011] Comprehensively analyze and process the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and import it into the optimization algorithm for iterative update. Output the stable crystal structure with the lowest total energy potential as the target crystal structure of the material to be predicted.

[0012] As an optimal technical solution, input the basic information of the material to be predicted into the structure prediction platform, generate the eigenvalue of the basic information of the material to be predicted and conduct a rationality evaluation. The specific process is as follows:

[0013] The basic information of the material to be predicted includes the total number of atomic species of the material to be predicted, the number of atoms of each species, the charge number of atoms of each species, the total number of chemical bond types, the number of chemical bonds, the bond lengths of chemical bonds of each species, and the bond angles of chemical bonds of each species.

[0014] Extract the reference bond angles and reference bond lengths corresponding to various chemical bonds from the structure prediction platform, conduct joint analysis and processing with the basic information of the material to be predicted to generate the eigenvalue of the basic information of the material to be predicted, and compare it with the rationality evaluation threshold to obtain the rationality evaluation result of the material to be predicted. If the eigenvalue of the basic information of the material to be predicted is greater than or equal to the rationality evaluation threshold, it is determined that the rationality evaluation result of the material to be predicted is reasonable; if the eigenvalue of the basic information of the material to be predicted is less than the rationality evaluation threshold, it is determined that the rationality evaluation result of the material to be predicted is unreasonable, and a reminder message is sent to the structure prediction platform.

[0015] As an optimal technical solution, based on the eigenvalue of the basic information of the material to be predicted, retrieve the crystal structure data sets that meet the material to be predicted from the structure prediction platform and obtain the structure eigenvalues of each crystal structure, specifically including:

[0016] Based on the eigenvalue of the basic information of the material to be predicted, perform mapping and matching with each crystal structure data set corresponding to the interval of each basic information eigenvalue pre-stored in the structure prediction platform to obtain each crystal structure data set of the material to be predicted. Each crystal structure data set of the material to be predicted includes the lengths of each side of the unit cell, the unit cell angles, the space group symbol, and the coordinates of each atom of each crystal structure.

[0017] Based on the space group symbol of each crystal structure, perform mapping and matching with the symmetry coefficients corresponding to each space group symbol pre-stored in the structure prediction platform to obtain the symmetry coefficients of each crystal structure. The symmetry coefficients are used to characterize the space group symmetry characteristics of the crystal structure.

[0018] Based on the coordinates of each atom in each crystal structure, analyze and process to obtain the correlation coefficient of each crystal structure. The correlation coefficient of each crystal structure is used to characterize the degree of order of each crystal structure.

[0019] The edge lengths of each side of the unit cell, the cell angles, the symmetry coefficients, and the correlation coefficients of each crystal structure are processed and analyzed to obtain the structure characteristic values of each crystal structure, and the structure characteristic values are used to describe the crystal structure characteristics.

[0020] As a preferred technical solution, the corresponding feature learning of the local environment between atoms of each crystal structure specifically includes:

[0021] Based on the structure characteristic values of each crystal structure, mapping and matching are performed with the local segmentation ratios corresponding to the intervals of the structure characteristic values pre-stored in the structure prediction platform to obtain the local segmentation ratios of each crystal structure.

[0022] Based on the local segmentation ratios of each crystal structure, corresponding local segmentation of each crystal structure is performed to obtain the local environments of each crystal structure.

[0023] Each crystal structure is transformed into a graph structure and input into the MEGNet model, and each atom of each crystal structure is denoted as each node to obtain the local environment interaction information and the overall environment interaction information of each node of each crystal structure. Each node is updated layer by layer through the message passing mechanism, and the feature representation vectors of each node of each crystal structure are output.

[0024] The interaction between each node is denoted as an edge vector, and based on the feature representation vectors of the two nodes connected by each edge vector, iterative update is performed through the graph convolutional layer, and the feature representation vectors of each edge of each crystal structure are output.

[0025] As a preferred technical solution, the output of the preliminary energy of each crystal structure specifically includes:

[0026] Based on the feature representation vectors of each node of each crystal structure and the feature representation vectors of each edge of each crystal structure, the feature representation values of each node of each crystal structure are obtained through processing by the graph convolutional layer. The feature representation values of each node are used to describe the local environment information and the interaction information of each node in the corresponding crystal structure.

[0027] The feature representation values of each node of each crystal structure are input into the neural network model, and through processing and analysis, the preliminary energy of each crystal structure is obtained, which is denoted as the neural network potential of each crystal structure.

[0028] As a preferred technical solution, the analysis and processing to obtain the evaluation results of the neural network potential of each crystal structure include:

[0029] The neural network potential threshold is extracted from the structure prediction platform. If the neural network potential of a certain crystal structure is greater than or equal to the neural network potential threshold, it is determined that the evaluation result of the neural network potential of this crystal structure is unstable.

[0030] If the neural network potential of a certain crystal structure is less than the neural network potential threshold, it is determined that the evaluation result of the neural network potential of the crystal structure is stable, and the crystal structures with stable evaluation results of the neural network potential are extracted and recorded as each stable crystal structure.

[0031] As a preferred technical solution, calculating the short-range interaction information between atoms of each stable crystal structure and obtaining the empirical potential of each stable crystal structure through cumulative processing specifically includes:

[0032] Statistical coordinates of each atom of each stable crystal structure, analyzing and processing to obtain the spatial distance between each atom, calculating the short-range interaction information between atoms of each stable crystal structure using the Buckingham potential model, and accumulating the short-range interaction information of each atom pair and recording the result as the empirical potential of each stable crystal structure.

[0033] As a preferred technical solution, comprehensively analyzing and processing the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and importing it into the optimization algorithm for iterative update, and outputting the stable crystal structure with the lowest total energy potential as the target crystal structure of the material to be predicted, specifically including:

[0034] Based on the neural network potential and empirical potential of each stable crystal structure, analyzing and processing to obtain the total energy potential of each stable crystal structure, and comparing and analyzing it with the preset total energy potential threshold in the structure prediction platform, and counting the stable crystal structures with the total energy potential lower than the total energy potential threshold and recording them as each predicted crystal structure.

[0035] Import the total energy potential of each predicted crystal structure into the Bayesian optimization algorithm for optimization processing. When the Bayesian optimization algorithm converges, output the predicted crystal structure with the lowest total energy potential as the target crystal structure of the material to be predicted.

[0036] As a preferred technical solution, the analyzing and processing to obtain the total energy potential of each stable crystal structure, the specific processing conditions are:

[0037] ;

[0038] Among them, is the total energy potential of the th stable crystal structure, is the neural network potential of the th stable crystal structure, is the empirical potential of the th stable crystal structure, is the energy balance coefficient, is the number of the stable crystal structure, , is the total number of stable crystal structures.

[0039] In addition, the present invention also provides a crystal structure prediction system based on a hybrid potential deep learning model, which is characterized by including:

[0040] A crystal structure feature module, which is used to input the basic information of the material to be predicted on the structure prediction platform, generate the feature values of the basic information of the material to be predicted and conduct a rationality evaluation. After obtaining the rationality evaluation result, based on the feature values of the basic information of the material to be predicted, retrieve various crystal structure data sets that match the material to be predicted from the structure prediction platform and obtain the structure feature values of various crystal structures.

[0041] A neural network potential calculation module, which is used to conduct corresponding feature learning on the local environment between atoms of each crystal structure based on the structure feature values of each crystal structure, output the preliminary energy of each crystal structure, denoted as the neural network potential of each crystal structure, and analyze and process to obtain the neural network potential evaluation result of each crystal structure.

[0042] An empirical potential calculation module, which is used to screen out various stable crystal structures of the material to be predicted based on the neural network potential evaluation results of each crystal structure, calculate the short-range interaction information between atoms of each stable crystal structure, and obtain the empirical potential of each stable crystal structure through cumulative processing.

[0043] A target crystal structure output module, which is used to comprehensively analyze and process the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and import it into an optimization algorithm for iterative update, and output the stable crystal structure with the lowest total energy potential as the target crystal structure of the material to be predicted.

[0044] Compared with the prior art, the embodiments of the present invention at least have the following beneficial effects:

[0045] (1) The present invention provides a crystal structure prediction method and system based on a hybrid potential deep learning model, which combines the deep learning model MEGNet and the Buckingham potential, can accurately describe the local structure and interaction between atoms, and thus significantly improve the accuracy of crystal structure energy prediction. By introducing the Bayesian optimization algorithm, the search process of the crystal structure is optimized, which not only improves the calculation efficiency, but also avoids the local minimum problem that the traditional method may fall into, ensuring that the global optimal or near-optimal stable structure can be found. The method of the present invention has significant advantages in large-scale material screening and new material discovery, can maintain high-precision prediction ability while reducing the calculation cost. In addition, the technical solution of the present invention has wide applicability and can be applied to the crystal structure prediction of various material systems.

[0046] (2) By comprehensively considering the rationality of the atomic composition and the chemical bond composition of the material to be predicted, the present invention identifies whether the material to be predicted exists, and then predicts the crystal structure of the material to be predicted, improving the prediction accuracy, reducing unnecessary resource waste, and improving the prediction efficiency.

[0047] (3) By obtaining the structure characteristic values of each crystal structure, the present invention provides a standardized method to describe and compare different crystal structures, so as to quickly determine the position of the crystal structure in the material space, obtain more detailed local environment information, and help the model better understand and predict the interactions and information in the crystal structure.

[0048] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of the crystal structure prediction method based on the hybrid potential deep learning model of the present invention.

[0050] Figure 2 is a comparison diagram of the MEGNet-Buckingham algorithm, the predicted energy by DFT, and the DFT-PSO method.

[0051] Figure 3 is the crystal structure of CaS corresponding to mp-22862 in the Materials Project material library.

[0052] Figure 4 is the crystal structure with the lowest energy searched by the method proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] 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 the 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.

[0054] In the description of the present invention, it should be understood that the terms "open hole", "upper", "lower", "thickness", "top", "middle", "length", "inner", "periphery", etc. indicating the orientation or position relationship are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the components or elements referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0055] Please refer to Figure 1As shown in the figure, an embodiment of the present invention provides a crystal structure prediction method based on a hybrid potential deep learning model, including:

[0056] Input the basic information of the material to be predicted into the structure prediction platform, generate the eigenvalue of the basic information of the material to be predicted and conduct a rationality evaluation. After obtaining the rationality evaluation result, retrieve the crystal structure data sets that meet the material to be predicted from the structure prediction platform based on the eigenvalue of the basic information of the material to be predicted, and merge them to obtain the structure eigenvalue of each crystal structure.

[0057] The process of inputting the basic information of the material to be predicted into the structure prediction platform, generating the eigenvalue of the basic information of the material to be predicted and conducting a rationality evaluation is as follows:

[0058] The basic information of the material to be predicted includes the total number of atomic species of the material to be predicted, the number of atoms of each species, the charge number of atoms of each species, the total number of chemical bond types, the number of chemical bonds, the bond length of each type of chemical bond, and the bond angle of each type of chemical bond, which are obtained through a chemical resource library, a chemical material library or input and upload by the user in the embodiment of the present invention.

[0059] It should be noted that the atomic species refers to the types of different elements present in the material. The atomic species determines the electronic structure and potential chemical properties of the material. The chemical activity, valence electron number and atomic radius of different elements are different, all of which will affect the overall properties of the material.

[0060] The number of atoms of each species refers to the number of atoms of each element in the material. The number of atoms determines the stoichiometric ratio and affects the crystal structure and phase state of the material. For example, different ratios of the same element can form different compounds.

[0061] The charge number of atoms of each species refers to the charge state of each atom in the material (usually considering the oxidation state). The charge number affects the chemical reactivity and electron arrangement of the atom, and thus affects the electronic characteristics, conductivity and magnetism of the material.

[0062] The chemical bond type refers to different types of chemical bonds formed between atoms, such as covalent bonds, ionic bonds, metallic bonds and hydrogen bonds, etc. The type of chemical bond directly affects the mechanical strength, melting point, conductivity and solubility of the material.

[0063] The number of chemical bonds refers to the total number of each type of chemical bond in the material. The number of chemical bonds is related to the stability and structural integrity of the material. For example, more covalent bonds usually mean higher hardness and melting point.

[0064] The bond length of each type of chemical bond refers to the average distance between the nuclei of two bonded atoms. The bond length affects the density and lattice parameters of the material and is a key factor determining the mechanical and electronic properties of the material.

[0065] The bond angle of various chemical bonds refers to the angle between two adjacent chemical bonds in a molecule. The bond angle affects the geometric shape of the molecule, and thus affects the physical and chemical properties of the molecule, such as polarity, reactivity, and intermolecular forces.

[0066] Extract the reference bond angles and reference bond lengths corresponding to various chemical bonds from the structure prediction platform, and jointly analyze and process them with the basic information of the material to be predicted to generate the characteristic values of the basic information of the material to be predicted, specifically including:

[0067] ;

[0068] Among them, is the characteristic value of the basic information of the material to be predicted, is the number of the th type of atom, is the charge number of the th type of atom, is the bond length of the th type of chemical bond, is the bond angle of the th type of chemical bond, is the reference bond length of the th type of chemical bond, is the reference bond angle of the th type of chemical bond, is the atomic species number, , is the total number of atomic species, is the chemical bond type number, , is the total number of chemical bond types, is the weight factor of the electro-neutrality characteristic unit, is the weight factor of the bond length and bond angle comprehensive characteristic unit.

[0069] It should be noted that the value ranges of the electro-neutrality characteristic unit weight factor and the bond length and bond angle comprehensive characteristic unit weight factor are both between 0 and 1, and satisfy , The unit weight factor of the electro-neutrality feature is the influence factor corresponding to the basic information feature value of the material to be predicted preset in the structure prediction platform, representing the numerical value of the influence degree of the electro-neutrality feature on the basic information feature value of the material to be predicted. The unit weight factor of the comprehensive bond length and bond angle feature is the influence factor corresponding to the basic information feature value of the material to be predicted preset in the structure prediction platform, representing the numerical value of the influence degree of the comprehensive bond length and bond angle feature on the basic information feature value of the material to be predicted. When in use, the unit weight factor of the electro-neutrality feature and the unit weight factor of the comprehensive bond length and bond angle feature can be directly obtained from the structure prediction platform, and their corresponding relationship is a pre-set mapping relationship. For example: input the basic information of the material to be predicted involved in the embodiments of the present invention, such as the total number of atomic types of the material to be predicted, the number of atoms of each type, the charge number of atoms of each type, the total number of chemical bond types, the number of chemical bonds, the bond lengths of chemical bonds of each type, and the bond angles of chemical bonds of each type, into the mapping set for mapping and matching to obtain the unit weight factor of the electro-neutrality feature and the unit weight factor of the comprehensive bond length and bond angle feature of the material to be predicted involved in the embodiments of the present invention, and the mapping relationship therein is one-to-one.

[0070] It should also be noted that there is a certain correlation between the electro-neutrality feature and the comprehensive bond length and bond angle feature of the material to be predicted. The electro-neutrality feature refers to the balance state of positive and negative charges in the material, which is crucial for the stability and chemical properties of the material. The comprehensive bond length and bond angle feature describes the spatial arrangement and connection mode between atoms in the material, and these features directly affect the physical and chemical properties of the material.

[0071] The charge number of an atom affects the electron cloud distribution between atoms, thereby affecting the bond length of the chemical bond. For example, an atom with a higher charge number may attract the electron cloud more tightly, resulting in a shorter bond length. The bond angle of an electro-neutral material is usually determined by the charge balance between atoms. If the charge distribution is uneven, it may lead to the distortion of the bond angle, thereby affecting the structure of the material. Different types of chemical bonds have different requirements for electro-neutrality. For example, ionic bonds require strict charge balance, while covalent bonds may allow a certain amount of charge offset. Electro-neutrality is the basis for the stability of the material, while bond length and bond angle determine the geometric structure of the material. A stable material usually has a balanced charge distribution and optimized bond length and bond angle. The atomic type and charge number determine the electronegativity difference between atoms, thereby affecting the bond length and bond angle. The type and number of chemical bonds determine the overall structure of the material and the distribution of bond length and bond angle. Bond length and bond angle directly reflect the spatial arrangement and interaction between atoms, and these are all related to the charge state of the atoms. In summary, the correlation between the electro-neutrality feature and the comprehensive bond length and bond angle feature is reflected in jointly determining the stability and properties of the material.

[0072] Compare the eigenvalue of the basic information of the material to be predicted with the rationality evaluation threshold to obtain the rationality evaluation result of the material to be predicted. If the eigenvalue of the basic information of the material to be predicted is greater than or equal to the rationality evaluation threshold, it is determined that the rationality evaluation result of the material to be predicted is reasonable. If the eigenvalue of the basic information of the material to be predicted is less than the rationality evaluation threshold, it is determined that the rationality evaluation result of the material to be predicted is unreasonable, and a reminder message is sent to the structure prediction platform.

[0073] The processing method of the rationality evaluation threshold is as follows: In a specific embodiment, the rationality evaluation threshold is directly set in the structure prediction platform during the development of the crystal structure prediction system based on the hybrid potential deep learning model involved in the embodiments of the present invention. There are various methods for setting the rationality evaluation threshold. For example, it is obtained through statistical analysis methods, including statistically analyzing the eigenvalue of the basic information in the reasonable state of historical materials (for example, when the electric neutrality is 0), and performing mean processing on the eigenvalues of the basic information obtained multiple times to obtain the rationality evaluation threshold.

[0074] Retrieve the crystal structure data sets that match the material to be predicted from the structure prediction platform based on the eigenvalue of the basic information of the material to be predicted, and obtain the structure eigenvalues of each crystal structure. Specifically, it includes:

[0075] Based on the basic information of the material to be predicted, match it with the crystal structure library corresponding to each basic information pre-stored in the structure prediction platform, and then perform mapping matching according to the eigenvalue of the basic information of the material to be predicted and each crystal structure data set in the corresponding crystal structure library to obtain each crystal structure data set of the material to be predicted. Each crystal structure data set of the material to be predicted includes the edge lengths of each side of the unit cell, the unit cell angles, the space group symbol, and the atomic coordinates of each atom.

[0076] It should be noted that the unit cell is the smallest repeating unit in the crystal structure. The edge lengths of each side of the unit cell refer to the lengths of the unit cell in three spatial dimensions. The unit cell edge lengths determine the linear dimensions and proportions of the crystal, and they are the basis for calculating crystal density, lattice energy, and other physical properties.

[0077] The unit cell angles refer to the angles between the three sides of the unit cell, and they describe the spatial orientation of the unit cell. The unit cell angles affect the shape and symmetry of the crystal and are important factors determining the crystal's spatial structure.

[0078] The space group symbol is a classification of the atomic arrangement symmetry in the crystal. It describes all possible symmetry operations in the crystal, such as rotation, reflection, and inversion. The space group symbol provides a complete description of the crystal symmetry. It determines the possible positions of atoms in the crystal and the physical properties of the crystal, such as optical and magnetic properties. For example, the space group symbol of a certain crystal is P2 1 / c, where P indicates that the space group belongs to the Paraiso class (class P), i.e., there is no mirror symmetry. 2 1 It indicates that the space group has a two-fold screw axis with the axis direction [1 0 0], and the screw rotation angle is 180 degrees divided by 2, i.e., 90 degrees. c indicates that the space group has a mirror plane perpendicular to the screw axis.

[0079] The atomic coordinates refer to the position of each atom within the unit cell, which is described by a set of coordinates that are normalized with respect to the starting point and side lengths of the unit cell. Atomic coordinates are a direct way to describe the atomic arrangement within a crystal and are crucial for understanding crystal structure, calculating interatomic distances and angles, and predicting crystal properties.

[0080] These parameters are the basic descriptors of the crystal structure, which can precisely define a crystal structure and make it uniquely determined in three-dimensional space. Through the space group symbol, the symmetry of the crystal can be understood, and the unit cell parameters and atomic coordinates are the basis for calculating the physical properties of the crystal.

[0081] Based on the space group symbols of each crystal structure, mapping and matching are performed with the symmetry coefficients corresponding to the space group symbols pre-stored in the structure prediction platform to obtain the symmetry coefficients of each crystal structure, and the symmetry coefficients are used to characterize the space group symmetry characteristics of the crystal structure.

[0082] Based on the atomic coordinates in each crystal structure, the correlation coefficients of each crystal structure are obtained through analysis and processing. The correlation coefficients of each crystal structure are used to characterize the degree of order of each crystal structure, and the specific calculation formula is:

[0083] ;

[0084] ;

[0085] where is the correlation coefficient of the th crystal structure, is the spatial horizontal coordinate value of the rd atom in the th crystal structure, is the spatial front-back coordinate value of the rd atom in the th crystal structure, is the spatial up-down coordinate value of the rd atom in the th crystal structure, is the average value of the spatial horizontal coordinate values of the atoms in the th crystal structure, is the average value of the spatial front-back coordinate values of the atoms in the th crystal structure. is the average of the spatial up and down coordinate values of the atoms of the th crystal structure, is the covariance matrix of the th crystal structure, used to calculate the correlation coefficient of each atom in the th crystal structure, is the normalization factor, is the crystal structure number, , is the total number of crystal structures, is the atom number, , is the atom type number, , is the total number of atom types.

[0086] Process and analyze the edge lengths of each side of the unit cell, the unit cell angles, the symmetry coefficient, and the correlation coefficient of each crystal structure to obtain the structural characteristic values of each crystal structure, and the structural characteristic values are used to describe the crystal structure characteristics.

[0087] ;

[0088] ;

[0089] Among them, is the structural characteristic value of the th crystal structure, is the unit cell volume of the th crystal structure, is the length of the a side of the unit cell of the th crystal structure, is the length of the b side of the unit cell of the th crystal structure, is the length of the c side of the unit cell of the th crystal structure, is the angle between the b side and the c side of the unit cell of the th crystal structure, is the angle between the a side and the c side of the unit cell of the th crystal structure, is the angle between the a side and the b side of the unit cell of the th crystal structure, is the symmetry coefficient of the th crystal structure, is the correlation coefficient of the th crystal structure, is the characteristic factor of the unit volume of the crystal structure, is the crystal structure number, , is the total number of crystal structures, is the natural constant.

[0090] It should be noted that the characteristic factor of the unit volume of the unit cell can be directly extracted from the structure prediction platform during use. In the structure prediction platform, a preset mapping set is formed between the unit cell volume and the characteristic factor of the unit volume of the unit cell, and their mapping relationship is one-to-one correspondence. The unit cell volume involved in the embodiments of the present invention is input into the mapping set for corresponding mapping matching to obtain the characteristic factor of the unit volume of the unit cell involved in the embodiments of the present invention.

[0091] It should also be noted that there is a certain correlation between the unit cell volume of each crystal structure, the symmetry coefficient of each crystal structure, and the correlation coefficient of each crystal structure. The unit cell volume refers to the volume of the smallest repeating unit (unit cell) in the crystal structure. It is determined by the three side lengths of the unit cell and reflects the density and spatial occupancy of the arrangement of atoms or molecules in the crystal. The symmetry coefficient describes the richness of the symmetry operations in the crystal structure. The higher the crystal symmetry, the higher its symmetry coefficient. Symmetry includes point symmetry, plane symmetry, axial symmetry, etc. The correlation coefficient is usually used to describe the positional relationship between different atoms or molecules in the crystal structure, or to describe the similarity between different crystal structures. The unit cell volume may affect the symmetry of the crystal structure. For example, changes in volume may cause changes in the unit cell parameters, thereby changing the symmetry of the crystal. In some cases, an increase or decrease in volume may lead to a decrease or increase in symmetry. The unit cell volume is related to the crystal density, and density changes may affect the stability of the crystal structure, thereby affecting symmetry. The symmetry of the crystal structure may affect the correlation of its physical properties. For example, a highly symmetric crystal structure may have a more regular property distribution, so the correlation coefficient is higher. Crystal structures with high symmetry tend to have simpler correlations because symmetry can simplify the parameters for describing the crystal structure. Changes in the unit cell volume may affect the physical and chemical properties of the material, and the changes in these properties can be quantified by the correlation coefficient.

[0092] Based on the structure characteristic values of each crystal structure, corresponding feature learning is performed on the local environment between the atoms of each crystal structure, and the preliminary energy of each crystal structure is output, denoted as the neural network potential of each crystal structure, and the neural network potential evaluation result of each crystal structure is obtained through analysis and processing.

[0093] The corresponding feature learning on the local environment between the atoms of each crystal structure specifically includes:

[0094] Based on the structure characteristic values of each crystal structure, mapping matching is performed with the local segmentation ratios corresponding to the pre-stored structure characteristic value intervals in the structure prediction platform to obtain the local segmentation ratios of each crystal structure. In the embodiments of the present invention, the larger the structure characteristic value, the smaller the local segmentation ratio.

[0095] Perform corresponding local segmentation on each crystal structure based on the local segmentation ratio of each crystal structure to obtain the local environments of each crystal structure.

[0096] Convert each crystal structure into a graph structure based on each atomic coordinate and input it into the MEGNet model. Denote each atom of each crystal structure as each node, obtain the local environment interaction information and the overall environment interaction information of each node of each crystal structure, and perform layer-by-layer update on each node through the message passing mechanism to output the feature representation vectors of each node of each crystal structure.

[0097] The local environment interaction information refers to the interaction between each node and its directly adjacent nodes. In the graph structure, this is equivalent to the adjacency information of the node, including the types of adjacent atoms, distance (bond length), angle (bond angle), and charge distribution. The local environment interaction information is crucial for understanding the chemical properties of atoms because it reflects the direct chemical environment of the atoms, which directly affects the electronic state and reactivity of the atoms. The type of adjacent atom refers to the type of other atoms directly connected to the target atom. For example, in a molecule, a certain carbon atom may be connected to hydrogen atoms and oxygen atoms. The distance (bond length) refers to the straight-line distance between the target atom and its adjacent atom, which usually corresponds to the length of the chemical bond. The bond length can affect the chemical properties and reactivity of the atoms. The angle (bond angle) refers to the angle formed by the target atom and its two adjacent atoms, which describes the spatial structure in the molecule or crystal. The bond angle has an important impact on the geometric shape and properties of the molecule. The charge distribution refers to the distribution of electrons around the chemical bond, which affects the polarity of the atoms and the intermolecular forces.

[0098] The overall environmental interaction information refers to the comprehensive influence of the entire crystal structure on a single atom. It includes not only local interactions but also the indirect influence of atoms at greater distances in the crystal on a specific atom, as well as the influence of crystal symmetry and periodicity on atoms. The overall environmental interaction information helps to capture the long-range order and periodicity characteristics in the crystal structure. The overall environmental interaction information includes local interactions, indirect influences, crystal symmetry and periodicity, and long-range order. Although local interactions overlap with local environmental interaction information, in the overall environment, it is regarded as part of a larger-scale interaction. Indirect influence refers to the influence of atoms at a relatively far distance in the crystal on the electronic state and chemical properties of the target atom. These influences may be transmitted through the diffusion of the electron cloud, the modulation of the crystal field, etc. Crystal symmetry and periodicity refer to the regular arrangement and repeating patterns in the crystal structure, which have a long-term impact on the electronic state and physical properties of atoms. For example, certain atoms in the crystal may have specific energy level splittings due to symmetry. Long-range order refers to the regularity and consistency of the atomic arrangement throughout the crystal, which affects the macroscopic properties of the material, such as conductivity, magnetism, etc.

[0099] The message passing mechanism means that in a graph neural network, message passing is an algorithmic process where each node updates its own features based on the features of its neighbor nodes and the features of the edges. In the MEGNet model, this mechanism is achieved by propagating and aggregating information in the network. Each node receives "messages" from its neighbor nodes, and these messages contain the state information of the neighbor nodes.

[0100] Layer-by-layer update refers to repeating the message passing process in multiple layers of the network. In each layer, the feature vector of the node is updated based on the messages received from the neighbor nodes. This update is iterative, and each layer may capture higher-level abstract features. For example, the first layer may only capture the information of direct neighbors, while subsequent layers may be able to capture more distant interactions or more complex structural patterns.

[0101] After multiple layers of message passing and update, each node will have a feature representation vector that encodes the local and overall environmental information of the node in the crystal structure.

[0102] The interaction between each node is denoted as an edge vector. Based on the feature representation vectors of the two nodes connected by each edge vector, through the graph convolution layer, iterative update is performed to output the feature representation vectors of each edge of each crystal structure.

[0103] In the graph structure, the edge vector represents the interaction between two nodes (atoms). This vector contains various information, such as the type of chemical bond between the two atoms (covalent bond, ionic bond, etc.), bond length, bond angle, charge transfer, etc. Edge vectors are important elements in graph neural networks for representing the relationships between nodes, and they carry detailed information about the interactions between nodes.

[0104] The graph convolutional layer is the core component of a graph neural network, which mimics the convolutional operation in traditional convolutional neural networks for image processing but is applicable to graph-structured data. In the graph convolutional layer, the feature vector of each node is updated based on the feature vectors of its neighboring nodes and the edge vectors connecting them. This process is iterative, meaning that in each layer of the network, the feature vectors of the nodes are updated. During the iterative update process in the graph convolutional layer, not only are the feature vectors of the nodes updated, but the edge vectors are also updated accordingly based on the updates of the nodes. Eventually, each edge will have an updated feature representation vector, which reflects how the interaction of that edge affects the entire graph structure after considering the information of the surrounding nodes.

[0105] Each node (atom) has a feature vector that encodes the properties of the node, such as the atomic species, charge state, position, etc.

[0106] The MEGNet (Material Graphs Network) model is a deep learning framework for predicting material properties. The MEGNet model combines graph neural networks (Graph Neural Networks, GNNs) and physics-inspired methods, and is specifically designed to process data in materials science.

[0107] MEGNet uses graph neural networks to represent the crystal structure of materials. In graph neural networks, atoms are regarded as nodes and chemical bonds are regarded as edges. This representation method can effectively capture the three-dimensional structural information of materials. The MEGNet model not only considers the local atomic environment in the crystal structure but also integrates global properties such as unit cell parameters and space group information. This multi-scale representation helps to understand the properties of materials more comprehensively. Some physical information, such as atomic charge, bond length, and bond angle, is embedded in the model, which helps the model better understand the electronic structure and other related properties of materials.

[0108] MEGNet uses a message-passing mechanism to update the states of nodes (atoms) and edges (chemical bonds). In this process, the feature vector of each atom is updated based on the feature vectors of its neighboring atoms.

[0109] In graph neural networks, the readout layer is used to extract global features from a graph to generate a representation of the entire material. In MEGNet, this layer is used to generate the final feature vector for predicting material properties.

[0110] The advantage of the MEGNet model lies in its ability to learn highly abstract features from complex material structures, which are closely related to the physical and chemical properties of materials.

[0111] The preliminary energies of the crystal structures are output, specifically including:

[0112] Based on the feature representation vectors of the nodes and the feature representation vectors of the edges of each crystal structure, the feature representation values of the nodes of each crystal structure are obtained after being processed by the graph convolutional layer. The feature representation values of the nodes are used to describe the local environment information and interaction information of the nodes in the corresponding crystal structure.

[0113] The feature representation values of the nodes of each crystal structure are input into a neural network model, and through processing and analysis, the preliminary energy of each crystal structure is obtained, denoted as the neural network potential of each crystal structure.

[0114] ;

[0115] Where is the neural network potential of the th crystal structure, is the feature representation value of the th node of the th crystal structure, are the neural network parameters, is the expression of the neural network model, is the crystal structure number, , is the total number of crystal structures, is the node number, , is the total number of nodes, is the atomic species number, , is the total number of atomic species.

[0116] It should be noted that the neural network parameters refer to the coefficients connecting the neurons in the neural network. They determine the influence degree of the input features on the output results, that is, the importance of each feature in the prediction.

[0117] The analysis and processing obtain the evaluation results of the neural network potential of each crystal structure, including:

[0118] Extract the neural network potential threshold from the structure prediction platform. If the neural network potential of a certain crystal structure is greater than or equal to the neural network potential threshold, it is determined that the evaluation result of the neural network potential of this crystal structure is unstable.

[0119] If the neural network potential of a certain crystal structure is less than the neural network potential threshold, it is determined that the evaluation result of the neural network potential of this crystal structure is stable, and the crystal structures with stable evaluation results of the neural network potential are extracted and recorded as each stable crystal structure.

[0120] It should be noted that the neural network potential threshold is processed as follows: In a specific embodiment, the neural network potential threshold is directly set in the structure prediction platform during the development of the crystal structure prediction system based on the hybrid potential deep learning model involved in the embodiments of the present invention. There are various methods for setting the neural network potential threshold. For example, it is obtained through statistical analysis methods, including statistically analyzing the neural network potentials of historical stable crystal structures and performing mean processing on the neural network potentials obtained multiple times to obtain the neural network potential threshold.

[0121] Based on the evaluation results of the neural network potentials of each crystal structure, screen out each stable crystal structure of the material to be predicted, calculate the short-range interactions between atoms of each stable crystal structure, and obtain the empirical potential of each stable crystal structure through cumulative processing.

[0122] The calculation of the short-range interactions between atoms of each stable crystal structure, performing cumulative processing and recording it as the empirical potential of each stable crystal structure specifically includes:

[0123] The short-range interactions include a repulsive term, an attractive term, and an exponential decay.

[0124] The repulsive term refers to when atoms are very close, due to the repulsion of the electron clouds, a strong repulsive force will be generated between the atoms, and this term describes this repulsive effect.

[0125] The attractive term refers to as the distance between atoms increases, there will be van der Waals forces or other attractive effects between the atoms, and this term describes this attractive effect.

[0126] The exponential decay means that this term indicates that as the distance between atoms increases, the interaction force will rapidly weaken.

[0127] Statistical analysis of the atomic coordinates of each atom in each stable crystal structure to obtain the spatial distances between each atom, using the Buckingham potential model to calculate the short-range interactions between atoms of each stable crystal structure, and accumulating the short-range interactions of each atom pair and recording the result as the empirical potential of each stable crystal structure, specifically including:

[0128] ;

[0129] ;

[0130] Among them, is the empirical potential of the th stable crystal structure, represents the th crystal structure's th node and the th node's distance. A is the strength coefficient of the repulsive term, B is the constant parameter controlling the decay rate of the repulsive term, C is the strength coefficient of the attractive term, and are both node numbers, and d, , , is the number of the stable crystal structure, , is the total number of stable crystal structures, .

[0131] It should be noted that the strength coefficient of the repulsive term, the constant parameter controlling the decay rate of the repulsive term, and the strength coefficient of the attractive term are obtained by density functional theory (DFT) calculations.

[0132] It should also be noted that the Buckingham potential model is an empirical potential energy function widely used in molecular dynamics and materials science to describe the short-range interactions between atoms. This model takes into account the repulsive and attractive forces between atoms. When the atoms are very close, the repulsive force becomes very strong, which prevents the atoms from overlapping each other. When there is a certain distance between the atoms, the attractive force plays a major role, which reflects the van der Waals force or London dispersion force between the atoms. The Buckingham potential model mainly describes short-range interactions, that is, the forces when the atomic distances are relatively close.

[0133] The neural network potential and empirical potential of each stable crystal structure are comprehensively analyzed and processed to obtain the total energy potential of each stable crystal structure, and then imported into the optimization algorithm for iterative update. The stable crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted.

[0134] The process of comprehensively analyzing and processing the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and then importing it into the optimization algorithm for iterative update, and outputting the stable crystal structure with the lowest total energy potential as the target crystal structure of the material to be predicted specifically includes:

[0135] The process of analyzing and processing to obtain the total energy potential of each stable crystal structure, and the specific processing conditions are:

[0136] ;

[0137] wherein, is the total energy potential of the th stable crystal structure, is the neural network potential of the th stable crystal structure, is the empirical potential of the th stable crystal structure, is the energy balance coefficient, is the number of the stable crystal structure, , is the total number of the stable crystal structures.

[0138] It should be noted that, in the embodiments of the present invention, the energy balance coefficient is obtained by selecting a group of materials with known structures, calculating the energies of this group of materials using high-precision first-principles, and recording them as reference data. The energy balance coefficient in the hybrid model is adjusted using the reference data, and the optimal energy balance coefficient is obtained by minimizing the error function.

[0139] Based on the neural network potential and the empirical potential of each stable crystal structure, the total energy potential of each stable crystal structure is obtained through analysis and processing, and is compared and analyzed with the total energy potential threshold preset in the structure prediction platform. The stable crystal structures with total energy potential lower than the total energy potential threshold are counted and recorded as each predicted crystal structure.

[0140] As Figure 2 shown, it is a comparison of the energy calculation results using the hybrid potential energy model and DFT, as well as a comparison of the time with the DFT-PSO algorithm. The energy predicted by the method of the embodiments of the present invention is close to the DFT value, and the time for searching the crystal structure is better than the DFT-PSO algorithm, lower than three orders of magnitude. It should be noted that the Chinese names of the chemical formulas on the horizontal axis in the figure are sodium fluoride, potassium fluoride, rubidium fluoride, cesium fluoride, lithium chloride, sodium chloride, potassium chloride, rubidium chloride, cesium chloride, calcium oxide, strontium oxide, cadmium oxide, cobalt sulfide, zinc sulfide, carbon, silicon, gallium arsenide, and cadmium telluride in sequence.

[0141] DFT-PSO is a method that combines Density Functional Theory (DFT) and Particle Swarm Optimization (PSO) algorithm. DFT is a quantum mechanical method widely used to calculate the electronic structures of atoms, molecules, and solid materials. DFT approximately describes the behavior of many-electron systems by solving the Kohn-Sham equations. PSO is a computational method in which each "particle" represents a candidate solution in the problem space and moves in the solution space through an iterative process to find the optimal solution. The particles adjust their positions based on their own experience and the experience of neighboring particles. Combining DFT and PSO, namely DFT-PSO, means using the PSO algorithm to optimize the process of material structure, while DFT is used to evaluate the energy of each candidate structure.

[0142] It should be noted that the total energy potential threshold is processed as follows: In a specific embodiment, the total energy potential threshold is directly set in the structure prediction platform during the development of the crystal structure prediction system based on the hybrid potential deep learning model involved in the embodiments of the present invention. There are various methods for setting the total energy potential threshold. For example, it can be obtained through statistical analysis, including statistically analyzing the total energy potential of historical stable crystal structures and performing mean processing on the total energy potentials obtained multiple times to obtain the total energy potential threshold.

[0143] The total energy potentials of each predicted crystal structure are imported into the Bayesian optimization algorithm for optimization processing. When the Bayesian optimization algorithm converges, the predicted crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted.

[0144] The Bayesian optimization algorithm is an algorithm for global optimization, suitable for cases where the calculation cost of the objective function is high and the definition is not clear. In material design, the objective function is usually a material property prediction model, such as energy.

[0145] The Bayesian optimization algorithm searches for the input value that minimizes the objective function by iteratively updating the model of the objective function. In the embodiments of the present invention, the goal is to find the crystal structure with the lowest total energy potential. When the Bayesian optimization algorithm reaches a certain stopping condition, the algorithm is considered to have converged. When the algorithm converges, the crystal structure with the lowest predicted total energy is output.

[0146] It needs to be explained that the Bayesian optimization algorithm is an optimization method based on a probability model, which mainly consists of the following steps:

[0147] Probability model (Surrogate Model): The algorithm first establishes a probability model to approximate the objective function. Gaussian Processes (GPs) are the most commonly used probability models.

[0148] Acquisition Function: The acquisition function is used to determine the next sampling point. Based on the probability model, sampling is carried out in both the unknown region and the known optimal region. Common acquisition functions include Expected Improvement (EI), Probability of Improvement (PI), and Upper Confidence Bound (UCB).

[0149] In each iteration, the acquisition function is used to select the next sampling point. The total energy of the crystal structure is calculated at the selected sampling point. The new sampling point is added to the dataset, and the probability model is updated. Repeat the above iterative process until the stopping condition is met, such as achieving the preset algorithm convergence.

[0150] As Figure 3 and Figure 4 shown, Figure 3 Figure (a) is the crystal structure of CaS corresponding to mp-22862 in the Materials Project material library, Figure 4 and figure (b) is the crystal structure with the lowest energy searched by the method proposed in the present invention.

[0151] In this embodiment, the present invention provides a crystal structure prediction system based on a hybrid potential deep learning model, including:

[0152] A crystal structure feature module, which is used to input the basic information of the material to be predicted on the structure prediction platform, generate the characteristic values of the basic information of the material to be predicted and conduct a rationality evaluation. After obtaining the rationality evaluation result, the structural characteristic values of each crystal structure dataset that meets the material to be predicted are retrieved from the structure prediction platform based on the characteristic values of the basic information of the material to be predicted.

[0153] A neural network potential calculation module, which is used to perform corresponding feature learning on the local environment between atoms of each crystal structure based on the structural characteristic values of each crystal structure, output the preliminary energy of each crystal structure, denoted as the neural network potential of each crystal structure, and analyze and process to obtain the neural network potential evaluation result of each crystal structure.

[0154] An empirical potential calculation module, which is used to screen out each stable crystal structure of the material to be predicted based on the neural network potential evaluation results of each crystal structure, calculate the short-range interaction between atoms of each stable crystal structure, and obtain the empirical potential of each stable crystal structure through cumulative processing.

[0155] The target crystal structure output module is used to comprehensively analyze and process the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and import it into the optimization algorithm for iterative update. The stable crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted.

[0156] 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 such 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.

[0157] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, according to the content of this specification, many modifications and changes can be made. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications 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 prediction method based on a hybrid potential energy deep learning model, characterized in that: include: Input the basic information of the material to be predicted into the structure prediction platform, generate the basic information characteristic value of the material to be predicted and conduct a rationality evaluation, and after obtaining the rationality evaluation result, retrieve various crystal structure data sets that meet the material to be predicted from the structure prediction platform based on the basic information characteristic value of the material to be predicted, and merge to obtain the structural characteristic value of each crystal structure; Based on the structural characteristic values ​​of each crystal structure, the corresponding characteristic learning is performed on the local environment between atoms of each crystal structure, and the preliminary energy of each crystal structure is output, which is recorded as the neural network potential of each crystal structure, and the neural network potential evaluation results of each crystal structure are obtained by analysis and processing; Based on the neural network potential evaluation results of each crystal structure, the stable crystal structures of the material to be predicted are screened, and the short-range interaction information between atoms of each stable crystal structure is calculated. The empirical potential of each stable crystal structure is obtained through accumulation processing; The total energy potential of each stable crystal structure is obtained by comprehensive analysis of the neural network potential and empirical potential of each stable crystal structure, and then imported into the optimization algorithm for iterative update. The stable crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted. The corresponding feature learning of the local environment between atoms in each crystal structure specifically includes: Based on the structural characteristic values ​​of each crystal structure, mapping and matching are performed with the local segmentation ratios corresponding to the structural characteristic value intervals pre-stored in the structure prediction platform to obtain the local segmentation ratios of each crystal structure; Performing corresponding local segmentation on each crystal structure based on the local segmentation ratio of each crystal structure to obtain each local environment of each crystal structure; Each crystal structure is converted into a graph structure and input into the MEGNet model, and each atom of each crystal structure is recorded as a node, and the local environment interaction information and the overall environment interaction information of each node of each crystal structure are obtained. Each node is updated layer by layer through the message passing mechanism, and the feature representation vector of each node of each crystal structure is output; The interaction between each node is recorded as an edge vector. Based on the feature representation vectors of the two nodes connected by each edge vector, the feature representation vectors of each edge of each crystal structure are iteratively updated through the graph convolution layer. The calculating of the short-range interaction information between atoms of each stable crystal structure and obtaining the empirical potential of each stable crystal structure by accumulation processing specifically includes: The coordinates of each atom in each stable crystal structure are counted, and the spatial distances between atoms are obtained through analysis and processing. The short-range interaction information between atoms in each stable crystal structure is calculated using the Buckingham potential model. The short-range interaction information of each atomic pair is accumulated and the result is recorded as the empirical potential of each stable crystal structure.

2. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 1, characterized in that: The basic information of the material to be predicted is input into the structure prediction platform, the basic information characteristic value of the material to be predicted is generated and the rationality evaluation is performed. The specific process is as follows: The basic information of the material to be predicted includes the total number of atomic species, the number of atoms of each type, the charge number of atoms of each type, the total number of chemical bond species, the number of chemical bonds, the bond lengths of chemical bonds of each type, and the bond angles of chemical bonds of each type. The reference bond angles and reference bond lengths corresponding to various types of chemical bonds are extracted from the structure prediction platform, and are jointly analyzed and processed with the basic information of the material to be predicted to generate the basic information characteristic value of the material to be predicted, and compared with the rationality assessment threshold to obtain the rationality assessment result of the material to be predicted. If the basic information characteristic value of the material to be predicted is greater than or equal to the rationality assessment threshold, the rationality assessment result of the material to be predicted is determined to be reasonable; if the basic information characteristic value of the material to be predicted is less than the rationality assessment threshold, the rationality assessment result of the material to be predicted is determined to be unreasonable, and a reminder message is sent to the structure prediction platform.

3. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 1, characterized in that: The method of retrieving various crystal structure data sets that meet the material to be predicted from the structure prediction platform based on the basic information characteristic value of the material to be predicted and merging them to obtain the structural characteristic value of each crystal structure specifically includes: Based on the basic information characteristic values ​​of the material to be predicted, mapping and matching are performed with each crystal structure data set corresponding to each basic information characteristic value interval pre-stored in the structure prediction platform to obtain each crystal structure data set of the material to be predicted, wherein each crystal structure data set of the material to be predicted includes each side length of the unit cell of each crystal structure, unit cell angle, space group symbol and each atomic coordinate; Based on the space group symbol of each crystal structure, mapping and matching are performed with the symmetry coefficients corresponding to each space group symbol pre-stored in the structure prediction platform to obtain the symmetry coefficient of each crystal structure, wherein the symmetry coefficient is used to characterize the space group symmetry characteristics of the crystal structure; Based on the coordinates of each atom in each crystal structure, the correlation coefficient of each crystal structure is obtained by analysis and processing, and the correlation coefficient of each crystal structure is used to characterize the order degree of each crystal structure; The lengths of the sides of the unit cells of each crystal structure, the unit cell angles, the symmetry coefficients and the correlation coefficients are processed and analyzed to obtain the structural characteristic values ​​of each crystal structure, and the structural characteristic values ​​are used to describe the crystal structure characteristics.

4. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 3 is characterized in that: The outputting of the preliminary energy of each crystal structure specifically includes: Based on the feature representation vectors of each node of each crystal structure and the feature representation vectors of each edge of each crystal structure, a feature representation value of each node of each crystal structure is obtained after processing by a graph convolution layer, wherein the feature representation value of each node is used to describe the local environment information and interaction information of each node in the corresponding crystal structure; The characteristic representation value of each node of each crystal structure is input into the neural network model, and the preliminary energy of each crystal structure is obtained through processing and analysis, which is recorded as the neural network potential of each crystal structure.

5. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 1, characterized in that: The analysis and processing obtains the neural network potential evaluation results of each crystal structure, including: Extracting a neural network potential threshold from the structure prediction platform, if the neural network potential of a certain crystal structure is greater than or equal to the neural network potential threshold, then determining that the neural network potential evaluation result of the crystal structure is unstable; If the neural network potential of a certain crystal structure is less than the neural network potential threshold, the neural network potential evaluation result of the crystal structure is determined to be stable, and the crystal structures whose neural network potential evaluation results are stable are extracted and recorded as stable crystal structures.

6. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 1, characterized in that: The neural network potential and empirical potential of each stable crystal structure are comprehensively analyzed and processed to obtain the total energy potential of each stable crystal structure, and the total energy potential of each stable crystal structure is imported into the optimization algorithm for iterative update, and the stable crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted, specifically including: Based on the neural network potential and empirical potential of each stable crystal structure, the total energy potential of each stable crystal structure is obtained by analysis and processing, and compared and analyzed with the total energy potential threshold preset in the structure prediction platform, and the stable crystal structures with total energy potential lower than the total energy potential threshold are counted and recorded as each predicted crystal structure; The total energy potential of each predicted crystal structure is introduced into the Bayesian optimization algorithm for optimization. When the Bayesian optimization algorithm converges, the predicted crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted.

7. The crystal structure prediction method based on the hybrid potential energy deep learning model according to claim 1, characterized in that: The analysis process obtains the total energy potential of each stable crystal structure, and the specific processing conditions are: ； in, For the The total energy potential of a stable crystal structure, For the A neural network potential with a stable crystal structure, For the The empirical potential of a stable crystal structure, is the energy balance coefficient, is the number of stable crystal structures, , is the total number of stable crystal structures.

8. A system using the crystal structure prediction method based on the hybrid potential energy deep learning model according to any one of claims 1 to 7, characterized in that: include: A crystal structure characteristic module is used to input the basic information of the material to be predicted on the structure prediction platform, generate the basic information characteristic value of the material to be predicted and perform a rationality evaluation, and after obtaining the rationality evaluation result, retrieve various crystal structure data sets that meet the material to be predicted from the structure prediction platform based on the basic information characteristic value of the material to be predicted, and merge to obtain the structural characteristic value of each crystal structure; A neural network potential calculation module is used to perform corresponding feature learning on the local environment between atoms of each crystal structure based on the structural characteristic values ​​of each crystal structure, output the preliminary energy of each crystal structure, record it as the neural network potential of each crystal structure, and analyze and process to obtain the neural network potential evaluation results of each crystal structure; The empirical potential calculation module is used to screen out the stable crystal structures of the materials to be predicted based on the neural network potential evaluation results of each crystal structure, calculate the short-range interaction information between atoms of each stable crystal structure, and obtain the empirical potential of each stable crystal structure through accumulation processing; The target crystal structure output module is used to comprehensively analyze and process the neural network potential and empirical potential of each stable crystal structure to obtain the total energy potential of each stable crystal structure, and import it into the optimization algorithm for iterative update. The stable crystal structure with the lowest total energy potential is output as the target crystal structure of the material to be predicted.

Images