Stable crystal structure prediction method and system based on machine learning and target optimization
Through a machine learning and target optimization method, using graph neural network and Lennard-Jones potential function, combined with Bayesian optimization algorithm, the problem of long calculation time and insufficient stability measurement in crystal structure prediction is solved, and the efficient screening of stable crystal structures is achieved.
Patent Information
- Application Number
- CN202311343525.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-17
- Publication Date
- 2025-08-12
- Estimated Expiration
- 2043-10-17
AI Technical Summary
In the prior art, the crystal structure prediction method has the problem of high DFT calculation time cost and insufficient measurement of single energy stability, which leads to a large time-consuming search for crystal structures and may not necessarily find a stable crystal structure.
Using a method based on machine learning and target optimization, a graph neural network is used instead of DFT calculation, combined with the Lennard-Jones potential function and Bayesian optimization algorithm, a stable crystal structure is selected through the crystal formation target optimization function and potential energy.
It effectively reduces the DFT calculation time, improves the efficiency and accuracy of crystal structure prediction, and can find a new stable crystal structure.
Smart Images
Figure QLYQS_1 
Figure BDA0004498032870000051 
Figure HDA0004498032880000011
Abstract
Description
Technical Field
[0001] The present invention relates to the field of crystal structure technology, and in particular to a stable crystal structure prediction method and system based on machine learning and target optimization. Background Art
[0002] Crystal structure prediction is an important technology for discovering new materials and understanding the properties of materials. With the continuous improvement and perfection of theoretical methods and the rapid development of computer technology, it has become possible to use crystal structure prediction to discover new materials. The most mainstream method of crystal structure prediction is to use the energy calculated by DFT as a stability metric based on the chemical composition of the primary material, and then combine it with a search optimization algorithm to find a crystal structure that meets the requirements. Through this method, many structure prediction software have been developed, such as CALYPSO, USPEX, AIRSS, GAtor, etc. These software predict stable or metastable crystal structures under the chemical composition conditions of a given compound. They can be used to predict or determine clusters, two-dimensional layer structures, two-dimensional surfaces and three-dimensional crystal structures and design multifunctional materials. However, this method still has some problems in the structure search process: (1) The DFT calculation time cost is high, resulting in a long crystal structure search; (2) The stability of the crystal structure is affected by many factors. Selecting a single energy as a stability metric does not guarantee that a stable crystal structure will be found. Summary of the Invention
[0003] To address the shortcomings of the aforementioned background technology, the present invention aims to address the problem that the stability of crystal structures is affected by multiple factors, and that selecting a single energy as a stability metric does not guarantee that a stable crystal structure will be found. This invention provides a method and system for predicting stable crystal structures based on machine learning and target optimization. This method implements crystal energy prediction, structure search, and atomic bonding screening by running code, thereby obtaining new stable structures.
[0004] To achieve the above objectives, the present invention provides, in a first aspect, a method for predicting stable crystal structures based on machine learning and target optimization, comprising:
[0005] Obtain the chemical composition of the primary crystals;
[0006] Set the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtain the atomic coordinates of the primary crystal;
[0007] According to the atomic coordinates of the primary crystal, the data set was obtained from the crystal database;
[0008] The graph neural network model is trained based on the acquired data set to obtain the formation energy of the primary crystal;
[0009] Use potential function to obtain the potential energy of the primary crystal;
[0010] According to the formation energy and potential energy of the primary crystal, a target optimization function is set, and the primary crystal structure is optimized using the target optimization function to obtain a contact map of the optimized crystal structure;
[0011] Based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure, a stable crystal structure is screened out.
[0012] Preferably, the chemical composition of the primary crystal includes the types and numbers of atoms in the primary crystal.
[0013] Preferably, the data set includes stable structure data and metastable structure data of crystal formation energy.
[0014] Preferably, the dataset is divided into training set, test set and validation set in a ratio of 5:1 to train the graph neural network model.
[0015] Preferably, during the training of the graph neural network model, stable structure data and metastable structure data are used as input, and the formation energy and band gap of the crystal are used as output.
[0016] Preferably, the potential function used is the Lennard-Jones potential function.
[0017] Preferably, the objective optimization function is as follows:
[0018] f=ΔH+λ*|E|
[0019] Where f is the target optimization function, ΔH is the formation energy, λ is the weight coefficient, and its value range is [0-1]; E| is the absolute value of the potential energy between two molecules.
[0020] Preferably, the spatial characteristics of the crystal include space group, symmetry, and Wyckoff position.
[0021] A second aspect of the present invention provides a stable crystal structure prediction system based on machine learning and target optimization, comprising:
[0022] A data acquisition module is used to obtain the chemical composition of the primary crystal; set the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtain the atomic coordinates of the primary crystal; and obtain a data set from a crystal database based on the atomic coordinates of the primary crystal;
[0023] The crystal energy acquisition module is used to train the graph neural network model based on the acquired data set to obtain the formation energy of the primary crystal; use the potential function to obtain the potential energy of the primary crystal; set the target optimization function based on the formation energy and potential energy of the primary crystal, and use the target optimization function to optimize its primary crystal structure to obtain the contact map of the optimized crystal structure;
[0024] The atomic bonding screening module is used to screen out stable crystal structures based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] The present invention provides a method and system for predicting a stable crystal structure based on machine learning and target optimization. The method only requires the operator to provide the chemical composition and space group parameters of the predicted crystal, and the energy prediction, structure search, and atomic bonding screening of the crystal are realized by running the code, so that a new stable structure can be obtained on this basis. Compared with the traditional crystal structure prediction framework, this method uses a graph neural network instead of DFT calculation, improves the objective function of the optimization algorithm, and takes into account thermodynamic stability and kinetic stability. In addition, on this basis, an atomic bonding screening module based on a contact graph is added. Through this method, the calculation time of DFT can be effectively reduced, and a new stable crystal structure can be obtained. The method is improved according to the mainstream crystal structure prediction process, and the efficiency of structure calculation and screening is improved through deep learning and target optimization technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A flow chart of a stable crystal structure prediction method based on deep learning and target optimization provided by the present invention.
[0028] Figure 2 is a representation of BPO4 crystals.
[0029] Figure 3 This is the result diagram of crystal formation energy prediction.
[0030] Figure 4 Results of a crystal structure prediction method based on deep learning and target optimization. DETAILED DESCRIPTION
[0031] In order to enable those skilled in the art to better understand and implement the technical solution of the present invention, the present invention is further described below with reference to specific embodiments and drawings, but the embodiments are not intended to limit the present invention.
[0032] The present invention provides a stable crystal structure prediction method based on machine learning and target optimization, see Figure 1 As shown, including:
[0033] S1. Obtain the chemical composition of the primary crystal;
[0034] The chemical composition of the primary crystal includes the types and numbers of atoms in the primary crystal.
[0035] In this embodiment, the chemical composition of the primary material is first introduced. The chemical composition of the primary material is the type and number of atoms in the primary crystal.
[0036] S2. Setting the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtaining the atomic coordinates of the primary crystal;
[0037] The spatial characteristics of the crystal include space group, symmetry, and Wyckoff position.
[0038] In this embodiment, the representation of crystal structure characteristics determines the space group, symmetry, and Wyckoff position of the crystal based on the chemical composition of the crystal, thereby obtaining the atomic coordinates of the random primary crystal;
[0039] Based on the imported crystal chemical composition, the space group range is set, and the appropriate symmetry and Wyckoff position are selected to obtain the coordinates of the atoms in the crystal structure. Figure 2 As shown, it represents the crystal structure diagram, crystal parameter information, and crystal contact diagram of crystal BPO4, where the crystal parameter information mainly includes crystal composition, lattice, atomic position, and symmetry.
[0040] S3. Obtain a data set from a crystal database based on the atomic coordinates of the primary crystal;
[0041] The data set includes stable structure data and metastable structure data of crystal formation energies.
[0042] In this embodiment, a crystal data set is constructed by downloading stable and metastable structure data containing crystal formation energy from a crystal database, and the data set is divided into a training set and a validation set in a ratio of 5:1 to train a graph neural network model.
[0043] During the training of the graph neural network model, stable and metastable structure data are used as input, and the crystal formation energy and band gap are used as output. Specifically, the model input is stable and metastable crystal structure data from 60,000 Materials Projects. This dataset primarily includes crystal energies and band gaps calculated using DFT. Of this data, 48,000 are used for model training, and the remaining data is used for testing and validation. The crystal formation energy and band gap are used as outputs, and the trained graph neural network model is used to predict the crystal formation energy and band gap.
[0044] S4. Based on the acquired data set, a graph neural network model is trained to predict the formation energy.
[0045] The graph neural network model is trained based on the acquired data set to predict the crystal formation energy. Figure 3The prediction performance of the graph neural network model for formation energy is shown in Figure 2. The blue dots are the predicted values, the purple solid line is the fitting function of the predicted values, and the red dashed line is the fitting function of the formation energy calculated using DFT. The purple solid line and the red dashed line basically coincide with each other, indicating that the predicted results are very close to the calculated results.
[0046] S5. Using the potential function to obtain the potential energy of the primary crystal;
[0047] It should be noted that the prediction of material properties uses a constructed dataset to train a graph neural network model to predict the formation energy of the primary crystal. The potential function is calculated using the Lennard-Jones potential function formula to calculate the potential energy corresponding to the selected crystal structure.
[0048] The potential function used is the Lennard-Jones potential function. The specific potential function is calculated using the Lennard-Jones potential function formula to calculate the potential energy corresponding to the selected crystal structure;
[0049]
[0050] Where E is the potential energy between two molecules, ε represents the strength of the attraction term, σ represents the effective diameter between molecules, and r is the distance between the two molecules. ij ) 12 It represents the long-range attraction, which gradually decreases as the distance r between the two molecules increases. The second term (σ / r ij ) 6 It indicates the short-range repulsive force, which increases rapidly as the distance r decreases.
[0051] S6. Setting a target optimization function according to the formation energy and potential energy of the primary crystal;
[0052] In this embodiment, the objective function is set to construct an objective function including the formation energy and the Lennard-Jones potential function to perform search optimization. The objective function is modified in the optimization algorithm, and the optimization parameters are set.
[0053] The objective optimization function is as follows:
[0054] f=ΔH+λ*|E| (2)
[0055] Where f is the target optimization function, ΔH is the formation energy, λ is the weight coefficient, and its value range is [0-1]; |E| is the absolute value of the potential energy between two molecules.
[0056] S7, iteratively optimizing the primary crystal structure according to the set target optimization function;
[0057] According to formula (2), the Bayesian optimization algorithm is used for multiple iterations to obtain a stable crystal structure that satisfies the minimum formation energy and the intermolecular potential energy approaches 0. Using the Bayesian optimization algorithm to continuously iterate the initial crystal structure can predict more stable crystal structures.
[0058] S8. Based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure, a stable crystal structure is screened. Figure 4 As shown in the figure, it is the crystal structure diagram of Cu2As3 and its structure optimization process. Figure 4 The red circles in (A) represent the crystals found during the 5000-step iterative optimization process. Figure 4 The label in (A) indicates that the crystal formation energy obtained at step 3702 is -2.13596, which is the lowest crystal formation energy. Figure 4 (B) is the crystal structure diagram of Cu2As3 corresponding to the number 1245668 in the Materials Project material library. Figure 4 (B) is the Cu2As3 crystal structure diagram predicted using the method of the present invention, and the corresponding crystal has the lowest formation energy.
[0059] In this embodiment, the contact map of the optimized crystal structure is analyzed to determine the atomic bonding conditions of the searched crystal and select a more stable crystal structure.
[0060] The method provided by the present invention only requires the operator to provide the chemical composition and space group parameters of the predicted crystal, and by running the code to realize the energy prediction, structure search, and atomic bonding screening of the crystal, a new stable structure can be obtained on this basis. Compared with the traditional crystal structure prediction framework, this solution uses a graph neural network instead of DFT calculation, improves the objective function of the optimization algorithm, and takes into account thermodynamic stability and kinetic stability. In addition, on this basis, an atomic bonding screening module based on a contact graph is added. Through this solution, the calculation time of DFT can be effectively reduced, and a new stable crystal structure can be obtained.
[0061] This method is improved according to the mainstream crystal structure prediction process, and the efficiency of structure calculation and screening is improved through deep learning and target optimization technology. Figure 1 The flowchart of the stable crystal structure prediction method based on deep learning and target optimization is elaborated in detail:
[0062] (1) First, a crystal formation energy prediction dataset was constructed. This dataset greatly affects the accuracy of crystal formation energy prediction. 60,000 stable and metastable crystal structure data from the Material Project were selected to train and validate the graph neural network model, of which 48,000 crystal structure data were used for model training and 12,000 crystal structure data were used for validation. This dataset contains the chemical composition parameters, lattice parameters, and formation energy of the crystal structure, with the goal of training a crystal formation energy prediction model.
[0063] (2) A crystal formation energy prediction model is constructed. The model mainly consists of a MEGNet layer to update the matrices {vi} and {ek}, and two set2set layers to learn a representation vector from the matrices {vi} and {ek} respectively. These vectors are then combined using a concatenate layer, and then traversed through a fully connected layer composed of multiple dense layers to obtain the formation energy.
[0064] (3) The chemical composition of the predicted crystal is input. A symmetry is selected from 230 space groups through random screening. Then, lattice parameters are randomly generated based on the selected symmetry. Similarly, given the number of atoms and symmetry, appropriate Wyckoff positions are selected. Based on this, the corresponding atomic coordinates are obtained using the lattice parameters and Wyckoff positions.
[0065] (4) According to the crystal parameters obtained in (3), the crystal structure can be expressed by the vector v. i and e k Represent atomic properties and atomic bond properties, respectively, where i∈(1, ..., N), k∈(1, ..., M), N is the total number of atoms, and M is the total number of atom pairs.
[0066] (5) Based on the input crystal parameters, the primary crystal structure can be obtained. The formation energy of the primary crystal structure is calculated through the crystal formation energy prediction model, and the potential energy of the primary crystal is calculated according to the Lennard-Jones potential function formula as the objective function of the optimization algorithm.
[0067] (6) An optimization algorithm is used to iterate continuously to generate new crystal structures, so as to form energy and Lennard-Jones potential functions to construct the objective function as the search metric and obtain a crystal structure that meets the requirements.
[0068] (7) Compare the contact diagrams of the searched crystal structures, analyze the atomic bonding inside the crystal structures, remove the crystals that do not meet the requirements (i.e., the atoms in the crystals are not bonded), retain the crystal structures in which all atoms are bonded, and update the initial structure through optimization iteration, and continuously search for more stable crystals.
[0069] The present invention provides a stable crystal structure prediction system based on machine learning and target optimization, comprising:
[0070] A data acquisition module is used to obtain the chemical composition of the primary crystal; set the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtain the atomic coordinates of the primary crystal; and obtain a data set from a crystal database based on the atomic coordinates of the primary crystal;
[0071] The crystal energy acquisition module is used to train the graph neural network model based on the acquired data set to obtain the formation energy of the primary crystal; use the potential function to obtain the potential energy of the primary crystal; set the target optimization function based on the formation energy and potential energy of the primary crystal, and use the target optimization function to optimize its primary crystal structure to obtain the contact map of the optimized crystal structure;
[0072] The atomic bonding screening module is used to screen out stable crystal structures based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure.
[0073] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A stable crystal structure prediction method based on machine learning and target optimization, characterized in that: include: Obtain the chemical composition of the primary crystals; Set the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtain the atomic coordinates of the primary crystal; According to the atomic coordinates of the primary crystal, the data set was obtained from the crystal database; The graph neural network model is trained based on the acquired data set to obtain the formation energy of the primary crystal; Use potential function to obtain the potential energy of the primary crystal; According to the formation energy and potential energy of the primary crystal, a target optimization function is set, and the primary crystal structure is optimized using the target optimization function to obtain a contact map of the optimized crystal structure; Based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure, a stable crystal structure is screened out.
2. The stable crystal structure prediction method based on machine learning and target optimization according to claim 1, characterized in that The chemical composition of the primary crystal includes the types and numbers of atoms in the primary crystal.
3. The stable crystal structure prediction method based on machine learning and target optimization according to claim 1, characterized in that The data set includes stable structure data and metastable structure data of crystal formation energies.
4. The method for predicting stable crystal structures based on machine learning and target optimization according to claim 3, wherein: The dataset is divided into training set and validation set in a ratio of 5:1 to train the graph neural network model.
5. The method for predicting stable crystal structures based on machine learning and target optimization according to claim 1, wherein: During the training of the graph neural network model, stable structure data and metastable structure data are taken as input, and the formation energy and band gap of the crystal are taken as output.
6. The stable crystal structure prediction method based on machine learning and target optimization according to claim 1, characterized in that The potential function used is the Lennard-Jones potential function.
7. The method for predicting stable crystal structures based on machine learning and target optimization according to claim 1, wherein: The objective optimization function is as follows: ； Where, is the target optimization function, To form energy, is the weight coefficient, and its value range is [0~1]; is the absolute value of the potential energy between two molecules.
8. The method for predicting stable crystal structures based on machine learning and target optimization according to claim 1, wherein: The spatial characteristics of the crystal include space group, symmetry, and Wyckoff position.
9. A stable crystal structure prediction system based on machine learning and target optimization, characterized in that: include: A data acquisition module is used to obtain the chemical composition of the primary crystal; set the spatial characteristics of the crystal according to the chemical composition of the primary crystal and obtain the atomic coordinates of the primary crystal; and obtain a data set from a crystal database based on the atomic coordinates of the primary crystal; The crystal energy acquisition module is used to train the graph neural network model based on the acquired data set to obtain the formation energy of the primary crystal; use the potential function to obtain the potential energy of the primary crystal; set the target optimization function based on the formation energy and potential energy of the primary crystal, and use the target optimization function to optimize its primary crystal structure to obtain the contact map of the optimized crystal structure; The atomic bonding screening module is used to screen out stable crystal structures based on the contact map of the optimized crystal structure and the bonding of atoms in the crystal structure.