A method for rapid crystal structure prediction based on machine learning potential

Abstract

The present invention discloses a method for rapidly predicting crystal structure based on machine learning potential, and specifically relates to a method for predicting crystal structure by introducing a graph neural network machine learning potential method M3Gnet to predict energy. The method first constructs an input file and sets initial input parameters, such as chemical composition, element coordination, space group, and the number of generated structures; then, according to the corresponding structure generation command, a series of crystal structures that meet the requirements of the input file are generated, and chemical information such as bond length is continuously checked during the generation process to ensure the authenticity of the structure; then, the structure energy obtained by rapid prediction through graph neural network machine learning is evaluated for the synthesizability, stability, and other properties of the crystal structure. The present invention replaces the traditional DFT method with the machine learning potential, which can accelerate the energy calculation time in the crystal structure prediction process and improve the efficiency of crystal structure prediction.

Description

Technical Field

[0001] The present invention relates to the field of material science and technology, and in particular to a method for rapidly predicting crystal structure based on machine learning potential. Background Art

[0002] In recent years, with the continuous advancement and development of computer technology, modern crystal structure prediction methods based on structure generation algorithms and first-principles calculations have played a vital role in the design of new materials, and have achieved great success in many functional materials fields, such as semiconductors, superconductors, and nonlinear optics.

[0003] However, there are at least the following problems in the prior art:

[0004] As the system size and the number of atoms continue to increase, the time cost of existing methods also gradually increases, because their success mainly depends on efficient structure sampling and accurate evaluation of the energy of the sampled structures. Summary of the invention

[0005] The purpose of the present invention is to provide a method for rapid prediction of crystal structure based on machine learning potential in view of the deficiencies in the prior art. The method adopts the construction of an initial input file and saves it in a .yaml format file, sets the initial crystal structure prediction parameters, inputs the corresponding running command for the generated structure, introduces the PyXtal software package for crystal structure generation, and continuously checks the chemical environment information such as bond length during the generation process; inputs the corresponding running command for energy calculation, calculates and evaluates the energy and other information of the generated structure, introduces the M3Gnet graph neural network machine learning potential function to accurately predict the energy; inputs the corresponding running command for result analysis, and integrates the prediction information. The generated structure is analyzed by the Pymatgen software package for space group, coordination number and other material information, and finally saved as an output file in the form of txt, and all crystal structure information is converted into a file in the form of POSCAR for easy processing; selects the structure with lower energy in the prediction result as the research object, and performs phonon spectrum curve and thermodynamic stability analysis on the research symmetry; further analyzes the electronic structure and optical properties of the research object to obtain performance analysis results. The method improves the computational cost problem in the crystal structure prediction based on the first principle prediction method, and achieves the reliable, fast and accurate technical effect of predicting the crystal structure.

[0006] The method for rapidly predicting crystal structure based on machine learning potential of the present invention is carried out according to the following steps:

[0007] a. Construct the initial input file and save it in .yaml format. Set the initial crystal structure prediction parameters, including chemical composition, space group, number of generated structures, coordination number and accuracy information. The initial crystal structure prediction is TiO 2 Crystal structure, MgAl 2 O 4 Crystal structure, BaBOF 3 Crystal structure or LiGaS 2 Crystal structure;

[0008] b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process;

[0009] c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential method;

[0010] d. Input the corresponding run command for result analysis, integrate the prediction information, analyze the space group and coordination number material information of the generated structure by using the Pymatgen software package, and finally save it into a txt output file, and convert all the crystal structure information into a POSCAR file;

[0011] e. Selecting structures with lower energy in the prediction results as research objects, and performing phonon spectrum curve and thermodynamic stability analysis on the research objects;

[0012] f. Further analyze the electronic structure and optical properties of the research object to obtain performance analysis results.

[0013] In step a, TiO is predicted 2 The crystal structure is I 4 1 / amd structure and Pnma structure;MgAl 2 O 4 The crystal structure is and Pnma structures; BaBOF 3 The crystal structure is Cc structure; LiGaS 2 The crystal structure of Pna2 1 , Pmc2 1 as well as structure.

[0014] In step e, the predicted crystal structure is based on energy, and some low-energy structures are selected to further calculate the phonon spectrum, electronic structure and optical properties, where the parameters are set as:

[0015] The CASTEP module in Materials Studio software was used to analyze and calculate the properties. The full electron projection augmented wave method and the Perdew-Burke-Emzerhof form in the generalized gradient approximation were used. The plane wave cutoff energy was set to 880 eV, and the K-point meshing ratio was

[0016] The birefringence and frequency doubling coefficients obtained from the optical properties of the object in step f, wherein:

[0017] The calculation of the birefringence of the research object is obtained by the following formula:

[0018] ε(ω)=ε 1 (ω)+ε 2 (ω)

[0019] where ε 1 (ω) is the real part of the dielectric function, ε 2 (ω) is the imaginary part of the dielectric function

[0020]

[0021] Where Ω is the volume of the unit cell, v and c represent the valence band and conduction band respectively, and u is the vector defining the polarization of the incident optical electric field;

[0022] The frequency multiplication coefficient of the research object is obtained by the following formula:

[0023]

[0024] where α, β, γ are Cartesian components, vv′ represents the valence band (VB), cc′ represents the conduction band (CB), P(αβγ) represents complete substitution, and the band difference and momentum matrix elements are denoted by ω ij and P ij .

[0025] The invention discloses a method for rapidly predicting crystal structure based on machine learning potential, which uses graph neural network machine learning potential to predict structural energy. In addition, the method further comprises: using the CASTEP module in Materials Studio software to perform property analysis and calculation, using the full electron projection augmented wave method, the Perdew-Burke-Emzerhof form in the generalized gradient approximation, setting the plane wave cutoff energy to 880 eV, and the K point grid division ratio to

[0026] Compared with the existing crystal structure prediction, the present invention has the following advantages:

[0027] It can quickly generate crystal structures with reasonable chemical bonds and chemical environments. It only needs information such as chemical composition, coordination number of each atom, number of generated structures and space group to which the desired crystal structure belongs, and a series of potential crystal structures can be quickly obtained.

[0028] Low cost: The present invention does not use first-principles calculations to perform energy calculations on a large number of crystal structures obtained, but instead uses machine learning potential functions based on graph neural networks to perform preliminary optimization and energy evaluation of the crystal structures;

[0029] Simple operation: The present invention has corresponding operation commands in the three main parts, and no additional manual operation is required;

[0030] Good reliability of results: The present invention can generate stable and metastable crystal structures, and can serve as an auxiliary tool for global structure prediction in the system screening of large-scale new materials, thereby accelerating the screening process of global crystal structure prediction.

[0031] No pollution: The present invention does not involve any experiment or drug operation during the entire design and use process.

[0032] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand.

[0033] The present invention proposes and constructs a method for quickly predicting crystal structure based on machine learning potential, thereby improving the computational cost problem in crystal structure prediction based on first-principles prediction methods, and achieving the technical effect of predicting crystal structure reliably, quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flow chart of a method for predicting crystal structure based on machine learning potential according to the present invention;

[0035] Figure 2 The TiO 2 MgAl 2 O 4 Schematic diagram of the crystal structure, where A is TiO 2 Schematic diagram of crystal structure (from Materials Project database); B is TiO 2 Schematic diagram of crystal structure (derived from Example 1 of the present invention); C is TiO 2 Schematic diagram of crystal structure (from Materials Project database); D is TiO 2 Crystal structure diagram (derived from Example 1 of the present invention); E is MgAl 2 O4 Schematic diagram of the crystal structure (from the Materials Project database); F is MgAl 2 O 4 Schematic diagram of the crystal structure (derived from Example 2 of the present invention); G is MgAl 2 O 4 Schematic diagram of the crystal structure (from the Materials Project database); H is MgAl 2 O 4 Structural schematic diagram of the crystal structure (derived from Example 2 of the present invention);

[0036] Figure 3 BaBOF of the present invention 3 Schematic diagram of the crystal structure, where A is BaBOF 3 Schematic diagram of crystal structure (from Materials Project database); B is BaBOF 3 Structural schematic diagram of the crystal structure (derived from Example 3 of the present invention);

[0037] Figure 4 The TiO 2 Schematic diagram of the comparison results of unit cell parameters, where A is TiO 2 Schematic diagram of the comparison of unit cell parameters (space group is I4 1 / amd); B is TiO 2 Schematic diagram of the comparison results of unit cell parameters (space group is Pnma); C is MgAl 2 O 4 Schematic diagram of the comparison of unit cell parameters (space group is ); D is MgAl 2 O 4 Schematic diagram of the comparison results of unit cell parameters (space group is Pnma);

[0038] Figure 5 The LiGaS predicted in the present invention 2 The phonon spectrum (Pna2 1 , Pmc2 1 as well as structure). DETAILED DESCRIPTION

[0039] The present invention is further illustrated by the following examples, but the protection scope of the present invention is not limited to the following examples.

[0040] Example 1

[0041] TiO 2 The structure prediction of TiO2 The crystal structure is I4 1 / amd structure and Pnma structure, but is not limited to this crystal structure (such as Figure 1 ).

[0042] a. Build the initial input file and save it in .yaml format, set the initial crystal structure prediction parameters, specifically the chemical composition Ti8O16, space group 1-230, the number of generated structures 300 / 400 / 500, the coordination number and precision information Ti and O;

[0043] b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process;

[0044] c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential function;

[0045] d. Input the corresponding run command for result analysis, integrate the prediction information, and use the Pymatgen software package to analyze the space group and coordination number material information of the generated structure. Finally, save it into a txt output file and output the predicted TiO 2 The crystal structure information is converted into POSCAR format files;

[0046] According to the implementation steps of Example 1, Figure 2 and attached Figure 4 The predicted structure in Example 1 and the cell parameter comparison information of the corresponding structure are displayed. The predicted structure is compared with the existing materials database (Materials Project) and some structures reported in the literature. The comparison results show that the method designed by the present invention has high reliability and accuracy.

[0047] Example 2

[0048] MgAl 2 O 4 The structural prediction of MgAl 2 O 4 The crystal structure is and Pnma structures, but are not limited to this crystal structure (e.g. Figure 1 ).

[0049] a. Build the initial input file and save it in .yaml format, set the initial crystal structure prediction parameters, specifically the chemical composition Mg4Al8O16, space group 1-230, the number of generated structures 300 / 400 / 500, the coordination number and accuracy information Mg, Al and O;

[0050] b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process;

[0051] c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential function;

[0052] d. Input the corresponding run command for result analysis, integrate the prediction information, and use the Pymatgen software package to analyze the space group and coordination number material information of the generated structure. Finally, save it into a txt output file and convert the obtained MgAl 2 O 4 Crystal structure information is converted into POSCAR format files;

[0053] According to the implementation steps of Example 2, Figure 2 and attached Figure 4 The predicted structure in Example 2 is shown, and the predicted structure is compared with some structures in the existing material database (Materials Project). The comparison results show that the method designed by the present invention has high reliability and accuracy.

[0054] Example 3

[0055] BaBOF 3 The structure prediction of the test case is BaBOF 3 The crystal structure is Cc structure, but it is not limited to this crystal structure (such as Figure 1 ).

[0056] a. Build the initial input file and save it in .yaml format, set the initial crystal structure prediction parameters, specifically the chemical composition Ba2B2O2F6, space group 1-230, the number of generated structures 300 / 400 / 500, the coordination number and accuracy information Ba, B, O and F;

[0057] b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process;

[0058] c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential function;

[0059] d. Input the corresponding run command for result analysis, integrate the prediction information, and use the Pymatgen software package to analyze the space group and coordination number material information of the generated structure. Finally, save it into a txt output file and convert the obtained BaBOF 3 The crystal structure information is converted into POSCAR format files;

[0060] According to the implementation steps of Example 3, Figure 3 The predicted structure in Example 3 is shown, and the predicted structure is compared with some structures reported in the literature. The comparison results show that the method designed by the present invention has high reliability and accuracy.

[0061] Example 4

[0062] LiGaS 2 The structural prediction of LiGaS is a test case. 2 The crystal structure of Pna2 1 , Pmc2 1 as well as structure, but is not limited to this crystal structure (such as Figure 1 );

[0063] a. Build the initial input file and save it in .yaml format, set the initial crystal structure prediction parameters, specifically the chemical composition LiGaS2 / Li2Ga2S4 / Li3Ga3S6 / Li4Ga4S8, space group 1-230, the number of generated structures 500, coordination number and accuracy information Li, Ga and S;

[0064] b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process;

[0065] c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential function;

[0066] d. Input the corresponding run command for result analysis, integrate the prediction information, and use the Pymatgen software package to analyze the space group and coordination number material information of the generated structure. Finally, save it as an output file in txt format and output the obtained LiGaS 2The crystal structure information is converted into POSCAR format files;

[0067] e. LiGaS obtained in step d 2 The crystal structure is based on energy, and some low-energy structures are selected to further calculate the phonon spectrum, electronic structure and optical properties. The parameters are set as follows:

[0068] The CASTEP module in Materials Studio software was used to analyze and calculate the properties. The full electron projection augmented wave method and the Perdew-Burke-Emzerhof form in the generalized gradient approximation were used. The plane wave cutoff energy was set to 880 eV, and the K-point meshing ratio was

[0069] f. LiGaS obtained in step d 2 The crystal structure was further analyzed for electronic structure and optical properties, and the birefringence and frequency doubling coefficients were obtained from the optical properties, among which:

[0070] The LiGaS 2 The calculation of the birefringence of the crystal structure is obtained by the following formula:

[0071] ε(ω)=ε 1 (ω)+ε 2 (ω)

[0072] where ε 1 (ω) is the real part of the dielectric function, ε 2 (ω) is the imaginary part of the dielectric function

[0073]

[0074] Where Ω is the volume of the unit cell, v and c represent the valence band and conduction band respectively, and u is the vector defining the polarization of the incident optical electric field;

[0075] The LiGaS 2 The harmonic generation coefficient of the crystal structure is obtained by the following formula:

[0076]

[0077] where α, β, γ are Cartesian components, vv′ represents the valence band (VB), cc′ represents the conduction band (CB), P(αβγ) represents complete substitution, and the band difference and momentum matrix elements are denoted by ω ij and P ij .

[0078] According to the implementation steps of Example 4, Figure 5The phonon spectra of the predicted structure in Example 4 are shown, from which it can be seen that the phonon spectra of the four crystal structures predicted by the present invention in Example 4 do not generate imaginary frequencies, indicating that the present invention can obtain a stable or metastable structure that is experimentally synthesizable, wherein Pna2 1 The structure has the lowest energy after geometric optimization and is the most stable structure. Pmc2 1 as well as The structural energy increases in turn. In addition, Pna2 1 The structure is basically consistent with that in the ICSD database (Table 1), which further illustrates the reliability, accuracy and rationality of the method of the present invention;

[0079] Table 1

[0080]

[0081]

[0082] Table 2

[0083]

[0084] The results in Table 2 show that the method proposed in the present invention can discover new nonlinear optical crystal structures with excellent performance.

[0085] It can be seen from the examples that the method for rapid prediction of crystal structure established by the present invention can accurately, reliably and quickly obtain the crystal structure. The total time consumption of the three embodiments 1-3 is less than two hours. The time consumption of the 2000 crystal structures involved in Example 4 is only 3 hours from structure generation to structure optimization; the time for crystal structure prediction is greatly accelerated. The embodiments respectively select binary, ternary and quaternary crystal structures of different systems, indicating that the present invention is not limited to a certain type of material and has a wide range of applicability. It saves the time-consuming problem of traditional DFT calculation as the number of atoms in the system increases.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit it. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for rapid prediction of crystal structure based on machine learning potential, characterized in that: The method proceeds in the following steps: a. Construct the initial input file and save it in yaml format, set the initial crystal structure prediction parameters, specifically the chemical composition, space group, number of generated structures, coordination number and accuracy information, where the initial crystal structure is predicted to be TiO2 crystal structure, MgAl2O4 crystal structure, BaBOF3 crystal structure or LiGaS2 crystal structure; b. Enter the corresponding run command to generate the structure, introduce the PyXtal software package to generate the crystal structure, and continuously check the bond length chemical environment information during the generation process; c. Input the corresponding operation command of energy prediction, calculate and evaluate the energy information of the generated structure, and accurately predict the energy by introducing the M3Gnet graph neural network machine learning potential method; d. Input the corresponding run command for result analysis, integrate the prediction information, analyze the space group and coordination number material information of the generated structure by using the Pymatgen software package, and finally save it into a txt output file, and convert all the crystal structure information into a POSCAR file; e. Selecting a structure with lower energy in the prediction results as a research object, and performing phonon spectrum curve and thermodynamic stability analysis on the research object; f. Further analyze the electronic structure and optical properties of the research object to obtain performance analysis results.

2. The method for rapidly predicting crystal structure based on machine learning potential as claimed in claim 1, characterized in that: The crystal structure of TiO2 predicted in step a is I 41 / amd structure and Pnma structure; the crystal structure of MgAl2O4 is and Pnma structures; the crystal structure of BaBOF3 is Cc structure; the crystal structure of LiGaS2 is Pna21, Pmc21 and structure.

3. The method for rapidly predicting crystal structure based on machine learning potential as claimed in claim 1, characterized in that: In step e, the predicted crystal structure is based on energy, and some low-energy structures are selected to further calculate the phonon spectrum, electronic structure and optical properties, where the parameters are set as: The CASTEP module in Materials Studio software was used to analyze and calculate the properties. The full electron projection augmented wave method and the Perdew-Burke-Emzerhof form in the generalized gradient approximation were used. The plane wave cutoff energy was set to 880 eV, and the K-point meshing ratio was 4. The method for rapidly predicting crystal structure based on machine learning potential as claimed in claim 1, characterized in that: The birefringence and frequency doubling coefficients obtained from the optical properties of the object in step f, wherein: The calculation of the birefringence of the research object is obtained by the following formula: ε(ω)=ε1(ω)+ε2(ω) Where ε1(ω) is the real part of the dielectric function and ε2(ω) is the imaginary part of the dielectric function where Ω is the volume of the unit cell, v and c represent the valence band and conduction band, respectively, and u is the vector defining the polarization of the incident optical field. The frequency multiplication coefficient of the research object is obtained by the following formula: where α, β, γ are Cartesian components, vv′ represents the valence band (VB), cc′ represents the conduction band (CB), P(αβγ) represents complete substitution, and the band difference and momentum matrix elements are represented by ω ij and P ij .

Images