A method for training machine learning interatomic potentials for high-throughput semi-heusler materials

Abstract

The application discloses a method for training machine learning interatomic potential (MLIP) of high-throughput semi-Heusler material. The calculation is mainly based on the first principle MatHub-3d database. There are 273 semi-Heusler materials in the database, and 130 stable semiconductor semi-Heusler structures are screened out according to their band structure and phonon spectrum in the database, wherein the phonon spectrum has no virtual frequency and the band gap is greater than 0.1 eV. Then, a double adaptive sampling (DAS) method is applied to train three MLIP models for each semi-Heusler material, and the accuracy of the models is verified by comparing with the results of density functional theory (DFT) calculation. The method can improve the efficiency through high-throughput training, achieve the purpose of quickly training the MLIP model of the target material system, and is favorable for providing a theoretical reference for the design of thermoelectric materials.

Description

Technical Field

[0001] This invention belongs to the field of computational materials for training interatomic potentials in machine learning, and specifically relates to a method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning. Background Technology

[0002] Research on interatomic potentials began in the 1980s, when researchers introduced many-body potentials into metallic systems, achieving breakthroughs in reducing fracture stress in metallic Ni. Subsequently, more interatomic potentials covering various chemical elements and multi-element systems were developed to predict the energy and atomic forces of materials. While traditional interatomic potentials can calculate the energy and forces of large-scale systems, their simulations are time-consuming. In the last decade, research on constructing interatomic potentials using machine learning methods has developed rapidly. Machine Learning Interatomic Potential (MLIP) is not based on physical formulas but rather uses numerical interpolation to predict potentials among known reference data. It combines the computational speed advantage of traditional interatomic potentials with simulation accuracy comparable to density functional theory. It excels particularly in accurate predictions at the atomic scale and is often used to accelerate the simulation of thermal transport properties in thermoelectric materials.

[0003] Semi-Hassler materials, as ternary intermetallic compounds, are usually represented by the chemical formula ABX. Their crystal structure can be considered as a combination of rock salt and zincblende structures. This material possesses both excellent electrical transport properties and stable thermal properties, resulting in a high thermoelectric figure of merit. It is a promising thermoelectric material and has received extensive research attention in recent years.

[0004] In recent years, thanks to the rapid advancements in high-throughput computing and machine learning algorithms, machine learning-based interatomic potentials have been widely used in the discovery and design of new materials. Automated processes for training machine learning-based interatomic potentials on a large number of materials facilitate rapid screening and evaluation of their performance, significantly accelerating the discovery of new materials.

[0005] This invention develops a high-throughput method for training interatomic potentials in semi-Hassler materials using machine learning, aiming to more quickly calculate the energy and atomic forces of semi-Hassler materials in order to explore semi-Hassler thermoelectric materials with good thermal transport properties. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method for high-throughput training of interatomic potentials in semi-Hassler materials using machine learning. By providing an initial cell structure, the MLIP training process can be automated and is easy to operate.

[0007] To rapidly obtain the interatomic potential of materials for machine learning, the objective of this invention can be achieved through the following technical solutions:

[0008] A method for high-throughput training of interatomic potentials in semi-Hassler materials using machine learning includes the following steps:

[0009] S1, based on specific screening criteria, the target semi-Hasler material is obtained;

[0010] S2, structural optimization of the target semi-Hassler material is performed to obtain the optimized crystal structure;

[0011] S3, using the optimized crystal structure, first-principles molecular dynamics AIMD simulations based on density functional theory (DFT) were performed to obtain DFT energy and atomic forces;

[0012] S4 extracts atomic coordinates, energy, and atomic forces from DFT energy and atomic forces as the initial training set;

[0013] S5. Using the optimized crystal structure and the initial training set, the dual adaptive sampling (DAS) method is applied to train the high-throughput machine learning interatomic potential (MLIP) model, resulting in the trained MLIP model.

[0014] S6. Perform molecular dynamics (MD) simulations on the trained MLIP model to obtain several configurations containing atomic coordinates, energies, and atomic forces. Compare these configurations with the energies and atomic forces calculated by VASP software to obtain the mean absolute error (MAE) of the energies and atomic forces, thereby verifying the accuracy of the trained MLIP model.

[0015] Preferably, in S2, when the target semi-Hassler material undergoes structural optimization, the lattice parameters of the crystal structure are a, b, and c, the high-symmetry K-points are set to 60 / a+1, 60 / b+1, and 60 / c+1, and the energy convergence criterion is 10. -8 eV, cutoff energy is 520eV.

[0016] Preferably, in S3, the AIMD simulation constructs a 4×4×4 supercell structure with an energy convergence criterion of 10. -4 eV, cutoff energy of 360eV, ensemble of NVT, temperature of 300K, step size of 20 steps.

[0017] Preferably, in S4, the atomic coordinates, energy, and atomic forces of the material are extracted at two-step intervals, and the initial training set contains 10 configurations.

[0018] Preferably, in S5, the training process of the DAS method is as follows:

[0019] (a) Based on the initial training set of each target semi-Hasler material, train three MLIP models with identical functional forms simultaneously using different randomly initialized parameters;

[0020] (b) Within a small temperature range, MD simulation sampling was performed using the MLIP model and the optimized crystal structure in LAMMPS software. Based on the stress divergence threshold, atomic coordinates, energies and atomic forces of several configurations with stress divergence greater than the threshold were obtained.

[0021] (c) Using VASP software, calculate the energy and atomic forces of the configuration based on the atomic coordinates obtained in step (b), and add the atomic coordinates, energy, and atomic forces of the configuration to the training set;

[0022] (d) Using the updated training set from step (c), train and update the MLIP model and update the force divergence threshold;

[0023] (e) Repeat steps (b), (c) and (d) until all configurations sampled in step (b) are less than the stress divergence threshold, and no new configurations are selected.

[0024] (f) Repeat step (e) within each small temperature range.

[0025] Preferably, in step (a), the descriptor form of the MLIP model is a moment tensor potential, and the cutoff radius is set to... The calculation formula is as follows:

[0026]

[0027] Where E is the interatomic potential, and the function V passes through a set of basis functions B. α Linear expansion, n i It is the atomic environment of the i-th atom, determined by its atom type z. i Its nearest neighbor atom type z j and the position r of the nearest neighbor relative to the i-th atom ij Composition, ξ={ξ α The parameters are found by fitting the training set, and the basis function B is... α The functional form is the moment tensor descriptor M. μ,v (n i ).

[0028] Preferably, in step (b), the ensemble for the MD simulation is NVT, with two small temperature ranges set to 250-400K and 650-800K respectively, spaced 50K apart. The MD simulation lasts for 2.5 ps, with a time step of 0.5 fs. Configurations are sampled every 20 steps, with a maximum of ten configurations selected each time. The force divergence threshold refers to the deviation of different models in predicting the forces on atoms of the same configuration. The initial force divergence threshold is set to... The calculation formula is as follows:

[0029]

[0030] in It is the force divergence threshold, N m f is the number of MLIP models. j,m The force on atom j is predicted by model m.

[0031] Preferably, in step (c), the high-symmetric K-point calculated by VASP is set to 1×1×1, and the energy convergence criterion is 10. - 5 eV, cutoff energy is 400eV.

[0032] Preferably, in S6, the MD simulation of the validated MLIP model is first performed for 0.1 ns, and then a configuration is selected every 0.5 ps, for a total of ten configurations. The temperatures are set to 300 K and 700 K, and the ensemble is NVT. The high-symmetry K-point calculated by the VASP software is set to 1×1×1, and the energy convergence criterion is 10. -5 eV, cutoff energy is 400eV.

[0033] Preferably, in S6, the calculation formula for MAE is as follows:

[0034]

[0035] Where n is the number of data points, y i It is the MLIP prediction value of the i-th data point. It is the DFT prediction value of the i-th data point.

[0036] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0037] (1) The present invention adopts a high-throughput computing method, which can quickly calculate the inter-atomic potential of the target system through machine learning, which is efficient and accurate.

[0038] (2) In this invention, the energy, atomic forces and interatomic interaction constants of the semi-Hassler material can be obtained quickly in theory. The accurate high-order interatomic interaction constants obtained quickly can accelerate the simulation of thermal transport properties and provide a reference for screening good thermoelectric materials.

[0039] (3) The method of the present invention does not involve experimental operations or use chemicals in the whole process, which is in line with the concept of green environmental protection; and it is simple to operate, low in cost, and suitable for promotion and application to other semiconductor materials. Attached Figure Description

[0040] Figure 1 This is a flowchart of a method for high-throughput training of interatomic potentials in semi-Hassler materials using machine learning, as described in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram comparing the atomic forces of TiCoSb in Embodiment 2 of the present invention, specifically the DFT and MLIP.

[0042] Figure 3 This is a schematic diagram of the absolute error distribution of atomic forces on TiCoSb in Embodiment 2 of the present invention;

[0043] Figure 4 This is a schematic diagram comparing the atomic forces of CdLiP and MLIP in Embodiment 3 of the present invention.

[0044] Figure 5 This is a schematic diagram of the absolute error distribution of atomic forces on CdLiP in Embodiment 3 of the present invention. Detailed Implementation

[0045] This invention develops a high-throughput method for training machine learning interatomic potentials (MLIPs) for semi-Hassler materials. The calculations are primarily based on the first-principles MATLAB-3d database. The database contains 273 semi-Hassler materials. Based on their band structure and phonon spectra in the database, 130 stable semiconductor semi-Hassler structures with no imaginary frequencies in their phonon spectra and a band gap greater than 0.1 eV were selected. Then, the dual adaptive sampling (DAS) method was applied to train three MLIP models for each semi-Hassler material. The accuracy of the models was verified by comparing the results with those calculated using density functional theory (DFT). This method, through high-throughput training, improves efficiency and achieves the goal of quickly training MLIP models for the target material system, which is beneficial for providing theoretical reference for the design of thermoelectric materials.

[0046] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0047] Example 1:

[0048] Figure 1 This is a flowchart illustrating a method for high-throughput training of interatomic potentials in semi-Hassler materials using machine learning, as described in an embodiment of the present invention.

[0049] like Figure 1 As shown, this embodiment provides a method for high-throughput training of interatomic potentials in semi-Hassler materials using machine learning, including the following steps:

[0050] S1, based on specific screening criteria, yields the target semi-Hasler material.

[0051] S2, the target semi-Hassler material is structurally optimized to obtain the optimized crystal structure.

[0052] When optimizing the structure of the target semi-Hassler material, the lattice parameters of the crystal structure are a, b, and c, and its high-symmetry K-points are set to 60 / a+1, 60 / b+1, and 60 / c+1, with an energy convergence criterion of 10. -8 eV, cutoff energy is 520eV.

[0053] S3 utilizes the optimized crystal structure and performs first-principles molecular dynamics AIMD simulations based on density functional theory (DFT) to obtain DFT energy and atomic forces.

[0054] S4 extracts atomic coordinates, energy, and atomic forces from the DFT energy and atomic forces to serve as the initial training set.

[0055] S5. Using the optimized crystal structure and the initial training set, the Dual Adaptive Sampling (DAS) method is applied to train the target semi-Hassler material using high-throughput machine learning interatomic potential (MLIP) to obtain the trained MLIP model.

[0056] The training process for the Dual Adaptive Sampling (DAS) method is as follows:

[0057] (a) Based on the initial training set of each half-Hasler material, train three MLIP models with identical functional forms simultaneously using different randomly initialized parameters.

[0058] In step (a), the descriptor form of the MLIP model is a moment tensor potential, and the cutoff radius is set to... The calculation formula is as follows:

[0059]

[0060] Where E is the interatomic potential, and the function V passes through a set of basis functions B. α Linear expansion, n i It is the atomic environment of the i-th atom, determined by its atom type z. i Its nearest neighbor atom type z j and the position r of the nearest neighbor relative to the i-th atom ij Composition, ξ={ξ α The parameters are found by fitting the training set, and the basis function B is... α The functional form is the moment tensor descriptor M. μ,v (n i ).

[0061] (b) Within a small temperature range, MD simulation sampling was performed using the MLIP model and the optimized crystal structure in LAMMPS software. Based on the stress divergence threshold, atomic coordinates, energies and atomic forces of several configurations with stress divergence greater than the threshold were obtained.

[0062] The force divergence threshold refers to the deviation between different models in predicting the forces on atoms with the same configuration. The initial force divergence threshold is set to... The calculation formula is as follows:

[0063]

[0064] in It is the force divergence threshold, N m f is the number of MLIP models. j,m The force on atom j is predicted by model m.

[0065] (c) Using VASP software, calculate the energy and atomic forces of the configuration based on the atomic coordinates obtained in step (b), and add the atomic coordinates, energy, and atomic forces of the configuration to the training set;

[0066] (d) Using the updated training set from step (c), train and update the MLIP model and update the force divergence threshold;

[0067] (e) Repeat steps (b), (c) and (d) until all configurations sampled in step (b) are less than the stress divergence threshold, and no new configurations are selected.

[0068] (f) Repeat step (e) within each small temperature range.

[0069] S6. Perform molecular dynamics (MD) simulations on the trained MLIP model to obtain several configurations containing atomic coordinates, energy, and atomic forces. Compare these configurations with the energy and atomic forces calculated by VASP software to obtain the mean absolute error (MAE) of energy and atomic forces, which is used to verify the accuracy of the trained MLIP model.

[0070] The formula for calculating MAE is as follows:

[0071]

[0072] Where n is the number of data points, y i It is the MLIP prediction value of the i-th data point. It is the DFT prediction value of the i-th data point.

[0073] Example 2:

[0074] This embodiment employs the high-throughput training method for interatomic potentials in semi-Hassler materials from Embodiment 1 above. Specifically, for the semi-Hassler material TiCoSb, the following steps are included:

[0075] S1. Based on the band gap of the band structure being greater than 0.1 eV and the absence of imaginary frequencies in the phonon spectrum from the MATLAB-3d database, 273 semi-Hasler materials were screened, resulting in 130 semi-Hasler materials.

[0076] S2, structural optimization of TiCoSb, a semi-Hassler material among 130 semi-Hassler materials, was performed. The lattice parameters of the crystal structure were as follows: The high-symmetry K-point for structural optimization is set to 15×15×15, and the energy convergence criterion is 10. -8 eV, cutoff energy is 520eV.

[0077] S3. First-principles molecular dynamics (AIMD) simulations of TiCoSb were performed using the optimized crystal structure and VASP software based on density functional theory (DFT). A 4×4×4 supercell structure was constructed, and the energy convergence criterion was 10. -4 eV, cutoff energy of 360eV, ensemble of NVT, temperature of 300K, step size of 20 steps.

[0078] S4, from the results obtained from AIMD, the atomic coordinates, energies and atomic forces are extracted every 2 steps as the initial training set for the semi-Hassler material TiCoSb, which contains a total of 10 configurations;

[0079] S5. Using the optimized crystal structure and the initial training set, the DAS method was applied to train TiCoSb MLIP, resulting in the training of three MLIP models.

[0080] Furthermore, in S5, the training process for the DAS method is as follows:

[0081] (a) Based on the initial training set of TiCoSb, train three MLIP models with identical functional forms simultaneously using different randomly initialized parameters;

[0082] (b) Within a small temperature range, MD simulations were performed using the MLIP model and the optimized crystal structure in LAMMPS software. Two small temperature ranges were set: 250-400K and 650-800K, with an interval of 50K. The MD simulation lasted 2.5 ps with a time step of 0.5 fs. Configurations were sampled every 20 steps, with a maximum of ten configurations selected each time. Based on a stress divergence threshold, atomic coordinates, energies, and atomic forces were obtained for several configurations with stress divergence greater than the threshold. The initial stress divergence threshold was set to...

[0083] (c) The high-symmetric K-point calculated using VASP software is set to 1×1×1, and the energy convergence criterion is 10. -5 eV, cutoff energy is 400eV;

[0084] (d) Using the updated training set from step (c), train and update the MLIP model and update the force divergence threshold;

[0085] (e) Repeat steps (b), (c) and (d) until all configurations sampled in step (b) are less than the stress divergence threshold, and no new configurations are selected.

[0086] (f) Repeat step (e) within each small temperature range.

[0087] S6. MD simulations were performed on the trained MLIP model using LAMMPS software. Initially, a 0.1 ns simulation was performed, followed by a 0.5 ps simulation, selecting ten configurations in total. Temperatures were set to 300 K and 700 K, and the ensemble was NVT. Ten configurations containing atomic coordinates, energies, and atomic forces were obtained and compared with the energies and atomic forces calculated by VASP software. The high-symmetry K-point calculated by VASP software was set to 1×1×1, and the energy convergence criterion was 10. -5 eV, cutoff energy of 400 eV, to verify the accuracy of the trained MLIP model. The comparison results of atomic forces are shown in [link to relevant documentation]. Figure 2 and Figure 3 At 300K and 700K, the energy MAE is 1.910 meV·atom. -1 and 1.197meV·atom -1 The MAE values ​​of the atoms are respectively and

[0088] Example 3:

[0089] This embodiment employs the high-throughput training method for interatomic potentials in semi-Hassler materials from Embodiment 1 above. Specifically, for the semi-Hassler material CdLiP, the following steps are included:

[0090] S1. Based on the band gap of the band structure being greater than 0.1 eV and the absence of imaginary frequencies in the phonon spectrum from the MATLAB-3D database, 273 semi-Hassler materials were screened, resulting in 130 semi-Hassler materials.

[0091] S2, structural optimization of CdLiP, a semi-Hassler material among 130 semi-Hassler materials, was performed. The lattice parameters of the crystal structure were as follows: The high-symmetry K-point for structural optimization is set to 15×15×15, and the energy convergence criterion is 10. -8 eV, cutoff energy is 520eV.

[0092] S3. AIMD simulations of CdLiP were performed using the optimized crystal structure and the DFT-based VASP software, constructing a 4×4×4 supercell structure with an energy convergence criterion of 10. -4 eV, cutoff energy of 360eV, ensemble of NVT, temperature of 300K, step size of 20 steps.

[0093] S4, from the results obtained from AIMD, extracts atomic coordinates, energies and atomic forces every 2 steps as the initial training set for the semi-Hassler material CdLiP, which contains a total of 10 configurations.

[0094] S5. Using the optimized crystal structure and the initial training set, the DAS method is applied to train CdLiP for MLIP, and three MLIP models are trained.

[0095] Furthermore, in S5, the training process for the DAS method is as follows:

[0096] (a) Based on the training set of the initial structure of CdLiP, train three MLIP models with identical functional forms simultaneously using different randomly initialized parameters;

[0097] (b) Within a small temperature range, MD simulations were performed using the MLIP model and the optimized crystal structure in LAMMPS software. Two small temperature ranges were set: 250-400K and 650-800K, with an interval of 50K. The MD simulation lasted 2.5 ps with a time step of 0.5 fs. Configurations were sampled every 20 steps, with a maximum of ten configurations selected each time. Based on a stress divergence threshold, atomic coordinates, energies, and atomic forces were obtained for several configurations with stress divergence greater than the threshold. The initial stress divergence threshold was set to...

[0098] (c) The high-symmetric K-point calculated using VASP software is set to 1×1×1, and the energy convergence criterion is 10. -5 eV, cutoff energy is 400eV;

[0099] (d) Using the updated training set from step (c), train and update the MLIP model and update the force divergence threshold;

[0100] (e) Repeat steps (b), (c) and (d) until all configurations sampled in step (b) are less than the stress divergence threshold, and no new configurations are selected.

[0101] (f) Repeat step (e) within each small temperature range.

[0102] S6. MD simulations were performed on the obtained MLIP model using LAMMPS software. Initially, a 0.1 ns simulation was performed, followed by a 0.5 ps simulation, selecting ten configurations in total. Temperatures were set to 300 K and 700 K, and the ensemble was NVT. Ten configurations containing atomic coordinates, energies, and forces were obtained and compared with the energies and atomic forces calculated by VASP software. The high-symmetry K-point calculated by VASP software was set to 1×1×1, and the energy convergence criterion was 10. -5 eV, cutoff energy of 400 eV, to verify the accuracy of the MLIP model. The comparison results of atomic forces are shown in [reference needed]. Figure 4 and Figure 5 At 300K and 700K, the MAE is 0.142 meV·atom. -1 and 0.161 meV·atom -1 The MAE values ​​of the atoms are respectively and

[0103] Using the high-throughput training method for interatomic potentials of semi-Hassler materials in Example 1 above, the energy and atomic force MAE results for other semi-Hassler materials are shown in Table 1.

[0104] Table 1 shows the energy and atomic force (MAE) results for some semi-Hassler materials.

[0105]

[0106]

[0107]

[0108] The methods described above overcome the problems of long development time and high cost in traditional material research and development, and can quickly screen and obtain machine learning interatomic potentials for structurally stable systems. The use of an automated process in training machine learning interatomic potentials enables the systematic acquisition of machine learning interatomic potentials for semi-Hassler materials, facilitating the rapid determination of their energy, atomic forces, and interatomic interaction constants, thus providing a theoretical reference for simulating the thermal transport properties of thermoelectric materials.

[0109] In summary, this invention presents a high-throughput method for training interatomic potentials in semi-Hassler materials using machine learning. Utilizing the MatHub-3d materials database, 130 structurally stable semi-Hassler materials were selected through screening based on phonon spectra and band structures. The DAS method was applied to train the MLIP model, and the results were compared with those from DFT calculations to verify the accuracy of the MLIP model. This invention is simple to operate, easy to promote, and provides guidance for simulating the thermal transport properties of thermoelectric materials.

[0110] In the description of this specification, references to terms such as "an embodiment," "example," and "specific example" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0111] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning, characterized in that, Includes the following steps: S1, based on specific screening criteria, the target semi-Hasler material is obtained; S2, perform structural optimization on the target semi-Hassler material to obtain an optimized crystal structure; S3. Using the optimized crystal structure, first-principles molecular dynamics AIMD simulations were performed based on density functional theory (DFT) to obtain the DFT energy and atomic forces. S4. Extract the atomic coordinates, energy, and atomic force from the DFT energy and atomic force to form the initial training set; S5. Using the optimized crystal structure and the initial training set, apply the dual adaptive sampling (DAS) method to perform high-throughput machine learning interatomic potential (MLIP) training to obtain the trained MLIP model. S6. Perform molecular dynamics (MD) simulation on the trained MLIP model to obtain several configurations containing atomic coordinates, energy, and atomic forces. Compare these configurations with the energy and atomic forces calculated by VASP software to obtain the mean absolute error (MAE) of the energy and atomic forces, thereby verifying the accuracy of the trained MLIP model. In S5, the training process of the DAS method is as follows: (a) Based on the initial training set of each target semi-Hasler material, train three MLIP models with identical functional forms simultaneously using different randomly initialized parameters; (b) Within a small temperature range, MD simulation sampling was performed using the MLIP model and the optimized crystal structure in LAMMPS software. Based on the stress divergence threshold, atomic coordinates, energies and atomic forces of several configurations with stress divergence greater than the threshold were obtained. (c) Using VASP software, calculate the energy and atomic forces based on the atomic coordinates of the configuration obtained in step (b), and add the atomic coordinates, energy, and atomic forces of the configuration to the training set; (d) Using the updated training set from step (c), train and update the MLIP model and update the force divergence threshold; (e) Repeat steps (b), (c) and (d) until all configurations sampled in step (b) are less than the stress divergence threshold, and no new configurations are selected. (f) Repeat step (e) within each small temperature range; In step (a), the descriptor form of the MLIP model is a moment tensor potential, with a cutoff radius set to 6.0 Å, and the calculation formula is shown below: in It is an interatomic potential, a function Through a set of basis functions Linear expansion, It is the first i The atomic environment of an atom, determined by its atom type Its nearest neighbor atom type and nearest neighbor relative to the first i The position of each atom constitute, The parameters and basis functions are found by fitting the training set. The functional form is a moment tensor descriptor. ; In step (b), the ensemble for the MD simulation is NVT, with two smaller temperature ranges set at 250-400 K and 650-800 K, with an interval of 50 K. The MD simulation lasts for 2.5 ps, with a time step of 0.5 fs. Configurations are sampled every 20 steps, with a maximum of ten configurations selected each time. The force divergence threshold refers to the deviation of different models in predicting the forces on atoms of the same configuration. The initial force divergence threshold is set to 0.01 eV / Å, and the calculation formula is shown below: in It is the force divergence threshold. It represents the number of MLIP models. It is a model m Predicted atoms j The forces acting on it.

2. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In S2, when the target semi-Hassler material undergoes structural optimization, the lattice parameters of the crystal structure are a, b, and c, the high-symmetry K-point is set to 60 / a+1, 60 / b+1, and 60 / c+1, and the energy convergence criterion is 10. -8 eV, cutoff energy is 520 eV.

3. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In S3, the AIMD simulation constructs a 4×4×4 supercell structure, with an energy convergence criterion of 10. -4 eV, cutoff energy is 360 eV, ensemble is NVT, temperature is 300 K, step size is 20 steps.

4. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In step S4, the atomic coordinates, energy, and atomic forces of the material are extracted at two-step intervals, and the initial training set contains 10 configurations.

5. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In step (c), the high-symmetric K-point calculated by VASP is set to 1×1×1, and the energy convergence criterion is 10. -5 eV, cutoff energy is 400 eV.

6. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In step S6, the MD simulation of the trained MLIP model is verified. First, a 0.1 ns interval is performed, then a configuration is selected every 0.5 ps, for a total of ten configurations. The temperatures are set to 300 K and 700 K, and the ensemble is NVT. The high-symmetry K-point calculated by VASP software is set to 1×1×1, and the energy convergence criterion is 10. -5 eV, cutoff energy is 400 eV.

7. The method for high-throughput training of interatomic potentials in semi-Hassler materials for machine learning according to claim 1, characterized in that, In step S6, the calculation formula for MAE is as follows: Where n is the number of data points. It is the first i MLIP prediction values ​​for each data point It is the first i DFT prediction values ​​for each data point.

Images