Multi-scale automatic analysis method for atomic structure of amorphous material

Abstract

The invention relates to the technical field of computers, in particular to a multi-scale automatic analysis method for an amorphous material atomic structure. Comprising the following steps: S1, taking a reactant as an initial unit, and obtaining a cluster unit of a reaction product based on first principle molecular dynamics simulation; s2, carrying out unit cell construction on the cluster unit, and optimizing unit cell parameters and atomic coordinates of the cluster unit through calculation according to a first principle to obtain unit cell units; s3, spatial stacking is carried out on the unit cell units, unit cell parameters and atomic coordinates of the unit cell units are optimized through calculation according to a first principle, and supercell units are obtained; s4, performing first principle molecular dynamics simulation on the supercell unit to obtain an accurate machine learning potential function; s5, carrying out cell expansion construction on the supercell unit, carrying out molecular dynamics simulation by using an accurate machine learning potential function, and obtaining an amorphous material atomic structure; and S6, packaging the steps S1 to S5 into an automatic workflow. The method is high in efficiency and small in error.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more particularly, to a method for multi-scale automatic analysis of the atomic structure of amorphous materials. Background Art

[0002] Due to the lack of long-range ordered structure, amorphous materials exhibit unique physical and chemical properties and show broad application prospects in fields such as new energy and aerospace. However, their complex atomic arrangement characteristics also make their structure analysis extremely challenging. Currently, the mainstream methods for analyzing the atomic structure of amorphous materials mainly include Reverse Monte Carlo (RMC) simulation and Molecular Dynamics (MD) simulation after high-temperature annealing.

[0003] Reverse Monte Carlo simulation uses the pair distribution function (PDF) data obtained from experiments and combines the Monte Carlo algorithm to randomly adjust atomic coordinates to match the experimental results. However, this method has significant limitations: First, RMC is an empirical algorithm, and the atomic structure of the amorphous material generated is not verified by first-principles or kinetic simulations, resulting in a lack of theoretical support for the rationality of the structure; Second, this method cannot capture the dynamic behavior during the formation of amorphous materials and is difficult to explain the relationship between local order and macroscopic properties of amorphous materials.

[0004] Although the high-temperature annealing molecular dynamics simulation can overcome some of the above problems of the reverse Monte Carlo simulation, it is limited by the initial crystal structure, and the MD simulation results often deviate from the atomic structure characteristics of the amorphous material observed in experiments. For example, traditional MD simulations initially require setting a crystal or ordered structure as a starting point, and the complex kinetic behaviors during the formation of amorphous materials (such as lattice distortion and defect diffusion) are difficult to fully characterize through the conventional simulation time scale (usually < 100 picoseconds).

[0005] In summary, the reverse Monte Carlo simulation cannot directly analyze new materials without experimental data because it relies on experimental PDF data; while the high-temperature annealing molecular dynamics simulation is limited by the initial crystal structure and is difficult to obtain the amorphous structure synthesized in experiments. The deficiencies of the existing technologies in material theoretical calculations can be summarized as the following key problems: (1) Contradiction between computational efficiency and scale: Traditional first-principles methods are only applicable to small-scale systems (such as less than 200 atoms) and are difficult to directly simulate large amorphous systems (usually more than 1000 atoms), resulting in unrepresentative and unreasonable analysis results.

[0006] (2)Kinetic time scale limitation: The process of amorphous formation involves complex kinetic behaviors, and it is difficult for conventional molecular dynamics simulations to span a sufficient time scale (>100 picoseconds) to capture the details of structural evolution.

[0007] (3)Potential function accuracy limitation: Empirical potential functions (such as Lennard-Jones, Born-Mayer, embedded-atom model EAM, etc.) cannot accurately describe the complex interactions in multi-component systems (such as amorphous alloys), resulting in significant errors in structure prediction.

[0008] (4)Inefficiency of manual intervention: Existing methods rely on manual intervention for data transfer and parameter adjustment, leading to cumulative errors and a process that takes up to several weeks.

[0009] In addition, existing solutions mostly adopt single-scale simulations (such as relying only on experimental data) or manually intervened data transfer (such as manually adjusting parameters), which have problems such as low efficiency and error accumulation, and it is difficult to achieve the automation of analyzing the atomic structure of amorphous materials from the atomic scale. In the prior art, there is no method that combines the high precision of first-principles with the high efficiency of machine learning potential functions and simultaneously realizes the full automation of analyzing amorphous structures from atomic clusters to large-scale amorphous structures. Summary of the Invention

[0010] The technical problem to be solved by the present invention is that the existing methods for analyzing the atomic structure of amorphous materials are inefficient and error-prone, and it is difficult to achieve the automation of analyzing the atomic structure of amorphous materials from the atomic scale. To overcome the above defects of the prior art, the present invention provides a multi-scale automatic analysis method for the atomic structure of amorphous materials.

[0011] The present invention provides a multi-scale automatic analysis method for the atomic structure of amorphous materials, including the following steps: Step S1: Using the reactants for forming the amorphous material as the initial units, obtaining cluster units generated by the reactants based on first-principles molecular dynamics simulation; Step S2: Based on the cluster units, optimizing their unit cell parameters and atomic coordinates through first-principles calculation to obtain unit cell units; Step S3: Stacking the unit cell units in space, and optimizing their unit cell parameters and atomic coordinates through first-principles calculation to obtain supercell units; Step S4: Conducting first-principles molecular dynamics simulation on the supercell units, sampling the simulation trajectories to form a dataset, and training and obtaining an accurate machine learning potential function using the dataset; Step S5: Conducting cell expansion construction on the supercell units, and performing machine learning potential function-driven molecular dynamics simulation based on the machine learning potential function to obtain the atomic structure of the amorphous material. Step S6: Package the steps S1 to S5 into an automated workflow for automatically parsing the atomic structure of amorphous materials.

[0012] Compared with the prior art, the present invention has the following advantages: By constructing a multi-scale modeling system of chemical reaction → cluster unit → unit cell unit → supercell unit → atomic structure of amorphous material, deeply integrating machine learning algorithms with an automated calculation process, combining first-principles calculations, high-precision machine learning potential functions, and an automated workflow engine, the present invention systematically overcomes the long-standing cross-scale modeling bottleneck and the accuracy-efficiency balance problem in the parsing of the atomic structure of amorphous materials. The technical solution of the present invention improves the accuracy of parsing the atomic structure of amorphous materials and is of great significance for promoting the application of amorphous material science and engineering.

[0013] In a possible implementation manner, step S1 specifically includes the following steps: Step S1.1: Construct the minimum cluster unit of the reactant, where the minimum cluster unit is the minimum molecular composition in the reactant; Step S1.2: Determine the reaction ratio between each of the minimum cluster units, construct each of the minimum cluster units in the same unit cell, with the vacuum layer thickness in each direction being greater than 1 nm and the atomic spacing between adjacent minimum cluster units being greater than 1 nm, to form a cluster reaction model; Step S1.3: Perform first-principles molecular dynamics simulation on the cluster reaction model, using the NVT ensemble with constant number of particles, volume, and temperature, with the simulation duration being greater than 3 picoseconds, to generate cluster units.

[0014] In the above solution, first-principles MD is suitable for theoretical calculations of systems with a small size (<100 atoms) and a short time scale (<100 ps). It can achieve high-precision simulation of chemical reaction processes without relying on empirical parameters, can dynamically capture atomic trajectories and energy evolution, thereby accurately characterizing bond-breaking / formation events in the reaction path, and obtaining chemical reaction products in a real environment. Therefore, the cluster units generated based on first-principles molecular dynamics simulation are accurate.

[0015] In a possible implementation manner, step S2 specifically includes the following steps: Step S2.1: Expand the cluster unit into a periodic cluster unit, where the vacuum layer thickness of the periodic cluster unit does not exceed 1 nm; Step S2.2: Based on first-principles calculations, perform geometric optimization of the unit cell parameters and atomic coordinates of the periodic cluster unit to obtain unit cell units.

[0016] In the above solution, for the unit cell with periodic characteristics, ab initio calculations can perform high-precision simulation calculations without empirical parameters. Therefore, accurate unit cells can be simulated using ab initio calculations.

[0017] In a possible implementation, in step S2.2, the specific parameters for optimizing the unit cell parameters and atomic coordinates of the periodic cluster unit are as follows: the unit cell shape and symmetry are variable, and the external pressure is 1 - 1000 atmospheres.

[0018] In a possible implementation, step S3 specifically includes the following steps: Step S3.1: Stack the unit cells in space in a close-packed manner to form a supercell, and the size of the supercell is at least doubled in all three XYZ directions relative to the unit cell. Step S3.2: Perform geometric optimization of the unit cell parameters and atomic coordinates of the supercell based on ab initio calculations to obtain a supercell unit.

[0019] In a possible implementation, in step S3.2, the specific parameters for optimizing the unit cell parameters and atomic coordinates of the supercell are as follows: the unit cell shape and symmetry are variable, and the external pressure is 1 - 1000 atmospheres.

[0020] In a possible implementation, step S4 specifically includes the following steps: Step S4.1: Perform ab initio molecular dynamics simulation on the supercell unit, using the NVT ensemble with constant number of particles, volume, and temperature, and the simulation duration is greater than 3 picoseconds. Step S4.2: Extract the data of atomic coordinates, unit cell parameters, structural energy, atomic forces, and unit cell stress from the simulation trajectory of the ab initio molecular dynamics simulation in step S4.1, and construct a high-dimensional training dataset. Step S4.3: Train a machine learning potential function based on the high-dimensional training dataset. When predicting the structural energy, atomic forces, and unit cell stress, the coefficient of determination between the observed data of each physical quantity and the test set data should be higher than 0.98 to obtain a well-trained machine learning potential function.

[0021] In a possible implementation, step S5 specifically includes the following steps: Step S5.1: Expand the supercell unit to construct an initial amorphous material atomic structure, expanding at least 2 times in all three XYZ directions. Step S5.2: Perform molecular dynamics simulation driven by the machine learning potential function on the initial amorphous material atomic structure using the accurate machine learning potential function obtained in step S4 to obtain the final amorphous material atomic structure.

[0022] In the above solution, the high-precision machine learning potential function constructed based on first-principles calculations provides a reliable energy-force field description for the calculation of the large-scale long-range disordered structure of amorphous materials. When dealing with such ultra-large systems containing more than thousands of atoms, the molecular dynamics simulation driven by the machine learning potential function can significantly improve the calculation efficiency while maintaining a reasonable description of the atomic-scale bonding and relaxation processes. This method effectively balances the calculation accuracy and resource consumption while ensuring the physical rationality of the structural evolution, making the reliability of the predicted atomic structure of amorphous materials reach the experimental characterization level.

[0023] In a possible implementation manner, step S5.2 specifically includes the following processes: Step S5.2.1: Perform molecular dynamics simulation driven by the machine learning potential function on the initial atomic structure of the amorphous material. Select the NPT ensemble with a constant number of particles, pressure, and temperature. Limit the pressure to 1000 - 10000 atmospheres, the temperature to be greater than 300 Kelvin, and the simulation duration to be greater than 50 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material. Step S5.2.2: Perform molecular dynamics simulation driven by the machine learning potential function for heating and cooling on the atomic structure of the pressure-equilibrated amorphous material obtained in step S5.2.1. Select the NVT ensemble with a temperature range of 300 - 400 Kelvin and a heating / cooling rate greater than 1 picosecond per Kelvin to obtain the final atomic structure of the amorphous material.

[0024] In a possible implementation manner, step S6 specifically includes the following steps: Step S6.1: Package step S1 into a reactant input module and a cluster unit module; Step S6.2: Package step S2 into a unit cell module; Step S6.3: Package step S3 into a supercell unit module; Step S6.4: Package step S4 into a machine learning potential function training module; Step S6.5: Package step S5 into an amorphous material atomic structure simulation module; Step S6.6: Integrate the modules obtained in steps S6.1 to S6.5.

[0025] The beneficial effects of the present invention are as follows: 1. The present invention constructs a multi-scale modeling framework of chemical reaction principle → cluster unit → unit cell → supercell → atomic structure of amorphous material, and drives the structural evolution through chemical reaction kinetic mechanisms (such as diffusion, nucleation, phase change). Compared with the existing methods, the accuracy and rationality of the amorphous structure are improved and guaranteed.

[0026] 2. The present invention adopts an accurately trained machine learning potential function, which can accurately and efficiently simulate complex amorphous materials (such as amorphous materials containing more than 5 elements). Compared with existing solutions, it has the advantages of accuracy and efficiency.

[0027] 3. The present invention adopts an analytical method from the chemical reaction principle between reactants → cluster unit → unit cell unit → supercell unit → atomic structure of amorphous material, which can capture the atomic rearrangement mechanism (such as the transition from local order to long-range disorder), the defect diffusion path and the phase transition critical point, and observe the specific process of the formation of the atomic structure of amorphous materials. Compared with existing solutions, it has the advantage of accurately and dynamically restoring the atomic structure of amorphous materials.

[0028] 4. The analytical method starting from the chemical reaction principle between reactants adopted by the present invention can predict the amorphous materials formed by the reaction of various reactants, and is applicable to complex systems containing metals / non-metals, main group / transition group elements (such as amorphous high-entropy alloys). Compared with existing solutions, it has advantages such as good robustness and prediction ability.

[0029] 5. The present invention realizes the full-process automation from reactant input to the output of the atomic structure of amorphous materials through code algorithm encapsulation and a modular workflow engine. Compared with existing solutions, it can improve the efficiency of amorphous material structure analysis and reduce the R & D cost of amorphous materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a technical solution flowchart of the multi-scale automatic analysis method for the atomic structure of amorphous materials of the present invention; Figure 2 It is a schematic diagram of reactants in Embodiment 1 of the present invention; Figure 3 It is a schematic diagram of cluster units in Embodiment 1 of the present invention; Figure 4 It is a schematic diagram of unit cell units in Embodiment 1 of the present invention; Figure 5 It is a schematic diagram of supercell units in Embodiment 1 of the present invention; Figure 6 It is a schematic diagram of the accuracy of the structure energy prediction of the machine learning potential function in Embodiment 1 of the present invention; Figure 7 It is a schematic diagram of the accuracy of the atomic force prediction of the machine learning potential function in Embodiment 1 of the present invention; Figure 8 It is a schematic diagram of the accuracy of the unit cell stress prediction of the machine learning potential function in Embodiment 1 of the present invention; Figure 9 It is a schematic diagram of the initial amorphous structure in Embodiment 1 of the present invention; Figure 10Schematic diagram of the atomic structure of the final amorphous material in Embodiment 1 of the present invention; Figure 11 Automation flowchart of module encapsulation in the multi-scale automatic analysis method of the atomic structure of the amorphous material of the present invention; Figure 12 Comparison diagram of the theoretical pair distribution function and the experimental sample pair distribution function of the atomic structure of the final amorphous material in Embodiment 1 of the present invention; Figure 13 Schematic diagram of the atomic structure of the amorphous material obtained in Embodiment 2 of the present invention; Figure 14 Comparison of the theoretical pair distribution function and the experimental sample pair distribution function of the atomic structure of the final amorphous material in Embodiment 2 of the present invention; Figure 15 Schematic diagram of the atomic structure of the amorphous material obtained in Embodiment 3 of the present invention. Detailed implementation manners

[0031] To make the above objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below. It should be noted that the following embodiments are only used to illustrate the implementation methods and typical parameters of the present invention, and are not used to limit the parameter range described in the present invention. Reasonable changes derived therefrom are still within the protection scope of the claims of the present invention.

[0032] It should be noted that the endpoints and any values in the ranges disclosed in this article are not limited to the exact ranges or values. These ranges or values should be understood to include values close to these ranges or values. For numerical ranges, the endpoint values of each range, between the endpoint values of each range and individual point values, and between individual point values can be combined with each other to obtain one or more new numerical ranges, and these numerical ranges should be regarded as specifically disclosed in this article.

[0033] Unless otherwise defined, all terms, symbols and other scientific terms used in this article are intended to have the same meaning as commonly understood by those skilled in the art to which the present invention belongs. In some cases, terms with conventional understood meanings are defined in this article for the purpose of clarification or easy reference. Such definitions in this article should not be understood as indicating a significant difference from the conventional understanding in the art. The technical methods described or cited in this article are generally well understood by those skilled in the art and are adopted by conventional methods.

[0034] Combined with the drawings and embodiments of the present invention, the technical solutions of the present invention will be elaborated in detail below. It should be noted that the specific forms of each structure in the embodiments described in this specification are only exemplary descriptions, aiming to facilitate the understanding of the core technical concept of the present invention. The scope of protection of a method for multi-scale automatic analysis of the atomic structure of amorphous materials proposed by the present invention is not limited to the specific structural forms clearly described in the embodiments. Other embodiments obtained by those skilled in the art through conventional technical means improvement, parameter adjustment or equivalent replacement based on the existing technical knowledge should be included in the protection scope of the present invention.

[0035] A method for multi-scale automatic analysis of the atomic structure of amorphous materials of the present invention includes the following steps: Step S1: Taking the reactants forming the amorphous material and their stoichiometric ratios as the initial objects, based on first-principles molecular dynamics simulation of the chemical reaction process between the reactants, obtaining cluster units with aperiodic arrangement characteristics of short-range order and long-range disorder formed by the reaction of the reactants; Step S2: Based on the cluster units, geometrically optimizing the lattice parameters and atomic coordinates of the cluster units through first-principles calculation to obtain lattice cell units with periodic arrangement characteristics of short-range order and long-range order; Step S3: Expanding and stacking the lattice cell units in space, and geometrically optimizing the lattice parameters and atomic coordinates through first-principles calculation to obtain supercell units with periodic arrangement characteristics of short-range order and long-range order; Step S4: Conducting first-principles molecular dynamics simulation on the supercell units with a simulation duration of not less than 10 picoseconds to obtain the first-principles molecular dynamics simulation trajectory. By extracting data such as atomic coordinates, lattice parameters, structural energy, atomic forces, and lattice cell stress in the simulation trajectory, constructing a high-dimensional training dataset, and using it as the training set, validation set, and test set to train the machine learning potential function to obtain a machine learning potential function with high-precision prediction ability; Step S5: Based on the supercell units, perform periodic cell expansion construction to obtain an initial amorphous structure in a mechanically metastable state without structural optimization. Subsequently, based on the trained machine learning potential function, conduct molecular dynamics simulation driven by the machine learning potential function to finally obtain the atomic structure of the amorphous material with long-range disorder characteristics; Step S6: Package the steps S1 to S5 into an automated workflow for automatically analyzing the atomic structure of amorphous materials.

[0036] In steps S1 - S6, the specific software packages for first - principles calculations and first - principles molecular dynamics simulations can be as follows: When selecting the commercial software VASP, the basis set is selected as the plane - wave basis set, the pseudopotential is set as the PBE - PAW pseudopotential, and the cutoff energy is set to 1.3 times the maximum cutoff energy in the pseudopotential file; when selecting the open - source software CP2K, the basis set is selected as the MOLOPT - DZVP Gaussian basis set, the pseudopotential is set as the GTH - PBE pseudopotential, and the cutoff energies CUTOFF and REAL_CUTOFF are respectively set to be above 600 Ry and 60 Ry; when selecting the domestic software ABACUS, the basis set is selected as the numerical orbital DZP basis set, the cutoff radius of the basis set is selected as 8 Å, the pseudopotential is set as the Dojo - NC - SR pseudopotential, and the real - space cutoff energy is set to be above 100 Ry.

[0037] Step S1 specifically includes the following steps: Step S1.1: Construct the minimum cluster unit of the reactant, and the minimum cluster unit is the minimum molecular composition in the reactant; Step S1.2: Determine the reaction ratio between each of the minimum cluster units, construct each of the minimum cluster units in step S1.1 within the same unit cell, with the vacuum layer thickness in each direction greater than 1 nm and the atomic distance between adjacent minimum cluster units greater than 1 nm, to form a cluster reaction model; Step S1.3: Perform first - principles molecular dynamics simulation on the cluster reaction model in step S1.2, adopt the NVT ensemble with constant number of particles, volume, and temperature, select the exchange - correlation functional as PBE or R 2 SCAN, set the dispersion interaction as D3 or D4 description, the self - consistent field energy convergence criterion is less than 1×10 -5 eV, do not adopt the k - point grid, set the simulation temperature as 400 K, the time step as 2 femtoseconds, the temperature control time as 20 - 50 times the time step, and the simulation duration is greater than 3 picoseconds to obtain cluster units with the characteristics of short - range order and long - range disorder and non - periodic arrangement.

[0038] Step S2 specifically includes the following steps: Step S2.1: Place the cluster units in step S1.3 into a periodic boundary box, with the requirement that the vacuum layer thickness does not exceed 1 nm, to obtain periodic cluster units; Step S2.2: Based on first - principles calculations, perform geometric optimization of the unit - cell parameters and atomic coordinates of the periodic cluster units in step S2.1 to obtain unit - cell units with the characteristics of short - range order and long - range order and periodic arrangement.

[0039] In step S2.2, the specific parameters for geometric optimization of the periodic cluster units for unit - cell parameters and atomic coordinates are: the unit - cell shape and symmetry are variable, and the exchange - correlation functional is selected as PBE or R2 SCAN, set D3 or D4 to describe the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV, and the atomic force convergence criterion is less than 1×10 -3 eV / Å. A uniform k-point mesh of 2×2×2 is adopted, and the external pressure is set to 1 - 1000 atmospheres.

[0040] The specific steps of step S3 are as follows: Step S3.1: Stack the unit cells described in step S2.2 in space in a close-packed manner to form a supercell, and the size of the supercell is at least doubled in all three XYZ directions relative to the unit cell. Step S3.2: Geometrically optimize the lattice parameters and atomic coordinates of the supercell in step S3.1 based on the first principles to obtain a supercell unit with periodic arrangement characteristics of short-range order and long-range order.

[0041] In step S3.2, the specific parameters for geometric optimization of the lattice parameters and atomic coordinates of the supercell are: the lattice shape and symmetry are variable, and the exchange-correlation functional is selected as PBE or R 2 SCAN, set D3 or D4 to describe the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV, and the atomic force convergence criterion is less than 1×10 -3 eV / Å. The k-point mesh is not adopted, and the external pressure is set to 1 - 1000 atmospheres.

[0042] The specific steps of step S4 are as follows: Step S4.1: Perform first-principles molecular dynamics simulation on the supercell unit in step S3.2, adopt the NVT ensemble with constant number of particles, volume and temperature, and the exchange-correlation functional is selected as PBE or R 2 SCAN, set D3 or D4 to describe the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV. The k-point mesh is not adopted, the simulation temperature is set to 400K, the time step is set to 2 femtoseconds, the temperature control time is set to 20 - 50 times the time step, and the simulation duration is not less than 10 picoseconds to obtain the first-principles molecular dynamics simulation trajectory. Step S4.2: Extract the data of atomic coordinates, lattice parameters, structural energy, atomic forces, and lattice stresses from the simulation trajectory of the first-principles molecular dynamics simulation in step S4.1 to obtain the sampling results of the first-principles dynamics simulation and construct a high-dimensional training dataset. Step S4.3: Divide the high-dimensional training dataset described in step S4.2 into a training set, a validation set, and a test set according to the ratio of 75:5:10, and train the machine learning potential function based on an artificial neural network (such as a graph neural network). After 150 - 200 training cycles, a machine learning potential function with high-precision prediction ability is obtained. When predicting the structural energy, atomic forces, and cell stress, the coefficient of determination between the observed data of each physical quantity and the test set data should be higher than 0.98.

[0043] In step S4.3, the specific parameters for training the machine learning potential function based on an artificial neural network (such as a graph neural network) are as follows: When using the graph neural network MACE, the graph neural network selects 128 equivariant features, two message passing layers, and three relevant orders. The angular resolution is set to 1, the cutoff radius is set to 6 Å, the number of radial basis functions is set to 8, and the training weights for structural energy, atomic forces, and cell stress are set to 1, 10, and 100 respectively.

[0044] Step S5 specifically includes the following steps: Step S5.1: Expand the supercell unit described in step S3.2 to construct a cell, expanding it by at least 2 times along the X, Y, and Z directions to obtain the atomic structure of the initial amorphous material. Step S5.2: Use the accurately trained machine learning potential function described in step S4.3 to perform machine learning potential function-driven molecular dynamics simulations on the atomic structure of the initial amorphous material described in step S5.1 to obtain the atomic structure of the amorphous material.

[0045] Step S5.2 specifically includes the following process: Step S5.2.1: Perform machine learning potential function-driven molecular dynamics simulations with constant particle number, pressure, and temperature on the atomic structure of the initial amorphous material described in step S5.1. Use the NPT ensemble, limit the pressure to 1000 - 10000 atmospheres, set the simulation temperature to 300 K, set the time step to 1 femtosecond, set the temperature control time to 20 - 50 times the time step, set the pressure control time to 100 times the time step, and set the simulation duration to be greater than 50 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material. Step S5.2.2: Perform machine learning potential function-driven molecular dynamics simulations of heating and cooling on the atomic structure of the pressure-equilibrated amorphous material obtained in step S5.2.1. Use the NVT ensemble, with a temperature range of 300 - 400 K, a heating / cooling rate greater than 1 picosecond / K, set the time step to 2 femtoseconds, set the temperature control time to 20 - 50 times the time step, and set the simulation duration to be more than 200 picoseconds.

[0046] Step S6: Package the steps from S1 to S5 into an automated workflow for automatically parsing the atomic structure of amorphous materials. If a Python module is used to package the automated workflow, step S6 specifically includes the following steps: Step S6.1: Package steps S1.1 to S1.3 into an automated reactant input Python module and a cluster unit generation Python module. The specific steps are as follows: Step S6.1.1: User interface. The user inputs the reactant molecular formulas (such as Li3PO4, ZrCl4) and their stoichiometric ratios (such as 1:2), and calls Python code to generate the cluster units of the reactants described in step S1.1, with the condition settings as described in step S1.1; Step S6.1.2: Cluster reaction unit construction. Call Python code to construct the cluster units described in step S6.1.1 within the same unit cell, with the atomic spacing between adjacent clusters selected to be greater than 1 nm, to obtain the cluster reaction unit of step S1.2, with the condition settings as described in step S1.2; Step S6.1.3: Call the ASE Python interface to simulate the cluster reaction unit of step S6.1.2 using first-principles molecular dynamics, and call ASE Python to obtain the cluster reaction products, with the condition settings as described in step S1.3.

[0047] Step S6.2: Package steps S2.1 to S2.2 into an automated unit cell optimization Python module. The specific steps are as follows: Step S6.2.1: Call the ASE Python interface to perform cell expansion on the cluster reaction products of step S6.1.3 to obtain a periodic cluster unit, with the condition settings as described in step S2.1; Step S6.2.2: Call the ASE Python interface to optimize the unit cell parameters and atomic coordinates of the periodic cluster unit of step S6.2.1 to obtain a unit cell, with the condition settings as described in step S2.2.

[0048] Step S6.3: Package steps S3.1 to S3.2 into an automated supercell unit construction Python module. The specific steps are as follows: Step S6.3.1: Call the ASE Python interface to perform cell expansion on the unit cell of step S6.1.3 to obtain a supercell structure, with the condition settings as described in step S3.1; Step S6.3.2: Call the ASE Python interface to optimize the unit cell parameters and atomic coordinates of the supercell structure of step S6.3.1 to obtain a supercell unit, with the condition settings as described in step S3.2.

[0049] Step S6.4: Package the steps from S4.1 to S4.3 into an automated Python module for training machine learning potential functions. The specific steps are as follows: Step S6.4.1: Invoke the ASE Python interface to perform first-principles molecular dynamics simulations on the supercell unit in Step S6.1.3, with the conditions set as described in Step S4.1; Step S6.4.2: Invoke the ASE Python interface to extract data such as atomic coordinates, unit cell parameters, structural energy, atomic forces, and unit cell stress from the trajectory of the first-principles molecular dynamics simulation in Step S6.4.1 into an extxyz format file, obtain the first-principles sampling results, and construct a high-dimensional training dataset, with the conditions set as described in Step S4.2; Step S6.4.3: Through the Python interface, optimize the machine learning potential function for the first-principles sampling results in Step S6.4.2 to obtain an accurate machine learning potential function, with the conditions set as described in Step S4.3.

[0050] Step S6.5: Package the steps from S5.1 to S5.2 into an automated Python module for simulating the atomic structure of amorphous materials. The specific steps are as follows: Step S6.5.1: Invoke the ASE Python interface to perform cell expansion on the supercell unit in Step S6.1.3 to obtain an initial amorphous structure, with the conditions set as described in Step S5.1; Step S6.5.2: Invoke the ASE Python interface to perform machine learning potential function-driven molecular dynamics simulations on the initial amorphous structure in Step S6.5.1. After the simulation ends, obtain the final atomic structure of the amorphous material, with the conditions set as described in Step S5.2.

[0051] Step S6.6: Integrate the modules obtained from the above Steps S6.1 to S6.5.

[0052] The encapsulation method of the automated workflow of the present invention can also be other computer module encapsulation forms, such as development and encapsulation using programming languages like C / C++, Ruby, Perl, or Java, and extension in combination with encapsulation forms such as program packaging, containerized deployment, virtual machines, or serverless architectures, which will not be elaborated here.

[0053] The following gives specific automated parsing methods for different types of amorphous materials. The following embodiments all use Python modules to package the automated workflow, and the detailed steps are the same as the specific descriptions in the above Step S6. In the embodiments, they are simply abbreviated as modules and will not be elaborated. Embodiment 1

[0054] Such as Figure 1As shown in the figure, a multi-scale automatic analysis method for the atomic structure of amorphous materials in this embodiment includes the following steps: In this embodiment, the specific software package for first-principles calculation and first-principles molecular dynamics simulation selects the open-source software CP2K, the basis set selects the MOLOPT-DZVP high-quality basis set, the pseudopotential is set to the GTH-PBE pseudopotential, and the cut-off energies CUTOFF and REAL_CUTOFF are respectively set to 800 Ry and 60 Ry.

[0055] First, in step S1, as Figure 2 shown, the reactants are selected as Li3PO4 and ZrCl4, and the reaction ratio between them is selected as 1:2. Then, the acquisition of the cluster units corresponding to the reaction system in step S1 specifically includes the following steps: Step S1.1: Construct the minimum cluster units Li3PO4 and ZrCl4 of the reactants; Step S1.2: Establish a cluster reaction model according to the reaction ratio of 1:2, construct each of the minimum cluster units in step S1.1 in the same unit cell, the vacuum layer thickness in each direction is 1.5 nm, and the atomic spacing between adjacent minimum cluster units is 1.5 nm to form a cluster reaction model; Step S1.3: Perform first-principles molecular dynamics simulation on the cluster reaction model in step S1.2, adopt the NVT ensemble with constant number of particles, volume and temperature, select the R 2 SCAN for the exchange-correlation functional, set the dispersion interaction to the D3 description, the self-consistent field energy convergence criterion is less than 1×10 -5 eV, do not adopt the k-point grid, the simulation temperature is set to 400 K, the time step is set to 2 femtoseconds, the temperature control time is set to 30 times the time step, and the simulation duration is set to 10 picoseconds to obtain cluster units with the characteristics of short-range order and long-range disorder and non-periodic arrangement, as Figure 3 shown.

[0056] Then, perform step S2, which specifically includes the following steps: Step S2.1: Place the cluster units in step S1.3 in a periodic boundary box, requiring the vacuum layer thickness not to exceed 1 nm to obtain periodic cluster units; Step S2.2: Geometrically optimize the unit cell parameters and atomic coordinates of the periodic cluster units in step S2.1 based on first-principles calculation, set the unit cell shape and symmetry to be variable, select the R 2 SCAN for the exchange-correlation functional, set the D3 description for the dispersion interaction, the self-consistent field energy convergence criterion is less than 1×10 -5 eV, and the atomic force convergence criterion is less than 1×10 -3eV / Å, a uniform k - point grid of 2×2×2 is adopted, the external pressure is set to 100 atmospheres, and a unit cell with the characteristics of periodic arrangement of short - range order and long - range order is obtained, as Figure 4 shown.

[0057] Then step S3 is carried out, which specifically includes the following steps: Step S3.1: Stack the unit cells described in step S2.2 in space in a close - packed manner to form a supercell, and the size of the supercell is enlarged by 2 times along the XYZ three directions relative to the unit cell; Step S3.2: Geometrically optimize the cell parameters and atomic coordinates of the supercell in step S3.1 based on the first - principles, set the cell shape and symmetry to be variable, select the exchange - correlation functional R 2 SCAN, set D3 to describe the dispersion effect, the self - consistent field energy convergence criterion is less than 1×10 -5 eV, the atomic force convergence criterion is less than 1×10 -3 eV / Å, no k - point grid is adopted, the external pressure is set to 100 atmospheres, and a supercell unit with the characteristics of periodic arrangement of short - range order and long - range order is obtained, as Figure 5 shown.

[0058] Then step S4 is carried out, which specifically includes the following steps: Step S4.1: Carry out first - principles molecular dynamics simulation on the supercell unit in step S3.2, adopt the NVT ensemble with constant number of particles, volume and temperature, select the exchange - correlation functional R 2 SCAN, set D3 to describe the dispersion effect, the self - consistent field energy convergence criterion is less than 1×10 -5 eV, no k - point grid is adopted, the simulation temperature is set to 400K, the time step is set to 2 femtoseconds, the temperature control time is set to 30 times the time step, and the simulation duration is set to 20 picoseconds to obtain the first - principles molecular dynamics simulation trajectory; Step S4.2: Take the data of atomic coordinates, cell parameters, structural energy, atomic force, and cell stress in the simulation trajectory of the first - principles molecular dynamics simulation in step S4.1 to obtain the first - principles dynamics simulation sampling results and construct a high - dimensional training data set; Step S4.3: Divide the high-dimensional training dataset described in step S4.2 into a training set, a validation set, and a test set in the ratio of 75:5:10. Train the graph neural network-based machine learning potential function MACE. The training parameters are set as follows: the graph neural network selects 128 equivariant features, two message passing layers, and three relevant orders; the angular resolution is set to 1, the cutoff radius is set to 6 Å, the radial basis function is set to 8; the training weights for the structural energy, atomic force, and cell stress are set to 1, 10, and 100 respectively. After 150 training epochs, a machine learning potential function with high-precision prediction ability is obtained. When predicting the structural energy, atomic force, and cell stress, as Figures 6 - 8 shown, the RMSE errors of the structural energy, atomic force, and lattice strain predicted by the machine learning potential function are 0.2 meV / atom, 9.2 meV / Å, and 364 bar respectively. The determination coefficients between the observed data of each physical quantity and the test set data are all higher than 0.98. Finally, a machine learning potential function with accurate training is obtained.

[0059] Then proceed to step S5, which specifically includes the following steps: Step S5.1: Expand the supercell unit described in step S3.2. Expand it by at least 2 times along the X, Y, and Z directions to obtain the atomic structure of the initial amorphous material, as Figure 9 shown; Step S5.2: Use the accurately trained machine learning potential function described in step S4.3 to perform machine learning potential function-driven molecular dynamics simulation on the atomic structure of the initial amorphous material described in step S5.1, as Figure 10 shown, to obtain the atomic structure of the amorphous material. The machine learning potential function-driven molecular dynamics simulation specifically includes the following processes: Step S5.2.1: Perform machine learning potential function-driven molecular dynamics simulation with constant particle number, pressure, and temperature on the atomic structure of the initial amorphous material described in step S5.1. Use the NPT ensemble, limit the pressure to 3000 atmospheres, set the simulation temperature to 300 K, set the time step to 1 femtosecond, set the temperature control time to 30 times the time step, set the pressure control time to 100 times the time step, and set the simulation duration to 100 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material; Step S5.2.2: Perform machine learning potential function-driven molecular dynamics simulation of heating and cooling on the atomic structure of the pressure-equilibrated amorphous material obtained in step S5.2.1. Use the NVT ensemble, with a temperature range of 300 - 400 K, set the heating / cooling rate to 2 picoseconds / K, set the time step to 2 femtoseconds, set the temperature control time to 30 times the time step, and set the simulation duration to 400 picoseconds to obtain the final atomic structure of the amorphous material.

[0060] Finally, step S6 is carried out, which specifically includes the following steps, as Figure 11 shown below: Step S6.1: Package the step S1 as a reactant input module and a cluster unit module; Step S6.2: Package the step S2 as a unit cell module; Step S6.3: Package the step S3 as a supercell unit module; Step S6.4: Package the step S4 as a machine learning potential function training module; Step S6.5: Package the step S5 as an amorphous material atomic structure simulation module; Step S6.6: Integrate the modules obtained from the steps S6.1 to S6.5.

[0061] To further illustrate the accuracy of the method for analyzing the atomic structure of amorphous materials in the present invention, as Figure 12 shown, the pair distribution function of the atomic structure of the amorphous material obtained in this embodiment is calculated. The result obtained by the present invention is basically consistent with the pair distribution function measured in the experiment, verifying the accuracy of the method for multi-scale automatic analysis of the atomic structure of an amorphous material in the present invention. Example 2

[0062] A multi-scale automatic analysis method for the atomic structure of an amorphous material in this embodiment includes the following steps: In this embodiment, for the first-principles calculation and the first-principles molecular dynamics simulation, the specific software package selects the commercial software VASP, the basis set selects the plane wave basis set, the pseudopotential is set as the PBE-PAW pseudopotential, and the cutoff energy is set to 1.3 times the maximum cutoff energy in the pseudopotential file.

[0063] First, in step S1, the reactants are selected as Li2SO4 and ZrCl4, and the reaction ratio between the two is selected as 1:2. Then, the acquisition of the cluster unit in the step S1 corresponding to this reaction system specifically includes the following steps: Step S1.1: Construct the minimum cluster units of the reactants Li2SO4 and ZrCl4; Step S1.2: Establish a cluster reaction model according to the reaction ratio of 1:2. Construct each of the minimum cluster units in step S1.1 in the same unit cell, with a vacuum layer thickness of 1.5 nm in each direction and an atomic spacing of 1.5 nm between adjacent minimum cluster units to form a cluster reaction model; Step S1.3: Perform first-principles molecular dynamics simulation on the cluster reaction model in step S1.2, using the NVT ensemble with constant number of particles, volume, and temperature, and selecting the exchange-correlation functional as R 2SCAN, the description of the dispersion effect is set to D3, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV. The k-point grid is not used. The simulation temperature is set to 400 K, the time step is set to 2 femtoseconds, the temperature control time is set to 30 times the time step, and the simulation duration is set to 10 picoseconds to obtain cluster units with the characteristics of short-range order and long-range disorder and an aperiodic arrangement.

[0064] Then comes step S2, which specifically includes the following steps: Step S2.1: Place the cluster units described in step S1.3 in a periodic boundary box, requiring the vacuum layer thickness not to exceed 1 nm to obtain periodic cluster units; Step S2.2: Based on first-principles calculations, perform geometric optimization of the unit cell parameters and atomic coordinates of the periodic cluster units described in step S2.1. The unit cell shape and symmetry are set to be variable, and the exchange-correlation functional is selected as R 2 SCAN, set the D3 description of the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV, and the atomic force convergence criterion is less than 1×10 -3 eV / Å. A uniform k-point grid of 2×2×2 is used, and the external pressure is set to 100 atmospheres to obtain unit cell units with the characteristics of short-range order and long-range order and a periodic arrangement.

[0065] Then comes step S3, which specifically includes the following steps: Step S3.1: Stack the unit cell units described in step S2.2 in space in a close-packed manner to form a supercell, and the size of the supercell is enlarged by 2 times in the XYZ three directions relative to the unit cell units; Step S3.2: Based on first principles, perform geometric optimization of the unit cell parameters and atomic coordinates of the supercell described in step S3.1. The unit cell shape and symmetry are set to be variable, and the exchange-correlation functional is selected as R 2 SCAN, D3 description of the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV, and the atomic force convergence criterion is less than 1×10 -3 eV / Å. The k-point grid is not used, and the external pressure is set to 100 atmospheres to obtain supercell units with the characteristics of short-range order and long-range order and a periodic arrangement.

[0066] Then comes step S4, which specifically includes the following steps: Step S4.1: Perform first-principles molecular dynamics simulation on the supercell units described in step S3.2, using the NVT ensemble with constant number of particles, volume, and temperature. The exchange-correlation functional is selected as R 2 SCAN, set the D3 description of the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5eV. Without using a k-point mesh, the simulation temperature is set to 400 K, the time step is set to 2 femtoseconds, the temperature control time is set to 30 times the time step, and the simulation duration is set to 20 picoseconds to obtain the first-principles molecular dynamics simulation trajectory; Step S4.2: Take the data of atomic coordinates, unit cell parameters, structural energy, atomic forces, and unit cell stress in the first-principles molecular dynamics simulation trajectory described in Step S4.1 to obtain the first-principles dynamics simulation sampling results and construct a high-dimensional training dataset; Step S4.3: Divide the high-dimensional training dataset described in Step S4.2 into a training set, a validation set, and a test set according to the ratio of 75:5:10, and train the graph neural network-based machine learning potential function MACE. The training parameters are set as follows: the graph neural network selects 128 equivariant features, two layers of message passing layers, and three relevant orders, the angular resolution is set to 1, the cutoff radius is set to 6 Å, the radial basis function is set to 8, and the training weights of structural energy, atomic forces, and unit cell stress are set to 1, 10, and 100 respectively. After 150 training epochs, a machine learning potential function with high-precision prediction ability is obtained. When predicting structural energy, atomic forces, and unit cell stress, the RMSE errors of the structural energy, atomic forces, and lattice strain predicted by the machine learning potential function are 0.3 meV / atom, 10.1 meV / angstrom, and 301 bar respectively. The determination coefficients between the observed data of each physical quantity and the test set data are all higher than 0.98, and finally a machine learning potential function with accurate training is obtained.

[0067] Then comes Step S5, which specifically includes the following steps: Step S5.1: Expand the supercell unit described in Step S3.2 by at least 2 times along the X, Y, and Z directions to obtain the atomic structure of the initial amorphous material; Step S5.2: Use the accurately trained machine learning potential function described in Step S4.3 to perform machine learning potential function-driven molecular dynamics simulation on the atomic structure of the initial amorphous material described in Step S5.1 to obtain the atomic structure of the amorphous material, as Figure 13 shown. The machine learning potential function-driven molecular dynamics simulation specifically includes the following process: Step S5.2.1: Perform machine learning potential function-driven molecular dynamics simulation with constant number of particles, pressure, and temperature on the atomic structure of the initial amorphous material described in Step S5.1. Use the NPT ensemble, limit the pressure to 3000 atmospheres, set the simulation temperature to 300 K, the time step to 1 femtosecond, the temperature control time to 30 times the time step, the pressure control time to 100 times the time step, and the simulation duration to 100 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material; Step S5.2.2: Perform molecular dynamics simulations driven by machine learning potential functions for heating and cooling the atomic structure of the pressure-balanced amorphous material obtained in Step S5.2.1. Use the NVT ensemble, with a temperature range of 300 - 400 K, a heating / cooling rate set to 2 picoseconds / K, a time step set to 2 femtoseconds, a temperature control time set to 30 times the time step, and a simulation duration set to 400 picoseconds to obtain the atomic structure of the final amorphous material, as Figure 13 shown.

[0068] Finally, there is Step S6, which specifically includes the following steps: Step S6.1: Package the said Step S1 into a reactant input module and a cluster unit module; Step S6.2: Package the said Step S2 into a unit cell module; Step S6.3: Package the said Step S3 into a supercell unit module; Step S6.4: Package the said Step S4 into a machine learning potential function training module; Step S6.5: Package the said Step S5 into an amorphous material atomic structure simulation module; Step S6.6: Integrate the various modules obtained in Steps S6.1 to S6.5.

[0069] To further illustrate the accuracy of the method for analyzing the atomic structure of amorphous materials in the present invention, as Figure 14 shown, the pair distribution function of the atomic structure of the amorphous material obtained in this embodiment was calculated. The results obtained in the present invention are basically in agreement with the pair distribution function tested in the experiment, verifying the accuracy of a multi-scale automatic analysis method for the atomic structure of amorphous materials in the present invention. Example 3

[0070] A multi-scale automatic analysis method for the atomic structure of an amorphous material in this embodiment includes the following steps: In this embodiment, for the first-principles calculation and the first-principles molecular dynamics simulation, the specific software package selects the commercial software VASP, the basis set selects the plane wave basis set, the pseudopotential is set to the PBE-PAW pseudopotential, and the cutoff energy is set to 1.3 times the maximum cutoff energy in the pseudopotential file.

[0071] First, in Step S1, the reactants are selected as LiCl and ZrCl4, and the reaction ratio between the two is selected as 2:1. Then, the acquisition of the cluster unit in Step S1 corresponding to this reaction system specifically includes the following steps: Step S1.1: Construct the minimum cluster units of the reactants, LiCl and ZrCl4; Step S1.2: Establish a cluster reaction model according to a reaction ratio of 2:1, construct each of the minimum cluster units in Step S1.1 within the same unit cell, with a vacuum layer thickness of 1.5 nm in each direction and an atomic spacing of 1.5 nm between adjacent minimum cluster units to form a cluster reaction model; Step S1.3: Perform first-principles molecular dynamics simulation on the cluster reaction model in Step S1.2, using the NVT ensemble with constant number of particles, volume, and temperature, select the R 2 SCAN for the exchange-correlation functional, set the dispersion interaction to D3 description, with the self-consistent field energy convergence criterion less than 1×10 -5 eV, do not use the k-point mesh, set the simulation temperature to 400 K, the time step to 2 femtoseconds, the temperature control time to 30 times the time step, and the simulation duration to 10 picoseconds to obtain cluster units with the characteristics of short-range order and long-range disorder and non-periodic arrangement.

[0072] Then comes Step S2, which specifically includes the following steps: Step S2.1: Place the cluster units in Step S1.3 into a periodic boundary box, with the requirement that the vacuum layer thickness does not exceed 1 nm, to obtain periodic cluster units; Step S2.2: Based on first-principles calculations, perform geometric optimization of the cell parameters and atomic coordinates of the periodic cluster units in Step S2.1, set the cell shape and symmetry to be variable, select the R 2 SCAN for the exchange-correlation functional, set the D3 description for the dispersion interaction, with the self-consistent field energy convergence criterion less than 1×10 -5 eV, and the atomic force convergence criterion less than 1×10 -3 eV / Å, use a 2×2×2 uniform k-point mesh, and set the external pressure to 100 atmospheres to obtain cell units with the characteristics of short-range order and long-range order and periodic arrangement.

[0073] Then comes Step S3, which specifically includes the following steps: Step S3.1: Stack the cell units in Step S2.2 in a close-packed manner in space to form a supercell, and the size of the supercell is enlarged by 2 times along the XYZ three directions relative to the cell units; Step S3.2: Based on first-principles, perform geometric optimization of the cell parameters and atomic coordinates of the supercell in Step S3.1, set the cell shape and symmetry to be variable, select the R 2 SCAN for the exchange-correlation functional, D3 description for the dispersion interaction, with the self-consistent field energy convergence criterion less than 1×10 -5 eV, and the atomic force convergence criterion less than 1×10 -3eV / Å, without using a k-point mesh, the external pressure is set to 100 atmospheres, and a supercell unit with periodic arrangement characteristics of short-range order and long-range order is obtained.

[0074] Then comes step S4, which specifically includes the following steps: Step S4.1: Perform ab initio molecular dynamics simulation on the supercell unit described in step S3.2. Use the NVT ensemble with constant number of particles, volume, and temperature. The exchange-correlation functional is selected as R 2 SCAN, set D3 to describe the dispersion effect, and the self-consistent field energy convergence criterion is less than 1×10 -5 eV, without using a k-point mesh, the simulation temperature is set to 400 K, the time step is set to 2 femtoseconds, the temperature control time is set to 30 times the time step, and the simulation duration is set to 20 picoseconds to obtain the ab initio molecular dynamics simulation trajectory; Step S4.2: Take the data of atomic coordinates, unit cell parameters, structural energy, atomic forces, and unit cell stresses in the simulation trajectory of the ab initio molecular dynamics simulation described in step S4.1 to obtain the ab initio dynamics simulation sampling results and construct a high-dimensional training dataset; Step S4.3: Divide the high-dimensional training dataset described in step S4.2 into a training set, a validation set, and a test set according to the ratio of 75:5:10. Train the graph neural network-based machine learning potential function MACE. The training parameters are set as follows: the graph neural network selects 128 equivariant features, two layers of message passing layers, and three correlation orders. The angular resolution is set to 1, the cutoff radius is set to 6 Å, the radial basis function is set to 8, and the training weights of structural energy, atomic forces, and unit cell stresses are set to 1, 10, and 100 respectively. After 150 training epochs, a machine learning potential function with high-precision prediction ability is obtained. When predicting structural energy, atomic forces, and unit cell stresses, the RMSE errors of the structural energy, atomic forces, and lattice strain predicted by the machine learning potential function are 0.3 meV / atom, 9.5 meV / Å, and 350 bar respectively. The determination coefficients between the observed data of each physical quantity and the test set data are all higher than 0.98, and finally a machine learning potential function with accurate training is obtained.

[0075] Then comes step S5, which specifically includes the following steps: Step S5.1: Expand the supercell unit described in step S3.2, expand it by at least 2 times along the XYZ three directions to obtain the atomic structure of the initial amorphous material; Step S5.2: Use the accurately trained machine learning potential function described in step S4.3 to perform machine learning potential function-driven molecular dynamics simulation on the atomic structure of the initial amorphous material described in step S5.1 to obtain the atomic structure of the amorphous material, as Figure 15As shown, the molecular dynamics simulation driven by the machine learning potential function specifically includes the following processes: Step S5.2.1: Perform a molecular dynamics simulation driven by the machine learning potential function with constant number of particles, pressure, and temperature on the atomic structure of the initial amorphous material described in Step S5.1. Use the NPT ensemble, limit the pressure to 3000 atmospheres, set the simulation temperature to 300 K, set the time step to 1 femtosecond, set the temperature control time to 30 times the time step, set the pressure control time to 100 times the time step, and set the simulation duration to 100 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material; Step S5.2.2: Perform a molecular dynamics simulation driven by the machine learning potential function for heating and cooling on the atomic structure of the pressure-equilibrated amorphous material obtained in Step S5.2.1. Use the NVT ensemble, with a temperature range of 300 - 400 K, set the heating / cooling rate to 2 picoseconds / K, set the time step to 2 femtoseconds, set the temperature control time to 30 times the time step, and set the simulation duration to 400 picoseconds to obtain the atomic structure of the final amorphous material, as Figure 15 shown.

[0076] Finally, there is Step S6, which specifically includes the following steps: Step S6.1: Package the said Step S1 into a reactant input module and a cluster unit module; Step S6.2: Package the said Step S2 into a unit cell module; Step S6.3: Package the said Step S3 into a supercell unit module; Step S6.4: Package the said Step S4 into a machine learning potential function training module; Step S6.5: Package the said Step S5 into an amorphous material atomic structure simulation module; Step S6.6: Integrate the various modules obtained from Step S6.1 to Step S6.5.

[0077] As described above, this is the specific implementation manner of the present application. However, the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-scale automatic analysis method for the atomic structure of amorphous materials, characterized in that It includes the following steps: Step S1: Taking the reactants for forming the amorphous material as the initial units, obtaining the cluster units generated by the reactants based on the first-principles molecular dynamics simulation; Step S2: Based on the cluster units, optimizing their unit cell parameters and atomic coordinates by first-principles calculation to obtain the unit cell units; Step S3: Stacking the unit cell units in space, optimizing their unit cell parameters and atomic coordinates by first-principles calculation to obtain the supercell units; Step S4: Conducting first-principles molecular dynamics simulation on the supercell units, sampling the simulation trajectories for dataset, and training with the dataset to obtain an accurate machine learning potential function; Step S5: Conducting cell expansion construction on the supercell units, and performing machine learning potential function-driven molecular dynamics simulation based on the machine learning potential function to obtain the atomic structure of the amorphous material; Step S6: Encapsulating Steps S1 to S5 into an automated workflow for automatically analyzing the atomic structure of the amorphous material.

2. The method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 1, wherein The specific steps of Step S1 include the following steps: Step S1.1: Constructing the minimum cluster units of the reactants, where the minimum cluster units are the minimum molecular compositions in the reactants; Step S1.2: Determining the reaction ratios between the minimum cluster units, constructing all the minimum cluster units in the same unit cell, with the vacuum layer thickness in all directions being greater than 1 nm and the atomic spacing between adjacent minimum cluster units being greater than 1 nm to form a cluster reaction model; Step S1.3: Conducting first-principles molecular dynamics simulation on the cluster reaction model, using the NVT ensemble with constant number of particles, volume and temperature, and the simulation duration being greater than 3 picoseconds to generate the cluster units.

3. The multi-scale automatic analysis method for the atomic structure of an amorphous material according to claim 1, wherein The specific steps of Step S2 include the following steps: Step S2.1: Expanding the cluster units into periodic cluster units, where the vacuum layer thickness of the periodic cluster units does not exceed 1 nm; Step S2.2: Conducting geometric optimization of the unit cell parameters and atomic coordinates of the periodic cluster units by first-principles calculation to obtain the unit cell units.

4. The multi-scale automatic analysis method for the atomic structure of an amorphous material according to claim 3, wherein, In Step S2.2, the specific parameters for optimizing the unit cell parameters and atomic coordinates of the periodic cluster units are: the unit cell shape and symmetry are variable, and the external pressure is 1 - 1000 atmospheres.

5. A method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 1, characterized in that, The specific steps of Step S3 include the following steps: Step S3.1: Stacking the unit cell units in space in a close-packed manner to form a supercell, where the size of the supercell is at least doubled along the XYZ three directions relative to the unit cell units; Step S3.2: Conducting geometric optimization of the unit cell parameters and atomic coordinates of the supercell by first-principles calculation to obtain the supercell units.

6. The method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 5, wherein, In Step S3.2, the specific parameters for optimizing the unit cell parameters and atomic coordinates of the supercell are: the unit cell shape and symmetry are variable, and the external pressure is 1 - 1000 atmospheres.

7. A method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 1, characterized in that The specific steps of Step S4 include the following steps: Step S4.1: Conducting first-principles molecular dynamics simulation on the supercell units, using the NVT ensemble with constant number of particles, volume and temperature, and the simulation duration being greater than 3 picoseconds; Step S4.2: Extract the data of atomic coordinates, unit cell parameters, structural energy, atomic forces, and unit cell stress from the simulation trajectory of the first-principles molecular dynamics simulation in Step S4.1, and construct a high-dimensional training dataset; Step S4.3: Train a machine learning potential function based on the high-dimensional training dataset. When predicting the structural energy, atomic forces, and unit cell stress, the coefficient of determination between the observed data of each physical quantity and the test set data should be higher than 0.98 to obtain an accurate machine learning potential function.

8. A method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 1, characterized in that The specific steps of Step S5 are as follows: Step S5.1: Expand the supercell unit to construct it, expanding at least 2 times along the XYZ three directions to obtain the atomic structure of the initial amorphous material; Step S5.2: Use the accurate machine learning potential function obtained in Step S4 to perform molecular dynamics simulation driven by the machine learning potential function on the atomic structure of the initial amorphous material to obtain the atomic structure of the final amorphous material.

9. The method for multi-scale automatic analysis of the atomic structure of an amorphous material according to claim 8, wherein, Step S5.2 specifically includes the following process: Step S5.2.1: Perform molecular dynamics simulation driven by the machine learning potential function on the atomic structure of the initial amorphous material. Select the NPT ensemble with constant number of particles, pressure, and temperature. Limit the pressure to 1000 - 10000 atmospheres, the temperature to be greater than 300 Kelvin, and the simulation duration to be greater than 50 picoseconds to obtain the atomic structure of the pressure-equilibrated amorphous material; Step S5.2.2: Perform molecular dynamics simulation driven by the machine learning potential function for heating and cooling on the atomic structure of the pressure-equilibrated amorphous material obtained in Step S5.2.

1. Select the NVT ensemble with a temperature range of 300 - 400 Kelvin and a heating / cooling rate greater than 1 picosecond per Kelvin to obtain the atomic structure of the final amorphous material.

10. A method for multi-scale automatic analysis of the atomic structure of an amorphous material according to any one of claims 1-9, characterized in that, The specific steps of Step S6 are as follows: Step S6.1: Package Step S1 into a reactant input module and a cluster unit module; Step S6.2: Package Step S2 into a unit cell module; Step S6.3: Package Step S3 into a supercell unit module; Step S6.4: Package Step S4 into a machine learning potential function training module; Step S6.5: Package Step S5 into an atomic structure simulation module of amorphous materials; Step S6.6: Integrate the various modules obtained in Steps S6.1 to S6.5.

Images