Artificial neural network potential energy function development method for simulating dynamic process of molecular sieve packaging platinum cluster in oxygen environment

Abstract

The invention provides an artificial neural network potential energy function development method for simulating a dynamic process of a molecular sieve packaged platinum cluster in an oxygen environment, which comprises the following steps of: performing potential energy surface sampling on a system of a Silicalites molecular sieve packaged platinum nanocluster and O2 and Pt-O structures by using a random surface walking global optimization method of LASP software; a part of structures are randomly extracted from a data set obtained through sampling, VASP software is used for structural optimization, and the obtained structures and data such as energy and force of the structures are added into a DFT training set. Neural network training is carried out on the DFT data set by means of machine learning, and an artificial neural network potential energy function G-NN is obtained; the artificial neural potential energy function obtained by training not only meets the expectation in calculation precision, but also has the advantage of magnitude order in calculation speed, and overcomes the defects of traditional molecular dynamics in calculation systems and calculation speed.

Description

Technical Field

[0001] The present invention relates to the development of an artificial neural network potential energy function (General Neural Network, abbreviated as G-NN), and specifically to the development of an artificial neural network potential energy function by combining traditional density functional theory (DFT) methods with artificial neural networks through machine learning. Background Art

[0002] Metal nanoparticles hold enormous potential in catalysis, energy production, and environmental protection. However, high temperatures or redox reaction environments can cause sintering, which reduces catalytic activity and even leads to catalyst deactivation. An effective solution is to embed metal nanoparticles into porous materials with defined channel structures, such as zeolites. Zeolites have high hydrothermal stability and are ideal supports for stabilizing metal catalysts. Among known zeolites, metal clusters encapsulated in MFI-type zeolites are frequently used in various hydrogenation, dehydrogenation, and oxidation reactions. Previous studies have demonstrated the importance of atmospheric conditions on the size and performance of metal catalysts in zeolite-type zeolites. However, theoretical simulations of platinum clusters encapsulated in MFI-type zeolites are rare. To explore the theoretical mechanisms underlying the stability, adsorption capacity, and migration of metal catalysts within the pores of zeolite zeolites under reactive atmospheres, first-principles DFT calculations of such systems are essential. However, zeolite systems are large and complex, making structural simulations of large systems difficult using traditional theoretical calculations. Understanding the interactions between platinum clusters and the zeolite under varying reaction conditions requires significant computational time and resources. In recent years, machine learning has advanced rapidly, and numerous artificial neural network potentials have been developed to describe various architectures. Remarkably, even for large architectures, these potentials can maintain high accuracy and computational speed. Therefore, training an artificial neural network potential function that can simulate the dynamics of molecular sieve-encapsulated platinum clusters in an oxygen environment is highly desirable, and its application is of great significance for theoretical research on zeolite molecular sieve-encapsulated metal catalysts. Summary of the Invention

[0003] The purpose of this invention is to train an artificial neural network potential energy function to simulate the dynamic process of molecular sieve-encapsulated platinum clusters in an oxygen environment, thereby revealing the adsorption behavior, stability, and migration mechanism of metal nanoclusters in oxygen-containing zeolites.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] The development method of artificial neural network potential energy function for simulating the dynamic process of molecular sieve encapsulated platinum clusters in oxygen environment includes:

[0006] (1) Use Material Studio software to construct a series of typical periodic and non-periodic chemical structure files;

[0007] (2) Using LASP software to call the first principles calculation software VASP, calculate the energy, force and stress of the structure and construct an initial training set;

[0008] (3) Combine first-principles calculations with global optimization of random surface walks to generate more structure files and expand the training set;

[0009] (4) Develop an artificial neural network potential energy function suitable for the O2-Pt-MFI system using LASP software.

[0010] Optionally, the Material Studio software is used to construct a series of typical periodic and aperiodic chemical structure files, specifically including:

[0011] According to the training requirements, construct a variety of representative chemical structures composed of a single element or two different elements;

[0012] Construct typical structural units with the characteristics of the target system to enhance the diversity and representativeness of the training set.

[0013] Optionally, the method of using the LASP software to call the first principles calculation program VASP to calculate the energy, force and stress of the structure and construct an initial training set specifically includes:

[0014] Set calculation parameters, including exchange-correlation functional, plane wave cutoff energy, and k-point sampling method;

[0015] Set the number of sampling steps and structure output level of the random surface walk global optimization method;

[0016] Set the convergence criteria and select the generated structure;

[0017] Perform high-precision first-principles calculations on selected structures to obtain energy, force, and stress information for all structures;

[0018] Power-finger structural descriptors are used to distinguish the chemical environments of atoms in training structures.

[0019] Optionally, the first principles calculation is combined with a random surface walk global optimization method to generate more structure files and expand the training set:

[0020] Repeat the above training process and randomly select the structure after each training;

[0021] Remove structures that are duplicated in the training set.

[0022] Optionally, the artificial neural network potential energy function suitable for the O2-Pt-MFI system is developed by LASP software, specifically including:

[0023] Prepare input files, including structure files, parameter input files, and artificial neural network parameter configuration files;

[0024] Set the weight ratio of energy, force and stress during training;

[0025] The energy and force of the structural unit were calculated using the initial artificial neural network potential function and compared with the DFT results;

[0026] If the error is large, new structural units are continuously added to the training set until the error is reduced to a reasonable range;

[0027] During the training process, when the standard mean square error of energy and force tends to be stable, multiple candidate artificial neural network potential energy functions are tested, and the artificial neural network potential energy function that is most consistent with the DFT optimization results in energy, force and structure is finally selected as the final function;

[0028] The fitted artificial neural network potential energy function is used to calculate some basic chemical structures, and the calculation results are compared with the DFT results. If the error is small, the accuracy of the function can be determined.

[0029] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0030] The present invention utilizes Material Studio software to construct a series of typical periodic and aperiodic chemical structure files; LASP software is used to calculate the energy, force, and stress of the structures to construct an initial training set; then, first-principles calculations are combined with random surface walk global optimization to generate more structure files and expand the training set; finally, LASP software is used to develop an artificial neural network potential energy function suitable for the O2-Pt-MFI system. Based on the first-principles DFT method and combining machine learning with artificial neural networks, the present invention develops an O2-Pt-MFI artificial neural network potential energy function (G-NN) for studying the kinetic behavior of metal clusters encapsulated in molecular sieves, demonstrating the feasibility of machine learning in the theoretical calculation of inorganic crystalline microporous materials such as zeolites. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following

[0032] A brief introduction is required using the accompanying drawings.

[0033] Figure 1It is a statistical table of the number of atoms in the structure file during the potential energy function fitting process in the embodiment.

[0034] Figure 2 ] is the root mean square error of the energy and force of the training set structure calculated by G-NN and DFT in the embodiment, where Figure (a) is the root mean square error of energy and Figure (b) is the root mean square error of force.

[0035] Figure 3 The adsorption energies of oxygen on three low-index crystal planes calculated by G-NN and DFT in the embodiment are shown in the second and third columns, respectively.

[0036] Figure 4 The dissociation energy barriers of oxygen on three low-index crystal planes calculated by G-NN and DFT in the embodiment are shown in the second and third columns, respectively. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] The purpose of this invention is to develop an artificial neural network potential energy function to describe the O2-Pt-MFI system.

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0040] The three low-index crystal planes of platinum are usually called (100), (110) and (111) crystal planes. These crystal planes have different atomic arrangements and surface properties in the face-centered cubic (FCC) structure of platinum. The activation of O2 on the metal surface is a key step in heterogeneous catalysis and material oxidation. It is a major challenge to use experimental characterization to explore its activation mechanism. There are even fewer studies that reveal the adsorption behavior, stability and migration mechanism of metal nanoclusters in zeolites under the influence of oxygen atmosphere. Therefore, it is very necessary to develop an artificial neural network potential energy function that can describe the interaction between platinum clusters and molecular sieves in an oxygen environment. The specific steps include the following:

[0041] (1) Use Material Studio software to build a series of typical periodic and non-periodic chemical structure files.

[0042] Specifically, according to the chemical elements contained in the system to be trained, Material Studio software is used to build some chemical structures with target system characteristics composed of a single element or two different elements, such as Pt n Nanoclusters, metal oxides (Pt x O y ), alloy (Pt x Si y ), non-metallic oxides (Si x O y ) and their combinations, including a periodic structure with a fixed unit cell size and a non-periodic structure with a non-fixed unit cell size.

[0043] (2) Use LASP software to call the first-principles calculation software VASP to calculate the energy, force and stress of each generated structure and construct an initial training set.

[0044] Specifically, the first-principles calculation software VASP was called using the script provided by the LASP software to optimize the chemical structure mentioned above using the random surface walk (SSW) method. The exchange functional used for optimization was the spin-polarized GGA-PBE, and the plane wave cutoff energy was set to 400 eV. The first Brillouin zone k-point sampling used a 1×1×1 gamma-centered grid.

[0045] During the structure optimization process, the sampling step number is set to 100, that is, 100 stable structures similar to the existing structure and other unstable structures in the search path are collected on the potential energy surface. The structure output is set to the highest level 6 to avoid unnecessary sampling. The convergence accuracy of the energy and force of the electron density and structure optimization is 10 -7 eV and More than 10 5 Structural units, 5000-10000 structures were randomly selected for high-precision single-point calculations. The energy, force and stress of the calculated structures constituted the initial training set, and power-tipped structural descriptors (PTSDs) were used to distinguish the different chemical environments of atoms in each training structural unit. Each element has 297 two-body, 126 three-body and 4 four-body interaction descriptors.

[0046] (3) Combine first-principles calculations with random surface walk (SSW) global optimization to generate more structure files and expand the training set.

[0047] Specifically, the training of the potential energy function requires continuous iteration, that is, the above process needs to be repeated continuously, and the optimized structure is randomly selected each time. The number of structures selected each time is controlled between 5000-10000, and the structures repeated in the training set are deleted at the same time, finally obtaining a DFT data set with a large number of structures.

[0048] (4) Based on the prepared DFT data set, LASP software was used to perform potential energy fitting on the DFT-calculated energy, force, and stress;

[0049] Specifically, all the structures in the DFT dataset mentioned above are fitted after high-precision single-point calculation. When fitting, input files need to be prepared, including structure files, parameter input files, and artificial neural network parameter configuration files. The fitting setting is energy: force: stress equal to 200:1000:10; by continuously adding new structures to the training set, further fitting is carried out to reduce the error with the DFT calculation results; finally, the DFT dataset we used includes 72,541 structural data, such as Figure 1 As shown. The superiority of the artificial neural network potential cannot be completely judged by the root mean square error of energy and force, but it can be used as a reference condition. Insufficient training steps will lead to poor fitting effect and large error with the DFT calculation results; and too many training steps, although the error standards of energy and force are reduced, will lead to overfitting, affecting the predictive ability of the potential energy function. The best way is to select the potential energy function under the corresponding training steps for testing when the standard mean square error of energy and force does not change much for a long time during training. When the potential energy training is repeated for 4800 steps, the potential energy function under 3800 / 4200 / 4400 / 4600 / 4800 training steps is selected for comparison with the results optimized by the DFT method. Under the premise that the fitting training steps exceed 4800, the calculation results are exported and fitted using Origin software to obtain the root mean square error of energy and force, as shown below. Figure 2 As shown in Figure (a), the root mean square error of energy is 8.27 meV / atom, and the root mean square error of force is 8.27 meV / atom. All are within an acceptable range. Finally, the fitted artificial neural network potential energy function (G-NN) is used to calculate the oxygen adsorption and dissociation energy on the low-index crystal plane of platinum. The kinetic behavior of platinum metal in the molecular sieve encapsulated platinum metal system in an oxygen atmosphere involves the adsorption and dissociation process of oxygen, that is, oxygen will be adsorbed on the surface of platinum metal and dissociated into oxygen atoms. The structures of oxygen adsorbed on three low-index crystal planes of platinum are selected as the basic chemical structures. The adsorption energy and oxygen dissociation energy barrier are calculated using the fitted artificial neural network potential energy function and compared with the results based on first-principles calculations. The three low-index crystal planes of platinum are usually called (100), (110) and (111) crystal planes. The interaction energy between the platinum surface and oxygen is usually expressed as the adsorption energy E adsThe more negative the adsorption energy, the stronger the interaction between the two. For the calculation of the oxygen dissociation energy barrier, that is, the transition state, the density functional theory method uses the CL-NEB method using the VASP software, while the G-NN potential method uses the two-end walking method, using the artificial neural network potential energy function to calculate the adsorption energy of oxygen molecules on different Pt crystal surfaces. The adsorption energy E ads The calculation formula is E ads =E(O2 / surface)-E(surface)-E(O2), where E(O2 / surface), E(surface) and E(O2) represent the energy of oxygen adsorption on the platinum surface, the platinum surface in vacuum and O2, respectively. The calculated results are compared with the DFT computer results, such as Figure 3 As shown in the figure, the second and third columns are the results of G-NN and DFT calculations respectively. The energy calculated by the artificial neural network potential energy function method is less than 0.2eV compared with the energy calculated by the DFT method. The calculated results of the oxygen dissociation energy barrier are as follows: Figure 4 As shown in the figure, the second and third columns are the results of the G-NN and DFT calculations, respectively. The error between the two is also no more than 0.2eV. The size of the error is within a reasonable range, which confirms the importance of the artificial neural network function.

[0050] In summary, this invention develops the O2-Pt-MFI artificial neural network potential energy function based on the first-principles DFT method and combines machine learning with artificial neural networks. This potential energy function can be used to study the kinetic behavior of metal clusters encapsulated in molecular sieves, revealing the adsorption behavior, stability and migration mechanism of metal nanoclusters in zeolites in an oxygen environment, etc., which is of great significance to the theoretical research on zeolite molecular sieve encapsulated metal catalysts.

Claims

1. A method for developing an artificial neural network potential energy function to simulate the kinetics of molecular sieve-encapsulated platinum clusters in an oxygen environment, characterized in that: The following steps are involved: Use Material Studio software to build a series of typical periodic and non-periodic chemical structure files; Using LASP software to call the first principles calculation program VASP, the energy, force and stress of the structure are calculated to construct an initial training set; First-principles calculations are combined with a random surface walk global optimization method to generate more structure files and expand the training set; An artificial neural network potential energy function suitable for the O2-Pt-MFI system was developed using LASP software.

2. The development method according to claim 1, characterized in that: The construction of the structure file specifically includes: selecting a plurality of different atomic structures according to the training system; According to the training requirements, construct a variety of representative chemical structures composed of a single element or two different elements; Construct typical structural units with the characteristics of the target system to enhance the diversity and representativeness of the training set.

3. The development method according to claim 1, characterized in that: LASP software was used to calculate the energy, force, and stress of the structure and construct an initial training set, including: Set calculation parameters, including exchange-correlation functional, plane wave cutoff energy, and k-point sampling method; Set the number of sampling steps and structure output level of the random surface walk global optimization method; Set the convergence criteria and select the generated structure; Perform high-precision first-principles calculations on selected structures to obtain energy, force, and stress information for all structures; Power-finger structural descriptors are used to distinguish the chemical environments of atoms in training structures.

4. The development method according to claim 1, characterized in that: By combining first-principles calculations with a global optimization method based on random surface walks, we generate more structure files and expand the training set. Specifically, we: Repeat the above training process and randomly select the structure after each training; Remove structures that are duplicated in the training set.

5. The development method according to claim 1, characterized in that: The artificial neural network potential energy function for the O2-Pt-MFI system was developed using LASP software, specifically including: Prepare input files, including structure files, parameter input files, and neural network parameter configuration files; Use power-finger structural descriptors to distinguish the chemical environments of atoms in training structures; Set the weight ratio of energy, force and stress during training; The energy and force of the structural unit are calculated using the initial neural network potential function and compared with the DFT results; If the error is large, new structural units are continuously added to the training set until the error is lower than the set threshold; During the training process, when the standard mean square error of energy and force tends to be stable, multiple candidate artificial neural network potential energy functions are tested, and the artificial neural network potential energy function that is most consistent with the DFT optimization results in energy, force and structure is finally selected as the final function; The fitted artificial neural network potential energy function is used to calculate some basic chemical structures, and the calculation results are compared with the DFT results. If the error is small, the accuracy of the function can be determined.

Images