Potential function model acquisition method and device for conductor / insulator interface simulation
Through a machine learning method combining the maximum localized Wannier function center and classical electrostatic algorithm, the problem of long-range interaction and dielectric response description in electrochemical interface simulation is solved, and the calculation efficiency and accuracy of conductor/insulator interface simulation is improved. .
Patent Information
- Application Number
- CN202410806242.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-21
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-06-21
AI Technical Summary
The prior art is difficult to accurately describe long-range interactions and dielectric responses in electrochemical interface simulations, especially in conductor/insulator interfaces, resulting in simulation results errors and computational efficiency.
Through a machine learning model based on the maximum localized Wannier function center, the charge spatial distribution of conductors and insulators is obtained, and combined with classic electrostatic algorithms and short-range machine learning potential functions, a machine learning potential function containing long-range information is constructed for description The overall charge distribution and stress information of the system are suitable for the given external electric field boundary conditions.
It realizes accurate description of long-range interactions and dielectric responses in conductor/insulator interface simulation, improves computational efficiency and accuracy, and can accelerate electrochemical interface simulation.
Smart Images

Figure CN118888015A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of computational chemistry and physics, and in particular to a method and device for acquiring a potential function model for conductor / insulator interface simulation. Background Art
[0002] Computational simulation can provide microscopic information of electrochemical interfaces and is an important research method for understanding the characteristics and mechanisms of electrochemical interface reactions. However, in order to achieve accurate simulation of electrochemical interfaces, it is necessary to simultaneously meet the conditions of sufficient sampling and electronic structure information. In principle, molecular simulation methods based on first principles, such as first principles molecular dynamics, can be used to simulate the system to meet the above two conditions. However, due to the limitation of computing power, molecular simulation methods based on first principles can only be applied to simple model systems of thousands of atoms to simulate processes of hundreds of picoseconds. In practice, electrochemical interfaces usually involve larger time and space scales, which poses a huge challenge to existing computational simulation methods. In recent years, the rise of artificial intelligence methods has become a breakthrough in solving this problem. It can improve the computational efficiency by a million times while ensuring the computational accuracy, making it possible to simulate millions of atoms at the nanosecond level and maintain the first principles accuracy. However, there are still two main difficulties in the application of existing artificial intelligence methods in the simulation of electrochemical interfaces: describing the long-range interactions in the system and the dielectric response of the system. Among the machine learning potential functions that are widely used now, descriptors based on local environment construction are mostly used to directly predict the potential energy surface of the system, which makes it impossible for long-range interactions to be described by machine learning potential functions. The lack of long-range interactions will cause significant errors in the simulation results when there are ions in the system, so it is necessary to include them in the simulation of electrochemical interfaces. The existing methods that include long-range interactions of the system are mainly divided into two categories: long-range and short-range separation of the effects in the system, using machine learning models to predict the charge distribution of the system, calculating long-range electrostatic interactions based on charge distribution based on classical electrostatic algorithms, and calculating short-range interactions based on short-range machine learning potential functions; or constructing descriptors containing long-range information to describe the potential energy surface of the system. There is no significant difference in accuracy and efficiency between these two methods. In addition, most of the machine learning potential functions that are widely used now do not include the dielectric response of the system and cannot be used for simulations under given external electric field boundary conditions. Even the few existing potential function models that include the dielectric response of the system cannot simultaneously describe the dielectric response of conductors and insulators, so they cannot be applied to metal / insulator interface simulations such as electrochemical interfaces. Summary of the invention
[0003] In view of the above technical problems, the technical solution adopted by the present invention is:
[0004] According to one aspect of the present invention, a method for obtaining a potential function model for conductor / insulator interface simulation is provided, the method comprising the following steps:
[0005] S100, based on the initial atomic structure of a given designated system, molecular simulation is performed on the designated system to obtain n atomic structures, and m atomic structures are extracted from the n atomic structures as an initial data set, wherein the designated system includes conductors and insulators in contact with each other, and n>m>1.
[0006] S200, using the first principles to calculate the total energy, total atomic force information and the maximum localized Wan Niel function center corresponding to each atomic structure corresponding to the current data set, and obtaining the maximum localized Wan Niel function center machine learning model based on the maximum localized Wan Niel function center; the initial value of the current data set is the initial data set; the total atomic force information includes the total atomic force corresponding to all atoms corresponding to the corresponding atomic structure.
[0007] S300, based on the maximum localized Wannier function center machine learning model, obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set, and apply the same external electric field as the external electric field applied in the first principles calculation process to each atomic structure corresponding to the current data set, and obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; obtain the overall charge space distribution corresponding to each atomic structure corresponding to the current data set.
[0008] S400, constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and using the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; the long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure.
[0009] S500, based on the short-range energy data set and the short-range atomic force data set, obtain N system short-range machine learning potential functions, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein each energy in the short-range energy data set is equal to the difference between the total energy corresponding to the corresponding atomic structure and the long-range energy, and each atomic force in the short-range atomic force data set is formed by the difference between the total atomic force information corresponding to the corresponding atomic structure and the atomic force in the long-range atomic force information, and N≥2.
[0010] S600, based on one system machine learning potential function among N system machine learning potential functions as a reference system machine learning potential function, molecular simulation is performed on the current k source atomic structures with a corresponding set external electric field applied to obtain k*n1 atomic structures as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures.
[0011] S700, based on N system machine learning potential functions, obtain k*n1 maximum atomic force deviations corresponding to the atomic structure for the current test. If there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, the corresponding atomic structure is used as a candidate data set, and m1 atomic structures are obtained from the candidate data set and added to the current data set, and k atomic structures are obtained from the m1 atomic structures added to the data set as the current k source atomic structures, and execute S200; otherwise, execute S800.
[0012] S800, taking the current data set as a target data set, and obtaining a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
[0013] According to a second aspect of the present invention, there is provided a potential function model acquisition device for conductor / insulator interface simulation, the device comprising:
[0014] The first molecular simulation module is used to perform molecular simulation on a designated system based on an initial atomic structure of the given designated system to obtain n atomic structures, and extract m atomic structures from the n atomic structures as an initial data set, wherein the designated system includes conductors and insulators in contact with each other, and n>m>1.
[0015] The first calculation module is used to calculate the total energy, total atomic force information and the maximum localized Wan Niehl function center corresponding to each atomic structure corresponding to the current data set using the first principles, and obtain the maximum localized Wan Niehl function center machine learning model based on the maximum localized Wan Niehl function center; the initial value of the current data set is the initial data set; the total atomic force information includes the total atomic force corresponding to all atoms corresponding to the corresponding atomic structure.
[0016] The charge space distribution acquisition module is used to obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set based on the maximum localized Wannier function center machine learning model, and to apply an external electric field identical to the external electric field applied in the first principles calculation process to each atomic structure corresponding to the current data set, and to obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; and to obtain the overall charge space distribution corresponding to each atomic structure corresponding to the current data set.
[0017] The second calculation module is used to construct a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and use the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; the long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure.
[0018] A machine learning potential function acquisition module is used to acquire N system short-range machine learning potential functions based on a short-range energy data set and a short-range atomic force data set, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein each energy in the short-range energy data set is equal to the difference between the total energy corresponding to the corresponding atomic structure and the long-range energy, and each atomic force in the short-range atomic force data set is formed by the difference between the total atomic force information corresponding to the corresponding atomic structure and the atomic force in the long-range atomic force information, and N≥2.
[0019] The second molecular simulation module is used to perform molecular simulation on the current k source atomic structures with a corresponding set external electric field based on one of the N system machine learning potential functions as a reference system machine learning potential function, and obtain k*n1 atomic structures as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures.
[0020] A judgment module is used to obtain k*n1 maximum atomic force deviations corresponding to the current test atomic structure based on N system machine learning potential functions. If there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, the corresponding atomic structure is used as a candidate data set, and m1 atomic structures are obtained from the candidate data set and added to the current data set, and k atomic structures are obtained from the m1 atomic structures added to the data set as the current k source atomic structures. Otherwise, the current data set is used as the target data set.
[0021] A target machine learning potential function acquisition module is used to acquire a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
[0022] The present invention has at least the following beneficial effects:
[0023] The technical solution provided by the present invention starts from a single structure, automatically explores the workflow of the system potential energy surface and the collection structure, trains machine learning potential functions that can be used for conductor / insulator interface simulation under given external electric field boundary conditions, and can accelerate electrochemical interface simulation.
[0024] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0026] Figure 1 A flow chart of obtaining a potential function model for conductor / insulator interface simulation provided by an embodiment of the present invention;
[0027] Figures 2a to 3 It is a schematic diagram of the experimental effect of the present invention. DETAILED DESCRIPTION
[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 those skilled in the art without creative work are within the scope of protection of the present invention.
[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.
[0030] It should be noted that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the steps as sequential processes, many of the steps can be implemented in parallel, concurrently or simultaneously. In addition, the order of the steps can be rearranged. The process can be terminated when its operation is completed, but it can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.
[0031] Embodiment 1:
[0032] This embodiment provides a method for obtaining a potential function model for conductor / insulator interface simulation. This method addresses the shortcoming that existing machine learning potential function methods cannot simultaneously describe the dielectric responses of conductors and insulators. The dielectric responses of conductors and insulators are processed separately in two subframes and then merged into the overall dielectric response of the system. Based on this hybrid description of the dielectric response, the charge distribution of the system under given external electric field boundary conditions can be obtained. Furthermore, the potential energy surface of the system is described by separating long-range and short-range interactions, and finally a machine learning potential energy surface model including the dielectric response of the conductor / insulator interface is obtained. Specifically, as Figure 1 As shown, the method may include the following steps:
[0033] S100, based on the initial atomic structure of a given designated system, molecular simulation is performed on the designated system to obtain n atomic structures, and m atomic structures are extracted from the n atomic structures as an initial data set, wherein the designated system includes a conductor and an insulator in contact with each other, that is, a conductor / insulator interface system, wherein n>m>1.
[0034] In the embodiment of the present invention, the designated system can be set based on actual needs and can be any system used for molecular simulation. In the embodiment of the present invention, the specific objects corresponding to the conductor and the insulator can be determined based on actual needs.
[0035] In an embodiment of the present invention, molecular simulation can be performed based on molecular dynamics simulation or Monte Carlo simulation. Specifically, a computational configuration file that can be used for molecular dynamics simulation or Monte Carlo simulation can be configured in the molecular dynamics simulation software based on the atomic structure of the specified system, and the simulation is performed to obtain the corresponding simulation trajectory. n can be set based on actual needs. In an illustrative embodiment, n can be tens of thousands, that is, tens of thousands of atomic movements are simulated. m can be set based on actual needs, and the order of magnitude can be hundreds. The m atomic structures can be user-specified structures or randomly extracted structures.
[0036] It is known to those skilled in the art that atomic structure may include atomic species and atomic positions.
[0037] S200, using the first principles to calculate the total energy, total atomic force information and the maximum localized Wannier function center corresponding to each atomic structure corresponding to the current data set, and obtaining the maximum localized Wannier function center machine learning model based on the maximum localized Wannier function center; the initial value of the current data set is the initial data set; the total atomic force information includes the total atomic force corresponding to all atoms corresponding to the corresponding atomic structure, that is, the total atomic force information of a certain atomic structure includes the atomic force of each atom in the atomic structure.
[0038] Those skilled in the art know that any method that uses first principles to calculate the total energy, total atomic force information and the maximum localized Wannier function center corresponding to each atomic structure corresponding to the current data set falls within the scope of protection of the present invention.
[0039] In an embodiment of the present invention, the maximum localized Van nier function center machine learning model can be obtained based on the DeepMD-kit open source software. Specifically, the current data set and the corresponding maximum localized Van nier function center can be used as a training set to train the initial maximum localized Van nier function center machine learning model to obtain the maximum localized Van nier function center machine learning model, that is, the current data set is used as the input of the model, and the maximum localized Van nier function center is used as the output of the model to train the model. The initial maximum localized Van nier function center machine learning model can be a deep learning neural network model. It is known to those skilled in the art that any method of using the current data set and the corresponding maximum localized Van nier function center as a training set to train the initial maximum localized Van nier function center machine learning model to obtain the maximum localized Van nier function center machine learning model belongs to the protection scope of the present invention.
[0040] S300, based on the maximum localized Wannier function center machine learning model, obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set, and apply the same external electric field as the external electric field applied in the first principles calculation process to each atomic structure corresponding to the current data set, and obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; obtain the overall charge space distribution corresponding to each atomic structure corresponding to the current data set.
[0041] Since the type of each atom of the insulator is known, the charge size of each atom can be known, and since the position of each atom is known, the charge space distribution of the insulator of each atomic structure can be known, that is, the charge space distribution includes the charge size and position. It is known to those skilled in the art that any method for obtaining the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set based on the maximum localized Wannier function center machine learning model belongs to the protection scope of the present invention.
[0042] In an embodiment of the present invention, the polarizable electrode model in the ELECTRODE module in the LAMMPS open source software can be used to obtain the charge space distribution corresponding to the conductor in the atomic structure to which an external electric field is applied. The specific acquisition method can be the existing technology.
[0043] S400, constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and using the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; the long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure.
[0044] In an embodiment of the present invention, the system long-range machine learning potential function is constructed based on a classical electrostatic interaction algorithm such as the Ewald summation algorithm. It is known to those skilled in the art that any method of constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and using a classical electrostatic interaction algorithm to obtain the corresponding overall charge space distribution to construct a system long-range machine learning potential function, all fall within the scope of protection of the present invention.
[0045] Those skilled in the art know that any method of using the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set falls within the scope of protection of the present invention.
[0046] S500, based on the short-range energy data set and the short-range atomic force data set, obtain N system short-range machine learning potential functions, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function, that is, add each system short-range machine learning potential function and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein, each energy in the short-range energy data set is equal to the difference between the total energy and the long-range energy corresponding to the corresponding atomic structure, and each atomic force in the short-range atomic force data set is formed by the difference between the total atomic force information corresponding to the corresponding atomic structure and the atomic force in the long-range atomic force information, that is, the atomic force difference of each atom in each atomic structure is equal to the difference obtained by subtracting the corresponding long-range atomic force from the corresponding total atomic force, that is, the atomic force difference △P corresponding to the rth atom in a certain atomic structure r =(PC r -PL r ), PC r is the total atomic force corresponding to the r-th atom in the atomic structure, and PL is the long-range atomic force corresponding to the r-th atom in the atomic structure.
[0047] In the embodiment of the present invention, N≥2. The specific value of N can be determined based on actual needs.
[0048] In an embodiment of the present invention, the system short-range machine learning potential function can be obtained based on the DeepMD-kit open source software. Specifically, the current data set and the corresponding energy and atomic force information can be used as a training set to train the initial system short-range machine learning potential function model to obtain the system short-range machine learning potential function, that is, the current data set is used as the input of the model, and the energy and atomic force information are used as the output of the model to train the model. It is known to those skilled in the art that any method that uses the current data set and the corresponding energy and atomic force information as a training set to train the initial system short-range machine learning potential function to obtain the system short-range machine learning potential function belongs to the protection scope of the present invention.
[0049] S600, based on one of the N system machine learning potential functions as the reference system machine learning potential function, performs molecular simulation on the current k source atomic structures with the corresponding set external electric field, explores the system potential energy surface, generates corresponding simulation trajectories, and obtains k*n1 atomic structures as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures.
[0050] In an embodiment of the present invention, the reference system machine learning potential function can be obtained randomly.
[0051] In the embodiment of the present invention, the value of k can be set based on actual needs, 1<k<m. n1 can be equal to n, or not equal to n.
[0052] In the embodiment of the present invention, the initial atomic structure may belong to the n atomic structures obtained in S100, or may not belong to them, and may be an atomic structure specified by a user.
[0053] In the embodiment of the present invention, the external electric field corresponding to the current k source atomic structures is greater than or equal to the external electric field corresponding to the previous k source atomic structures, that is, the external electric field is applied in an incremental manner. That is, the external electric field strength is extrapolated from small to large during the exploration process, wherein the external electric field condition can be achieved by controlling the external electric field or controlling the amount of electricity.
[0054] In an embodiment of the present invention, k*n1 atomic structures can be realized by modifying and integrating the ELECTRODE module in the open source software LAMMPS and the LAMMPS interface in the open source software DeepMD-kit. The LAMMPS interface in DeepMD-kit is used to perform molecular dynamics simulations, and the ELECTRODE module in LAMMPS applies the polarizable electrode method to molecular dynamics simulations. It is known to those skilled in the art that the modification and integration method of the ELECTRODE module in the open source software LAMMPS and the LAMMPS interface in the open source software DeepMD-kit may belong to the prior art.
[0055] S700, based on N system machine learning potential functions, obtain k*n1 maximum atomic force deviations corresponding to the atomic structure for the current test. If there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, the corresponding atomic structure is used as a candidate data set, and m1 atomic structures are obtained from the candidate data set and added to the current data set, and k atomic structures are obtained from the m1 atomic structures added to the data set as the current k source atomic structures, and execute S200; otherwise, that is, there is no maximum atomic force deviation within the preset error range, execute S800.
[0056] In an embodiment of the present invention, the preset error range may be [a, b], where a is a set lower limit value, b is a set upper limit value, b=f*a, f≥2. b may be an empirical value, and in an exemplary embodiment, b may be 0.2 millielectronvolts per angstrom. In an exemplary embodiment, 2≤f≤3.
[0057] Further, in the embodiment of the present invention, in S700, the maximum atomic force deviation corresponding to each atomic structure corresponding to the atomic structure currently tested satisfies the following conditions:
[0058] Pf=max(Pf1,Pf2,...,Pf i , ..., Pf h ), Pf i is the maximum atomic force deviation of the i-th atom in the corresponding atomic structure, i ranges from 1 to h, and h is the number of atoms in the corresponding atomic structure; P ij is the atomic force of the i-th atom obtained based on the machine learning potential function of the j-th system, is the mean value of the atomic force on the ith atom obtained based on the machine learning potential function of N systems, and max() means taking the maximum value.
[0059] In an embodiment of the present invention, if the maximum atomic force deviation is less than b, it means that the structure is a structure that can be accurately predicted by machine learning, and the training error of the machine learning potential function is usually taken. If it is greater than a, it means that the structure is a structure where the machine learning prediction error is too large, and there may be an unreasonable structure. If it falls between a and b, it is a structure with a moderate machine learning prediction error, and further learning is required.
[0060] In the embodiment of the present invention, m1 can be set based on actual needs, and m1 can be equal to m or not equal to m. It is known to those skilled in the art that if the number of atomic structures in the candidate data set is less than or equal to m1, all atomic structures in the candidate data set can be added to the current data set to update the current data set.
[0061] S800, taking the current data set as a target data set, and obtaining a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
[0062] Furthermore, in the embodiment of the present invention, S800 may specifically include:
[0063] S810, using the first principles to calculate the total energy, total atomic force information and the maximum localized Wan Niehl function center corresponding to each atomic structure of the target data set, and based on the maximum localized Wan Niehl function center, obtain the maximum localized Wan Niehl function center machine learning model.
[0064] S820, based on the maximum localized Wannier function center machine learning model in S810, obtain the charge spatial distribution corresponding to the insulator in each atomic structure corresponding to the target data set, and apply an external electric field identical to the external electric field applied in the first principles calculation process in S810 to each atomic structure corresponding to the target data set, and obtain the charge spatial distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; obtain the overall charge spatial distribution corresponding to each atomic structure corresponding to the target data set.
[0065] S830, constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the target data set, and using the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the target data set.
[0066] S840, based on the short-range energy data set and the short-range atomic force data set, obtain a system short-range machine learning potential function, and add the system short-range machine learning potential function and the system long-range machine learning potential function to obtain the target system machine learning potential function as a potential function model for conductor / insulator interface simulation.
[0067] The specific implementation of S810 to S840 may be substantially the same as the implementation of S200 to S500 described above, except that, in the process of training using the data set in S810 to S840, the number of steps set is greater than the number of steps set in S200 to S500.
[0068] The potential function model for conductor / insulator interface simulation provided by the embodiment of the present invention can be realized by controlling the electric field or charge of the simulation system according to actual research problems.
[0069] Embodiment 2:
[0070] This embodiment provides a potential function model acquisition device for conductor / insulator interface simulation, the device comprising:
[0071] The first molecular simulation module is used to perform molecular simulation on a designated system based on an initial atomic structure of the given designated system to obtain n atomic structures, and extract m atomic structures from the n atomic structures as an initial data set, wherein the designated system includes conductors and insulators in contact with each other, and n>m>1.
[0072] A first calculation module is used to calculate the total energy, total atomic force information and the maximum localized Wannier function center corresponding to each atomic structure corresponding to the current data set by using the first principles, and obtain the maximum localized Wannier function center machine learning model based on the maximum localized Wannier function center; the initial value of the current data set is the initial data set; the total atomic force information includes the total atomic force corresponding to all atoms corresponding to the corresponding atomic structure;
[0073] The charge space distribution acquisition module is used to obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set based on the maximum localized Wannier function center machine learning model, and to apply an external electric field identical to the external electric field applied in the first principles calculation process to each atomic structure corresponding to the current data set, and to obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; and to obtain the overall charge space distribution corresponding to each atomic structure corresponding to the current data set.
[0074] The second calculation module is used to construct a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and use the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; the long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure.
[0075] A machine learning potential function acquisition module is used to acquire N system short-range machine learning potential functions based on a short-range energy data set and a short-range atomic force data set, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein each energy in the short-range energy data set is equal to the difference between the total energy corresponding to the corresponding atomic structure and the long-range energy, and each atomic force in the short-range atomic force data set is equal to the difference between the total atomic force information corresponding to the corresponding atomic structure and the long-range atomic force information, and N≥2.
[0076] The second molecular simulation module is used to perform molecular simulation on the current k source atomic structures with a corresponding set external electric field based on one of the N system machine learning potential functions as a reference system machine learning potential function, and obtain k*n1 atomic structures as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures.
[0077] A judgment module is used to obtain k*n1 maximum atomic force deviations corresponding to the current test atomic structure based on N system machine learning potential functions. If there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, the corresponding atomic structure is used as a candidate data set, and m1 atomic structures are obtained from the candidate data set and added to the current data set, and k atomic structures are obtained from the m1 atomic structures added to the data set as the current k source atomic structures. Otherwise, the current data set is used as the target data set.
[0078] A target machine learning potential function acquisition module is used to acquire a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
[0079] The device can be used to perform Figure 1 Therefore, for the functions that can be realized by each functional module of the device, please refer to Figure 1 The description of the illustrated embodiment is not repeated in detail.
[0080] In a specific application of this embodiment, the conductor / insulator interface system can be a Pt(111) / KF solution interface system. The present invention takes the Pt(111) / KF solution interface system as an example, and uses the method provided by the present invention to obtain the corresponding machine learning potential energy model. The error of the obtained machine learning potential energy model can be as follows: Figure 2a to Figure 2b As shown. Figure 2a to Figure 2b As shown in Figure 2, the maximum localized Vanier function center root mean square error of the machine learning potential model on the training data set and the test data set is Order of magnitude, such as Figure 2a and Figure 2b As shown in Figure 2, the root mean square error of atomic forces in the machine learning potential energy model on the training and test datasets is Order of magnitude, such as Figure 2c and Figure 2d This shows that the accuracy of the machine learning potential energy model obtained by the method of the present invention is relatively high.
[0081] Figure 3The figure is a schematic diagram of the calculation results of the differential capacitance of the Pt(111) / KF solution interface. In this embodiment, the PBE functional and Grimme D3 correction are used to describe the potential energy surface. Using the machine learning potential energy function model obtained by the present invention, the charge-potential relationship of the Pt(111) / KF solution interface can be obtained by constant charge simulation (curve 1), and the interface differential capacitance curve (curve 2) can be obtained by taking the first derivative of the charge with respect to the potential. Figure 3 It can be seen that the differential capacitance calculation results are quantitatively consistent with the data obtained from the previous first-principles calculations and experiments, which shows that the accuracy of the machine learning potential energy model obtained by the method of the present invention is relatively high.
[0082] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.
[0083] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for obtaining a potential function model for conductor / insulator interface simulation, characterized in that: The method comprises the following steps: S100, based on the given initial atomic structure of the designated system, performing molecular simulation on the designated system to obtain n atomic structures, and extracting m atomic structures from the n atomic structures as an initial data set, wherein the designated system includes a conductor and an insulator in contact with each other, and n>m>1; S200, using first principles to calculate the total energy, total atomic force information and maximum localized Wannier function center corresponding to each atomic structure corresponding to the current data set, and obtaining a maximum localized Wannier function center machine learning model based on the maximum localized Wannier function center; the initial value of the current data set is the initial data set; The total atomic force information includes the total atomic forces corresponding to all atoms of the corresponding atomic structure; S300, based on the maximum localized Wannier function center machine learning model, obtaining the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set, applying an external electric field identical to the external electric field applied in the first principles calculation process to each atomic structure corresponding to the current data set, and obtaining the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; Get the overall charge space distribution corresponding to each atomic structure of the current data set; S400, constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and using the constructed system long-range machine learning potential function to obtain long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; The long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure; S500, based on the short-range energy data set and the short-range atomic force data set, obtain N system short-range machine learning potential functions, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein each energy in the short-range energy data set is equal to the difference between the total energy corresponding to the corresponding atomic structure and the long-range energy, and each atomic force in the short-range atomic force data set is formed by the difference between the total atomic force information corresponding to the corresponding atomic structure and the atomic force in the long-range atomic force information, and N≥2; S600, based on one of the N system machine learning potential functions as a reference system machine learning potential function, a molecular simulation is performed on the current k source atomic structures with a corresponding set external electric field applied, and k*n1 atomic structures are obtained as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures; S700, based on the N system machine learning potential functions, obtain k*n1 maximum atomic force deviations corresponding to the current test atomic structure, if there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, then use the corresponding atomic structure as a candidate data set, and obtain m1 atomic structures from the candidate data set and add them to the current data set, and obtain k atomic structures from the m1 atomic structures added to the data set as the current k source atomic structures, and execute S200; otherwise, execute S800; S800, taking the current data set as a target data set, and obtaining a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
2. The method according to claim 1, characterized in that In S700, the maximum atomic force deviation corresponding to each atomic structure corresponding to the atomic structure currently tested satisfies the following conditions: Pf=max(Pf1,Pf2,...,Pf i , ..., Pf h ), Pf i is the maximum atomic force deviation of the i-th atom in the corresponding atomic structure, i ranges from 1 to h, and h is the number of atoms in the corresponding atomic structure; P ij is the atomic force of the i-th atom obtained based on the machine learning potential function of the j-th system, is the mean value of the atomic force on the ith atom obtained based on the machine learning potential function of N systems, and max() means taking the maximum value.
3. The method according to claim 1, characterized in that The preset error range is [a, b], where a is a set lower limit value, b is a set upper limit value, b=f*a, f≥2.
4. The method according to claim 1, characterized in that: The external electric field corresponding to the current k source atomic structures is greater than or equal to the external electric field corresponding to the previous k source atomic structures.
5. The method according to claim 1, characterized in that S800 specifically includes: S810, using first principles to calculate the total energy, total atomic force information and the maximum localized Wannier function center corresponding to each atomic structure corresponding to the target data set, and obtaining a maximum localized Wannier function center machine learning model based on the maximum localized Wannier function center; S820, based on the maximum localized Wannier function center machine learning model in S810, obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the target data set, apply an external electric field identical to the external electric field applied in the first principles calculation process in S810 to each atomic structure corresponding to the target data set, and obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; obtain the overall charge space distribution corresponding to each atomic structure corresponding to the target data set; S830, constructing a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the target data set, and using the constructed system long-range machine learning potential function to obtain long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the target data set; S840, based on the short-range energy data set and the short-range atomic force data set, obtain a system short-range machine learning potential function, and add the system short-range machine learning potential function and the system long-range machine learning potential function to obtain the target system machine learning potential function as a potential function model for conductor / insulator interface simulation.
6. The method according to claim 1, characterized in that In S100 , molecular simulation is performed based on molecular dynamics simulation or Monte Carlo simulation.
7. The method according to claim 1, characterized in that The maximum localized Wannier function center machine learning model and the system short-range machine learning potential function are obtained based on the DeepMD-kit open source software.
8. The method according to claim 1, characterized in that The long-range machine learning potential function of the system is constructed based on the classical electrostatic interaction algorithm.
9. A potential function model acquisition device for conductor / insulator interface simulation, characterized in that: The device comprises: A first molecular simulation module is used to perform molecular simulation on a designated system based on an initial atomic structure of the designated system to obtain n atomic structures, and extract m atomic structures from the n atomic structures as an initial data set, wherein the designated system includes a conductor and an insulator in contact with each other, and n>m>1; A first calculation module is used to calculate the total energy, total atomic force information and the maximum localized Wannier function center corresponding to each atomic structure corresponding to the current data set by using the first principles, and obtain the maximum localized Wannier function center machine learning model based on the maximum localized Wannier function center; the initial value of the current data set is the initial data set; the total atomic force information includes the total atomic force corresponding to all atoms corresponding to the corresponding atomic structure; A charge space distribution acquisition module is used to obtain the charge space distribution corresponding to the insulator in each atomic structure corresponding to the current data set based on the maximum localized Wannier function center machine learning model, and to apply an external electric field identical to the external electric field applied in the first-principles calculation process to each atomic structure corresponding to the current data set, and to obtain the charge space distribution corresponding to the conductor in the atomic structure to which the external electric field is applied based on the polarizable electrode model; and to obtain the overall charge space distribution corresponding to each atomic structure corresponding to the current data set; The second calculation module is used to construct a system long-range machine learning potential function based on the overall charge space distribution corresponding to all atomic structures corresponding to the current data set, and use the constructed system long-range machine learning potential function to obtain the long-range energy and long-range atomic force information corresponding to each atomic structure corresponding to the current data set; the long-range atomic force information includes the long-range atomic forces corresponding to all atoms corresponding to the corresponding atomic structure; A machine learning potential function acquisition module, used to acquire N system short-range machine learning potential functions based on a short-range energy data set and a short-range atomic force data set, and add the N system short-range machine learning potential functions and the system long-range machine learning potential function to obtain N system machine learning potential functions; wherein each energy in the short-range energy data set is equal to the difference between the total energy corresponding to the corresponding atomic structure and the long-range energy, and each atomic force in the short-range atomic force data set is formed by the difference between the total atomic force information corresponding to the corresponding atomic structure and the atomic force in the long-range atomic force information, and N≥2; The second molecular simulation module is used to perform molecular simulation on the current k source atomic structures with a corresponding set external electric field based on one of the N system machine learning potential functions as a reference system machine learning potential function, and obtain k*n1 atomic structures as the current test atomic structures; n1>1; the initial values of the current k source atomic structures are the specified k initial atomic structures; A judgment module is used to obtain k*n1 maximum atomic force deviations corresponding to the current test atomic structure based on N system machine learning potential functions. If there is a maximum atomic force deviation within a preset error range among the k*n1 maximum atomic force deviations, the corresponding atomic structure is used as a candidate data set, and m1 atomic structures are obtained from the candidate data set and added to the current data set, and k atomic structures are obtained from the m1 atomic structures added to the data set as the current k source atomic structures. Otherwise, the current data set is used as the target data set; A target machine learning potential function acquisition module is used to acquire a target machine learning potential function based on the target data set as a potential function model for conductor / insulator interface simulation.
Citation Information
Patent Citations
Solid-liquid interface heat transfer model construction method based on molecular dynamics
CN114547994A
Numerical simulation method of short-channel negative-capacitance fence field effect transistor
CN117973298A
Construction method of ionic liquid machine learning potential function considering special hydrogen bonds
CN118098440A
Method, device, and program for simulating NANO substance in electric field
US20090299706A1
Molecular modeling with machine-learned universal potential functions
US20220238191A1
