A correction method and system for machine learning interatomic interaction potential

Abstract

The present invention discloses a method and system for correcting the machine learning interatomic interaction potential. The method proposed in the present invention uses molecular dynamics simulation to calculate the free energy of each phase of the material, and uses the equality of the two-phase free energy involved in the phase boundary in the experimentally measured phase diagram as the correction basis to correct the training data set required for the machine learning interatomic interaction potential. The machine learning interatomic interaction potential is then trained using the corrected training data set, thereby obtaining a more accurate and reliable machine learning interatomic interaction potential, providing more accurate and reliable data support and technical support for the simulation calculation of material properties.

Description

Technical Field

[0001] The present invention belongs to the technical field of material molecular dynamics simulation, and specifically relates to a method and system for correcting the interatomic interaction potential of machine learning. Background Art

[0002] During the preparation and service of materials, various transformations such as solid-liquid phase transitions and solid-solid phase transitions will occur under specific temperature and pressure conditions. Some phase transitions need to be avoided, such as the embrittlement of tin caused by phase transitions at low temperatures; other phase transitions can be utilized, such as refrigeration achieved by solid-liquid phase transitions and shape memory properties achieved by solid-solid phase transitions. Understanding the microscopic mechanism of material phase transitions requires the use of molecular dynamics simulations. This method uses the interatomic interaction potential and the initial relative spatial positions of atoms as input. Based on Newtonian mechanics, it uses numerical calculation techniques to solve the atomic dynamics equations to obtain information such as the position and velocity of atoms at different times. By taking statistical averages of the thermodynamic ensemble it describes, the free energy of each phase can be obtained, and information such as the material phase transition temperature and pressure can be calculated.

[0003] Molecular dynamics simulations using machine learning to calculate interatomic interaction potentials have been widely used in the research and analysis of materials, including phase diagram calculations, crystal structure searches, and phase-transition-induced shape memory materials. Traditional techniques use first-principles calculations based on density functional theory to establish a training set. This training set is then used to train parameters in interatomic interaction potentials in the form of neural networks, enabling molecular dynamics simulations to achieve the same accuracy as first-principles calculations. Existing research results indicate that the phase transition temperatures and pressures calculated using traditional first-principles molecular dynamics simulations based on density functional theory often deviate from experimentally measured results. In some cases, this deviation is quite large, and this deviation is primarily caused by inaccuracies in the interatomic interaction potentials.

[0004] It has also been proposed that, based on conventional techniques, machine-learning atomic interaction potentials for next-generation materials simulations could surpass the accuracy of first-principles calculations based on density functional theory, provided the training set is established using more precise quantum mechanical methods such as quantum chemistry. However, quantum mechanical calculations are significantly more expensive than first-principles calculations based on density functional theory and are currently only applicable to very small systems. Limited by current computing power, it is not yet possible to generate the training data required to achieve machine-learning atomic interaction potentials exceeding the accuracy of density functional theory in widespread practical applications. Summary of the Invention

[0005] In order to solve the problem of inaccurate interatomic interaction potential of existing machine learning, the present invention provides a method and system for correcting the interatomic interaction potential of machine learning. The present invention utilizes the property of molecular dynamics simulation that can calculate the free energy of each phase of the material, and corrects the machine learning interatomic interaction potential based on the fact that the phase boundary in the experimentally measured phase diagram implies that the free energy of the two phases is equal under this condition, thereby improving the accuracy and reliability under interatomic interaction, and further ensuring the accuracy and reliability of the molecular dynamics simulation calculation results.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for correcting a machine learning interatomic interaction potential, the correction method comprising:

[0008] A training set is obtained by first-principles calculations, and the initial machine learning interatomic interaction potential is obtained by training with the training set; each small system in the training set can be distinguished as belonging to each phase involved in the phase transition, and a small system can only belong to one phase;

[0009] Using the initial machine learning interatomic interaction potential, molecular dynamics simulation calculations are performed to obtain the function of the free energy correction value of each phase with respect to the material density;

[0010] Using the function of the free energy correction value of each phase with respect to the material density, each small system corresponding to the phase in the training set is corrected to establish a corrected training set;

[0011] The machine learning interatomic interaction potential is obtained by retraining using the revised training set;

[0012] Determine whether the retrained machine learning interatomic interaction potential meets the correction target. If so, complete the correction to obtain the corrected machine learning interatomic interaction potential. Otherwise, return the retrained machine learning interatomic interaction potential as the initial machine learning interatomic interaction potential to execute the molecular dynamics simulation calculation step.

[0013] The traditional method of using a training set obtained by first-principles calculations and using this training set to train the machine learning interatomic interaction potential is not accurate enough, and the method of using more accurate quantum mechanical calculation methods such as quantum chemistry to calculate the training data required for the machine learning interatomic interaction potential is too computationally expensive and has limited application. The method proposed in the present invention uses molecular dynamics simulation to calculate the free energy of each phase of the material, and uses the equality of the two-phase free energy involved in the phase boundary in the experimentally measured phase diagram as the correction basis to correct the training data set required for the machine learning interatomic interaction potential. The machine learning interatomic interaction potential is then trained using the corrected training data set, thereby obtaining a more accurate and reliable machine learning interatomic interaction potential, providing more accurate and reliable data support and technical support for the simulation calculation of material properties.

[0014] As a preferred embodiment, the method for determining the correction target of the present invention is: based on the experimentally measured phase diagram expressed with temperature and pressure as coordinates, several phase boundary points are randomly selected to determine the correction target that the free energy of the two phases involved at the selected phase boundary points should be equal.

[0015] As a preferred embodiment, the training set of the present invention consists of the atomic energy, atomic force and system stress of several small systems.

[0016] As a preferred embodiment, the molecular dynamics simulation calculation process of the present invention specifically includes:

[0017] Substitute the initial machine learning interatomic interaction potential into the molecular dynamics simulation to calculate the two phase free energies involved in several selected phase boundary points;

[0018] Calculate the difference in free energy between two phases;

[0019] Based on the difference between the free energies of the two phases, a correction value of the free energy of each phase is determined;

[0020] According to the material density of each phase at the phase boundary point, the corrected value of the free energy of each phase is calculated as a function of the material density.

[0021] As a preferred embodiment, the present invention determines the corrected value of each phase free energy based on the difference between the two phase free energies, specifically:

[0022] By determining the corrected free energy value of the first phase at the phase boundary point;

[0023] The free energy correction value of the first phase at the phase boundary point is summed with the difference between the free energies of the two phases to obtain the free energy correction value of the second phase at the phase boundary point.

[0024] As a preferred embodiment, the free energy correction value of the present invention as a function of material density is a constant function or a polynomial function of material density; the function parameters are determined by fitting the free energy correction values of the selected phase boundary points.

[0025] As a preferred embodiment, the fitting method of the present invention adopts the least squares method.

[0026] As a preferred embodiment, the correction process of the present invention specifically includes:

[0027] According to the function of the free energy correction value of each phase to the material density, the phase to which each small system belongs and the material density corresponding to each small system, the energy and system stress correction values of each small system are obtained respectively;

[0028] Add the energy and system stress correction values of each small system to the original energy and system stress in the training set to obtain the corrected energy and system stress of each small system;

[0029] The corrected energy, uncorrected atomic forces, and corrected system stresses constitute the corrected training set.

[0030] As a preferred embodiment, the present invention determines whether the retrained machine learning atomic interaction potential meets the correction target, specifically:

[0031] The retrained machine learning interatomic interaction potential is substituted into the molecular dynamics simulation calculation to obtain the free energy of the two phases involved at the selected phase boundary point, and the difference in the free energy of the two phases is calculated. If the difference is less than the preset value, the correction target is met, otherwise it does not meet the correction target.

[0032] On the other hand, the present invention also proposes a correction system for machine learning interatomic interaction potential, the correction system comprising:

[0033] An initialization unit, wherein the initialization unit uses first-principles calculations to obtain a training set, and uses the training set to train and obtain an initial machine learning interatomic interaction potential; each small system in the training set can be distinguished as belonging to each phase involved in the phase transition, and a small system can only belong to one phase;

[0034] a dynamics simulation calculation unit, wherein the dynamics simulation calculation unit performs molecular dynamics simulation calculations using the initial machine learning interatomic interaction potential to obtain a function of the free energy correction value of each phase with respect to the material density;

[0035] A correction unit, wherein the correction unit uses a function of the material density of the free energy correction value of each phase to correct each small system of the corresponding phase in the training set, establishes a corrected training set, and uses the corrected training set to retrain to obtain a machine learning interatomic interaction potential;

[0036] and a verification unit, which determines whether the retrained machine learning interatomic interaction potential meets the correction target. If so, the correction is completed to obtain the corrected machine learning interatomic interaction potential. Otherwise, the retrained machine learning interatomic interaction potential is input into the dynamic simulation calculation unit as the initial machine learning interatomic interaction potential.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] The method and system proposed in the present invention utilize the free energies of each phase calculated by molecular dynamics simulation, and use the experimentally determined phase diagram, under which the phase boundary implies the equality of the free energies of the two phases, as a correction basis to revise the training set required for machine learning interatomic interaction potentials. The machine learning interatomic interaction potentials are then retrained using the revised training set, thereby obtaining a more accurate and reliable machine learning interatomic interaction potential.

[0039] At the same time, the present invention further ensures the accuracy and reliability of the obtained machine learning interatomic interaction potential by testing and iteratively training the retrained machine learning interatomic interaction potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0041] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention.

[0042] Figure 2 2 is a system principle block diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0044] Example 1:

[0045] The existing method of using first principles to calculate the training data required for machine learning interatomic interaction potential is not accurate enough; and the method of using quantum chemistry and other quantum mechanical calculation methods to calculate the training data required for machine learning interatomic interaction potential is too computationally expensive and has limited application. Based on this, this embodiment proposes a correction method for machine learning interatomic interaction potential. The method proposed in this embodiment uses the property of molecular dynamics simulation that can calculate the free energy of each phase of the material, and uses the phase boundary in the experimentally measured phase diagram implying the equality of the free energy of the two phases under this condition as the correction basis to correct the training data required for machine learning interatomic interaction potential. The machine learning interatomic interaction potential is trained using the corrected training data, thereby obtaining a more accurate and reliable machine learning interatomic interaction potential, so as to more accurately simulate material properties.

[0046] First-principles calculations can determine the energy E, atomic forces f, and system stress S of small systems composed of a few to thousands of atoms. Formally, these can be expressed as functions of atomic position x. In practice, the energy, atomic forces, and system stress of each of these systems can be calculated from first principles for tens to millions of small systems with different atomic positions. Since atomic positions can be expressed as functions x(i) of the small system number i, the energy, atomic forces, and system stress can also be expressed as functions E(i), f(i), and S(i), respectively. The energy, atomic forces, and system stress of these small systems can be used as the training set {E(i), f(i), S(i)} for the machine-learning interatomic interaction potential. This training can be used to determine the parameters of the neural network and other parameters contained therein. The trained interatomic interaction potential with determined parameters is the machine-learning interatomic interaction potential. Molecular dynamics simulations using this machine-learning interatomic interaction potential can calculate the free energy F of each phase at arbitrary temperature T and pressure P. Different phases have different free energies, so free energy can be expressed as a function of the phase number I, temperature, and pressure, F(I,P,T). In a phase diagram with temperature and pressure as coordinate points, the free energies of the two phases should be equal at any boundary point (P,T) where the two phases meet. When the free energy functions calculated by molecular dynamics simulations at experimentally determined phase boundaries are significantly different, the machine learning interatomic interaction potentials need to be corrected so that the free energies of the two phases at the experimentally determined phase transition points, calculated based on molecular dynamics simulations, are equal, thereby more accurately simulating material properties.

[0047] Specific as Figure 1 As shown, the correction method proposed in this embodiment includes the following steps:

[0048] Step 1: Use first-principles calculations to obtain a training set, and use this training set to train and obtain the initial machine learning interatomic interaction potential; and each small system in the training set can be distinguished as belonging to each phase involved in the phase transition, and a small system belonging to a certain phase cannot belong to other phases at the same time.

[0049] Step 2: Use the initial machine learning interatomic interaction potential to perform molecular dynamics simulation calculations to obtain the corrected free energy values of each phase as a function of material density.

[0050] Step 3: Use the function of the free energy correction value of each phase with respect to the material density to correct each small system of the corresponding phase in the training set, and establish a corrected training set.

[0051] Step 4: Use the corrected training set to retrain and obtain the machine learning interatomic interaction potential.

[0052] Step 5: Test the retrained machine learning interatomic interaction potential. If the retrained machine learning interatomic interaction potential meets the correction target, the correction is completed to obtain the corrected machine learning interatomic interaction potential. Otherwise, the retrained machine learning interatomic interaction potential is used as the initial machine learning interatomic interaction potential and returned to step 2.

[0053] Specifically, the correction target determination method is as follows: based on the experimentally measured phase diagram expressed with temperature and pressure as coordinates, one or more phase boundary points (P, T) are randomly selected to determine the correction target. At the selected phase boundary point, the free energy of the two phases involved should be equal, that is, F(I, PT) = F(I', PT), where I and I' are the numbers of the two phases involved, respectively.

[0054] In an optional implementation method, step 1 is specifically as follows: obtaining the atomic energy, atomic force and system stress of tens to millions of small systems through first-principles calculations as a training set {E(i), f(i), S(i)}, wherein E(i) represents the atomic energy of the i-th small system, f(i) represents the atomic force of the i-th small system, and S(i) represents the system stress of the i-th small system; in the training set, each small system can be clearly divided into whether it belongs to each phase involved in the phase change, and a small system belonging to a certain phase cannot belong to other phases at the same time.

[0055] In an optional implementation manner, the specific steps of step 2 are as follows:

[0056] Step 21, the initial machine learning interatomic interaction potential is substituted into molecular dynamics simulation software (e.g., LAMMPS), and the two-phase free energies F(I, P, T) and F(I', P, T) of one or more selected phase boundary points are obtained by calculation;

[0057] Step 22, calculate the difference in free energy between the two phases, i.e., F(I,P,T)-F(I',P,T);

[0058] Step 23, determining a correction value of the free energy of each phase based on the difference between the free energies of the two phases;

[0059] For the case where the correction target contains only one phase boundary point, since the observable physical quantity depends only on the rate of change and difference of free energy and not on the absolute magnitude of free energy, the free energy correction value of one of the phases can be simply selected to be zero. For the case where the correction target contains multiple phase boundary points, it is necessary to fix the free energy correction value of one phase at each phase boundary point to zero or determine its correction value by other methods. Specifically, assuming that the free energy correction value of the first phase at a phase boundary point (P, T) is given as dF(I, P, T), then the free energy correction value of the second phase at that phase boundary point is the sum of the calculated free energy difference and the free energy correction value of the first phase, that is, dF(I', P, T) = dF(I, P, T) + F(I, P, T) - F(I', P, T).

[0060] In step 24, based on the material density of each phase at the phase boundary point, a correction value for the free energy of each phase is calculated as a function of the material density dF(I,ρ). It should be noted that this function can be a constant function, a polynomial function, or another function of the material density ρ. The function parameters are determined by fitting the correction values for the free energies of the selected phase boundary points. The fitting method can be, but is not limited to, the least squares method.

[0061] In an optional implementation manner, step 3 specifically includes:

[0062] Step 31: Calculate the energy and stress corrections for each small system based on the correction value as a function of material density, the phase to which each small system belongs, and the material density corresponding to each small system. Specifically, for each small system in the training set that is a phase involved in a phase transition, the energy and stress corrections are performed using the free energy correction value as a function of material density. Specifically, assuming that the free energy correction function is dF(I, ρ) for the phase and material density ρ corresponding to small system i, the energy correction dE(i) for that small system is calculated as follows: dE(i) = dF(I, ρ). The stress correction dS(i) is calculated as the derivative of the free energy function with respect to density: dS(i) = dF'(I, ρ). Here, dF' is the derivative of dF with respect to its independent variable. Since free energy is an extensive quantity, its specific value depends on the size of the system. Therefore, when applied to each small system, the specific value of the correction should be multiplied by a parameter related to the system size.

[0063] Step 32: Add the corrected values of the small system energies and system stresses to the original energies and system stresses in the training set to obtain the corrected energy E'(i) = E(i) + dE(i) and system stress S'(i) = S(i) + dS(i) for each small system. Since atomic forces do not require correction, the corrected energies, uncorrected atomic forces, and corrected system stresses constitute the corrected training set {E'(i), F(i), S'(i)}.

[0064] In an optional implementation, step 4 uses the corrected training set to retrain to obtain the machine learning interaction potential. The training method is the same as the training method in step 1, and both can adopt the existing machine learning interaction potential training method, which will not be repeated here.

[0065] In an optional implementation, the verification process in step 5 is specifically as follows: substituting the retrained machine learning interatomic interaction potential into the molecular dynamics simulation calculation to obtain the free energy of the two phases involved at the selected phase boundary point, and calculating the difference in the free energy of the two phases. If the difference is less than the preset value, the correction target is met, otherwise the iterative correction continues.

[0066] This embodiment also proposes a correction system for machine learning interatomic interaction potential, specifically as follows: Figure 2 As shown, the correction system includes: an initialization unit, a dynamic simulation calculation unit, a correction unit and a verification unit.

[0067] Among them, the initialization unit uses first-principles calculations to obtain a training set, and uses this training set to train to obtain the initial machine learning interatomic interaction potential; and each small system in the training set can be distinguished as whether it belongs to each phase involved in the phase change, and a small system belonging to a certain phase cannot belong to other phases at the same time.

[0068] The dynamics simulation calculation unit uses the initial machine learning interatomic interaction potential to perform molecular dynamics simulation calculations to obtain the function of the free energy correction value of each phase with respect to the material density.

[0069] The correction unit uses the function of the material density of the free energy correction value of each phase to correct each system of the corresponding phase in the training set, establishes a corrected training set, and uses the corrected training set to retrain to obtain the machine learning interatomic interaction potential.

[0070] The verification unit verifies the retrained machine learning interatomic interaction potential. If the retrained machine learning interatomic interaction potential meets the correction target, the correction is completed to obtain the corrected machine learning interatomic interaction potential; otherwise, the retrained machine learning interatomic interaction potential is used as the initial machine learning interatomic interaction potential and input into the dynamics simulation calculation unit.

[0071] Example 2:

[0072] The method and system proposed in the above embodiment are used to correct the machine learning atomic interaction potential of the phase transition involving two phases, namely phase one and phase two. The specific process is as follows:

[0073] The first-principles calculations are used to obtain the energy, atomic forces, and system stresses of several small systems as a training set. The training set is used to train the initial machine learning interatomic interaction potential. Each small system in the training set belongs to phase one and phase two respectively, and the same small system can only belong to one of them. At the same time, based on the experimentally measured phase diagram expressed in temperature and pressure as the coordinate system, multiple phase boundary points (P, T) about phase one and phase two are randomly selected to determine the correction target: at the selected phase boundary point, the free energy of the two phases involved should be equal, that is, (F(1,P,T)=F(2,P,T)).

[0074] Molecular dynamics simulations were performed using the initial machine learning interatomic interaction potential to obtain the free energy correction values of phase one and phase two as functions of material density F(1, P, T) and F(2, P, T).

[0075] For each small system belonging to phase one in the original training set, assuming that its corresponding material density is ρ, then the correction value dE of its energy is taken as dE=dF(1,ρ), and the correction value dS of the system stress is taken as dS=dF'(1,ρ), where dF' is the derivative of dF with respect to density ρ; the same is true for each small system belonging to phase two. Since free energy is an extensive quantity, its specific value is related to the size of the system. For each small system, the specific value of the correction should be multiplied by the parameter related to the size of the system.

[0076] The correction amount of the small system energy and system stress is added to the original energy and system stress to obtain the corrected energy and system stress of each small system, thereby generating a corrected training set.

[0077] The machine learning interatomic interaction potential is obtained by retraining using the corrected training set.

[0078] Finally, it is checked whether the correction effect reaches the correction target. If so, the correction is completed. Otherwise, the retrained machine learning interatomic interaction potential is used as the initial machine learning interatomic interaction potential for iterative processing until the correction target is reached.

[0079] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0080] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0081] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0082] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0083] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for correcting the potential of atomic interactions by machine learning, characterized in that: The correction method includes: A training set is obtained by first-principles calculations, and the initial machine learning interatomic interaction potential is obtained by training with the training set; each small system in the training set can be distinguished as belonging to each phase involved in the phase transition, and a small system can only belong to one phase; Using the initial machine learning interatomic interaction potential, molecular dynamics simulation calculations are performed to obtain a function of the corrected free energy of each phase with respect to the material density; Using the function of the free energy correction value of each phase with respect to the material density, each small system corresponding to the phase in the training set is corrected to establish a corrected training set; The machine learning interatomic interaction potential is obtained by retraining using the revised training set; Determine whether the retrained machine learning interatomic interaction potential meets the correction target. If so, complete the correction to obtain the corrected machine learning interatomic interaction potential. Otherwise, return the retrained machine learning interatomic interaction potential as the initial machine learning interatomic interaction potential to execute the molecular dynamics simulation calculation step. The molecular dynamics simulation calculation process specifically includes: Substitute the initial machine learning interatomic interaction potential into the molecular dynamics simulation software to calculate the two phase free energies involved in several selected phase boundary points; Calculate the difference in free energy between two phases; Based on the difference between the free energies of the two phases, a correction value of the free energy of each phase is determined; According to the material density of each phase at the phase boundary point, the corrected value of the free energy of each phase is calculated as a function of the material density.

2. The method for correcting the machine learning interatomic interaction potential according to claim 1, characterized in that: The correction target is determined by randomly selecting a number of phase boundary points based on an experimentally measured phase diagram expressed with temperature and pressure as coordinates, and determining that the correction target is that the free energies of the two phases involved at the selected phase boundary points should be equal.

3. The method for correcting the interatomic interaction potential of machine learning according to claim 1, characterized in that: The training set consists of atomic energies, atomic forces and system stresses of several small systems.

4. The method for correcting the interatomic interaction potential of machine learning according to claim 1, characterized in that: Based on the difference between the free energies of the two phases, the correction value of the free energy of each phase is determined, specifically: By determining the corrected free energy value of the first phase at the phase boundary point; The free energy correction value of the first phase at the phase boundary point is summed with the difference between the free energies of the two phases to obtain the free energy correction value of the second phase at the phase boundary point.

5. The method for correcting the machine learning interatomic interaction potential according to claim 1, characterized in that: The free energy correction value as a function of the material density is a constant function or a polynomial function of the material density; the function parameters are determined by fitting the free energy correction values of the selected phase boundary points.

6. The method for correcting the machine learning interatomic interaction potential according to claim 5, characterized in that: The fitting method adopts the least square method.

7. The method for correcting the interatomic interaction potential of machine learning according to claim 1, characterized in that: The revision process specifically includes: According to the function of the free energy correction value of each phase to the material density, the phase to which each small system belongs and the material density corresponding to each small system, the energy and system stress correction values of each small system are obtained respectively; Add the energy and system stress correction values of each small system to the original energy and system stress in the training set to obtain the corrected energy and system stress of each small system; The corrected energy, uncorrected atomic forces, and corrected system stresses constitute the corrected training set.

8. The method for correcting the machine learning interatomic interaction potential according to claim 1, characterized in that: Determine whether the retrained machine learning atomic interaction potential meets the correction target, specifically: The retrained machine learning interatomic interaction potential is substituted into the molecular dynamics simulation calculation to obtain the free energy of the two phases involved at the selected phase boundary point, and the difference in the free energy of the two phases is calculated. If the difference is less than the preset value, the correction target is met, otherwise it does not meet the correction target.

9. A correction system for machine learning interatomic interaction potential, characterized in that: The correction system comprises: An initialization unit, wherein the initialization unit uses first-principles calculations to obtain a training set, and uses the training set to train and obtain an initial machine learning interatomic interaction potential; each small system in the training set can be distinguished as belonging to each phase involved in the phase transition, and a small system can only belong to one phase; A dynamics simulation calculation unit is used to perform molecular dynamics simulation calculation using the initial machine learning interatomic interaction potential to obtain a function of the free energy correction value of each phase with respect to the material density. The molecular dynamics simulation calculation process specifically includes: Substitute the initial machine learning interatomic interaction potential into the molecular dynamics simulation software to calculate the two phase free energies involved in several selected phase boundary points; Calculate the difference in free energy between two phases; Based on the difference between the free energies of the two phases, a correction value of the free energy of each phase is determined; Based on the material density of each phase at the phase boundary point, the corrected value of the free energy of each phase is calculated as a function of the material density; A correction unit, wherein the correction unit uses a function of the material density of the free energy correction value of each phase to correct each small system of the corresponding phase in the training set, establishes a corrected training set, and uses the corrected training set to retrain to obtain a machine learning interatomic interaction potential; and a verification unit, which determines whether the retrained machine learning interatomic interaction potential meets the correction target. If so, the correction is completed to obtain the corrected machine learning interatomic interaction potential. Otherwise, the retrained machine learning interatomic interaction potential is input into the dynamic simulation calculation unit as the initial machine learning interatomic interaction potential.

Images