Recommended methods, equipment, and terminals for predicting the thermodynamically stable structure of disordered materials

Through machine learning clustering and density functional theory optimization, the problem of huge number of non-redundant structures in disordered structural materials is solved, and fast and reliable thermodynamic stable structure prediction is achieved, computing resources are saved, and energy similar Structural set, supporting experimental characterization and R&D of disordered materials.

CN113963758BActive Publication Date: 2024-11-05INST OF COAL CHEM CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111357893.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-16
Publication Date
2024-11-05
Estimated Expiration
2041-11-16

AI Technical Summary

Technical Problem

In the prior art, when dealing with disordered structural materials, the number of non-redundant structures is huge as supercell size increases or in special substitution/defect/co-occupying situations, resulting in high computing resources and time costs, and relying solely on thermodynamics. The experimental spectrum information of the structure explanation is not comprehensive, and multiple structures with similar energy are required to be statistically averaged.

Method used

The machine learning clustering algorithm is used to reduce the number of structures, and the structure is recommended through the principles of energy priority and diversity, combined with density functional theory optimization and machine learning potential model training, quickly find a thermodynamic stable structure set and save computing resources.

Benefits of technology

It significantly reduces computational costs and time, provides a reliable thermodynamic stable structure and structure set with similar energy, helping to experimental characterization and R&D process of disordered materials.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113963758B_ABST
    Figure CN113963758B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of disordered structural materials, and discloses a method, device, and terminal for predicting and recommending thermodynamically stable structures of disordered materials. The method uses an existing program to generate all possible non-redundant structures under a specified substitution / defect / co-occupancy state; clusters the structures using a clustering algorithm in machine learning; recommends a structure to be optimized using density functional theory from each clustering result; optimizes the recommended structure using a density functional theory method; prepares a training set from the trajectory of the optimized structure; trains a machine learning potential model; optimizes the structure without relaxation using the trained machine learning potential model; uses a multi-generation operation method to determine whether the set termination condition is met; after the termination condition is met, a thermodynamically stable structure set is recommended based on the structure set optimized by density functional theory and the structure set converged after machine learning potential optimization. The present invention is suitable for predicting and recommending thermodynamically stable structures of disordered materials and crystal structure prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of disordered structural materials, and in particular relates to a method, device and terminal for predicting and recommending the thermodynamically stable structure of disordered materials. Background Art

[0002] At present, when two (or more) atoms or ions occupy a certain position in a crystal structure, if their distribution relative to each other is random, that is, the probability of them occupying any such position is the same, then this structure is called a disordered structure. There are many types of disordered structures (including substitution / defect / co-occupancy, etc.). Disordered structural materials are widely used in semiconductors, high-temperature superconductors, metal alloys, ceramics, and zeolite catalysts due to their unique properties. Studying the structure of disordered materials is of great value in understanding the properties of disordered materials.

[0003] In experiments, a variety of methods can be used to characterize the structure of materials. Diffraction techniques can provide average long-range structural information, and spectra such as solid-state nuclear magnetic resonance, Raman, infrared or X-ray absorption near-edge structure can provide local structural information. However, the local structural information provided by such spectra is usually difficult to explain clearly by experiments alone. This makes molecular modeling essential for a deep understanding of such systems. Supercell approximation is one of the most commonly used methods for dealing with disordered structures. By constructing a large periodic unit cell, the local structural characteristics of the disordered system are reflected as much as possible within its boundaries: composition, coordination, etc. At present, many programs (SOD, supercell, enumlib, disorder, etc.) can realize the creation of supercell structure models under different disordered states (substitution / defect / co-occupancy), thereby effectively reducing the supercell structures that need to be considered. However, with the increase of supercell size or in special substitution / defect / co-occupancy cases, the number of non-redundant structures obtained by these programs is still very large. How to find thermodynamically stable structures from these non-redundant structures is a very important issue. The most common practices are as follows: First, a random method is used to randomly select several structures and use density functional theory optimization, and the optimized structures are used as candidate structures; second, a relatively cheap method (such as empirical potential function) is used to calculate the energy of all structures, sort them from low to high according to energy, select the first few structures and use density functional theory optimization, and the optimized structures are used as candidate structures; third, an exhaustive method is used to optimize all non-redundant structures using density functional theory, and then select the structure with the lowest energy as the candidate structure. The common practices either provide unreliable candidates (methods 1 and 2) or require high computing resources and time costs (method 3). In the process of interpreting experimental spectral information, sometimes it is not comprehensive to rely solely on the most thermodynamically stable structure found. Often, many structures with energies close to the energy of the most thermodynamically stable structure are required to be statistically averaged to obtain results consistent with the experimental spectral information. Therefore, it is very necessary to not only find the most thermodynamically stable structure from the non-redundant structures, but also find many structures whose energies are close to the energy of the most stable structure.

[0004] Through the above analysis, the problems and defects of the prior art are as follows:

[0005] (1) As the supercell size increases or in special substitution / defect / co-occupancy cases, the number of non-redundant structures obtained by programs (SOD, supercell, enumlib, disorder, etc.) remains large.

[0006] (2) The candidates provided by the existing random methods and relatively cheap methods are unreliable; the existing exhaustive methods require high computing resources and time costs.

[0007] (3) In the process of interpreting experimental spectral information, sometimes it is not comprehensive to rely solely on the most thermodynamically stable structure found. Often, it is necessary to statistically average many structures with energies close to the energy of the most thermodynamically stable structure to obtain results consistent with the experimental spectral information.

[0008] The difficulty in solving the above problems and defects is that it takes a very long time and a lot of computing resources to optimize a large number of non-redundant structures using density functional theory. For example, there is O / F co-occupancy in BaScO2F (O / F: 0.667 / 0.333). In order to obtain the most favorable distribution configuration of O and F in BaScO2F, when the supercell size is 2×2×2, the number of non-redundant structures is 2664. It takes an average of about 0.5h to use density functional theory to optimize a structure with 40 atoms in a single cell on a computer with 72 cores. It takes about 55 days to optimize 2664 structures. The commercial supercomputer charges 0.10 yuan / core / hour, which requires 95,904 yuan; when the supercell size is 2×2×3, the number of non-redundant structures increases to 6,849,807, which requires greater time cost and computing resources.

[0009] The significance of solving the above problems and defects is: it greatly saves time and money costs, quickly and reliably provides predictions and recommendations of thermodynamically stable structures of disordered materials from a large number of non-redundant structures, helps the process of experimental characterization of disordered material structures, and accelerates the process of disordered material research and development. Specifically, the time cost of 55 days and the money cost of 95,904 yuan required to use density functional theory to optimize 2664 structures in the above example of BaScO2F 2×2×2 supercell can be reduced to only 7 days and the money cost of 209 yuan required to use density functional theory to optimize 29 structures. Summary of the invention

[0010] In view of the problems existing in the prior art, the present invention provides a method, device and terminal for predicting and recommending the thermodynamically stable structure of disordered materials.

[0011] The present invention is implemented as follows: a method for predicting and recommending thermodynamically stable structures of disordered materials, the method for predicting and recommending thermodynamically stable structures of disordered materials comprising the following steps:

[0012] Step 1: Use existing programs (such as the supercell program) to generate all possible non-redundant structures under specified substitution / defect / co-occupancy states; use symmetry to remove a large number of symmetry-equivalent structures, greatly reducing the number of configurations that need to be considered, and quickly obtain all non-redundant structures under specified substitution / defect / co-occupancy states, providing a total structure set for subsequent steps.

[0013] Step 2: Use clustering algorithms in machine learning to cluster the structures. Clustering can effectively achieve differential sampling, which can not only ensure the diversity of recommendation results, but also make the training set cover a richer sampling space and improve the model prediction ability.

[0014] Step three, recommend the structures to be optimized using density functional theory from each clustering result, and the recommendation principles follow energy priority and diversity priority; the recommendation principle that takes both energy and diversity into consideration greatly improves the predicted recommendation probability of the recommended candidate structure becoming a thermodynamically stable structure of disordered materials.

[0015] Step 4: Use density functional theory to optimize the recommended structure. Using density functional theory to optimize the structure can not only make the configuration more reasonable and closer to the low-energy region of the potential energy surface, but also provide an online training set for the establishment of a machine learning potential model.

[0016] Step 5: Prepare a training set from the trajectory of the optimized structure; using the optimized trajectory to prepare the training set can effectively capture the information in the structure optimization process, making the trained model more suitable for subsequent structural relaxation using machine learning potentials.

[0017] Step six: train the machine learning potential model; the integrated approach can effectively improve the accuracy and robustness of the model. The trained machine learning potential model can replace the high-precision density functional theory to calculate the energy and force of the structure, thereby greatly saving time and reducing computational costs.

[0018] Step seven, use the trained machine learning potential model to optimize the unrelaxed structure; using the trained machine learning potential model to optimize the unrelaxed structure can make the configuration more reasonable, approaching the low-energy region of the potential energy surface, and increasing the probability of the configuration becoming a local minimum.

[0019] Step eight, use a multi-generational operation method to determine whether the set termination condition is met. If not, repeat steps two to eight until the termination condition is met. The multi-generational operation method can continuously increase the training set, update the machine learning model, and continuously improve the model's prediction ability, effectively improving the prediction and recommendation capabilities.

[0020] Step 9: After the termination condition is met, a thermodynamically stable structure set is recommended based on the structure set optimized by density functional theory and the structure set that converges after machine learning potential optimization. This not only effectively improves the probability of finding the most thermodynamically stable structure, but also effectively provides a group of structures with energy close to the most thermodynamically stable structure.

[0021] Furthermore, in step 2, clustering the structures using a clustering algorithm in machine learning includes:

[0022] Set the number of clusters. During the clustering process, each structure is converted into a feature vector as the input of the clustering. The crystal structure is converted into a feature vector using the atomic central symmetry function ACSFs; the atomic central symmetry function includes two parts: radial and angular functions. The radial symmetry function G2 describing the radial environment of atom i is The function G2 is the sum of a Gaussian function multiplied by a cutoff function. The width of the Gaussian function is defined by the parameter η, and the center of the Gaussian function is defined by the parameter R s Moved to a certain radial distance, the shifted G2 function is suitable for describing a spherical shell around the reference atom, and the truncation function takes the following form:

[0023]

[0024] Among them, R ij is the distance between atom i and atom j. If R ij Greater than the cutoff radius R c , the cutoff function and its derivative are zero; the angular function of the central atom i is the angle The sum of the cosine functions of is expressed as:

[0025]

[0026] where λ is either +1 or 1, and the angular resolution is provided by the parameter ζ. The larger the ζ value, the narrower the range of non-zero symmetry function values. Therefore, a set of angular functions with different ζ values ​​is used to obtain the distribution of angles centered on each reference atom. The angular distribution is determined by choosing appropriate η and R. c Determine and use it to control the radial part; by setting different parameter values, a series of functions G2 and G4 can be used to convert the local environment of the central atom into a characteristic vector; the atomic characteristic vectors of the same element are added together to obtain the characteristic vector of the same element, and then the characteristic vectors of different elements are sequentially concatenated. By calculating the distance between the characteristic vectors of two structures, the similarity of the two structures can be quantitatively described.

[0027] Furthermore, in step three, the strategy for optimizing the structure adopts a clustering method to cluster the structure set, predicts the energy of the structure in each cluster according to the machine learning potential model trained in step six, and sorts them from low to high according to energy, taking into account both energy priority and diversity principles when selecting the structure to be optimized.

[0028] Furthermore, in step five, the training set is derived from each frame structure in each structure optimization process in step four, while taking into account the energy difference between the previous and next two frame structures during selection. When the energy difference is less than a set value, the current frame structure is not added to the training set.

[0029] Further, in step six, the training of the machine learning potential model includes:

[0030] An integrated approach is adopted to simultaneously train multiple machine learning potential models, and the average value of the predictions of all models is used as the prediction result. During the training process of each machine learning potential model, the energy and force corresponding to the structure are simultaneously trained; wherein the machine learning potential model adopts a back propagation neural network BPNN, and other machine learning potential models are also applicable to the method of the present invention.

[0031] Furthermore, in step seven, the trained machine learning potential model is used to optimize the structure without relaxation; the structure that converges after optimization is saved in the structure set that converges after optimization by the machine learning potential model for use in step nine; the structure that does not converge after optimization is used as the source of the clustering structure in step two.

[0032] Furthermore, in step nine, after the termination condition is met, the structure set that converges after optimization by the machine learning potential model is sorted by energy, and the top K structures are recommended for verification using the more reliable density functional theory method in step four; the verified K structures are added to the structure set optimized using density functional theory, and finally the structures in the structure set optimized using density functional theory are sorted from low to high by energy, and the top N structures are selected as the recommended thermodynamically stable structure set.

[0033] Another object of the present invention is to provide a system for predicting and recommending the thermodynamically stable structure of disordered materials using the method for predicting and recommending the thermodynamically stable structure of disordered materials, wherein the system for predicting and recommending the thermodynamically stable structure of disordered materials comprises:

[0034] A non-redundant structure determination module is used to generate all possible non-redundant structures under specified substitution / defect / co-occupancy states using existing programs (such as supercell programs);

[0035] The structural clustering module is used to cluster the structure using the clustering algorithm in machine learning;

[0036] The module for obtaining structures to be optimized is used to recommend structures to be optimized that need to be optimized using density functional theory from each clustering result. The recommendation principles follow energy priority and diversity priority.

[0037] Recommended structure optimization module, used to optimize the recommended structure using density functional theory method;

[0038] A training set acquisition module, used to prepare training sets from the trajectory of the optimized structure;

[0039] Model training module, used to train machine learning potential models;

[0040] Unrelaxed Structure Optimization module, used to optimize unrelaxed structures using trained machine learning potential models;

[0041] The multi-generation operation module is used to use the multi-generation operation method to determine whether the set termination condition is reached;

[0042] The structure set recommendation module is used to recommend thermodynamically stable structure sets based on the structure sets optimized by density functional theory and the structure sets converged after machine learning potential optimization after meeting the termination conditions.

[0043] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0044] (1) Use existing programs (such as supercell programs) to generate all possible non-redundant structures under specified substitution / defect / co-occupancy states;

[0045] (2) Clustering structures using clustering algorithms in machine learning;

[0046] (3) Recommend the structures to be optimized using density functional theory from each clustering result, with the recommendation principles following energy priority and diversity priority;

[0047] (4) Density functional theory methods were used to optimize the recommended structures;

[0048] (5) preparing a training set from the trajectory of the optimized structure;

[0049] (6) Training machine learning potential models;

[0050] (7) Optimizing the structure without relaxation using a trained machine learning potential model;

[0051] (8) Using a multi-generation operation method, determine whether the set termination condition is met. If not, repeat steps (2) to (8) until the termination condition is met;

[0052] (9) After the termination conditions are met, a thermodynamically stable structure set is recommended based on the structure set optimized by density functional theory and the structure set that converges after machine learning potential optimization.

[0053] Another object of the present invention is to provide an information data processing terminal, which is used to implement the prediction and recommendation system for the thermodynamically stable structure of disordered materials.

[0054] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows: the method for predicting and recommending thermodynamically stable structures of disordered materials provided by the present invention takes disordered structural materials as the research object, adopts currently available programs (such as supercell programs) to obtain all non-redundant structures under specified substitution / defect / co-occupancy states, and recommends a set of thermodynamically stable structures among these non-redundant structures based on machine learning with very low computational and time costs. The recommended set of thermodynamically stable structures includes not only the most thermodynamically stable structures, but also many structures with energy similar to the most thermodynamically stable structures, thereby providing a reliable theoretical model for experiments. Therefore, the method provided by the present invention overcomes the defects of existing disordered structure programs (SOD, enumlib, supercell, disorder), and fills the gap in the field of disordered materials; it breaks the traditional practice of enumerating and calculating all non-redundant structures, and changes the view that a large number of density functional theory calculations are required. Only very small computational costs and time costs are required to find the most thermodynamically stable structure, and many structures with similar energy to the most thermodynamically stable structure can also be predicted and recommended; this invention can be used as a theoretical characterization instrument to collaborate with experimental characterization instruments, quickly provide candidate structures for experiments online, and assist in structural characterization, and can also be adapted to discover materials with excellent functions based on skeleton template modification, with broad and valuable commercial application scenarios. At the same time, the present invention is not only applicable to the prediction and recommendation of disordered structure materials, but can also be combined with sampling technology to be applicable to a wider range of crystal structure predictions for the discovery of new materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. 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.

[0056] Figure 1 It is a flow chart of a recommended method for predicting the thermodynamically stable structure of disordered materials provided by an embodiment of the present invention.

[0057] Figure 2 It is a schematic diagram of a recommended method for predicting the thermodynamically stable structure of disordered materials provided by an embodiment of the present invention.

[0058] Figure 3 is a schematic diagram of the structure of a prediction and recommendation system for thermodynamically stable structures of disordered materials provided by an embodiment of the present invention;

[0059] In the figure: 1. Non-redundant structure determination module; 2. Structure clustering module; 3. Structure acquisition module for optimization; 4. Recommended structure optimization module; 5. Training set acquisition module; 6. Model training module; 7. Unrelaxed structure optimization module; 8. Multi-generation operation module; 9. Structure set recommendation module.

[0060] Figure 4 This is a graph showing the structural label names and energies of 2664 non-redundant structures of BaScO2F (ICSD: 150171) after 2×2×2 cell expansion and optimization using VASP, provided by an embodiment of the present invention.

[0061] Figure 5 It is a relationship diagram between the number of running generations provided by an embodiment of the present invention and the energy of the structure optimized by density functional theory in each generation.

[0062] Figure 6 It is a schematic diagram of the total energy and the most stable structure of 10496 non-redundant structures in the example ∈-Fe2C 2×2×3 supercell structure provided by an embodiment of the present invention after VASP optimization.

[0063] Figure 7 It is a relationship diagram between the number of running generations provided by an embodiment of the present invention and the energy of the structure optimized by density functional theory in each generation. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0065] In view of the problems existing in the prior art, the present invention provides a method, device and terminal for predicting and recommending the thermodynamically stable structure of disordered materials. The present invention is described in detail below in conjunction with the accompanying drawings.

[0066] like Figure 1 As shown, the method for predicting and recommending the thermodynamically stable structure of disordered materials provided by the embodiment of the present invention comprises the following steps:

[0067] S101, using an existing program (such as a supercell program) to generate all possible non-redundant structures under a specified substitution / defect / co-occupancy state;

[0068] S102, clustering the structure using a clustering algorithm in machine learning;

[0069] S103, recommending the structures to be optimized that need to be optimized by density functional theory from each clustering result, with the recommendation principles following energy priority and diversity priority;

[0070] S104, optimization of the proposed structure using density functional theory methods;

[0071] S105, preparing a training set from the trajectory of the optimized structure;

[0072] S106, training a machine learning potential model;

[0073] S107, optimization of the structure without relaxation using the trained machine learning potential model;

[0074] S108, using a multi-generation operation method to determine whether the set termination condition is met, if not, repeat S102 to S108 until the termination condition is met;

[0075] S109, after the termination conditions are met, a thermodynamically stable structure set is recommended based on the structure set optimized by density functional theory and the structure set converged after machine learning potential optimization.

[0076] The principle diagram of the recommended method for predicting the thermodynamically stable structure of disordered materials provided by the embodiment of the present invention is as follows: Figure 2 shown.

[0077] like Figure 3 As shown, the prediction and recommendation system for thermodynamically stable structures of disordered materials provided by the embodiment of the present invention includes:

[0078] A non-redundant structure determination module 1 is used to generate all possible non-redundant structures under a specified replacement / defect / co-occupancy state using an existing program supercell;

[0079] The structure clustering module 2 is used to cluster the structure using the clustering algorithm in machine learning;

[0080] The module 3 for obtaining structures to be optimized is used to recommend structures to be optimized that need to be optimized by density functional theory from each clustering result, and the recommendation principle follows energy priority and diversity priority;

[0081] Recommended structure optimization module 4, used to optimize the recommended structure using density functional theory method;

[0082] A training set acquisition module 5, for preparing a training set from the trajectory of the optimized structure;

[0083] Model training module 6, used for training machine learning potential model;

[0084] An unrelaxed structure optimization module 7, for optimizing the unrelaxed structure using a trained machine learning potential model;

[0085] The multi-generation operation module 8 is used to use the multi-generation operation mode to determine whether the set termination condition is reached;

[0086] The structure set recommendation module 9 is used to recommend a thermodynamically stable structure set based on the structure set optimized by density functional theory and the structure set converged after machine learning potential optimization after the termination condition is met.

[0087] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0088] Example 1

[0089] The method for predicting and recommending the thermodynamically stable structure of disordered materials provided by the embodiment of the present invention includes:

[0090] Step 1: Use existing programs (such as supercell programs) to generate all possible non-redundant structures under specified substitution / defect / co-occupancy states;

[0091] Step 2: Cluster the structure using clustering algorithm in machine learning;

[0092] Step 3: Recommend the structure that needs to be optimized by density functional theory in the next step from each clustering result. The recommendation principle follows the energy priority and diversity priority.

[0093] Step 4: Optimize the recommended structure using density functional theory methods;

[0094] Step 5: Prepare a training set from the trajectory of the optimized structure;

[0095] Step 6: Train the machine learning model;

[0096] Step 7: Optimize the structure without relaxation using the trained machine learning potential model;

[0097] Step 8: Determine whether the set termination condition is met. If not, repeat steps 2 to 8 until the termination condition is met.

[0098] Step 9: After the termination conditions are met, a thermodynamically stable structure set is recommended based on the structure set optimized by density functional theory and the structure set that converges after machine learning potential optimization.

[0099] The strategy described in steps 2 and 3 provided in the embodiment of the present invention for recommending the structure to be optimized in step 4 uses a clustering method to cluster the structure set, predicts the energy of the structure in each cluster according to the machine learning potential model trained in step 6, and sorts them from low to high according to energy, and takes into account both energy priority and diversity principles when selecting the structure to be optimized, thereby ensuring that the structure recommended to be optimized in step 4 has a greater probability of becoming a thermodynamically stable structure.

[0100] Step 4 provided in the embodiment of the present invention uses density functional theory optimization, a reliable method, to optimize the structure recommended in step 3.

[0101] The training set in step five provided in the embodiment of the present invention is derived from each frame structure in each structure optimization process in step four, while taking into account the energy difference between the previous and next frame structures during selection. When the energy difference is less than the set value, the current frame structure is not added to the training set.

[0102] Step 6 provided in the embodiment of the present invention adopts an integrated approach to simultaneously train multiple machine learning potential models, and adopts the average value of all model predictions as the prediction result. During the training process of each machine learning potential model, the energy and force corresponding to the structure are simultaneously trained. In this example, the machine learning potential model is used. The present invention adopts BPNN (back propagation neural network), but is not limited to BPNN. Other machine learning potential models are also applicable to this method.

[0103] In step 7 provided in the embodiment of the present invention, the trained machine learning potential model is used to optimize the structure without relaxation. The structure that converges after optimization can be saved to the structure set that converges after the machine learning potential model is optimized for use in step 9; the structure that does not converge after optimization is used as the source of the clustering structure in step 2.

[0104] The method for predicting and recommending thermodynamically stable structures of disordered materials provided in an embodiment of the present invention adopts a multi-generation operation mode, and when the termination condition in step eight is not met, steps two to eight are repeated. Through the multi-generation operation mode, on the one hand, the training samples can be increased, the model training is more reliable, and the structure recommended to be optimized in step four has a greater probability of becoming a thermodynamically stable structure; on the other hand, the number of converged structures in the relaxed structure optimized by the machine learning potential model continues to increase, which greatly improves the probability of finding a thermodynamically stable structure.

[0105] The method for predicting and recommending thermodynamically stable structures of disordered materials provided by the embodiment of the present invention uses step nine to sort the structure set that converges after optimization by the machine learning potential model after the termination condition is met, recommends the first K structures to be verified by the more reliable density functional theory method in step four, then adds the verified K structures to the structure set optimized by density functional theory, and finally sorts the structures in the structure set optimized by density functional theory from low to high by energy, and selects the first N as the recommended thermodynamically stable structure set. Using this recommendation method, not only can the most thermodynamically stable structure be found, but also many structures with energies close to the energy of the most thermodynamically stable structure can be found.

[0106] Example 2

[0107] According to the present invention, a method for predicting and recommending the thermodynamically stable structure of disordered materials is implemented by the following specific technical solutions:

[0108] Step 1: Use existing programs such as the supercell program to generate all possible non-redundant structures under specified substitution / defect / co-occupancy states. In this process, the supercell size needs to be specified.

[0109] Step 2: Cluster the structures using clustering algorithms in machine learning (such as K-means). The structures that need to be clustered here refer to the non-redundant structures generated in step 1 for the first generation, and to the structures that did not converge after optimization by the previous generation machine learning potential model for the second generation and later. The number of clusters is generally set small (the recommended value is 3 or 5). During the clustering process, each structure needs to be converted into a feature vector as the input for clustering. The crystal structure is converted into a feature vector using atomic centrosymmetric functions (ACSFs). The atomic centrosymmetric function consists of two parts: radial and angular functions. The radial symmetry function G2 that describes the radial environment of atom i is The function G2 is the sum of a Gaussian function multiplied by a cutoff function. The width of the Gaussian function is defined by the parameter η, and the center of the Gaussian function can be determined by the parameter R s Moved to a certain radial distance, these shifted G2 functions are suitable for describing a spherical shell around a reference atom, and the cutoff function takes the form:

[0110]

[0111] Where R ij is the distance between atom i and atom j. If R ij Greater than the cutoff radius R c , the cutoff function and its derivative are zero. The angular function of the central atom i is the angle The sum of the cosine functions of is expressed as: The values ​​of λ can be between +1 and 1, and the angular resolution is provided by the parameter ζ. The larger the value of ζ, the narrower the range of non-zero symmetry function values. Therefore, a set of angular functions with different values ​​of ζ can be used to obtain a distribution of angles centered on each reference atom. In addition, this angular distribution can be obtained by choosing appropriate η and R. c To determine, they control the radial part. By setting different parameter values, a series of functions G2 and G4 can be used to convert the local environment of the central atom into a characteristic vector. Further, if the characteristic vector of the entire structure is needed, the atomic characteristic vectors of the same element can be added to obtain the characteristic vector of the same element, and then the characteristic vectors of different elements can be spliced ​​in sequence. By calculating the distance between the characteristic vectors of two structures, the similarity of the two structures can be quantitatively described.

[0112] Step three, recommend a structure from each clustering result in step two for the next step of density functional theory optimization. The recommendation principle takes into account both energy priority and diversity priority. Since the first generation did not have a trained machine learning potential, a random strategy was used to randomly select a structure from each clustering result. Starting from the second generation, the existing machine learning potential is used to sort the structures in each cluster from low to high according to energy, starting with the structure with the lowest energy, while taking into account the similarity between the selected structure and the structure currently optimized by density functional theory. The similarity calculation is calculated using the method described in step two. When the similarity is greater than the set value, the structure is selected. Otherwise, the structure ranked second in energy is selected for similarity judgment. Follow this step until a structure that meets the requirements is selected.

[0113] Step 4: Use density functional theory to optimize the structure selected in step 3, and save the optimized structure into a set of structures optimized using density functional theory.

[0114] Step 5, prepare the training set, extract each frame structure (including its corresponding energy and force) in the optimization process of each structure in step 4, and take into account the energy difference principle when extracting each frame structure, that is, it is added to the training set only when the energy of the previous frame structure is greater than the set value. The optimization trajectory of each generation of optimized structure is extracted according to the above method and then added to the training set. As the number of optimized structures increases, the number of training sets will continue to increase. The maximum training set value T can be set. When the number of training sets is greater than the set value T, the training sets are sorted from low to high by energy, and the top T structures are selected for training. At the same time, the optimized structure of each generation is saved for the convenience of step 3 and analysis results.

[0115] Step 6, training the machine learning potential model. The conversion of the structure space to the feature space is performed using the atomic central symmetry functions (ACSFs) described in step 2. The machine learning potential model uses BPNN (back propagation neural network), which is trained using the pytorch library. During the training process, the energy and force of the structure are trained simultaneously. In order to ensure the reliability of the prediction model, the present invention adopts an integrated BPNN method to train multiple BPNN models at the same time, and the average results predicted by multiple models are used as the final prediction results.

[0116] Step seven, use the machine learning potential model trained in step six to optimize the unrelaxed structure. The first generation of unrelaxed structures refers to the structure generated in step one, minus the structure optimized using density functional theory. The second generation and later structures that have not been relaxed refer to the structures that have not converged after optimization by the previous generation of machine learning potential model, minus the structure optimized by density functional theory of the current generation. The machine learning potential model uses BFGS to optimize atomic positions, and the convergence condition is when the force on the structure is less than the force set value or the energy error predicted by the integrated BPNN model is greater than the set value. After the optimization is completed, extract the last frame structure in the optimization process and save them in the structure set where the machine learning potential optimization converges and the structure where the current generation of machine learning potential model optimization does not converge. Among them, the structure where the machine learning potential model optimizes converges is used for further verification using density functional theory at the end of step nine, and the structure where the machine learning potential model optimizes does not converge is used for clustering in step two.

[0117] Step 8: Determine whether the set termination condition is met. There are two ways to determine the termination condition. The first is to determine whether the set number of generations is reached. If it is reached, the program terminates. If not, jump to step 2 and repeat until the termination condition is met. The second is to determine the number of generations of the best individual in history. If the number of generations of the best individual in history reaches the set number of survival generations, the program terminates. If not, jump to step 2 and repeat until the termination condition is met.

[0118] Step nine, after the termination conditions are met, the structures in the structure set that converges after machine learning potential optimization are sorted from low to high according to energy, and the top K structures are recommended for verification using the more reliable density functional theory method in step four, and then the verified K structures are added to the structure set optimized using density functional theory. Finally, the structures in the structure set optimized using density functional theory are sorted from low to high according to energy, and the top N are selected as the recommended thermodynamically stable structure set.

[0119] Example 3

[0120] In the BaScO2F (ICSD: 150171) structure, O atoms and F atoms co-occupy the same site, with O / F occupancy ratios of 0.667 / 0.333, 5 atoms in a unit cell, and 40 atoms in a 2×2×2 supercell, corresponding to 2664 non-redundant structures. This example recommends a set of thermodynamically stable structures from 2664 structures based on machine learning potential.

[0121] In step 1, the supercell program was used to specify a 2×2×2 cell expansion method to generate all possible combination structures of BaScO2F, a total of 735,471 structures, and then non-redundant structures were obtained by removing symmetry-equivalent structures, a total of 2,664 structures.

[0122] Step 2: Use the K-means algorithm in scikit-learn to cluster the structures. The first generation of structures that need to be clustered are the 2664 structures in step 1. The structures that need to be clustered after the second generation refer to the structures that did not converge after the previous generation of machine learning potential optimization. The number of clusters is 3, and each structure is converted into a feature vector using ACSFs as K-means input. The parameters G2 and G4 corresponding to ACSFs are set as follows: g2_etas = [1], g2_Rses = [1, 2, 3], g4_etas = [1], g4_zetas = [1, 2], g4_lambdas = [1, -1].

[0123] Step three, select a structure from each clustering result and recommend it to the next step of density functional theory optimization. For the first generation, the present invention randomly selects two structures as the next step to be optimized. In this example, the randomly selected structure is the structure numbered 34,259; for the second generation and later, one is selected from each clustering result taking into account the energy priority and diversity priority principles, a total of three structures to be optimized, and the structure of each clustering result will be sorted from low to high according to the energy predicted by the previous generation machine learning potential model. When selecting, start with the structure with the lowest energy, and at the same time judge the similarity between the structure and the m structure with the lowest energy optimized by density functional theory in the previous generations for comparison. In this example, the m value is set to 1, the similarity is set to 0.5, and the similarity calculation adopts the method described in step two. When the similarity is less than 0.5, the structure ranked second in energy is considered until a candidate structure that meets the conditions is selected.

[0124] Step 4: Use VASP software to optimize the structure selected in step 3, and save the optimized structure into the structure set optimized by density functional theory.

[0125] Step five, prepare a training set from the optimization trajectory of the optimized structure. There are 128 frames of structure (including the energy and force of the corresponding structure) in the optimization process of the first generation of two optimized structures. The present invention sets the energy difference value to 0.01eV. The energy difference between the previous and next structures is greater than the set value before it is added to the training set. The number of training sets is 40, and T is set to 20000 during the training process.

[0126] Step 6: Use an integrated approach to train 10 groups of BPNN models, and use the average of the 10 training groups as the prediction result. The parameters corresponding to ACSFs are set as: g2_etas = [0.05, 4, 20], g2_Rses = [0], g4_etas = [0.005], g4_zetas = [1, 4], g4_lambdas = [-1, 1]. The neural network uses two hidden layers, each with 40 nodes, and the corresponding activation function is the tanh function. After the training meets the set energy convergence standard and force convergence standard, the trained model parameters are saved.

[0127] Step seven, use the machine learning potential model trained in step six to optimize the unrelaxed structures. The first generation of unrelaxed structures refers to the 2664 non-redundant structures generated in step one, minus the two structures optimized by density functional theory in step three, for a total of 2662 structures. The structures that have not been relaxed after the second generation refer to the structures that have not converged after the optimization by the previous generation of machine learning potential model, and the structures optimized by density functional theory in the current generation are removed. The BFGS algorithm is used to optimize the atomic positions. When the force is less than 0.05, the structure converges. The last frame structure of the converged structure (including the corresponding energy and force) is saved to the structure set that has converged with the machine learning potential optimization, and the last frame structure of the unconverged structure (including the corresponding energy and force) is saved to the structure that has not converged after the machine learning potential optimization. In this example, there are 13 structures that have converged and 2649 structures that have not converged in the first generation of machine learning potential model optimization. The 2649 structures are provided for clustering in the next generation of step two.

[0128] Step eight, determine whether the set termination condition is met. Here, the termination condition is set in the first way of the present invention, that is, the set running algebra. In this example, the algebra set by the present invention is 6, so repeat steps two to eight until the termination condition is met.

[0129] Step nine, after the termination conditions are met, the structures in the structure set that converged after machine learning potential optimization are sorted from low to high according to energy, and the top 12 structures are selected to be recommended for density functional theory optimization. The optimized structures are saved in the structure set optimized using density functional theory. Finally, the structures in the structure set optimized using density functional theory are sorted from low to high according to energy, and the top 20 structures are selected as the final recommended thermodynamically stable structure set.

[0130] Figure 4 For example, the structure label names and energy correspondence diagrams of 2664 non-redundant structures after 2×2×2 expansion of BaScO2F (ICSD: 150171) are optimized by VASP. The purpose of optimizing 2664 structures by density functional theory is to obtain a complete test set, so as to evaluate whether the method of the present invention can recommend a reliable thermodynamically stable structure set with a small computational cost.

[0131] Figure 5 This is a graph showing the relationship between the number of generations run and the energy of the structure optimized using density functional theory in each generation. The black hollow balls represent the total energy of the structure optimized using density functional theory in each generation, and the gray solid balls represent the energy of the lowest energy structure in the history (among all the generations run so far) optimized using density functional theory. In this example, a total of six generations were run, and the lowest energy structure in the fourth generation was still the lowest energy structure in the fifth and sixth generations. And this structure is Figure 4 The lowest energy structure among the 2664 structures calculated by the exhaustive method has an energy value of -298.1296 eV.

[0132] Table 1 shows the top 20 structural label names and corresponding energies of 2664 non-redundant structures after 2×2×2 cell expansion of the example BaScO2F (ICSD: 150171) using VASP optimization and ranked from low to high energy.

[0133] Table 1

[0134]

[0135]

[0136] Table 2 shows the top 20 structure tags and corresponding energies recommended by the method of the present invention after the termination of 6 generations of operation, as well as whether the recommended structures appear in the corresponding structures in Table 1. Table 2 shows that 16 of the 20 structures recommended by this method appear in Table 1, indicating that this method can use very few computing resources and time costs to recommend a reliable prediction recommendation set of thermodynamically stable structures of disordered materials. The exhaustive method requires the use of density functional theory to optimize all 2664 structures to find the most thermodynamically stable structure. This method only needs to use density functional theory to optimize 29 structures to find the most thermodynamically stable structure, and at the same time provides a group of structures with energies close to the energy of the most thermodynamically stable structure.

[0137] Table 2

[0138] recommend Structure Tag Name Total energy (eV) Is it in Table 1 1 000029 -298.1296 yes 2 001840 -298.1036 yes 3 000791 -298.0732 yes 4 000210 -298.0698 yes 5 001873 -298.0495 yes 6 001789 -298.0490 yes 7 000098 -298.0484 yes 8 000095 -298.0477 yes 9 000537 -298.0450 yes 10 000444 -298.0438 yes 11 001827 -298.0433 yes 12 000794 -298.0429 yes 13 001818 -298.0409 yes 14 000046 -298.0408 yes 15 000373 -298.0388 yes 16 000211 -298.0379 yes 17 000479 -298.0335 no 18 000475 -298.0324 no 19 001791 -298.0310 no 20 000230 -298.0305 no

[0139] Example 4

[0140] In order to obtain a thermodynamically stable ε-Fe2C structure, the ε-Fe2C structure is generated by creating defects (removing half of the C atoms in the system, with a defect ratio of 0.5). The number of atoms in the ε-FeC unit cell is 4, and the number of atoms in the 2×2×3 supercell is 48 (24 Fe, 24 C). After removing 12 C atoms, the number of corresponding non-redundant structures reaches 10496 structures. This example recommends a thermodynamically stable structure set from 10496 structures based on machine learning potential. The case source article is J.Phys.Chem.C 2017, 121, 39, 21390–21396.

[0141] In step 1, the supercell program was used to specify a 2×2×3 cell expansion method to generate all possible combination structures of ε-Fe2C, a total of 2704156 structures, and then non-redundant structures were obtained by removing symmetry-equivalent structures, a total of 10496 structures.

[0142] Step 2: Use the K-means algorithm in scikit-learn to cluster the structures. The first generation of structures that need to be clustered are the 10496 structures in step 1. The structures that need to be clustered after the second generation refer to the structures that did not converge after the previous generation of machine learning potential optimization. The number of clusters is 3, and each structure is converted into a feature vector using ACSFs as K-means input. The parameters G2 and G4 corresponding to ACSFs are set as follows: g2_etas = [1], g2_Rses = [1, 2, 3], g4_etas = [1], g4_zetas = [1, 2], g4_lambdas = [1, -1].

[0143] Step three, select a structure from each clustering result and recommend it to the next step of density functional theory optimization. For the first generation, the present invention randomly selects two structures as the next step to be optimized. In this example, the randomly selected structure is a structure numbered 0, 52; for the second generation and later, one is selected from each clustering result taking into account the energy priority and diversity priority principles, a total of three structures to be optimized, and the structure of each clustering result will be sorted from low to high according to the energy predicted by the previous generation machine learning potential model. When selecting, start with the structure with the lowest energy, and at the same time judge the similarity between the structure and the m structure with the lowest energy optimized by density functional theory in the previous generations for comparison. In this example, the m value is set to 1, the similarity is set to 0.5, and the similarity calculation adopts the method described in step two. When the similarity is less than 0.5, the structure ranked second in energy is considered until a candidate structure that meets the conditions is selected.

[0144] Step 4: Use VASP software to optimize the structure selected in step 3, and save the optimized structure into the structure set optimized by density functional theory.

[0145] Step five, prepare a training set from the optimization trajectory of the optimized structure. There are 74 frames of structure (including the energy and force of the corresponding structure) in the optimization process of the first generation of two optimized structures. The present invention sets the energy difference value to 0.01eV. The energy difference between the previous and next structures is greater than the set value before it is added to the training set. The number of training sets is 18, and T is set to 20000 during the training process.

[0146] Step 6: Use an integrated approach to train 10 groups of BPNN models, and use the average of the 10 training groups as the prediction result. The parameters corresponding to ACSFs are set as: g2_etas = [0.05, 4, 20], g2_Rses = [0], g4_etas = [0.005], g4_zetas = [1, 4], g4_lambdas = [-1, 1]. The neural network uses two hidden layers, each with 40 nodes, and the corresponding activation function is the tanh function. After the training meets the set energy convergence standard and force convergence standard, the trained model parameters are saved.

[0147] Step seven, use the machine learning potential model trained in step six to optimize the unrelaxed structures. The first generation of unrelaxed structures refers to the 10,496 non-redundant structures generated in step one, minus the two structures optimized by density functional theory in step three, for a total of 10,494 structures. The structures that have not been relaxed after the second generation refer to the structures that have not converged after the optimization of the previous generation of machine learning potential models, and the structures optimized by density functional theory in the current generation are removed. The BFGS algorithm is used to optimize the atomic positions. When the force is less than 0.05, the structure converges. The last frame structure of the converged structure (including the corresponding energy and force) is saved to the structure set that has converged with the machine learning potential optimization, and the last frame structure of the unconverged structure (including the corresponding energy and force) is saved to the structure that has not converged after the machine learning potential optimization. In this example, there are 0 structures that have converged with the optimization of the first generation of machine learning potential models, and there are 10,494 structures that have not converged. The 10,494 structures are provided for clustering in the next generation of step two.

[0148] Step eight, determine whether the set termination condition is met. Here, the termination condition is set in the first way of the present invention, that is, the set running algebra. In this example, the algebra set by the present invention is 10, so repeat steps two to eight until the termination condition is met.

[0149] Step nine, after the termination conditions are met, the structures in the structure set that converged after machine learning potential optimization are sorted from low to high according to energy, and the top 21 structures are selected to be recommended for density functional theory optimization. The optimized structures are saved in the structure set optimized using density functional theory. Finally, the structures in the structure set optimized using density functional theory are sorted from low to high according to energy, and the top 16 structures are selected as the final recommended thermodynamically stable structure set.

[0150] Figure 6 The total energy and the most stable structure of 10496 non-redundant structures in the example ∈-Fe2C 2×2×3 supercell structure after VASP optimization. The picture comes from the article J.Phys.Chem.C 2017, 121, 39, 21390–21396. In the article, the authors used density functional theory to optimize 10496 structures and obtained the most thermodynamically stable structure. This example can be used to evaluate whether the method of the present invention can recommend a reliable set of thermodynamically stable structures at a lower computational cost. It should be noted that the present invention did not use VASP to optimize these 10496 non-redundant structures, but directly used the original results for comparison. In the specific implementation process, the parameters used by the present invention to optimize the structure using VASP are not completely consistent with those of the original author. Under the calculation parameters of the present invention Figure 5 The energy of the most thermodynamically stable structure shown in is -312.4113 eV, which is a little lower than the energy calculated using the parameters adopted in the original text, but this has no effect on the conclusion of the present invention.

[0151] Figure 7 This is a graph showing the relationship between the number of generations run and the energy of the structure optimized using density functional theory in each generation. The black hollow spheres represent the total energy of the structure optimized using density functional theory in each generation, and the gray solid spheres represent the energy of the lowest energy structure in the history (among all the generations run so far) optimized using density functional theory. In this example, a total of ten generations were run, and the energy of the lowest energy configuration in the eighth generation was -310.7659 eV, which was 1.64 eV away from the target configuration energy of -312.4113 eV, and this configuration was maintained in the ninth and tenth generations.

[0152] Table 3 shows the top 16 structure labels and corresponding energies recommended by the method of the present invention after 10 generations of operation. Table 3 shows that the energies corresponding to the 16 structures recommended by the method are all located in Figure 5 The energy of the first recommended structure is -312.4113 eV, which is the most stable structure among the 10496 non-redundant structures. This shows that this method can use very little computing resources and time to recommend a reliable set of prediction recommendations for thermodynamically stable structures of disordered materials. The exhaustive method needs to use density functional theory to optimize all 10496 structures to find the most thermodynamically stable structure. This method only needs to use density functional theory to optimize 50 structures to find the most thermodynamically stable structure, and at the same time provides a group of structures with energies close to the energy of the most thermodynamically stable structure.

[0153] Table 3

[0154]

[0155]

[0156] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0157] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for predicting and selecting thermodynamically stable structures of disordered materials, characterized in that: The method for predicting and selecting the thermodynamically stable structure of disordered materials comprises the following steps: Step 1: Use the existing program supercell to generate all possible non-redundant structures under the specified replacement / defective / co-occupancy state; Step 2: Cluster the structure using clustering algorithm in machine learning; Step 3: Select the structure to be optimized by density functional theory from each clustering result, and the selection principle follows energy priority and diversity priority; Step 4, optimizing the structure selected in step 3 by using density functional theory method, and saving the optimized structure into a structure set optimized by density functional theory; Step 5, prepare a training set from the trajectory of the optimized structure in step 4; Step 6: Train the machine learning potential model; Step 7, using the trained machine learning potential model to optimize the structure without relaxation; Step 8: Use a multi-generation operation method to determine whether the set termination condition is met. If not, repeat steps 2 to 8 until the termination condition is met. Step nine, after the termination condition is met, a thermodynamically stable structure set is selected based on the structure set optimized by density functional theory and the structure set converged after machine learning potential optimization; In step 2, the clustering algorithm in machine learning is used to cluster the structure, including: setting the number of clusters, converting each structure into a feature vector as the input of clustering during the clustering process, and converting the crystal structure into a feature vector using the atomic central symmetry function ACSFs; the atomic central symmetry function includes radial and angular functions, and the radial symmetry function G2 describing the radial environment of atom i is Function G2 is the sum of a Gaussian function multiplied by a cutoff function; the width of the Gaussian function is defined by parameter η, and the center of the Gaussian function is defined by parameter R s Moved to a certain radial distance, the shifted G2 function is suitable for describing a spherical shell around the reference atom, and the truncation function takes the following form: Among them, R ij is the distance between atom i and atom j. If R ij Greater than the cutoff radius R c , the cutoff function and its derivative are zero; the angular function of the central atom i is the angle The sum of the cosine functions of is expressed as: where λ is either +1 or 1, and the angular resolution is provided by the parameter ζ. The larger the ζ value, the narrower the range of non-zero symmetry function values. Therefore, a set of angular functions with different ζ values ​​is used to obtain the distribution of angles centered on each reference atom. The angular distribution is determined by choosing appropriate η and R. c Determine, used to control the radial part; by setting different parameter values, a series of functions G2 and G4 are used to convert the local environment of the central atom into a characteristic vector; the characteristic vectors of atoms of the same element are added together to obtain the characteristic vector of the same element, and then the characteristic vectors of different elements are sequentially spliced ​​to obtain the similarity of the two structures by calculating the distance between the characteristic vectors of the two structures; Step three specifically includes: selecting a structure from each clustering result in step two for the next step of density functional theory optimization; the selection principle takes into account both energy priority and diversity priority, and the first generation adopts a random strategy to randomly select a structure from each clustering result; starting from the second generation, the existing machine learning potential is used to sort the structures in each cluster from low to high according to energy, and the selection starts from the structure with the lowest energy, while taking into account the similarity between the selected structure and the structure currently optimized by density functional theory. The similarity calculation is calculated using the method described in step two. When the similarity is greater than the set value, the structure is selected, otherwise, the structure ranked second in energy is selected for similarity judgment, and this step is followed until a structure that meets the requirements is selected; In step 6, the training of the machine learning potential model includes: training multiple machine learning potential models simultaneously in an integrated manner, using the average value of predictions of all models as the prediction result, and simultaneously training the energy and force corresponding to the structure during the training process of each machine learning potential model; wherein the machine learning potential model adopts a back propagation neural network BPNN; In step 7, the trained machine learning potential model is used to optimize the structure without relaxation; the structure that converges after optimization is saved to the structure set that converges after optimization by the machine learning potential model for use in step 9; the structure that does not converge after optimization is used as the source of the clustering structure in step 2; In step nine, after the termination condition is met, the structure set that converges after optimization by the machine learning potential model is sorted by energy, and the top K structures are selected for verification using the more reliable density functional theory method in step four; the verified K structures are added to the structure set optimized using density functional theory, and finally the structures in the structure set optimized using density functional theory are sorted from low to high by energy, and the top N structures are selected as the selected thermodynamically stable structure set.

2. The method for predicting and selecting thermodynamically stable structures of disordered materials according to claim 1, characterized in that: In step three, the strategy for optimizing the structure is to cluster the structure set in a clustering manner, predict the energy of the structure in each cluster according to the machine learning potential model trained in step six, and sort them from low to high according to energy, taking into account both energy priority and diversity principles when selecting the structure to be optimized.

3. The method for predicting and selecting thermodynamically stable structures of disordered materials according to claim 1, characterized in that: In step 5, the training set is derived from each frame structure in each structure optimization process in step 4, while taking into account the energy difference between the previous and next frame structures during selection. When the energy difference is less than a set value, the current frame structure is not added to the training set.

4. A system for implementing the method for predicting and selecting the thermodynamically stable structure of disordered materials according to any one of claims 1 to 3, characterized in that: The prediction and selection system of the thermodynamically stable structure of disordered materials comprises: A non-redundant structure determination module, for generating all possible non-redundant structures under a specified replacement / defective / co-occupied state using a used program supercell; The structural clustering module is used to cluster the structure using the clustering algorithm in machine learning; The module for obtaining structures to be optimized is used to select the structures to be optimized that need to be optimized using density functional theory from each clustering result. The selection principle follows energy priority and diversity priority. Select the structure optimization module, which is used to optimize the selected structure using density functional theory methods; A training set acquisition module, used to prepare training sets from the trajectory of the optimized structure; Model training module, used to train machine learning potential models; Unrelaxed Structure Optimization module, used to optimize unrelaxed structures using trained machine learning potential models; The multi-generation operation module is used to use the multi-generation operation method to determine whether the set termination condition is reached; The structure set selection module is used to predict and select thermodynamically stable structure sets based on the structure sets optimized by density functional theory and the structure sets converged after machine learning potential optimization after the termination conditions are met.

5. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for predicting and selecting the thermodynamically stable structure of disordered materials as described in any one of claims 1 to 3.

6. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the prediction and selection system of the thermodynamically stable structure of disordered materials as described in claim 4.