Data set generation method and device for simulating electron microscope image
By performing multidimensional geometric screening, periodic expansion, and crystallographic coordinate system transformation on the cell structure data, the problem of low computational efficiency in single-machine numerical simulation was solved, and efficient large-scale material imaging simulation was realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-04-03
AI Technical Summary
Single-machine numerical simulation suffers from low computational efficiency when dealing with the repeated stacking of non-orthogonal unit cells in the target material, making it difficult to meet the needs of large-scale material imaging simulation.
By performing multidimensional geometric screening and periodic expansion on the original unit cell structure data, it is converted into orthogonal unit cell structure data. Then, through crystallographic coordinate system transformation technology, non-orthogonal unit cells are mapped to orthogonal coordinate systems. Representative unit cell structure data are selected for imaging simulation, generating simulated images and storing them.
It improves computational efficiency, reduces the complexity of mathematical models, and reduces redundant CPU consumption in coordinate calculations, thus meeting the needs of large-scale material imaging simulation.
Smart Images
Figure CN121789843A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computational materials science, and in particular to a method and apparatus for generating datasets that simulate electron microscope images. Background Technology
[0002] Numerical computation for reproducing microscopic imaging is an important field in computational materials science. Its core objective is to reproduce the imaging process of materials under a microscope within a computer through mathematical modeling and numerical solutions. This technique can simulate the microscopic characteristics of materials under different imaging parameters without relying on actual microscopes and sample preparation.
[0003] Single-machine numerical simulation is one of the implementation paths of this technology. The process involves first establishing a mathematical model by combining the physical laws of imaging with the cell structure parameters of the target material and setting initial conditions. Finally, the solution algorithm is executed by a single-machine central processing unit (CPU) to obtain the distribution of physical quantities and convert it into a simulated microscopic image.
[0004] However, single-machine numerical simulations target the repeated stacking of some non-orthogonal unit cells in the target material, directly simulating all unit cells. This leads to complex coordinates, an increasing number of atoms to be simulated, and consequently, low computational efficiency, making it difficult to meet the needs of large-scale material imaging simulation. Summary of the Invention
[0005] The main objective of this application is to provide a method and apparatus for generating electron microscope image datasets, aiming to form a standard image dataset and solve the problem of low computational efficiency in numerical computation reproduction of microscope imaging technology.
[0006] In a first aspect, embodiments of this application provide a method for generating a dataset simulating electron microscope images. The method includes: performing multidimensional geometric filtering on an original crystal database to select a first dataset, the first dataset including multiple first unit cell structure data conforming to the definition of two-dimensional materials; periodically expanding the first unit cell structure data to obtain second unit cell structure data, multiple second unit cell structure data forming a second dataset, the second dataset including multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cells; and using crystallographic coordinate system transformation technology to transform the second unit cell structures describing non-orthogonal unit cells in the second dataset. Data is mapped to an orthogonal coordinate system to obtain orthogonal unit cell structure data, which is then used as the third unit cell structure data. The third unit cell structure data and the second unit cell structure data of the orthogonal unit cell structure form a third dataset. Similar structure clusters of orthogonal unit cell data in the third dataset are identified, and representative orthogonal unit cell data are selected from each similar structure cluster as the fourth unit cell structure data, resulting in a fourth dataset. Imaging simulation is performed on each of the fourth unit cell structure data according to preset electron microscope parameters to generate a first simulated image. The first simulated image is associated and stored with its corresponding first unit cell structure data to obtain the target dataset.
[0007] In related technologies, for the repeated stacking of some non-orthogonal unit cells in the target material, all unit cells are directly simulated, which leads to complex coordinates, an increasing number of atoms to be simulated, and thus low computational efficiency, making it difficult to meet the needs of large-scale material imaging simulation.
[0008] This application first periodically expands multiple first unit cell structure data to obtain a second dataset, ensuring the integrity of the non-orthogonal unit cell structure data (second unit cell structure data). Then, through crystallographic coordinate system transformation techniques, the difficult-to-process non-orthogonal unit cell structure data is transformed into more easily solvable orthogonal unit cell structure data (third unit cell structure data), thereby significantly improving computational efficiency. Furthermore, this application reduces the complexity of the mathematical model by selecting representative fourth unit cell structure data from the third unit cell structure data instead of simulating all units, thus reducing redundant CPU consumption in coordinate calculations and improving the execution efficiency of the solution algorithm, meeting the needs of large-scale materials imaging simulation.
[0009] In one possible implementation, imaging simulation is performed on multiple fourth unit cell structure data according to preset electron microscope parameters to generate multiple first simulated images. This includes: dividing the fourth dataset into multiple subsets, each subset including at least one fourth unit cell structure data; allocating the multiple subsets to multiple parallel computing units, each parallel unit performing imaging simulation on the fourth unit cell structure data in its corresponding subset according to the preset electron microscope parameters to generate the first simulated image.
[0010] In one possible implementation, identifying similar structural clusters of orthogonal cell data in the third dataset and selecting representative orthogonal cell data from each similar structural cluster as fourth cell structure data to obtain the fourth dataset includes: converting the orthogonal cell data in the third dataset into a first cell structure feature vector; performing dimensionality reduction on the first cell structure feature vector to obtain a second cell structure feature vector; grouping structurally similar cell structure data in the second cell structure feature vector into the same cluster based on a clustering algorithm; and selecting the fourth cell structure data that best represents the structural features of each cluster to obtain the fourth dataset.
[0011] In one possible implementation, converting the third unit cell structure data into a first unit cell structure feature vector includes: projecting the atomic coordinates in the third unit cell structure data onto a two-dimensional plane to obtain a two-dimensional atomic distribution map; and performing vector flattening and data standardization on the two-dimensional atomic distribution histogram to obtain the first unit cell structure feature vector.
[0012] In one possible implementation, after generating multiple first simulated images by imaging multiple fourth unit cell structure data according to preset electron microscope parameters, the method further includes: performing constrained geometric transformations on the multiple first simulated images to obtain multiple second simulated images; performing Gaussian blur processing on the multiple second simulated images to obtain multiple third simulated images; and associating and storing the first simulated images with their corresponding first unit cell structure data to obtain a target dataset, including: associating and storing the third simulated images with their corresponding first unit cell structure data to obtain the target dataset.
[0013] In one possible implementation, a constrained geometric transformation is performed on multiple first simulated images to obtain multiple second simulated images, including: extracting crystallographic features of the unit cell structure data corresponding to the first simulated images, the crystallographic features including rotational symmetry order and mirror symmetry axis position; rotating the first simulated images based on the rotational symmetry order to obtain the second simulated images; and / or flipping the first simulated images based on the mirror symmetry axis position to obtain the second simulated images.
[0014] In one possible implementation, a first dataset is selected by performing multidimensional geometric screening on the original crystal database, including: selecting the first dataset based on thickness parameters, vacuum layer parameters, element types, and in-plane nearest neighbor continuity.
[0015] In one possible implementation, after associating and storing the first simulated image with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: dividing the target dataset into a training set, a validation set, and a test set based on a preset ratio.
[0016] In one possible implementation, after associating and storing the first simulated image with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: identifying the material dosage in the first unit cell structure data and generating a label based on the material dosage. The first simulated image is associated with the label based on the correspondence between the first simulated image and the first unit cell structure data.
[0017] Secondly, embodiments of this application also provide a dataset generation device for simulating electron microscope images. The device includes: a filtering module, an expansion module, a coordinate transformation module, an imaging simulation module, and a storage module. The filtering module performs multidimensional geometric filtering on the original crystal database to select a first dataset, which includes multiple first unit cell structure data conforming to the definition of two-dimensional materials. The expansion module periodically expands the first unit cell structure data to obtain second unit cell structure data. Multiple second unit cell structure data constitute a second dataset, which includes multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cells. The coordinate transformation module maps the second unit cell structure data describing non-orthogonal unit cells in the second dataset to an orthogonal coordinate system using crystallographic coordinate system transformation technology to obtain orthogonal unit cell structures. The data is processed by a third set of orthogonal cell structure data and a fourth set of orthogonal cell structure data. A filtering module identifies similar structure clusters of orthogonal cell data in the third set of data and selects representative orthogonal cell data from each cluster as fourth set of data, thus obtaining a fourth set of data. An imaging simulation module performs imaging simulation on each fourth set of data according to preset electron microscope parameters, generating a first simulated image. A storage module associates and stores the first simulated image with the corresponding first set of cell structure data to obtain a target set of data.
[0018] The beneficial effects of this invention are as follows: By periodically expanding multiple first unit cell structure data, a second dataset is obtained, ensuring the integrity of the non-orthogonal unit cell structure data (second unit cell structure data). Furthermore, through crystallographic coordinate system transformation technology, the difficult-to-process non-orthogonal unit cell structure data is transformed into more easily solvable orthogonal unit cell structure data (third unit cell structure data), thereby significantly improving computational efficiency. Moreover, this application also reduces the complexity of the mathematical model by selecting representative fourth unit cell structure data from the third unit cell structure data instead of simulating all unit cells, thereby reducing redundant CPU consumption in coordinate calculations and improving the execution efficiency of the solution algorithm, which can meet the needs of large-scale material imaging simulation. Attached Figure Description
[0019] Figure 1 This is a schematic flowchart of a method for generating a dataset of simulated electron microscope images provided in an embodiment of this application; Figure 2 This is a flowchart illustrating another method for generating a dataset of simulated electron microscope images provided in an embodiment of this application; Figure 3 This is a schematic diagram of a dataset generation device for simulating electron microscope images provided in an embodiment of this application.
[0020] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0021] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the data and materials used in the following embodiments are commercially available unless otherwise specified.
[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.
[0023] It should be understood that the term "and / or" used herein is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects are in an "or" relationship. Various modifications and variations can be made to this application without departing from its spirit or scope, which will be obvious to those skilled in the art. Therefore, this application is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the implementation methods provided in the embodiments of this application can be combined with each other without contradiction.
[0024] Before describing the technical solutions provided in the embodiments of this application, in order to facilitate the understanding of the embodiments of this application, this application first provides a specific explanation of the problems existing in the related technologies.
[0025] In related technologies, scanning transmission electron microscopy (STEM) is often used to scan the target material in order to obtain high-resolution information on its microstructure. However, STEM experiments rely on expensive double aberration-corrected electron microscopes, and there are strict limitations on the sample thickness. Furthermore, the electron dose is limited by the equipment performance, making it difficult to flexibly adjust different dose conditions to obtain diverse data.
[0026] Based on this, those skilled in the art have proposed a single-machine numerical simulation method to simulate and generate an image of the target material under STEM using its crystal structure data. This solves the problem of the high cost of directly scanning the target material with STEM.
[0027] However, single-machine numerical simulations target the repeated stacking of some non-orthogonal unit cells in the target material, directly simulating all unit cells. This leads to complex coordinates, an increasing number of atoms to be simulated, and consequently, low computational efficiency, making it difficult to meet the needs of large-scale material imaging simulation.
[0028] To address the aforementioned issues, this application provides a method and apparatus for generating a dataset of simulated electron microscope images. The following section will first provide a detailed description of a method for generating a dataset of simulated electron microscope images.
[0029] Figure 1 This is a schematic flowchart illustrating a method for generating a dataset of simulated electron microscope images provided in an embodiment of this application. Figure 1 As shown, a method for generating a dataset of simulated electron microscope images includes: S110-S160.
[0030] S110. Perform multidimensional geometric screening on the original crystal database to select the first dataset, which includes multiple first cell structure data that conform to the definition of two-dimensional materials.
[0031] The original crystal library can be a publicly available high-throughput database, such as the Open Quantum Materials Database (QQMD). The original crystal library contains various crystallographic information files (CIF).
[0032] The purpose of S110 is to accurately select the cell structure data of two-dimensional materials, namely the first cell structure data, through multi-dimensional geometric screening. The first cell structure data exists in the form of a CIF file.
[0033] In one possible implementation, S110 includes S111.
[0034] S111. Based on thickness parameters, vacuum layer parameters, element types, and in-plane nearest neighbor continuity, perform multidimensional geometric screening on the original crystal database to select the first dataset.
[0035] For example, the thickness parameter of the two-dimensional material used for STEM should be 0.3-0.5 nm, the vacuum layer parameter should be 10 Å-20 Å, the number of elements should not exceed 3, and the nearest neighbor threshold for in-plane nearest neighbor continuity can be 5 Å.
[0036] S120. Periodically expand the first unit cell structure data to obtain the second unit cell structure data. Multiple second unit cell structure data constitute a second dataset. The second dataset includes multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cell structures.
[0037] In this process, the data of multiple first-cell structures are periodically expanded, which is equivalent to supercelling of the data of multiple first-cell structures.
[0038] Because imaging simulation of two-dimensional materials requires ensuring that the two-dimensional plane covers a sufficiently large real space in order to generate STEM images that conform to the experimental scale, it is necessary to hypercell multiple first-cell structure data.
[0039] In one possible implementation, multiple CIF files storing first cell structure data are hierarchically expanded.
[0040] For example, for CIF files ≤1KB, the scaling factor in the xy plane is 16×16. Through high scaling, small-sized cells can achieve an imaging scale of ≥20nm. For CIF files >3KB, the scaling factor in the xy plane is 4×4. Only a small amount of scaling is needed to meet the ≥20nm scale, minimizing memory consumption and balancing scale requirements with computational efficiency.
[0041] Hierarchical expansion can avoid memory waste caused by excessive expansion of large-sized cells, while ensuring that small-sized cells achieve an effective imaging scale after expansion.
[0042] S130. Using crystallographic coordinate system transformation technology, the second cell structure data describing non-orthogonal unit cells in the second dataset is mapped to an orthogonal coordinate system to obtain orthogonal unit cell structure data, and this orthogonal unit cell structure data is used as the third unit cell structure data; the third unit cell structure data and the second unit cell structure data of the orthogonal unit cell structure constitute the third dataset.
[0043] Because the expanded unit cell data includes non-orthogonal unit cell data, and the lattice of non-orthogonal unit cells is not perpendicular, if the unit cell structure data of non-orthogonal unit cells is directly used for simulation imaging, spatial misalignment will occur between the sampling grid and the actual arrangement of crystal atoms. The scanning pixels cannot accurately cover the scattering region of atoms, resulting in atomic signal offset, periodic blurring of the lattice, and even artifacts that do not conform to physical laws in the generated simulation image.
[0044] In this embodiment, S130, the above-mentioned problem is solved by orthogonalizing the second cell structure data of the second dataset, which contains non-orthogonal unit cells, through crystallographic coordinate system transformation technology.
[0045] For example, taking the second unit cell structure data as hexagonal unit cell (such as graphene, hexagonal boron nitride) structure data, the lattice angle of the hexagonal unit cell is γ≈60° or γ≈120°.
[0046] When γ≈60°, the hexagonal unit cell is transformed into an orthorhombic unit cell using the matrix [[1,1,0],[-1,1,0],[0,0,1]].
[0047] When γ≈120°, the hexagonal unit cell is transformed into an orthorhombic unit cell using the matrix [[1, -1, 0], [1, 1, 0], [0, 0, 1]]. This ensures that the sampling grid is perfectly aligned with the x and y axes of the unit cell, guaranteeing that pixel-by-pixel scanning can accurately capture the scattering signals of atoms, thus fundamentally improving the spatial accuracy of the image simulation.
[0048] The second unit cell structure data can be any non-orthogonal unit cell that can be converted into an orthogonal unit cell, and this application does not limit it.
[0049] S140. Identify similar structure clusters of orthogonal cell data in the third dataset, and select representative orthogonal cell data from each similar structure cluster as fourth cell structure data to obtain the fourth dataset. Among them, unit cells with similar atomic arrangement patterns belong to the same cluster, which can be called a similar structure cluster.
[0050] For example, the core criterion for judging the similarity of atomic arrangement patterns is the two-dimensional atomic distribution characteristics of the unit cell (such as the eigenvectors after flattening the atomic density histogram, the distribution pattern of interatomic spacing, etc.). If the difference in atomic distribution characteristics between two unit cells is less than a preset threshold, they are judged to be structurally similar and classified into the same cluster.
[0051] S140 selects representative fourth cell structure data from all similar structure clusters, retaining the characteristics of different types of cell structures in the third cell structure data. The data is more concise and the information is more representative, greatly accelerating the computational efficiency.
[0052] In one possible implementation, Figure 2 This is a flowchart illustrating another method for generating a dataset of simulated electron microscope images provided in an embodiment of this application, as shown below. Figure 2 As shown, S140 may include S141-S144.
[0053] S141. Convert the third unit cell structure data into the first unit cell structure feature vector.
[0054] In one possible implementation, S141, converting the third cell structure data into the first cell structure feature vector, includes: steps one to two.
[0055] Step 1: Project the atomic coordinates in the third unit cell structure data onto a two-dimensional plane to obtain a two-dimensional atomic distribution map.
[0056] Step 2: Perform vector flattening and data standardization on the two-dimensional atomic distribution map to obtain the first unit cell structure feature vector.
[0057] For example, the atoms in the third unit cell are projected onto the xy plane to obtain a two-dimensional atomic distribution map. Based on the two-dimensional atomic distribution map, a 50*50 two-dimensional atomic distribution map is generated. The two-dimensional atomic distribution map is normalized and flattened to obtain a one-dimensional vector, which includes 2500 elements.
[0058] For example, a 50x50 two-dimensional atomic distribution map includes 2500 cells. These 2500 cells can be normalized and flattened to obtain a one-dimensional vector. The one-dimensional vector includes 2500 elements, and each cell contains projections of different atoms. The first cell contains two atomic projections, and the computer records the value of the first cell as 010. The second cell contains three atomic projections, and the value of the second cell is recorded as 011. The first two cells are transformed into a vector n = [(001, 010), (010, 011)]. In (001, 010), 001 represents the first cell, and 010 represents the number of atomic projections contained in the first cell. In (010, 011), 010 represents the second cell, and 011 represents the number of atomic projections contained in the second cell.
[0059] S142. Dimensionally reduce the first unit cell structure feature vector to obtain the second unit cell structure feature vector.
[0060] In one possible implementation, the dimensionality reduction of the first unit cell structure eigenvectors is performed through normalization and principal component analysis (PCA). For example, a 2500-dimensional vector can be reduced to 500 dimensions through normalization and PCA.
[0061] Each dimension in the eigenvector of the first unit cell structure represents the number of atoms within a certain grid. Standardization addresses the issue of differences in the order of magnitude of atoms across different grids, ensuring that the influence of each grid dimension on subsequent analysis is fair and providing unbiased data for PCA.
[0062] PCA is based on standardized unbiased data. It can eliminate redundant dimensions that are repetitive and related, and compress high-dimensional vectors into low-dimensional ones. This reduces the computational burden of subsequent clustering algorithms, such as finding the optimal K value, while retaining the core information that distinguishes different cell structures (such as defect locations and atomic arrangement patterns).
[0063] S143. Based on the clustering algorithm, classify the structurally similar cell structure data in the second cell structure feature vector into the same cluster.
[0064] S144. Select the fourth cell structure data that best represents the structural characteristics of each cluster to obtain the fourth dataset.
[0065] The beneficial effects of clustering have been described in S140 and will not be repeated here.
[0066] S150. Based on preset electron microscope parameters, perform imaging simulation on multiple fourth unit cell structure data to generate multiple first simulation images.
[0067] In one possible implementation, S150 includes: S151-S152.
[0068] S151. Divide the fourth dataset into multiple sub-datasets, each sub-dataset including at least one fourth cell structure data.
[0069] S152. Distribute multiple subsets of data to multiple parallel computing units, so that the multiple parallel units can perform imaging simulation on the fourth unit cell structure data in the subsets according to preset electron microscope parameters, and generate a first simulated image.
[0070] For example, the fourth dataset is divided into four subsets, each containing several fourth cell structure data. These four subsets are then assigned to four parallel computing units. A single CIF file (the smallest unit of computation) cannot be processed in parallel by multiple GPUs, but a single GPU can process multiple CIF files simultaneously, even if they come from different subsets.
[0071] Each parallel computing unit can perform imaging simulation on the fourth unit cell structure data in the subset of data according to preset electron microscope parameters, generating a first simulated image. This process is multi-GPU simulated imaging; this example only illustrates four GPUs. Compared with existing single-machine simulated imaging, this process, with multiple GPUs performing imaging simulation simultaneously, greatly accelerates data processing efficiency.
[0072] S160. Associate and store the first simulated image with the corresponding first unit cell structure data to obtain the target dataset.
[0073] The association between the first simulated image and its corresponding first unit cell structure data means that each first unit cell structure data corresponds to one first simulated image. The first simulated image, the first unit cell structure data, and the correspondence between them are stored as a target dataset. This target dataset can be used to train a deep learning model, enabling the model to automatically generate a first simulated image based on the provided first unit cell structure data in the target material. This avoids the high cost associated with using a double-spherical-aberration-corrected electron microscope. Furthermore, the target dataset processed in this application has a low mathematical model complexity, meeting the needs of large-scale material imaging simulations.
[0074] In one possible implementation, after S150, the method further includes: S151-S152.
[0075] S151. Perform constrained geometric transformations on multiple first simulated images to obtain multiple second simulated images.
[0076] The core objective of S151 is to expand the target dataset and improve the generalization ability of subsequent deep learning model training by changing the first simulated image while maintaining crystallographic authenticity.
[0077] In one possible implementation, S151 includes: step three and step four.
[0078] Step 3: Extract the crystallographic features of the cell structure data corresponding to the first simulated image. The crystallographic features include rotational symmetry order and mirror symmetry axis position.
[0079] Step 4: Based on the rotational symmetry order, rotate the first simulated image to obtain the second simulated image.
[0080] For example, the first simulated image is rotated at angular intervals of 360° / n, where n is an integer value of the rotational symmetry order.
[0081] And / or, based on the position of the mirror symmetry axis, the first simulated image is flipped to obtain the second simulated image.
[0082] For example, the image is scaled in the range of 1.0–1.2 times to make the relative error of the lattice spacing ≤5%.
[0083] S152. Perform Gaussian blur processing on multiple second simulated graphics to obtain multiple third simulated images.
[0084] S152. The purpose is to smooth the obtained second simulated image.
[0085] S151-S152 are further optimization processes for the first simulated image, resulting in a second simulated image that more closely resembles the image under a microscope.
[0086] Based on S151-S152, S160, the first simulated image is associated with and stored with the corresponding first unit cell structure data to obtain the target dataset, including: S161.
[0087] S161. Associate and store the third simulated image with the corresponding first unit cell structure data to obtain the target dataset.
[0088] In some possible implementations, the obtained third simulated image may also be subjected to crystallographic preservation enhancement processing.
[0089] For example, adjusting the image brightness within the range of 0.8x to 1.2x while preserving the contrast features between atoms and the background yields a crystallographically preserved and enhanced image.
[0090] In one possible implementation, after S160, where the first simulated image is associated with and stored with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: S170 S170. Based on a preset ratio, the target dataset is divided into a training set, a validation set, and a test set.
[0091] The training set should comprise the largest portion of the dataset. Its core purpose is to allow the model to understand the correlation between the simulated image and the unit cell structure data through a large number of samples, serving as crucial data for model parameter optimization. The validation set is used solely during training to evaluate the model's generalization ability in real time, helping to identify overfitting and guide parameter adjustments. The test set is used to verify the model's final performance and determine whether it meets the needs of practical applications.
[0092] After S160, where the first simulated image is associated with and stored with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: S181-S182.
[0093] S181. Identify the material dose in the first unit cell structure data and generate a label based on the material dose.
[0094] Here, material dose refers to the electron beam dose parameter used when generating the first simulated image. The label is used to visually indicate the electron beam dose attribute of the first simulated image.
[0095] S182. Based on the correspondence between the first simulated image and the first unit cell structure data, associate the first simulated image with the label.
[0096] The purpose of S181-S182 is to convert material dose parameters into labels and associate them with images, making the originally implicit physical parameters searchable structured information. This not only improves the management efficiency of the dataset, but also provides data support for subsequent in-depth analysis of the impact of dose on imaging quality and the adaptability of the model to different doses.
[0097] This application first periodically expands multiple first unit cell structure data to obtain a second dataset, ensuring the integrity of the non-orthogonal unit cell structure data (second unit cell structure data). Then, through crystallographic coordinate system transformation techniques, the difficult-to-process non-orthogonal unit cell structure data is transformed into more easily solvable orthogonal unit cell structure data (third unit cell structure data), thereby significantly improving computational efficiency. Furthermore, this application reduces the complexity of the mathematical model by selecting representative fourth unit cell structure data from the third unit cell structure data instead of simulating all units, thus reducing redundant CPU consumption in coordinate calculations and improving the execution efficiency of the solution algorithm, meeting the needs of large-scale materials imaging simulation.
[0098] Figure 3 This is a schematic diagram of a data set generation device for simulating electron microscope images provided in an embodiment of this application, as shown below. Figure 3 As shown, the dataset generation device 300 for simulating electron microscope images includes: a screening module 310, an expansion module 320, a coordinate transformation module 330, an imaging simulation module 340, and a storage module 350.
[0099] The filtering module 310 is used to perform multidimensional geometric filtering on the original crystal database to filter out the first dataset, which includes multiple first cell structure data that conform to the definition of two-dimensional materials.
[0100] The extension module 320 is used to periodically extend the first unit cell structure data to obtain second unit cell structure data. Multiple second unit cell structure data constitute a second dataset. The second dataset includes multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cell structures.
[0101] The coordinate transformation module 330 is used to map the second cell structure data describing the non-orthogonal unit cell in the second dataset to an orthogonal coordinate system through crystallographic coordinate system transformation technology to obtain orthogonal unit cell structure data, and use the orthogonal unit cell structure data as the third unit cell structure data; the third unit cell structure data and the second unit cell structure data of the orthogonal unit cell structure constitute the third dataset.
[0102] The filtering module 310 is used to identify similar structure clusters of orthogonal cell data in the third dataset, and to filter representative orthogonal cell data from each similar structure cluster as fourth cell structure data to obtain the fourth dataset.
[0103] The imaging simulation module 340 is used to perform imaging simulation on each fourth unit cell structure data according to preset electron microscope parameters to generate a first simulated image.
[0104] The storage module 350 is used to associate and store the first simulated image with its corresponding first unit cell structure data to obtain the target dataset.
[0105] It should be noted that the above embodiments are device embodiments corresponding to the method for generating datasets of simulated electron microscope images. Therefore, the device embodiments have all the beneficial effects of the aforementioned method embodiments, and will not be repeated here.
[0106] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A method for generating a dataset simulating electron microscope images, characterized in that, include: A multidimensional geometric screening was performed on the original crystal database to select the first dataset, which includes multiple first cell structure data that conform to the definition of two-dimensional materials. The first unit cell structure data is periodically expanded to obtain the second unit cell structure data. Multiple second unit cell structure data constitute a second dataset. The second dataset includes multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cells. By using crystallographic coordinate system transformation technology, the second cell structure data describing the non-orthogonal unit cell in the second dataset is mapped to an orthogonal coordinate system to obtain orthogonal unit cell structure data, and this orthogonal unit cell structure data is used as the third unit cell structure data; The third unit cell structure data and the second unit cell structure data of the orthogonal unit cell structure constitute the third dataset; Identify similar structure clusters of orthogonal cell data in the third dataset, and select representative orthogonal cell data from each similar structure cluster as fourth cell structure data to obtain the fourth dataset; Based on preset electron microscope parameters, imaging simulation is performed on the data of each fourth unit cell structure to generate a first simulated image; The first simulated image is associated with and stored with its corresponding first unit cell structure data to obtain the target dataset.
2. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, The step involves imaging and simulating the multiple fourth unit cell structure data according to preset electron microscope parameters to generate multiple first simulated images, including: The fourth dataset is divided into multiple sub-datasets, each of which includes at least one fourth cell structure data set. The multiple subsets of data are distributed to multiple parallel computing units. Each parallel unit performs imaging simulation on the fourth unit cell structure data in its corresponding subset of data according to the preset electron microscope parameters, and generates the first simulated image.
3. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, The process involves identifying similar structure clusters of orthogonal unit cell data in the third dataset, and selecting representative orthogonal unit cell data from each similar structure cluster as fourth unit cell structure data to obtain the fourth dataset, which includes: The orthogonal unit cell data in the third dataset are transformed into the first unit cell structure feature vector; The first unit cell structure feature vector is reduced in dimension to obtain the second unit cell structure feature vector. Based on the clustering algorithm, the cell structure data with similar structures in the second cell structure feature vector are grouped into the same cluster; In each cluster, the fourth cell structure data that best represents the structural characteristics of the cluster is selected to obtain the fourth dataset.
4. The method for generating a dataset of simulated electron microscope images according to claim 3, characterized in that, The step of converting the orthogonal unit cell data in the third dataset into a first unit cell structure feature vector includes: Projecting the atomic coordinates of the orthogonal unit cell data in the third dataset onto a two-dimensional plane yields a two-dimensional atomic distribution map; The two-dimensional atomic distribution map is flattened and normalized to obtain the first unit cell structure feature vector.
5. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, After performing imaging simulation on the plurality of fourth unit cell structure data according to preset electron microscope parameters to generate a plurality of first simulated images, the method further includes: A constrained geometric transformation is performed on multiple first simulated images to obtain multiple second simulated images; Gaussian blurring is applied to the plurality of second simulated images to obtain a plurality of third simulated images; The step of associating and storing the first simulated image with the corresponding first unit cell structure data to obtain the target dataset includes: The third simulated image is associated with and stored with the corresponding first unit cell structure data to obtain the target dataset.
6. The method for generating a dataset of simulated electron microscope images according to claim 5, characterized in that, The process of performing constrained geometric transformations on multiple first simulated images to obtain multiple second simulated images includes: Extract crystallographic features from the orthogonal unit cell structure data corresponding to the first simulated image, the crystallographic features including rotational symmetry order and mirror symmetry axis position; Based on the rotational symmetry order, the first simulated image is rotated to obtain the second simulated image; And / or, based on the position of the mirror symmetry axis, the first simulated image is flipped to obtain the second simulated image.
7. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, The process of performing multidimensional geometric filtering on the original crystal database to select the first dataset includes: Based on thickness parameters, vacuum layer parameters, element types, and in-plane nearest neighbor continuity, the original crystal database is subjected to multidimensional geometric screening to select the first dataset.
8. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, After associating and storing the first simulated image with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: Based on a preset ratio, the target dataset is divided into a training set, a validation set, and a test set.
9. The method for generating a dataset of simulated electron microscope images according to claim 1, characterized in that, After associating and storing the first simulated image with the corresponding first unit cell structure data to obtain the target dataset, the method further includes: Identify the material dosage in the first unit cell structure data and generate a label based on the material dosage; Based on the correspondence between the first simulated image and the first unit cell structure data, the first simulated image is associated with the label.
10. A dataset generation device for simulating electron microscope images, characterized in that, include: The filtering module performs multidimensional geometric filtering on the original crystal database to filter out the first dataset, which includes multiple first cell structure data that conform to the definition of two-dimensional materials. An extension module periodically extends the first unit cell structure data to obtain second unit cell structure data. Multiple second unit cell structure data constitute a second dataset, which includes multiple second unit cell structure data describing non-orthogonal unit cells and second unit cell structure data describing orthogonal unit cells. The coordinate transformation module uses crystallographic coordinate system transformation technology to map the second cell structure data describing the non-orthogonal unit cell in the second dataset to an orthogonal coordinate system to obtain orthogonal unit cell structure data, and uses the orthogonal unit cell structure data as the third unit cell structure data. The third unit cell structure data and the second unit cell structure data of the orthogonal unit cell structure constitute the third dataset; The filtering module identifies similar structure clusters of orthogonal cell data in the third dataset, and filters representative orthogonal cell data from each similar structure cluster as fourth cell structure data to obtain the fourth dataset. An imaging simulation module is used to perform imaging simulation on each of the fourth unit cell structure data according to preset electron microscope parameters, and generate a first simulated image; The storage module is used to associate and store the first simulated image with the corresponding first unit cell structure data to obtain the target dataset.