A method for determining and retrieving crystal structure symmetry using electron diffraction

Abstract

The application relates to the field of crystal structure symmetry classification and retrieval, and specifically discloses a crystal structure symmetry determination and retrieval method using electron diffraction, which comprises the following steps: a multi-view selected area electron diffraction data input module is used to acquire a two-zone axial diffraction spectrum and perform pretreatment; a symmetry feature coding module is used to extract a single-view descriptor through a double-branch convolutional neural network, to obtain a material descriptor through view pooling fusion; a hierarchical symmetry classification module is used to realize hierarchical classification of crystal systems and space groups, and to guarantee the logic of the results through consistency constraints; a high-dimensional feature embedding and database construction module is used to standardize crystal structure files, to generate descriptors, and to construct a vector database; and a constrained structure retrieval module is used to combine the classification results and user constraints to filter candidate structures, and to complete retrieval through cosine similarity. The application solves the problems of poor generalization, insufficient multi-view feature fusion and inability to associate a crystal library in the prior art, and realizes full-process automation of crystal structure symmetry classification and retrieval.

Description

Technical Field

[0001] This invention relates to the field of crystal structure symmetry classification and retrieval technology, specifically to a method for determining and retrieving crystal structure symmetry using electron diffraction. Background Technology

[0002] Crystal symmetry, as the "structural DNA" of functional materials, determines their core functions such as tensor properties, optical selection rules, and topological states, making it a key research object in materials science and condensed matter physics. For example, the ferroelectricity of barium titanate, the piezoelectricity of gallium nitride (supporting high-efficiency blue LEDs and wide-bandgap power electronic devices), and the valley-selective optical transitions of monolayer transition metal chalcogenides all originate from specific crystal symmetries. In the band topology domain, rotational symmetry and non-degenerate symmetry can stabilize symmetry-protected band crossings, such as the three-dimensional Dirac nodes in Dirac half-metals. Therefore, determining the space group symmetry is not only the foundation for structural characterization but also a core prerequisite for predicting whether a material possesses spontaneous polarization, nonlinear optical response, valley degrees of freedom, or topologically protected electronic states.

[0003] Diffraction is a core technique for experimentally resolving crystal symmetry, transforming the abstract symmetry principles of crystallography into measurable physical signals. Among them, X-ray diffraction is suitable for bulk average analysis, electron backscattering diffraction is good at microscale surface texture and orientation mapping, and selected area electron diffraction (SEED) has become a key technique that traditional bulk methods cannot replace because it can detect local crystal structures in nanoscale, heterogeneous or multiphase systems. However, as a two-dimensional projection of the reciprocal lattice, the SEED pattern still faces serious bottlenecks in its analysis process: recovering the three-dimensional crystal structure from a single two-dimensional projection is essentially an ill-posed inverse problem, facing three core challenges: (1) the diffraction pattern is highly dependent on crystal orientation, and it is difficult to achieve reliable classification with arbitrary orientation input; (2) dynamic scattering will destroy diffraction intensity information, increasing the difficulty of symmetry feature extraction; (3) accurate positioning of space group symmetry requires the systematic extinction information of multi-zone axis diffraction. Therefore, traditional indexing methods based on manual annotation, hand-designed features or crystallographic assumptions are time-consuming, labor-intensive, and dependent on expert experience, and cannot match the massive data generation rate of modern high-throughput material discovery platforms.

[0004] Deep learning technology has provided a new path for the automation of diffraction pattern analysis. Existing research has developed a variety of deep learning-based analysis models, such as convolutional neural network classifiers based on electron backscatter diffraction, diffraction fingerprint convolutional networks, one-dimensional peak-shaped hierarchical models, and multi-view fusion frameworks [1-7], which are used to directly classify crystal systems or space groups from diffraction data.

[0005] However, existing research still has key shortcomings: (1) It can only identify fixed symmetry categories in the training set (such as 7 crystal systems or some space groups), and cannot generalize to arbitrary orientations, and cannot associate diffraction patterns with specific entries in large crystal structure databases; (2) It fails to effectively integrate the inherent multi-view characteristics of crystallography - the complementarity of diffraction information of different orientations is the key to solving symmetry ambiguity, but existing models are difficult to achieve effective integration of multi-view symmetry features, resulting in the inability to achieve robust and comprehensive space group identification from selected area electron diffraction patterns of arbitrary orientations.

[0006] In summary, existing technologies are insufficient to meet the demands of high-throughput materials research, necessitating a novel machine learning model with the following capabilities: (1) generalizability to arbitrary crystal orientations; (2) the ability to resolve fine-grained symmetry features to support complete space group determination; and (3) scalability to massive and continuously updated crystal structure databases. The development of such a model is of great significance for automating diffraction pattern analysis, breaking down the barriers between electron diffraction experiments and large-scale structure databases, and accelerating the process of high-throughput and autonomous materials discovery. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention provides a method for determining and retrieving crystal structure symmetry using electron diffraction. This method solves the problems of poor generalization, insufficient fusion of multi-view features, inability to associate with crystal libraries, inefficient analysis, and reliance on expert experience in existing technologies, thereby automating the entire process of crystal structure symmetry classification and retrieval.

[0008] To achieve the above objectives, this invention proposes a method for determining and retrieving crystal structure symmetry using electron diffraction. This method is implemented sequentially through a multi-view selected area electron diffraction data input module, a symmetry feature encoding module, a hierarchical symmetry classification module, a high-dimensional feature embedding and database construction module, and a constrained structure retrieval module, comprising the following steps: S1. Use the multi-view selected area electron diffraction data input module to obtain the simulated or experimental dual-axis selected area electron diffraction pattern, and perform standardized preprocessing operations such as grayscale normalization, background noise removal, and reflection point enhancement on the pattern. S2. The preprocessed dual-axis map is input into a dual-branch convolutional neural network encoder using a symmetric feature encoding module to extract a 1024-dimensional single-view descriptor. The dual-view descriptors are then fused using a view pooling strategy to obtain a material descriptor. S3, the hierarchical symmetry classification module realizes hierarchical classification of 7 crystal systems and 191 space groups based on the fused material descriptors, and ensures the logic of the classification results through a consistency constraint mechanism; S4, the high-dimensional feature embedding and database construction module performs symmetry standardization processing on the structure files in the crystal structure database, generates corresponding material descriptors, and constructs a vector database that stores the mapping relationship between the descriptors and crystal structure metadata; S5, the constrained structure retrieval module combines symmetry classification results with user-selectable constraints to filter candidate structures, and uses cosine similarity calculation to retrieve crystal structures and output the results.

[0009] Preferably, in S1, the simulated selected area electron diffraction pattern is generated by inputting the structure file of the crystal structure database into the multiple scattering simulator, and the experimental selected area electron diffraction pattern is the selected area electron diffraction pattern of any two non-collinear axes of the target material acquired by transmission electron microscopy.

[0010] Preferably, in S1, the grayscale normalization maps the spectral pixel values ​​to... In the interval, the background noise removal is achieved using Gaussian filtering, and the reflection point enhancement is achieved using threshold segmentation.

[0011] Preferably, in S2, the dual-branch convolutional neural network encoder adopts a shared weight architecture, and the network structures of the two branches are completely identical, each including 5 convolutional layers, 2 max pooling layers and 1 fully connected layer. The fully connected layer maps the extracted high-dimensional features into a fixed-dimensional single-view descriptor.

[0012] Preferably, in S2, the convolutional kernel size of the convolutional layer is 3×3, the stride is 1, and the padding is 1, and the pooling kernel of the max pooling layer is 2×2 with a stride of 2.

[0013] Preferably, in S2, the view pooling strategy is to take the maximum value of each feature channel of the dual view descriptors and fuse them in this way to obtain the final material descriptor.

[0014] Preferably, in S3, the hierarchical classification is achieved by connecting a cascaded classification head after the encoder. The cascaded classification head includes a crystal system classification branch and a space group classification branch, which output the probability distributions of the crystal system category and the space group category, respectively. The hierarchical classification adopts an end-to-end training method, and the total loss function is the weighted sum of the crystal system classification loss and the space group classification loss.

[0015] Preferably, in S3, the consistency constraint mechanism is as follows: if there is a logical conflict between the space group prediction result and the crystal system classification result, the feature extraction weights of the encoder are adjusted retrospectively, and the crystal system classification result is re-optimized.

[0016] Preferably, in S4, the symmetry normalization process includes unifying the lattice parameter format and unifying the atomic coordinate sorting. The simulated dual-axis map corresponding to the normalized crystal structure file is input into the trained encoder to generate a material descriptor for each crystal structure.

[0017] Preferably, in S5, the user-selectable constraints include one or more of chemical composition, element types, and lattice parameter ranges; the candidate structures are sorted in descending order of cosine similarity, and the output search results include crystal structure metadata and similarity scores.

[0018] Therefore, this invention proposes a method for determining and retrieving crystal structure symmetry using electron diffraction, which has the following advantages: (1) The classification and retrieval accuracy has been significantly improved. It can complete the high-precision determination of 7 crystal systems and 191 space groups. The accuracy of crystal system classification and space group prediction is much higher than that of the existing single-view model. It can also maintain a high recall rate in the structural retrieval of massive materials, effectively solving the industry pain points of orientation dependence and symmetry ambiguity. (2) The analysis efficiency is improved exponentially. The fully automated process only takes tens of seconds, which is more than 100 times more efficient than the traditional manual analysis method. It supports batch processing of diffraction patterns and can adapt to the massive data generation rate of high-throughput material experiments, freeing it from dependence on the experience of senior crystallography experts. (3) It has excellent generalization and anti-interference capabilities and is applicable to a variety of material systems such as metal oxides, semiconductors, and topological materials. It can still stably resolve experimental spectra with noise and missing reflection points. At the same time, it breaks down the barriers between electron diffraction experiments and crystal structure databases, providing rapid detection methods for material research and development, and empowering industrial applications in semiconductors, new energy and other fields.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of a method for determining and retrieving crystal structure symmetry using electron diffraction according to the present invention. Detailed Implementation

[0021] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, the present invention provides a method for determining and retrieving crystal structure symmetry using electron diffraction, which is implemented sequentially through a multi-view selected area electron diffraction data input module, a symmetry feature encoding module, a hierarchical symmetry classification module, a high-dimensional feature embedding and database construction module, and a constrained structure retrieval module, including the following steps: S1. Use the multi-view selected area electron diffraction data input module to obtain the simulated or experimental dual-axis selected area electron diffraction pattern, and perform standardized preprocessing operations such as grayscale normalization, background noise removal, and reflection point enhancement on the pattern. The simulated selected area electron diffraction pattern is generated by inputting the structure file of the crystal structure database into the multiple scattering simulator. The selected area electron diffraction pattern of the experiment is the selected area electron diffraction pattern of any two non-collinear axes of the target material acquired by transmission electron microscopy.

[0024] Gray-level normalization maps the pixel values ​​of the image to... Background noise removal is achieved using Gaussian filtering, while reflection point enhancement is achieved through threshold segmentation.

[0025] S2. The preprocessed dual-axis map is input into a dual-branch convolutional neural network encoder using a symmetric feature encoding module to extract a 1024-dimensional single-view descriptor. The dual-view descriptors are then fused using a view pooling strategy to obtain a material descriptor. The dual-branch convolutional neural network encoder adopts a shared weight architecture, with the two branches having completely identical network structures, each including 5 convolutional layers, 2 max pooling layers, and 1 fully connected layer. The fully connected layer maps the extracted high-dimensional features into a fixed-dimensional single-view descriptor.

[0026] The convolutional layer has a kernel size of 3×3, a stride of 1, and padding of 1. The max pooling layer has a kernel size of 2×2 and a stride of 2.

[0027] The view pooling strategy involves taking the maximum value of each feature channel of the two view descriptors and fusing them to obtain the final material descriptor.

[0028] S3, the hierarchical symmetry classification module realizes hierarchical classification of 7 crystal systems and 191 space groups based on the fused material descriptors, and ensures the logic of the classification results through a consistency constraint mechanism; Hierarchical classification is achieved by connecting a cascaded classification head after the encoder. The cascaded classification head includes a crystal system classification branch and a space group classification branch, which output the probability distributions of the crystal system category and the space group category, respectively. The hierarchical classification adopts an end-to-end training method, and the total loss function is the weighted sum of the crystal system classification loss and the space group classification loss.

[0029] The consistency constraint mechanism is as follows: if there is a logical conflict between the space group prediction result and the crystal system classification result, the feature extraction weight of the encoder is adjusted backtracking, and the crystal system classification result is re-optimized.

[0030] S4, the high-dimensional feature embedding and database construction module performs symmetry standardization processing on the structure files in the crystal structure database, generates corresponding material descriptors, and constructs a vector database that stores the mapping relationship between the descriptors and crystal structure metadata; Symmetry normalization includes unifying the lattice parameter format and the atomic coordinate sorting. The simulated dual-axis plot corresponding to the normalized crystal structure file is input into the trained encoder to generate a material descriptor for each crystal structure.

[0031] S5, the constrained structure retrieval module combines symmetry classification results with user-selectable constraints to filter candidate structures, and uses cosine similarity calculation to retrieve crystal structures and output the results.

[0032] User-selectable constraints include one or more of the following: chemical composition, element types, and lattice parameter range; candidate structures are sorted in descending order of cosine similarity, and the output search results include crystal structure metadata and similarity scores.

[0033] Example 1 This invention uses tungsten oxide and iron oxide as experimental subjects to verify the applicability of the framework in scenarios such as distinguishing polymorphs, resolving stoichiometric ambiguities, and generalizing experimental spectra. The specific implementation process is as follows: The experiment first prepared high-purity samples using the sol-gel method and hydrothermal synthesis method. Selected area electron diffraction (SAED) patterns of any two non-collinear axes of each sample were acquired using a JEOL JEM-2100F transmission electron microscope (200kV accelerating voltage, 500nm selected area aperture). After grayscale normalization, Gaussian filtering and reflection point enhancement preprocessing, the patterns were input into the deployed pre-trained model.

[0034] The model extracts 1024-dimensional single-view descriptors through a dual-branch encoder, fuses them into the final material descriptor through view pooling, and then completes the crystal system and space group determination through a hierarchical classification head (combined with bidirectional consistency constraints). Finally, it performs the determination based on cosine similarity in a dedicated vector database (including Word documents). 3 Structure retrieval can be performed on polymorphs and common phases in the Fe-O system.

[0035] Experimental results show that this scheme successfully distinguishes... and , and The model achieves 100% accuracy in crystal system classification and 87.5% accuracy in space group prediction, even with noise and missing reflection points in experimental spectra. The entire single-sample processing time is only 42 seconds, significantly improving efficiency compared to traditional manual methods. The model can generalize to real-world scenarios without experimental data training, and prediction ambiguities are limited to adjacent crystallographic space groups, aligning with expert judgment logic. This fully validates the advantages of the technical solution in accuracy, robustness, and efficiency, and it can be directly applied to routine diffraction pattern analysis and high-throughput materials research in the laboratory.

[0036] Example 2 This embodiment uses over two thousand crystal structures covering multiple crystal systems and space groups as test objects. Simultaneously, it selects actual samples from various material systems, including metal oxides, semiconductors, and topological materials, to conduct experiments, verifying the practical application effect of the crystal structure symmetry determination and retrieval method using electron diffraction of this invention. The specific implementation steps are as follows: Experimental Sample and Data Preparation: Metal oxides such as tungsten oxide and iron oxide, as well as actual samples such as semiconductors and topological materials, were selected and high-purity samples were prepared by sol-gel method and hydrothermal synthesis method. At the same time, more than 2,000 crystal structure files were retrieved from the crystal structure database, covering all crystal systems and 191 space groups. The corresponding simulated double-axis selected area electron diffraction patterns were generated using a multiple scattering simulator. In addition, crystal structure data of about 13,000 materials were collected for retrieval and performance testing.

[0037] Diffraction pattern acquisition and preprocessing: Selected area electron diffraction patterns of any two non-collinear axes were acquired for each actual sample using a transmission electron microscope. Standardized preprocessing was performed on both simulated and experimental patterns, including mapping pixel values ​​to... The process includes grayscale normalization of the interval, background noise removal using Gaussian filtering, and enhancement of reflection points through threshold segmentation.

[0038] Model training and deployment: Construct an overall framework including a multi-view selected area electron diffraction data input module, a symmetry feature encoding module, a hierarchical symmetry classification module, a high-dimensional feature embedding and database construction module, and a constrained structure retrieval module. The symmetry feature encoding module adopts a shared weight dual-branch convolutional neural network encoder. After end-to-end training and model optimization, it is deployed in practice by inputting the preprocessed diffraction pattern into the trained model.

[0039] Crystal structure symmetry classification and retrieval: The model extracts single-view descriptors through a dual-branch encoder, fuses them into material descriptors through a view pooling strategy, and then completes hierarchical classification of crystal systems and space groups through a hierarchical classification head. At the same time, a bidirectional consistency constraint mechanism is combined to ensure the classification logic. Based on the classification results and user constraints such as chemical composition, cosine similarity comparison is performed in the constructed crystal structure vector database to achieve accurate retrieval of crystal structures.

[0040] Experimental Results and Verification: In this embodiment, on a test set of more than 2,000 crystal structures, the crystal system classification accuracy reached 95.09% and the space group prediction accuracy reached 80.12%, which is a significant improvement over the existing single-view deep learning model and effectively solves the problems of orientation dependence and ambiguity in traditional methods; in the chemical formula constrained structure retrieval of about 13,000 materials, the recall rate reached 93.9%.

[0041] From inputting the dual-axis electron diffraction pattern to completing the classification and retrieval, the entire process takes only 30-50 seconds, which is more than 100 times more efficient than the traditional manual analysis method. Moreover, the model supports batch processing of diffraction patterns and can be adapted to the massive data generation rate of hundreds to thousands of patterns per day by autonomous experimental robots.

[0042] In experimental tests on various material systems such as metal oxides, semiconductors, and topological materials, the average accuracy of crystal system classification and structure retrieval remains above 88%. Even if there are problems such as noise or missing reflection points in the experimental spectra, stable analysis can still be achieved, breaking through the limitation of existing models that are only applicable to specific material types.

[0043] This embodiment fully verifies that the method of the present invention has the advantages of high precision, high efficiency, strong anti-interference and generalization in crystal structure symmetry classification and retrieval tasks. It can be directly applied to routine diffraction pattern analysis and high-throughput material research in the laboratory, reducing the dependence on senior crystallography experts, reducing the human cost of material structure characterization, and providing a rapid detection method for material quality control in fields such as semiconductors, new energy, and aerospace, thus accelerating the research and development process of functional materials.

[0044] Therefore, this invention provides a method for determining and retrieving crystal structure symmetry using electron diffraction. This method integrates complementary information from selected area electron diffraction patterns of two-zone axes and automates the entire process from symmetry classification to structure retrieval through five core modules. First, the simulated or experimental diffraction patterns of two-zone axes are standardized and preprocessed. Then, symmetry features are extracted and fused using a shared-weight dual-branch convolutional neural network. Hierarchical classification combined with consistency constraints achieves accurate determination of the crystal system to the space group. Simultaneously, a standardized crystal structure vector database is constructed. Finally, combining the classification results with user constraints, a similarity algorithm is used to achieve accurate crystal structure retrieval. This method effectively solves the problems of poor generalization, insufficient multi-view feature fusion, and inability to associate with crystal structure databases in existing technologies. It overcomes the shortcomings of traditional analytical methods, such as low efficiency and reliance on expert experience. It combines high detection accuracy, strong anti-interference and generalization capabilities, adapts to the massive data processing needs of high-throughput materials research, and can also reduce the cost of materials characterization, providing a fast and reliable detection method for materials research and development in multiple fields.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for determining and retrieving crystal structure symmetry using electron diffraction, characterized in that, The process is implemented sequentially through a multi-view selected area electron diffraction data input module, a symmetry feature encoding module, a hierarchical symmetry classification module, a high-dimensional feature embedding and database construction module, and a constrained structure retrieval module, including the following steps: S1. Use the multi-view selected area electron diffraction data input module to obtain the simulated or experimental dual-axis selected area electron diffraction pattern, and perform standardized preprocessing operations such as grayscale normalization, background noise removal, and reflection point enhancement on the pattern. S2. The preprocessed dual-axis map is input into a dual-branch convolutional neural network encoder using a symmetric feature encoding module to extract a 1024-dimensional single-view descriptor. The dual-view descriptors are then fused using a view pooling strategy to obtain a material descriptor. S3, the hierarchical symmetry classification module realizes hierarchical classification of 7 crystal systems and 191 space groups based on the fused material descriptors, and ensures the logic of the classification results through a consistency constraint mechanism; S4, the high-dimensional feature embedding and database construction module performs symmetry standardization processing on the structure files in the crystal structure database, generates corresponding material descriptors, and constructs a vector database that stores the mapping relationship between the descriptors and crystal structure metadata; S5, the constrained structure retrieval module combines symmetry classification results with user-selectable constraints to filter candidate structures, and uses cosine similarity calculation to retrieve crystal structures and output the results.

2. The method according to claim 1, characterized in that, In S1, the simulated selected area electron diffraction pattern is generated by inputting the structure file of the crystal structure database into the multiple scattering simulator, and the experimental selected area electron diffraction pattern is the selected area electron diffraction pattern of any two non-collinear axes of the target material acquired by transmission electron microscopy.

3. The method according to claim 1, characterized in that, In S1, the grayscale normalization maps the spectral pixel values ​​to... In the interval, the background noise removal is achieved using Gaussian filtering, and the reflection point enhancement is achieved using threshold segmentation.

4. The method according to claim 1, characterized in that, In S2, the dual-branch convolutional neural network encoder adopts a shared weight architecture. The network structures of the two branches are completely identical, each including 5 convolutional layers, 2 max pooling layers and 1 fully connected layer. The fully connected layer maps the extracted high-dimensional features into a fixed-dimensional single-view descriptor.

5. The method according to claim 4, characterized in that, In S2, the convolutional kernel size of the convolutional layer is 3×3, the stride is 1, and the padding is 1. The pooling kernel of the max pooling layer is 2×2 and the stride is 2.

6. The method according to claim 1, characterized in that, In S2, the view pooling strategy is to take the maximum value of each feature channel of the dual view descriptors and fuse them in this way to obtain the final material descriptor.

7. The method according to claim 1, characterized in that, In S3, the hierarchical classification is achieved by connecting a cascaded classification head after the encoder. The cascaded classification head includes a crystal system classification branch and a space group classification branch, which output the probability distributions of the crystal system category and the space group category, respectively. The hierarchical classification adopts an end-to-end training method, and the total loss function is the weighted sum of the crystal system classification loss and the space group classification loss.

8. The method according to claim 1, characterized in that, In S3, the consistency constraint mechanism is as follows: if there is a logical conflict between the space group prediction result and the crystal system classification result, the feature extraction weight of the encoder is adjusted backtrackingly, and the crystal system classification result is re-optimized.

9. The method according to claim 1, characterized in that, In S4, the symmetry normalization process includes unifying the lattice parameter format and unifying the atomic coordinate sorting. The simulated dual-axis map corresponding to the normalized crystal structure file is input into the trained encoder to generate a material descriptor for each crystal structure.

10. The method according to claim 1, characterized in that, In S5, the user-selectable constraints include one or more of the following: chemical composition, element type, and lattice parameter range; candidate structures are sorted in descending order of cosine similarity, and the output search results include crystal structure metadata and similarity score.

Images