A method and device for detecting crystal structure defects of a sapphire substrate

Abstract

This application provides a method and apparatus for detecting crystal structure defects in sapphire substrates, belonging to the field of material defect detection technology. The method includes: acquiring historical process information of the sapphire substrate and querying associated typical defect types and distribution characteristics from a defect knowledge base as expected defect information; performing a full-field scan of the sapphire substrate using a first detection technique to obtain a first distribution image; performing image analysis on the first distribution image based on the expected defect information to identify one or more candidate defect regions and performing local scanning using a second detection technique with a spatial resolution or spectral resolution greater than the first detection technique to obtain a second distribution image; spatially registering and fusing the first distribution image and the second distribution image of the candidate defect regions to generate multimodal fusion feature data and inputting it into a defect classification and quantification model to obtain defect information of the candidate defect regions; and finally generating a sapphire substrate defect detection report and quality control judgment results.

Description

Technical Field

[0001] This application relates to the technical field of material defect detection, and in particular to a method and apparatus for detecting crystal structure defects in sapphire substrates. Background Technology

[0002] In the detection of crystal structure defects in sapphire substrates, existing methods include two main approaches: one is direct observation of the sapphire substrate using equipment such as optical microscopes, relying on the experience of the inspectors to determine the presence and nature of defects; the other is a single-technology scanning method, which uses specific detection techniques to scan the entire sapphire substrate, acquiring corresponding distribution images and performing defect analysis. Among these existing methods, the direct observation method heavily relies on the inspectors' experience, is highly subjective, and cannot guarantee the accuracy of the results. The single-technology scanning method lacks full utilization of historical process information, cannot predict the types and distribution characteristics of defects in advance, leading to the omission of some potential defects during the inspection process. Furthermore, it cannot accurately classify and quantify complex defect situations, failing to meet the current industry's needs for sapphire substrate quality inspection.

[0003] Therefore, there is an urgent need for a method and device for detecting crystal structure defects in sapphire substrates. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method and apparatus for detecting crystal structure defects in sapphire substrates.

[0005] A first aspect of this application provides a method for detecting crystal structure defects in a sapphire substrate, comprising: Obtain the process history information of the sapphire substrate to be tested, and query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information; The sapphire substrate is scanned in its entirety using a first detection technique to obtain a first distribution image; Based on the expected defect information, image analysis is performed on the first distribution image to identify one or more candidate defect regions; For the candidate defect region, a second detection technique with a spatial resolution or spectral resolution greater than the first detection technique is used for local scanning to obtain a second distribution image; wherein, the type of the second detection technique is selected based on the defect type in the expected defect information; Spatial registration and data fusion are performed on the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region; The multimodal fusion feature data is input into the trained defect classification and quantization model to obtain the defect information of the candidate defect region; Based on the defect information of all the candidate defect regions, a defect detection report and quality control judgment result for the sapphire substrate are generated.

[0006] A second aspect of this application provides a crystal structure defect detection device for sapphire substrates, comprising: The expected defect module is used to obtain the process history information of the sapphire substrate to be tested, and to query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information. The first detection module is used to perform a full-field scan of the sapphire substrate using a first detection technology to obtain a first distribution image; The defect identification module is used to perform image analysis on the first distribution image based on the expected defect information to identify one or more candidate defect regions; The second detection module is used to perform local scanning of the candidate defect region using a second detection technique with a spatial resolution or spectral resolution greater than that of the first detection technique to obtain a second distribution image; wherein, the type of the second detection technique is selected based on the defect type in the expected defect information; The registration and fusion module is used to spatially register and fuse the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region. The defect classification module is used to input the multimodal fusion feature data into the trained defect classification and quantization model to obtain the defect information of the candidate defect region; The report results module is used to generate a defect detection report and quality control judgment results for the sapphire substrate based on the defect information of all the candidate defect regions.

[0007] The beneficial effects of the crystal structure defect detection method and apparatus for sapphire substrate provided in this application are as follows: This application obtains expected defect information based on process history information, scans distribution images according to different detection technologies, generates multimodal fusion feature data through image analysis, spatial registration and data fusion, obtains defect information using a trained defect classification and quantification model, and finally generates a detection report and quality control judgment results, thereby achieving accurate detection and evaluation of crystal structure defects in sapphire substrates. Attached Figure Description

[0008] Figure 1 A schematic flowchart of a method for detecting crystal structure defects in a sapphire substrate provided in an embodiment of this application; Figure 2 This is a structural block diagram of a crystal structure defect detection device for sapphire substrates provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application.

[0009] The attached diagram is labeled as follows: 20. Crystal structure defect detection device for sapphire substrates; 21. Expected defect module; 22. First detection module; 23. Defect identification module; 24. Second detection module; 25. Registration and fusion module; 26. Defect classification module; 27. Result reporting module; 300. Electronic device; 301. Processor; 302. Input device; 303. Output device; 304. Memory; 305. Communication bus. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0012] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a method for detecting crystal structure defects in a sapphire substrate according to an embodiment of this application. The method includes: S101: Obtain the process history information of the sapphire substrate to be tested, and query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information.

[0013] In this embodiment, the sapphire substrate to be tested is aluminum oxide (alumina) for detecting crystal structure defects. Single-crystal substrates are the base materials for devices such as LED epitaxy and power semiconductors. It is necessary to clearly define their basic specifications (e.g., dimensions, crystal orientation, thickness) and production stages (e.g., post-growth, post-polishing, post-cleaning). Process history information is a complete record of the production process parameters and procedures for the sapphire substrate from crystal growth to the finished product. This information includes the crystal growth method, thermal history, and surface processing parameters. It may also include auxiliary information such as the growth furnace model, processing equipment, and production batch, serving as a basis for predicting defects.

[0014] In this embodiment, the pre-built defect knowledge base is a pre-constructed and updatable database based on a large amount of experimental data, engineering experience, and process rules in the field of sapphire substrate inspection. It stores the correlation between process condition combinations, typical defect types, and defect distribution characteristics, and also includes auxiliary information such as defect size and generation mechanism. This serves as the basis for matching defects with process history information. The associated typical defect types are the types of crystal structure defects generated under those process conditions, obtained by matching the process history information of the sapphire substrate to be inspected within the defect knowledge base. Common types in sapphire substrates include dislocations, stacking faults, stress concentration regions, vacancy clusters, inclusion defects, and surface microcracks. The distribution characteristics are the spatial distribution patterns of the matched typical defects within the sapphire substrate crystal. These characteristics are used for defect prediction and include distribution location (e.g., edge concentration, center concentration, global distribution), distribution shape (e.g., linear, point, planar), and size range (e.g., 1-10 μm point defects, 50-200 μm edge defects). The expected defect information is the final judgment information related to defects in the sapphire substrate to be tested, which is determined by matching the process history information with the defect knowledge base. It consists of 1-N typical defect types and their corresponding distribution characteristics, and serves as the guiding basis for subsequent detection parameter configuration and defect area identification.

[0015] S102: The first detection technology is used to perform a full-field scan of the sapphire substrate to obtain the first distribution image.

[0016] In this embodiment, the first detection technology is a fundamental detection technology designed for rapid full-field screening of sapphire substrates. Its characteristics include high detection efficiency, comprehensive coverage, and spatial / spectral resolution that meets the preliminary identification requirements for defect scale in the expected defect information. It does not need to achieve extreme precision and serves as a pre-processing technology in the hierarchical detection strategy formed with the subsequent local high-resolution second detection technology. It is a commonly used, low-cost, and high-efficiency technology for sapphire substrate defect detection. The sapphire substrate to be inspected is a finished product (after polishing / grinding) with acquired process history information, and needs to be fixed on the sample stage of the detection equipment for scanning.

[0017] In this embodiment, full-field scanning is a comprehensive, blind-spot-free scan covering the physical boundaries of the sapphire substrate to be inspected. During the scanning process, the detection unit (e.g., laser head, lens) of the detection device moves relative to the substrate, ensuring that every position of the effective detection surface (epitaxy / processed surface) of the substrate is scanned and covered, with no missed areas. The first distribution image is a visualized image data converted from the physical signals (e.g., scattering intensity, optical density, temperature) of each position on the substrate after full-field scanning using the first detection technology. The pixel coordinates of the image correspond one-to-one with the physical spatial coordinates of the substrate, and the pixel value represents the detection signal intensity at the corresponding position. Due to signal differences, defect areas and normal crystal areas exhibit a clear distinction in grayscale / color / brightness in the image, serving as the basis for subsequent identification of candidate defect areas.

[0018] S103: Based on the expected defect information, perform image analysis on the first distribution image to identify one or more candidate defect regions.

[0019] In this embodiment, the expected defect information is the final determined predictive information, including the typical defect types that are likely to exist in the sapphire substrate to be tested (such as dislocations, stress concentration areas, stacking faults, etc.) and their corresponding distribution characteristics (distribution location, morphology, and scale range). It serves as the guiding basis for this image analysis and defect identification, clarifying the focus of the analysis, avoiding meaningless blind analysis of the entire area, and improving the efficiency and accuracy of identification. The first distribution image is the visualized image data obtained by full-field scanning using the first detection technology. The pixel coordinates correspond one-to-one with the physical spatial coordinates of the sapphire substrate, and the pixel value represents the detection signal intensity at the corresponding position on the substrate, such as laser scattering intensity, metallographic grayscale, and infrared temperature. The defect area and the normal crystal area are clearly distinguished due to signal differences, and it is the basic data carrier for this image analysis and candidate defect identification.

[0020] In this embodiment, image analysis is a specialized data processing procedure performed on the first distribution image. Based on the expected defect information, algorithms such as image quality optimization, feature extraction, and region segmentation are used to analyze the grayscale, texture, and signal intensity features of the first distribution image to distinguish between normal crystal regions and suspected defect regions. Essentially, this is a process of filtering suspected defects and excluding normal regions, providing support for subsequent candidate defect region identification. Candidate defect regions are local areas identified from the first distribution image through image analysis that are suspected of containing crystal structure defects. They are not confirmed defect regions and require further verification in subsequent steps. The characteristic of these regions is that their image features (grayscale, signal intensity, texture) highly match the distribution and signal features in the expected defect information. These regions can include one or more independent local areas, and their spatial coordinates and size range need to be recorded to provide location information for subsequent local scanning.

[0021] S104: For the candidate defect region, a second detection technique with a spatial resolution or spectral resolution greater than that of the first detection technique is used for local scanning to obtain a second distribution image; wherein, the type of the second detection technique is selected based on the defect type in the expected defect information.

[0022] In this embodiment, the candidate defect region is the identified suspected defect region (not a confirmed defect). It is a local area selected from the first distribution image that matches the expected defect information features. Its spatial coordinates and size range have been recorded. It is the only target area for this local scan, eliminating the need for a full-field scan of the substrate, focusing only on the suspected area, thus improving detection efficiency. Spatial resolution is the minimum distance between two adjacent points on the substrate that the detection technology can distinguish, measured in μm (micrometers). The smaller the resolution value, the higher the detection accuracy. For example, a spatial resolution of 0.5 μm is better than 5 μm. It is used to characterize the accuracy of the detection technology in identifying the spatial location and size of defects and is one of the indicators that distinguishes the first and second detection technologies. Spectral resolution is the minimum wavelength difference between two different wavelength spectral signals that the detection technology can distinguish, measured in nm (nanometers). (Wavenumber) The smaller the resolution value, the higher the resolution of the spectral signal; it is used to characterize the ability of detection technology to identify defect spectral features (such as Raman peaks and fluorescence peaks), and is only applicable to spectral detection technologies. It is another indicator to distinguish between the first and second detection technologies.

[0023] In this embodiment, the first detection technology employs a rapid full-field screening technique, such as laser scattering imaging or panoramic scanning with a metallurgical microscope. Its characteristics include high detection efficiency and comprehensive coverage, but its spatial / spectral resolution is relatively low, only sufficient for preliminary identification of candidate defect areas and unable to achieve precise defect characterization. The second detection technology is a high-precision detection technology designed to achieve precise characterization of candidate defect areas. Its characteristics include a spatial or spectral resolution greater than the first detection technology (higher accuracy), and its detection range is limited to the candidate defect area (local scanning). The type of detection must be selected based on the expected defect type (different defects are suited to different high-precision detection technologies), serving as a precise verification step in the graded detection strategy. Local scanning differs from full-field scanning; it uses only the spatial coordinates of the candidate defect area as a reference, with a scanning range slightly larger than the candidate defect area (5-10 μm beyond the area edge to avoid missed detections). The detection equipment performs point-by-point, high-precision scanning of the suspected defect area, aiming to acquire the microscopic and precise feature signals of the defect.

[0024] In this embodiment, the second distribution image is high-precision visualized image data generated by locally scanning the candidate defect region using the second detection technology. Its coordinate system is consistent with the first distribution image, facilitating subsequent registration and fusion. Pixel values ​​represent high-precision detection signals at various locations within the candidate defect region, such as Raman spectral intensity and photoluminescence signal intensity, clearly presenting the microscopic morphology and spectral characteristics of the defects. This provides data support for subsequent defect classification and quantification. The defect types in the expected defect information are defined typical defect types, such as dislocations, stress concentration regions, stacking faults, and inclusion defects. Different defect types have significantly different crystal structures and corresponding characteristic signals, such as spectra and microscopic morphology. Therefore, the type of second detection technology needs to be selected appropriately; for example, Raman spectroscopy is suitable for stress defects, and photoluminescence technology is suitable for luminescence-related defects.

[0025] S105: Spatial registration and data fusion are performed on the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region.

[0026] In this embodiment, the same candidate defect region refers to a single, independent candidate defect region identified, such as a dislocation candidate region or a stress concentration region candidate region. It corresponds to a local image of that region in the first distribution image and a second distribution image obtained by scanning that region. This is a one-to-one match, meaning that one second distribution image (a single candidate defect region) corresponds to a local segment of that region in the first distribution image, avoiding image confusion between different candidate defect regions. The first distribution image is visualized image data obtained through full-field scanning using a first detection technique (e.g., laser scattering imaging). Here, it refers to a local segment (not a full-field image) corresponding to the same candidate defect region. Its pixel coordinates correspond one-to-one with the physical space coordinates of the substrate, representing the macroscopic signal characteristics of the candidate defect region, such as scattering intensity and grayscale. It serves as the macroscopic feature carrier for data fusion. The second distribution image is high-precision visualized image data obtained through local scanning using a second detection technique (e.g., Raman microscopy). It is an image generated by scanning the same candidate defect region, sharing the same coordinate system as the first distribution image (the substrate position remains unchanged), and has higher spatial / spectral resolution. It represents the microscopic and spectral signal characteristics of the candidate defect region, such as Raman scattering intensity and luminescence signal intensity. It serves as the microscopic feature carrier for data fusion.

[0027] In this embodiment, spatial registration is a process of precisely aligning the spatial coordinates of the first and second distribution images of the same candidate defect region. It eliminates spatial offsets caused by differences in detection technology and equipment precision errors, such as pixel misalignment and positional deviations, ensuring that the pixel coordinates of the same physical location (the same defect site) in both images correspond perfectly. This lays the spatial consistency foundation for subsequent data fusion, with the goal of precise matching at the same location. Data fusion, after spatial registration, is a process of jointly integrating the signal features of the first and second distribution images of the same candidate defect region. It fuses the complementary features of the two images—the macroscopic signal features (comprehensive coverage) of the first distribution image and the microscopic and spectral features (high precision) of the second distribution image—avoiding deviations in defect characterization due to insufficient features from a single image.

[0028] In this embodiment, the multimodal fusion feature data is a comprehensive feature data that is generated after spatial registration and data fusion, representing the same candidate defect region. Multimodal means that it integrates two different types of signal features (macro + micro / spectral) corresponding to two detection technologies (first and second), rather than a single feature. The data form can be feature vector, matrix or visualized feature map, including comprehensive information such as the macro distribution, micro structure and spectral characteristics of the defect, which is the input data for subsequent defect classification and quantification.

[0029] In this embodiment, the method further includes calculating a first confidence level for the first distribution image based on the structural similarity index and gradient consistency of local regions in the first distribution image; calculating a second confidence level for the second distribution image based on the signal-to-noise ratio and spectral peak fitting residual of the spectral signal at each scanning point in the second distribution image; and assigning a weight positively correlated with the first confidence level to the features from the first distribution image and a weight positively correlated with the second confidence level to the features from the second distribution image when generating the fusion features for each spatial location point.

[0030] Specifically, the first confidence score is a signal reliability index calculated for the first distribution image (a local segment of the candidate defect region), with a value ranging from 0 to 1. The closer the value is to 1, the more reliable the image features of that region. The calculation is based on the structural similarity index and gradient consistency of the local region in the first distribution image, used to measure the credibility of the features of the first distribution image and provide a weighting basis for weighted fusion. The structural similarity index (SSIM) is a quantitative index used to evaluate the image quality and structural integrity of a local region in the first distribution image, with a value ranging from 0 to 1. It represents the structural similarity between the image of that region and the image of the normal crystal region. The higher the similarity, the closer the index is to 1, indicating that the image of that region has no obvious noise interference, a clear structure, and stronger feature reliability.

[0031] In this embodiment, gradient consistency is a quantitative index used to evaluate the continuity of edge / texture features in a local region of the first distribution image. Its value ranges from 0 to 1, representing the uniformity and continuity of the image gradient (edge ​​brightness variation) in that region. A higher gradient consistency, closer to 1, indicates clearer and more unbroken defect edges in that region, resulting in higher accuracy of feature extraction. The second confidence score is a signal reliability index calculated for the second distribution image (a high-precision image of the candidate defect region). Its value ranges from 0 to 1. A value closer to 1 indicates a more reliable spectral signal in that region. The calculation is based on the signal-to-noise ratio and peak fitting residual of the spectral signal at each scan point of the second distribution image, representing the credibility of the features in the second distribution image and providing a weighting basis for weighted fusion.

[0032] In this embodiment, the scan point spectral signal is a spectral correlation signal, such as Raman scattering signal or photoluminescence signal, collected point by point during the local scanning of the candidate defect region by the second detection technology. It is the data source of the second distribution image and the basis for calculating the second confidence level. The signal-to-noise ratio (SNR) is the ratio of the signal intensity to the noise intensity of the spectral signal at each scan point of the second distribution image. It quantitatively represents the purity of the spectral signal. The higher the SNR, the less interference the spectral signal is subjected to and the clearer it is, and the stronger the reliability of the corresponding signal (the SNR has a higher weight when calculating the second confidence level). The spectral peak fitting residual is the deviation value between the fitted curve and the original spectral signal after performing spectral peak fitting (e.g., Gaussian fitting) on ​​the spectral signal at each scan point of the second distribution image. It quantitatively measures the accuracy of spectral signal analysis. The smaller the residual value, the more accurate the spectral peak fitting and the more reliable the extracted spectral features (e.g., peak position, intensity). A spatial location point is a single physical location within the same candidate defect region (corresponding to a single pixel in the first and second distribution images). Each spatial location point has corresponding first distribution image features and second distribution image features. The weighted fusion features of this point need to be calculated separately and finally combined to form the multimodal fusion feature data of the entire candidate defect region.

[0033] In this embodiment, the fused feature is a single comprehensive feature obtained by weighted fusion of first and second distributed image features at a single spatial location point. The weight is positively correlated with the confidence level, with higher-confidence features contributing more to the fused feature. The weight is a coefficient representing the proportion of contribution of the first and second distributed image features during the fusion process, ranging from 0 to 1. The sum of the weights of the two types of features is 1. The rule is that it is positively correlated with the confidence level; that is, the higher the first confidence level, the greater the weight of the first distributed image feature; and the higher the second confidence level, the greater the weight of the second distributed image feature.

[0034] Specifically, a weighted fusion method is used when generating the fusion features for each spatial location point. Let the feature value of the first distribution image at this point be F1, and its first confidence level be C1; let the feature value of the second distribution image at this point be F2, and its second confidence level be C2. Then the formula for calculating the fusion feature F at this point is: F = F1 * [C1 / (C1 + C2)] + F2 * [C2 / (C1 + C2)]. Where C1 and C2 are normalized and their values ​​range from [0, 1]. This weighted strategy ensures that signals with higher confidence levels dominate the fusion features, improving the reliability of the fusion features.

[0035] S106: Input the multimodal fusion feature data into the trained defect classification and quantization model to obtain defect information of candidate defect regions.

[0036] In this embodiment, the trained defect classification and quantification model is an intelligent algorithm model based on historical data in the field of sapphire substrate defect detection. It has undergone complete training, validation, and optimization, and is capable of performing two main functions (classification and quantification), rather than an untrained initial model. The functions are divided into two categories: classification (determining the defect type) and quantification (calculating the specific parameters of the defect). This model serves as a carrier connecting feature data and defect information. The defect classification and quantification model is specifically designed for sapphire substrate crystal structure defects. Its function is to read the input multimodal fusion feature data, analyze the correlation between features and defects through internal algorithms, and ultimately output two results: defect type (classification) and specific defect parameters (quantification), replacing traditional manual analysis and improving the efficiency and accuracy of defect identification.

[0037] In this embodiment, the defect information is the complete information of the candidate defect area identified after defect classification and quantification model analysis. It includes two dimensions (corresponding to the functions of the defect classification and quantification model): Defect classification information: the confirmed defect type, such as dislocation, stress concentration area, stacking fault, etc., which corresponds to the expected defect information; Defect quantification information: the specific quantifiable parameters of the defect, such as defect density, lattice stress value, defect size, distribution range, etc., which are the basis for generating inspection reports and making quality control judgments.

[0038] S107: Based on the defect information of all candidate defect areas, generate a defect detection report and quality control judgment results for the sapphire substrate.

[0039] In this embodiment, the sapphire substrate is the aluminum oxide to be tested throughout the entire process. For single-crystal substrates, their specifications (size, crystal orientation, processing status) and process history information are clearly defined. Both the inspection report and quality control assessment revolve around the overall defect situation of the substrate, ultimately outputting the inspection conclusion. The defect inspection report is a standardized and comprehensive inspection document generated by organizing and summarizing defect information from all candidate defect areas. It serves as the deliverable of this sapphire substrate defect inspection, clearly presenting the overall defect situation and inspection process details for future traceability, quality control archiving, and process optimization reference. The content must be complete, rigorous, and conform to industrial inspection report standards. The quality control assessment result is based on the defect information in the defect inspection report and, according to pre-set sapphire substrate quality control standards, makes a clear judgment on the pass / fail status and grade of the sapphire substrate to be inspected, such as pass, fail, or require re-inspection, along with the corresponding quality control level. Its purpose is for post-production quality control screening of the substrate, determining whether it meets the requirements of subsequent application scenarios (e.g., LED epitaxy, semiconductor devices).

[0040] As can be seen from the above, this application obtains expected defect information based on process history information, scans distribution images according to different detection technologies, generates multimodal fusion feature data through image analysis, spatial registration and data fusion, obtains defect information using a trained defect classification and quantification model, and finally generates a detection report and quality control judgment results, thereby achieving accurate detection and evaluation of crystal structure defects in sapphire substrates.

[0041] In one embodiment of this application, typical defect types and distribution characteristics associated with the defect are queried from a preset defect knowledge base based on process history information as expected defect information, including: The crystal growth method, thermal history, and surface processing parameters in the process history information are obtained and converted into a structured data format to obtain structured process history data. By using rule-based reasoning or case retrieval algorithms, structured historical process data is matched with defect features corresponding to different process conditions stored in the defect knowledge base to obtain matching results. Based on the correlation, the defect types in the matching results are sorted, and based on the sorting results, the defect types and their spatial distribution patterns in the crystal are selected to determine the expected defect information.

[0042] In this embodiment, the crystal growth method is the preparation method of the sapphire substrate crystal (common types of sapphire substrates), such as the Czochralski method, the mold-guided method, and the Czochralski method. These are key process conditions affecting defect generation; different growth methods correspond to different typical defects. For example, the Czochralski method is prone to dislocations. Thermal history refers to the temperature-related parameters and process records during the sapphire substrate crystal growth process, including growth temperature, cooling rate, annealing temperature, and holding time. These are factors that cause crystal lattice distortion and the generation of defects (such as stress concentration areas). Surface processing parameters are the relevant parameters of the surface treatment process after the sapphire substrate crystal growth, including grinding grit size, polishing grit size, polishing pressure, and processing method (such as chemical mechanical polishing). These mainly affect the generation and distribution of surface / near-surface defects (such as surface microcracks).

[0043] In this embodiment, the structured data format is a standardized and normalized form of data organization (distinct from the original messy process records). It organizes scattered and non-standardized process history information, such as text and scattered values ​​in the original production ledger, into a unified format of category-parameter-value-unit, which facilitates matching and calculation with data in the defect knowledge base, such as in tabular form or key-value pair form, thereby improving matching efficiency and accuracy.

[0044] In this embodiment, the case retrieval algorithm is based on the similarity comparison of historical matching cases. That is, it retrieves the historical case with the highest similarity to the current structured process historical data process parameters from the defect knowledge base, and uses the defect type and distribution characteristics corresponding to the historical case as the matching result of the current substrate, which is suitable for accurate matching of non-standard and special processes.

[0045] Specifically, the case retrieval algorithm adopts a five-level modular hierarchical structure: input layer, feature extraction layer, matching layer, ranking layer, and output layer. Each layer connects and works collaboratively from top to bottom, with data transmission between layers using a structured format to improve retrieval efficiency and matching accuracy. The input layer receives structured historical process data and index information from the defect knowledge base, completing data preprocessing and format standardization. The feature extraction layer extracts features from the structured historical process data (crystal growth methods, thermal history parameters, surface processing parameters) and process features from cases in the defect knowledge base, generating a unified-dimensional feature vector. The matching layer uses a similarity calculation model to perform a dimensional comparison between the input features and the features of cases in the knowledge base, outputting initial matching results. The ranking layer ranks the initial matching results based on correlation calculation rules, filtering out highly correlated cases. The output layer extracts the defect types and spatial distribution patterns corresponding to the ranked highly correlated cases, forming standardized expected defect information, which is then transmitted to the subsequent image analysis stage. Each layer includes a data verification module to ensure the accuracy of data transmission at each step, avoiding feature loss or matching deviations.

[0046] In this embodiment, process conditions are combinations of process parameters associated with defect features stored in the defect knowledge base. These correspond to three types of information in the structured process history data (growth method, thermal history, and surface processing parameters). Each process condition corresponds to a unique / set of typical defect types and distribution characteristics. For example, mold guiding method + cooling rate of 10℃ / h + no annealing constitutes a process condition. The matching result is the initial association result obtained by matching the structured process history data with the process conditions and defect features in the defect knowledge base through rule reasoning or case retrieval algorithms. This initial association result includes multiple defect types and their corresponding distribution characteristics (not a single result) and needs to be sorted by association degree before filtering.

[0047] In this embodiment, correlation degree is used to represent the degree of matching between structured process history data and a certain process condition and defect type in the defect knowledge base. It is a quantitative indicator, ranging from 0 to 1. The closer the value is to 1, the higher the correlation degree, representing the probability of the corresponding defect type occurring under that process condition. A higher correlation degree indicates a greater probability of defect occurrence, serving as the basis for sorting and filtering defect types. Sorting is based on the magnitude of the correlation degree, arranging all defect types (and their corresponding distribution characteristics) in the matching results from high to low. The aim is to filter out the defect types with the highest probability of occurrence and the best match to the current process, avoiding irrelevant or low-probability defect types from interfering with the accuracy of expected defect information. Spatial distribution pattern is the specific manifestation of the defect distribution characteristics, referring to the spatial distribution law of typical defects within the sapphire substrate crystal, corresponding one-to-one with the defect type. For example, dislocations correspond to global surface distribution, and stress concentration areas correspond to edge ring distribution. This needs to be selected and determined based on the sorting results.

[0048] As can be seen from the above, this embodiment can effectively match the process history information with the defect features corresponding to different process conditions in the defect knowledge base by converting the process history information into a structured data format. The matching results are obtained through rule reasoning or case retrieval algorithms, and the defect types in the matching results are sorted based on the correlation degree. This can accurately determine the typical defect types associated with the crystal and their spatial distribution patterns, providing more accurate expected defect information for subsequent detection.

[0049] In one embodiment of this application, based on expected defect information, image analysis is performed on a first distribution image to identify one or more candidate defect regions, including: Calculate the global signal-to-noise ratio and local contrast distribution of the first distribution image, evaluate the image quality, and obtain the image quality evaluation result; Based on the image quality assessment results, the parameters of the feature extraction and image segmentation algorithms are adjusted. Specifically, when the signal-to-noise ratio is less than the first threshold or the contrast non-uniformity is greater than the second threshold, an adaptive filtering algorithm is used to preprocess the image for noise reduction, and the fusion weights of features at each scale are adjusted in the multi-scale decomposition. The similarity threshold of the adaptive optimization region growing algorithm and the connection distance of the clustering algorithm are optimized in order to make the algorithm stable in a noisy environment, resulting in the optimized region growing algorithm and clustering algorithm. Based on the defect spatial configuration in the expected defect information, a corresponding spatial feature enhancement filter is generated. The first distribution image is processed using an optimized region growing algorithm, clustering algorithm, and spatial feature enhancement filter to extract feature maps. Adaptive threshold segmentation and region clustering are then performed to merge connected regions with similar features and identify them as candidate defect regions.

[0050] In this embodiment, the global signal-to-noise ratio (SNR) is a quantitative indicator used to evaluate the overall noise interference level of the first distribution image. The local contrast distribution is an indicator used to evaluate the uniformity of brightness difference in local areas of the first distribution image. The image quality assessment result is a comprehensive evaluation conclusion made by calculating the global signal-to-noise ratio and the local contrast distribution, based on the overall sharpness, noise interference level, and local discrimination of the first distribution image.

[0051] In this embodiment, the first threshold is a preset critical value used to determine whether the global signal-to-noise ratio of the first distributed image is acceptable, for example, 20dB, and is one of the criteria for image quality assessment. When the global signal-to-noise ratio is less than the first threshold, it indicates that the overall image has too much noise and noise reduction preprocessing is required. Contrast non-uniformity is an indicator used to represent the difference in local contrast distribution in the first distributed image, with a value range of 0-1. The larger the value, the greater the contrast difference between different local areas of the image, with some areas being clear and others blurry, making defect feature extraction more difficult; conversely, the contrast distribution is uniform, and feature extraction is simpler. The second threshold is a preset critical value used to determine whether the local contrast distribution of the first distributed image is uniform, and is another criterion for image quality assessment. When the contrast non-uniformity is greater than the second threshold, it indicates that the local contrast difference in the image is too large, requiring adjustment of algorithm parameters or preprocessing.

[0052] Specifically, the method for determining the first threshold is as follows: First, based on the requirements for defect detection of sapphire substrates, the minimum signal-to-noise ratio (SNR) requirement for effectively identifying the smallest defects (microcracks / dislocations, with a size greater than or equal to 5 μm) is determined. Through derivation based on optical detection principles, when the SNR is less than 20 dB, the defect features will be completely masked by noise and cannot be effectively extracted. Second, the SNR distribution of full-field scanning images (first detection technology) of sapphire substrates in mass production batches is statistically analyzed. More than 95% of the qualified scanning images have an SNR greater than or equal to 20 dB. After verification through standard sample testing, when the SNR is greater than or equal to 20 dB, defect features can be accurately extracted without noise reduction preprocessing. Therefore, the first threshold is set to 20 dB. When the SNR is less than 20 dB, adaptive filtering noise reduction is triggered.

[0053] The second threshold is determined as follows: First, local contrast non-uniformity is defined as the coefficient of variation of the contrast of each local region (50×50 pixels) of the image, which characterizes the spatial fluctuation of the image contrast. Second, through experimental testing, when the contrast non-uniformity is greater than 15%, the feature fusion of multi-scale decomposition will result in the loss of local features, causing the segmentation deviation of the defect area. Finally, based on the optical characteristics of the sapphire substrate (the overall optical characteristics of the single crystal substrate are uniform, and the contrast non-uniformity of the normal scanned image is mostly less than 15%), the second threshold is set to 15%. When the contrast non-uniformity is greater than 15%, the weight of multi-scale feature fusion is adjusted.

[0054] In this embodiment, multi-scale decomposition involves breaking down the first distribution image into feature maps of different scales (e.g., coarse and fine scales). Different scales correspond to different details. By adjusting the fusion weights of features at each scale, features related to the expected defects are highlighted, and noise interference is suppressed. The fusion weights of features at each scale are coefficients used to adjust the contribution ratio of feature maps at different scales in the final feature extraction after multi-scale decomposition, thereby increasing the weight of fine-scale features and ensuring the accuracy of feature extraction.

[0055] In this embodiment, the region growing algorithm is an algorithm used for image segmentation and region merging. It can select several seed points (suspected defect points) from the image and, according to a preset similarity threshold, gradually merge pixels around the seed points that are similar to the seed point in features (such as gray level and signal strength) to form a connected region, which is used to initially identify suspected defect regions. The similarity threshold is determined as follows: First, the similarity threshold is defined as the critical value of the gray-level difference between adjacent pixels and the growing region during the region growth process, representing the similarity judgment standard of gray-level features between pixels. Second, through experimental testing, when the similarity threshold is greater than 15 gray levels, the defect edge pixels cannot be included in the growing region due to the excessively large gray-level judgment standard, resulting in insufficient growth of the defect region. When the similarity threshold is less than 8 gray levels, noise pixels will be misjudged as defect pixels due to the excessively small gray-level judgment standard, resulting in overgrowth of the defect region. Finally, based on the optical characteristics of the sapphire substrate (uniform gray-level background of single-crystal substrate, and the true gray-level difference between typical crystal structure defects and the background is concentrated in the 8-15 gray level range), the similarity threshold is set to 10 gray levels. When the similarity threshold deviates from this value, the threshold size is adaptively adjusted according to the image quality assessment results to match the defect recognition requirements.

[0056] Specifically, the region growing algorithm adopts a five-level modular hierarchical structure: input preprocessing layer, seed point selection layer, growth execution layer, growth termination judgment layer, and post-processing optimization layer, adapted to the candidate defect region identification scenario of the first distribution image of sapphire substrate. The input preprocessing layer receives the first distribution image after adaptive filtering and noise reduction and spatial feature enhancement, and completes image grayscale normalization and pixel value calibration to eliminate interference caused by uneven illumination; the seed point selection layer selects pixels that meet the defect grayscale and texture features as initial seed points based on the defect spatial configuration in the expected defect information, and is equipped with a seed point verification module; the growth execution layer is the core of the algorithm, responsible for expanding the growth region pixel by pixel based on the similarity threshold to achieve the initial delineation of the defect region; the growth termination judgment layer monitors the expansion status of the growth region in real time and triggers the termination condition to avoid over-growth or under-growth; the post-processing optimization layer fills holes and smooths the boundaries of the grown region, removes isolated growth points, corrects growth deviations in noisy environments, and finally outputs the growth region that meets the requirements, providing support for subsequent fusion with clustering algorithms to identify candidate defect regions.

[0057] In this embodiment, the clustering algorithm is used to classify and merge multiple small regions after image segmentation. It adopts a five-level modular hierarchical structure: input preprocessing layer, feature adaptation layer, clustering core layer, post-processing optimization layer, and output layer. Each layer works together to adapt to the defect recognition scenario of the first distribution image of the sapphire substrate. The input preprocessing layer receives the first distribution image after adaptive filtering and noise reduction, and the feature map after processing by the spatial feature enhancement filter, and completes feature normalization and initial outlier removal. The feature adaptation layer combines the defect spatial configuration in the expected defect information to perform dimension filtering and weight allocation on the input features to highlight the defect features. The clustering core layer is the core of the algorithm and adopts an improved K-means clustering architecture, including three sub-modules: cluster center initialization, distance calculation, and iterative update. The post-processing optimization layer merges connected regions and removes outliers from the clustering results to correct clustering bias in noisy environments. The output layer matches the optimized clustered regions with defect features and outputs preliminary candidate defect regions that meet the requirements. Each layer embeds a data verification module to ensure the stability of feature transmission and the clustering process.

[0058] In this embodiment, the defect spatial configuration refers to the spatial morphology and distribution pattern of typical defects within the sapphire substrate crystal, as part of the expected defect information. It serves as the basis for generating the spatial feature enhancement filter. The spatial feature enhancement filter is an image filtering tool designed based on the expected defect spatial configuration. Its function is to enhance features matching the expected defect spatial configuration, suppress noise features unrelated to the defect configuration, highlight the difference between the defect region and the normal region, and improve the accuracy of defect feature extraction. The feature map is an image extracted after processing the first distribution image using the spatial feature enhancement filter, containing only defect-related features (e.g., signal strength, spatial morphology).

[0059] In this embodiment, adaptive threshold segmentation adjusts the segmentation threshold based on the local features of the feature map. It sets appropriate segmentation thresholds for different local regions of the feature map, dividing it into suspected defect regions and normal regions. This adapts to the feature differences between different regions, reducing missed and incorrect segmentation. Region clustering is a process of feature comparison and merging of multiple small suspected regions after adaptive threshold segmentation using an optimized clustering algorithm. It merges connected regions with similar features and close proximity into a complete suspected defect region, eliminating isolated noisy small regions, ultimately obtaining a clear and complete candidate defect region. Feature-similar connected regions are those in the feature map that have the same or similar features, such as grayscale, signal strength, and spatial morphology, and whose pixels are interconnected (without breaks). These regions are likely different parts of the same defect; merging them can form a complete candidate defect region, avoiding the identification of the same defect as multiple small regions.

[0060] As can be seen from the above, this embodiment evaluates image quality by calculating the global signal-to-noise ratio and local contrast distribution of the first distribution image, thus understanding the quality of the first distribution image. Adjusting the parameters of the feature extraction and image segmentation algorithms based on the image quality evaluation results makes the algorithms more adaptable to image quality. When the signal-to-noise ratio is less than the first threshold or the contrast non-uniformity is greater than the second threshold, an adaptive filtering algorithm is used for noise reduction preprocessing, and the fusion weights of features at each scale in multi-scale decomposition are adjusted, which can improve image quality. Adaptively optimizing the region growing and clustering algorithms ensures their stability in noisy environments, improving algorithm stability. Generating corresponding spatial feature enhancement filters based on the defect spatial configuration in the expected defect information enhances defect features. Using the optimized algorithms and filters to process, segment, and cluster the first distribution image accurately identifies candidate defect regions.

[0061] In one embodiment of this application, adjusting the parameters of the feature extraction and image segmentation algorithms based on the image quality assessment results further includes: Establish a lookup table that includes the optimal algorithm parameters under different noise levels and contrast conditions; Based on the calculated image quality assessment results, the corresponding optimal parameter set is retrieved from the lookup table and loaded, and then applied to the analysis process of the current first distribution image.

[0062] In this embodiment, the noise level is a graded description representing the degree of noise interference in the first distribution image, corresponding to the numerical range of the global signal-to-noise ratio, and is one of the dimensions for dividing lookup table entries and matching optimal parameters. The contrast condition is a graded description representing the uniformity of the local contrast distribution in the first distribution image, and is another dimension for dividing lookup table entries and matching optimal parameters. The lookup table (LUT) is a pre-defined standardized parameter mapping table, which is a one-to-one correspondence table between image quality conditions (noise level + contrast condition) and the optimal algorithm parameter set. Its function is to realize the rapid retrieval and loading of parameters, replacing the traditional real-time parameter iterative optimization, improving the efficiency of first distribution image analysis, and improving the consistency and accuracy of parameter adjustment.

[0063] In this embodiment, the optimal parameter set is a complete and coordinated set of algorithm parameters that match a specific image quality condition (single noise level + single contrast condition), including feature extraction algorithm parameters and image segmentation algorithm parameters, rather than a single parameter, so that the two types of algorithms work together and adapt to the current image quality.

[0064] As can be seen from the above, by establishing a lookup table of optimal algorithm parameters under different noise levels and contrast conditions, this embodiment can quickly retrieve and load the corresponding optimal parameter set based on the image quality assessment results and apply it to the analysis process of the first distribution image. This can improve the adaptability and accuracy of feature extraction and image segmentation algorithms under different image qualities, thereby more accurately identifying candidate defect regions and improving the accuracy and reliability of defect detection in the entire sapphire substrate crystal structure.

[0065] In one embodiment of this application, the second detection technique is microscopic spectral imaging technique, which obtains a second distribution image, including: The candidate defect region is scanned using excitation light or an electron beam of a selected wavelength; Acquire the spectral signal generated at each scanning point; The spectral signal at each scanning point is analyzed to extract one or more quantitative parameters; A second distribution image is generated based on the quantitative parameters of all scan points.

[0066] In this embodiment, microscopic spectral imaging technology is the specific type of the second detection technology. It is a detection technology that combines the high spatial resolution of a microscope with the high sensitivity of spectral analysis. It uses an excitation source (excitation light / electron beam) to irradiate the candidate defect region, collects the spectral signal of each scanning point, extracts quantitative parameters by analyzing the spectral signal, and finally correlates the quantitative parameters with spatial coordinates to generate an image (second distribution image) that has both spatial location and spectral characteristics. It is suitable for the accurate detection of microscopic defects (dislocations, stacking faults, vacancy clusters, etc.) in sapphire substrates. Commonly used types include Raman microscopic spectral imaging and PL (photoluminescence) microscopic spectral imaging, both of which are suitable for sapphire detection.

[0067] In this embodiment, the excitation light or electron beam of the selected wavelength is the signal excitation source for microscopic spectral imaging technology. It is the carrier that triggers the generation of characteristic spectral signals in the candidate defect region. The wavelength / energy needs to be selected according to the candidate defect type and the characteristics of the sapphire substrate to accurately excite the characteristic spectrum of the defect and avoid generating interference signals in the normal crystal region. The scanning point is a tiny discrete point on the candidate defect region that is illuminated during the excitation light / electron beam scanning process. Each scanning point corresponds to a unique spatial coordinate, which can be registered with the spatial coordinates of the first distribution image. It also corresponds to a set of independent spectral signals and is the smallest unit for spectral acquisition and signal analysis.

[0068] In this embodiment, the spectral signal analysis of each scanning point is performed using a preset spectral analysis algorithm. This algorithm processes and analyzes the raw spectral signal of each scanning point to remove noise interference and extract quantitative parameters related to the defect. Specifically, this includes background noise removal, characteristic peak identification, and calculation of parameters such as peak intensity, peak position, and residuals. This ensures that the extracted quantitative parameters accurately characterize the defect properties. The quantitative parameters are numerical parameters (dimensionless or with defined units) extracted from the spectral signals of the scanning points that characterize the microscopic properties of the defect. They serve as the basis for generating the second distribution image. Different microscopic spectral imaging techniques correspond to different quantitative parameters.

[0069] As can be seen from the above, this embodiment uses excitation light or electron beam of selected wavelength to scan the candidate defect region, collect spectral signals, and analyze and extract quantitative parameters. It can generate a second distribution image based on the quantitative parameters of all scanning points, providing more accurate and detailed information for subsequent defect detection and analysis, and improving the accuracy and reliability of sapphire substrate crystal structure defect detection.

[0070] In one embodiment of this application, the spectral signal at each scan point is analyzed to extract one or more quantitative parameters, including: Acquire Raman scattering, photoluminescence, or cathodoluminescence signals from spectral signals; For Raman scattering signals, the peak position, full width at half maximum (FWHM), or intensity of characteristic peaks are analyzed to calculate indices such as lattice stress, lattice disorder, or defect density. For photoluminescence or cathodoluminescence signals, the intensity, wavelength or lifetime of characteristic peaks are analyzed, and indices such as luminescent center concentration, nonradiative recombination center density or carrier diffusion length are calculated. The quantitative parameters are one or more combinations of the above indicators.

[0071] In this embodiment, the characteristic peak is the most crucial component of the spectral signal. It is a peak with clear physical meaning (corresponding to a specific wavelength / wavenumber) appearing in the spectral curve and is key to interpreting the spectral signal. Different defect types and crystal states correspond to different characteristic peaks. Peak position characterizes the energy characteristics of the spectral signal and directly indicates the degree of lattice distortion. The full width at half maximum (FWHM) is the width of the peak corresponding to half the height of the characteristic peak, characterizing the sharpness of the characteristic peak and directly indicating the degree of lattice disorder. The more disordered the lattice, the wider the characteristic peak and the larger the FWHM value. Intensity is the peak height or integrated intensity of the characteristic peak, characterizing the strength of the spectral signal and directly indicating the number or activity of defects. For example, the higher the defect density, the lower the intensity of the characteristic peak in the Raman scattering signal (lattice distortion leads to a decrease in scattering efficiency), and the higher the intensity of the characteristic peak in the photoluminescence signal (more luminescence centers formed by defects).

[0072] In this embodiment, lattice stress is a key parameter characterizing the degree of lattice distortion, directly representing the extent of damage caused by defects to the crystal structure. The greater the lattice stress, the more unstable the crystal structure and the more severe the damage caused by defects. Lattice disorder represents the density of defects; the higher the lattice disorder, the more defects there are, the more irregular the atomic arrangement, and the worse the crystal quality. Defect density is the number of defects per unit area, calculated jointly from the intensity of Raman characteristic peaks and the full width at half maximum (FWHM). When the defect density exceeds a preset threshold, the substrate does not meet the application requirements.

[0073] In this embodiment, the luminescent center concentration is the number of defects per unit volume that can generate photoluminescence / cathofluorescence signals (i.e., defects that can induce electron transitions and emit light), representing the activity and distribution density of defects. A higher luminescent center concentration indicates more defects and stronger activity. The nonradiative recombination center density is the number of defects per unit volume that prevent photon release during electron transitions, representing the degree of influence of defects on electronic states. A higher nonradiative recombination center density results in lower photoluminescence / cathofluorescence signal intensity and shorter lifetime. The carrier diffusion length is the diffusion distance of electrons within the defect region after excitation, representing the degree to which defects hinder carrier transport. More defects result in a shorter carrier diffusion length and poorer photoelectric properties of the crystal.

[0074] As can be seen from the above, this embodiment obtains Raman scattering signals, photoluminescence signals, or cathodoluminescence signals by analyzing the spectral signals of the scanning points, and calculates relevant indicators by analyzing the characteristic peaks of these signals respectively. Using these indicators as quantitative parameters, the specific quantitative characteristics of crystal structure defects in sapphire substrates can be effectively obtained, providing a reliable basis for accurately judging the type and extent of defects.

[0075] In one embodiment of this application, spatial registration and data fusion are performed on a first distribution image and a second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region, including: Using the encoder coordinates of the sample stage as a reference, a spatial coordinate correspondence between the first distribution image and the second distribution image is established, and the first registration is performed. Based on the first registration, a second registration is performed between the first distribution image and the second distribution image, using mutual information maximization or feature similarity measurement algorithms. In the image space after the second registration, the feature values ​​of the first distribution image are correlated and combined with the quantitative parameters of the second distribution image to generate multimodal fusion feature data.

[0076] In this embodiment, the spatial coordinate correspondence is established by using the sample stage encoder coordinates as a reference to establish the mapping relationship between the pixel coordinates of the first distribution image and the scan point coordinates of the second distribution image, corresponding to the actual spatial coordinates (encoder coordinates) of the substrate. This clarifies the correspondence between a certain pixel in the first distribution image, the corresponding encoder coordinates of the substrate, and the corresponding scan point in the second distribution image, achieving a one-to-one correspondence between the spatial points of the two types of images. This is the output of the first registration.

[0077] In this embodiment, maximizing mutual information is used to further improve the registration accuracy of the two types of images. Mutual information is an indicator of the signal correlation in the spatially overlapping region of the two types of images. The higher the signal correlation at the same spatial point in the two types of images, the greater the mutual information value. By fine-tuning the spatial position of the second distribution image, the mutual information value of the overlapping region (candidate defect region) of the two types of images reaches its maximum value, at which point the spatial alignment accuracy of the two types of images is the highest.

[0078] In this embodiment, the feature similarity measurement algorithm extracts key feature points (e.g., defect edge points, center points) from the first distribution image and key feature points (e.g., spectral signal anomalies, quantitative parameter mutation points) from the second distribution image. By calculating the similarity (e.g., distance similarity, signal similarity) of the feature points in the two types of images, the algorithm fine-tunes the position of the feature points to achieve precise alignment between the feature points of the two types of images. The feature values ​​of the first distribution image are numerical values ​​extracted from the first distribution image that characterize the macroscopic features of the candidate defect region. The quantitative parameters of the second distribution image are indicators extracted by analyzing the spectral signal that characterize the microscopic properties of the candidate defect region.

[0079] As can be seen from the above, this embodiment can establish a spatial coordinate correspondence between the first distribution image and the second distribution image by performing the first registration based on the encoder coordinates of the sample stage. The second registration, based on the mutual information maximization or feature similarity measurement algorithm, can improve the registration accuracy. In the image space after the second registration, the feature values ​​of the first distribution image and the quantitative parameters of the second distribution image are associated and combined to generate multimodal fusion feature data of the candidate defect region, providing more comprehensive and accurate data support for subsequent defect detection.

[0080] In one embodiment of this application, a second registration is performed on the first distribution image and the second distribution image based on a mutual information maximization or feature similarity measurement algorithm, including: In the first distribution image, stress concentration boundaries, high-contrast spots, or specific texture regions are extracted as the first feature set through edge detection, corner detection, or region segmentation methods. In the second distribution image, the boundaries where spectral parameters change abruptly, abnormal peak sites, or regions of uniformity mismatch are extracted as the second feature set; Calculate the similarity measure between the first feature set and the second feature set, and iteratively optimize the spatial transformation parameters to maximize the overall matching degree between the two feature sets.

[0081] In this embodiment, the edge detection method is used to identify and extract defect edges or boundaries from the first distributed image. It delineates the boundary between defects and normal areas by detecting regions with abrupt changes in grayscale / intensity, such as the edges of stress concentration areas or the contours of dislocation regions. This method is suitable for macroscopic defect images with clear edges, such as metallographic micrographs and laser scattering images. The corner detection method is used to identify and extract defect corners or feature inflection points from the first distributed image. It detects points in the image with drastic grayscale changes and obvious inflection point features, such as the endpoints of surface microcracks or the edges of inclusion defects. This method is suitable for scenarios with irregular defect shapes and obvious inflection points. The region segmentation method is used to segment and extract defect regions with specific features from the first distributed image. It segments candidate defect regions from the background based on grayscale / intensity differences in the image and extracts the entire defect region as a feature.

[0082] In this embodiment, the stress concentration boundary is a component of the first feature set and represents the boundary between the stress concentration region and the normal crystal region in the first distribution image. High-contrast spots are also a component of the first feature set and are tiny dot-like regions in the first distribution image that exhibit significant differences in grayscale / intensity compared to the surrounding background. Specific texture regions are also a component of the first feature set and are defect regions in the first distribution image that possess unique texture morphology. The first feature set is a set of macroscopic defect features extracted from the first distribution image using edge detection, corner detection, or region segmentation methods for precise matching. It consists of multiple feature points, feature boundaries, or feature regions, such as a combination of [stress concentration boundaries, high-contrast spots, and specific texture regions]. Its function is to serve as a baseline feature for fine registration, matching it with the second feature set to ensure targeted registration.

[0083] In this embodiment, abnormal peak sites are components of the second feature set, and are the locations in the second distribution image where abnormal peak values ​​of the spectral quantitative parameters occur. Uniformity mismatch regions are also components of the second feature set, and are regions in the second distribution image where the spectral quantitative parameters are unevenly distributed and significantly different from the surrounding areas. The second feature set is a collection of microscopic defect features extracted from the second distribution image for precise matching. It consists of multiple spectral parameter-related feature points, feature boundaries, or feature regions, such as a combination of [spectral parameter abrupt change boundaries, abnormal peak sites, and uniformity mismatch regions]. Its function is to perform similarity matching with the first feature set, improving registration accuracy through the correspondence between microscopic and macroscopic features.

[0084] As can be seen from the above, this embodiment first extracts the first feature set and the second feature set from the first distribution image and the second distribution image respectively, then calculates the similarity measure between the two feature sets and iteratively optimizes the spatial transformation parameters, which can maximize the overall matching degree of the two feature sets, thereby improving the accuracy of the second registration of the first distribution image and the second distribution image, and thus improving the quality of the generated multimodal fusion feature data, providing a guarantee for obtaining the defect information of the candidate defect region more accurately.

[0085] In one embodiment of this application, the trained defect classification and quantification model is obtained and optimized in the following manner: Obtain a historical dataset including multimodal fusion feature data samples and manually annotated defect ground truth labels; Using historical datasets to train deep neural networks or ensemble learning models enables defect classification and quantification models to predict defect types and regress the quantitative attributes of defects based on input multimodal fusion feature data. Establish an online adaptive mechanism for the model: Calculate the statistical distribution of fusion feature data for the current batch of sapphire substrates in real time; When a deviation is detected between the current statistical distribution and the model training set distribution, temperature scaling or Platt scaling methods are used to recalibrate the probability output of the defect classification and quantization model and correct the output confidence. Alternatively, based on the dominant defect type in the expected defect information, the contributions of feature channels highly correlated with specific defect types can be weighted and fused at the output layer of the defect classification and quantification model to enhance the sensitivity of identifying such defects.

[0086] In this embodiment, the historical dataset serves as the data foundation for training the defect classification and quantization model. It is a structured dataset comprising multimodal fusion feature data samples and manually labeled ground truth defect values. Deep neural networks are one of the algorithmic frameworks for training defect classification and quantization models. Adapted to defect detection with complex features, they are deep learning models composed of multiple network layers (input layer, hidden layer, output layer), such as CNN (Convolutional Neural Network), MLP (Multilayer Perceptron), and Transformer. Their advantage lies in automatically extracting deep features from multimodal fusion feature data, adapting to the complex and high-dimensional characteristics of sapphire defects, and improving classification and quantization accuracy. Ensemble learning models are another algorithmic framework for training defect classification and quantization models. Adapted to defect detection with small sample sizes and high stability, they fuse the prediction results of multiple basic weak classifiers / regressors (e.g., decision trees, random forests, XGBoost) to obtain the final prediction result. Their advantages include strong generalization ability, good anti-interference performance, and fast training speed.

[0087] In this embodiment, the defect classification and quantification model training involves dividing the historical dataset into training, validation, and test sets. Multimodal fusion feature data samples are input into a deep neural network / ensemble learning model. Through backpropagation and loss function optimization, the defect classification and quantification model continuously learns the correspondence between feature data and ground truth defect labels until the classification accuracy and quantization regression precision of the model reach preset thresholds. Defect type prediction is the classification function of the defect classification and quantification model. Based on the input multimodal fusion feature data, the model identifies and outputs the specific category of defects in the sapphire substrate, such as dislocations, stacking faults, surface microcracks, vacancy clusters, and inclusion defects. This is the core of qualitative defect determination. Defect quantification attribute regression is the quantification function of the defect classification and quantification model. Based on the input multimodal fusion feature data, the model calculates and outputs specific quantitative indicators of defects through regression algorithms, such as defect density, lattice stress, luminescent center concentration, and carrier diffusion length. This is the core of quantitative defect determination and corresponds one-to-one with the preceding quantitative parameters.

[0088] In this embodiment, the online adaptive mechanism of the model is a continuous optimization mechanism for the defect classification and quantification model. It is a mechanism that allows the defect classification and quantification model to adjust model parameters / output strategies based on the distribution changes of the current batch of detection data or expected defect information without retraining during the actual detection process. Its function is to solve the model distribution offset (inconsistency between the distribution of actual detection data and training set data), improve the recognition sensitivity of specific defects, and ensure the accuracy and stability of the model under different detection scenarios.

[0089] In this embodiment, the model training set distribution is the statistical distribution of the multimodal fusion feature data of the training set during defect classification and quantization model training. It is the basic distribution for defect classification and quantization model learning, and the optimal accuracy of the defect classification and quantization model is only effective near this distribution. When the actual detection data distribution deviates from this distribution, adaptive calibration needs to be initiated. The distribution offset is the statistical distribution of the fusion feature data of the current batch of sapphire substrates, which differs from the statistical distribution of the defect classification and quantization model training set. For example, the feature mean deviation is greater than or equal to 10%, and the variance deviation is greater than or equal to 15%. This is caused by factors such as fluctuations in the production process, changes in the detection environment, and changes in the substrate type, which will lead to a decrease in model prediction confidence and an increase in classification / quantization error.

[0090] In this embodiment, temperature scaling involves introducing a learnable temperature parameter T to scale the original probability distribution output by the defect classification and quantization model. This adjusts the confidence level of the model's output probabilities, addressing the issue of overconfidence or underconfidence caused by distribution shift, and making the output confidence level more closely match actual prediction accuracy. Platt scaling involves performing logistic regression fitting on the original logit values ​​output by the defect classification and quantization model. This maps the original output of the defect classification and quantization model to a more realistic probability distribution, correcting the probability output bias caused by distribution shift and improving the accuracy of the confidence level.

[0091] In this embodiment, the expected defect information is based on prior information such as the sapphire substrate's manufacturing process, furnace batch, and preparation method, predicting the types of defects that are likely to be prevalent in this batch of substrates. The dominant defect type is the defect type with the highest proportion and the most prevalent occurrence in this batch of sapphire substrates from the expected defect information. For example, the dominant defect type of a certain batch of Czochralski-process substrates is dislocation, which is the object of feature channel weighted fusion in the online adaptive mechanism. The defect classification and quantization model output layer is the last layer of the defect classification and quantization model. It is the core layer that outputs the probability and quantification attributes of defect types, including multiple feature channels corresponding to defect features and defect types, and is the execution layer of the weighted fusion operation.

[0092] In this embodiment, the feature channel is a neuronal channel in the output layer of the defect classification and quantization model used to characterize specific defect features or those highly correlated with specific defect types. Each feature channel corresponds to a defect feature, such as a lattice stress feature channel or a defect density feature channel. Different defect types correspond to different combinations of feature channels.

[0093] As can be seen from the above, this embodiment trains the defect classification and quantification model using a historical dataset that includes multimodal fusion feature data samples and manually labeled defect ground truth values. This enables the defect classification and quantification model to accurately predict defect types and regress defect quantification attributes based on multimodal fusion feature data. The establishment of an online adaptive mechanism for the model, which calculates the statistical distribution of the fusion feature data of the current batch of sapphire substrates in real time, and recalibrates the model output probability and corrects the output confidence when the distribution shifts, using temperature scaling or Platt scaling methods, can improve the stability and accuracy of the model in different batches of inspection. Furthermore, by weighting and fusing the contributions of the dominant defect type in the expected defect information to the relevant feature channels of a specific defect type, the model's sensitivity to identifying this type of defect can be enhanced.

[0094] In one embodiment of this application, after generating a defect detection report and quality control judgment result for the sapphire substrate based on defect information from all candidate defect regions, the method further includes: Based on the correlation analysis between defect detection reports and process history information, the process steps that cause defects are located. Based on a pre-defined library of process optimization strategies, suggestions for adjusting parameters of process steps are generated.

[0095] In this embodiment, the process steps that cause defects are specific production process steps that are determined through correlation analysis, resulting in various defects (dislocations, stacking faults, etc.) in the sapphire substrate. The preset process optimization strategy library is a structured set of strategies built in advance based on sapphire substrate production experience, defect mechanism research, and process parameter debugging data. The content is stored in a categorized manner according to defect type, abnormal process step, optimization strategy, and parameter adjustment range. It includes solutions to potential process hazards corresponding to various defects and can be directly called to generate parameter adjustment suggestions.

[0096] As can be seen from the above, this embodiment, after generating a defect detection report and quality control judgment results for the sapphire substrate, can locate the process steps that cause defects based on the correlation analysis between the defect detection report and process history information, and generate adjustment parameter suggestions for the process steps according to the preset process optimization strategy library, which helps to improve the production process in a targeted manner and reduce crystal structure defects in the product.

[0097] In one embodiment of this application, before performing a full-field scan of the sapphire substrate using a first detection technology to obtain a first distribution image, an adaptive configuration process for the detection parameters is also included: Based on the size of the sapphire substrate and the defect scale in the expected defect information, the scanning resolution and scanning step size are determined: If the substrate is larger than the preset size threshold or is expected to be a large area defect, configure the first set of scanning parameters, which includes resolution and scanning step size. If the substrate size is less than or equal to the preset size threshold or is expected to be a point / micro-area defect, configure a second set of scanning parameters, which includes resolution and scanning step size. Based on the transparency and surface condition information of the sapphire substrate, the illumination and acquisition parameters of the optical detection unit are adjusted: For substrates with high transparency, optimize the polarization angle and incident light wavelength to enhance internal stress contrast; For substrates with rough surfaces or coatings, enable dark field illumination or confocal mode and configure the corresponding image filtering parameters to suppress surface scattering interference.

[0098] In this embodiment, the adaptive configuration process of the detection parameters is a process in which the first detection technology automatically matches and adjusts the detection parameters based on the substrate's own characteristics (size, transparency, surface condition) and expected defect information (defect scale) before performing a full-field scan of the sapphire substrate. Its purpose is to adapt the detection parameters to the object being detected, ensuring scanning efficiency while improving the clarity and accuracy of defect identification, and avoiding detection deviations caused by fixed parameters for different substrates / defects. The defect scale in the expected defect information is the spatial size and distribution range of the defect to be detected, predicted based on prior information such as the sapphire substrate's manufacturing process, furnace batch, and production experience. The scanning resolution is the number of pixels per unit area / number of detection points (unit: μm / pixel, μm / point) when the first detection technology scans the sapphire substrate. Higher resolution results in clearer image details and easier identification of minute defects, but lower scanning efficiency; lower resolution results in higher scanning efficiency.

[0099] In this embodiment, the scanning step size is the distance (unit: μm) that the scanning probe / acquisition unit of the detection device moves along the X / Y axis on the surface of the sapphire substrate in a single pass. A smaller step size results in finer scanning coverage and a lower defect miss rate; a larger step size results in faster scanning speed. The preset size threshold is a pre-set boundary value for the size of the sapphire substrate, such as 6 inches, which serves as the criterion for distinguishing between large and small substrates and is used as the basis for grouping and configuring scanning parameters. The first set of scanning parameters is a combination of scanning parameters adapted to large substrates (greater than the preset size threshold) or anticipated large-area defects, prioritizing the efficiency of full-field scanning. The second set of scanning parameters is a combination of scanning parameters adapted to small substrates (less than or equal to the preset size threshold) or anticipated point / micro-area defects, prioritizing scanning accuracy and detail clarity to ensure that small defects are not missed.

[0100] Specifically, the method for determining the preset size threshold is as follows: First, based on the industrial mass production specifications of sapphire substrates, the mainstream specifications are 2 inches, 4 inches, 6 inches, and 8 inches. Among them, 6 inches (152.4 mm) is the industry dividing point for large-size substrates, 8-inch substrates are ultra-large sizes, and 2 / 4 inch substrates are small sizes. Second, through experimental testing, when detecting large-area defects (such as grain boundaries and cracks) on large-size substrates (greater than 6 inches), low resolution and large step length can balance detection efficiency and full-field coverage. When detecting small-size substrates (less than or equal to 6 inches) or point / micro-area defects (such as dislocations and inclusions), high resolution and small step length are required to ensure detection accuracy. Finally, referring to the industry-standard size division for semiconductor substrate inspection, the preset size threshold is set to 6 inches (152.4 mm) (the first set of scanning parameters is configured when the substrate size is greater than 6 inches, and the second set of scanning parameters is configured when the substrate size is less than or equal to 6 inches).

[0101] In this embodiment, the illumination parameters are the operating parameters of the illumination module of the optical detection unit, including polarization angle, incident light wavelength, illumination mode (bright field / dark field), and light intensity, which directly affect the imaging contrast of substrate defects. The acquisition parameters are the operating parameters of the acquisition module of the optical detection unit, including image resolution, acquisition frame rate, exposure time, and filtering parameters, which work in conjunction with the illumination parameters to determine the quality of the final acquired first distribution image.

[0102] In this embodiment, the polarization angle is the angle between the vibration direction of the polarized illumination light and the surface / crystal orientation of the sapphire substrate. Adjusting this angle can change the propagation and scattering patterns of the incident light within the substrate, enhancing the imaging contrast between internal stress areas and normal areas, and making internal stress defects clearer. The incident light wavelength is the wavelength of the incident light illuminating the sapphire substrate from the optical detection unit. Different wavelengths of light have different penetration depths and scattering characteristics in the sapphire crystal. Optimizing the wavelength can improve the ability of the incident light to identify internal defects in the substrate and enhance the contrast of internal stress.

[0103] In this embodiment, dark-field illumination is an illumination mode for the optical detection unit. Unlike conventional bright-field illumination, it allows only light scattered / diffracted by the substrate defect area to enter the acquisition module, while direct light from the normal area is blocked. This suppresses interference from direct light on smooth surfaces, highlighting the signal from defect areas, and is suitable for substrates with rough surfaces or coatings. Confocal mode is a high-precision acquisition mode for the optical detection unit. It uses a focusing lens to focus incident light onto a specific depth layer of the substrate, acquiring only the signal from that layer and shielding it from scattered light interference from other layers. This improves vertical resolution and suppresses surface scattering interference, making it suitable for scenarios involving coated surfaces and micro-defect detection. Surface scattering interference occurs when the sapphire substrate surface is rough or coated, causing random scattering of incident light and generating a large number of stray signals. These signals, captured by the acquisition module, can overwrite the effective signals from internal substrate defects, blurring the first distribution image and reducing defect identification. This type of interference requires careful suppression during optical parameter adjustment.

[0104] As can be seen from the above, this embodiment can adaptively configure the detection parameters according to the size, expected defect size, transparency and surface condition information of the sapphire substrate, and select appropriate scanning resolution, scanning step size, illumination parameters and acquisition parameters for substrates with different conditions, thereby improving detection efficiency and accuracy. It can reasonably configure parameters for detection of large-area substrates or large-area defects, and can also accurately detect small-sized substrates or point / micro-area defects. It can enhance the internal stress contrast of substrates with high transparency and suppress surface scattering interference of substrates with rough surfaces or coatings.

[0105] In one embodiment of this application, when scanning the candidate defect region using a second detection technique, the method further includes: Real-time monitoring of the signal-to-noise ratio of the acquired spectral signals; When the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, the signal integration time or the number of accumulations at that scan point is increased until the signal-to-noise ratio meets the standard or the maximum allowable time is determined based on the preset maximum integration time for a single point or the overall scan cycle limit. Based on the expected defect information, a priority spectral range is preset for different defect types, and the quality of spectral data within this range is prioritized during the scanning process.

[0106] In this embodiment, the preset signal-to-noise ratio (SNR) threshold is a pre-set critical value used to determine whether the spectral signal quality meets the standard. It is the criterion for triggering adjustments to the signal integration time / accumulation count. If the SNR is greater than or equal to the preset SNR threshold, the signal quality is considered to meet the standard; if the SNR is less than the preset SNR threshold, it is considered to fail and requires adjustment. The signal integration time is the time (unit: ms, s) during which the spectrometer of the second detection technology continuously acquires and accumulates the spectral signal of a single scan point. Its function is to increase the effective signal intensity: the longer the integration time, the more effective signal the spectrometer receives, and the higher the SNR, but the longer the time taken for a single scan point; the shorter the integration time, the faster the scanning speed, but the SNR does not meet the standard, and it is one of the parameters for adjusting the SNR.

[0107] Specifically, the method for determining the preset signal-to-noise ratio (SNR) threshold is as follows: First, based on the spectral signal characteristics of crystal defects in sapphire substrates, the effective resolution of Raman / photoluminescence / cathofluorescence signals requires an SNR greater than or equal to 30 dB. When the SNR is less than this value, the peak fitting error is large, making it impossible to accurately extract quantitative parameters. Second, after testing with a microscopic spectral imaging instrument, when the SNR is greater than or equal to 30 dB, the fitting error of the peak position and full width at half maximum (FWHM) is less than or equal to 2%, meeting the defect quantification requirements. Finally, the preset SNR threshold is set to 30 dB (when the spectral signal SNR is less than 30 dB, the integration time / accumulation count is increased).

[0108] In this embodiment, the preset maximum integration time per point is a pre-set maximum signal integration time allowed for a single scan point, such as 500ms. This is determined by a combination of overall scan cycle requirements, detection efficiency needs, and equipment lifespan. It avoids situations where the integration time for a single scan point is too long, causing the scanning time for the entire candidate defect area to exceed the allowable range for industrial detection. This serves as the upper limit for adjusting the integration time. The maximum allowable time is the longest time a single scan point can be used to adjust the integration time / accumulation count (i.e., the time cannot exceed this limit when adjusting the integration time / accumulation count) calculated based on the preset maximum integration time per point and the overall scan cycle limit. Its purpose is to balance signal quality and scanning efficiency, ensuring the signal-to-noise ratio meets the standard while avoiding impacting the overall scanning progress. The expected defect information is the predicted defect type in the candidate defect area based on prior information such as the sapphire substrate's fabrication process, furnace batch, and production experience. Examples of defects include dislocations, stacking faults, and inclusion defects. The preset priority acquisition spectral range is a pre-set spectral interval that represents the core characteristics of different defect types.

[0109] As can be seen from the above, this embodiment improves the quality of the spectral signal by real-time monitoring of the signal-to-noise ratio of the acquired spectral signal and increasing the signal integration time or the number of accumulations at the scanning points when the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, thus making the detection results more accurate. Based on the expected defect information, it presets the priority acquisition spectral range for different defect types and prioritizes their data quality, enabling the targeted acquisition of key spectral data, improving detection efficiency and the ability to detect specific defects.

[0110] Corresponding to the crystal structure defect detection method for sapphire substrates in the above embodiments, Figure 2 This is a structural block diagram of a crystal structure defect detection device for sapphire substrates according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The crystal structure defect detection device 20 for the sapphire substrate includes: a defect prediction module 21, a first detection module 22, a defect identification module 23, a second detection module 24, a registration and fusion module 25, a defect classification module 26, and a result reporting module 27.

[0111] Among them, the expected defect module 21 is used to obtain the process history information of the sapphire substrate to be inspected, and to query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information. The first detection module 22 is used to perform a full-field scan of the sapphire substrate using a first detection technology to obtain a first distribution image; Defect identification module 23 is used to perform image analysis on the first distribution image based on expected defect information to identify one or more candidate defect regions; The second detection module 24 is used to perform local scanning of the candidate defect region using a second detection technique with a spatial resolution or spectral resolution greater than that of the first detection technique to obtain a second distribution image; wherein, the type of the second detection technique is selected based on the defect type in the expected defect information; The registration and fusion module 25 is used to spatially register and fuse the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region. Defect classification module 26 is used to input multimodal fusion feature data into the trained defect classification and quantization model to obtain defect information of candidate defect regions; The report results module 27 is used to generate a defect detection report and quality control judgment results for the sapphire substrate based on the defect information of all candidate defect areas.

[0112] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the expected defect module 21, the first detection module 22, the defect identification module 23, the second detection module 24, the registration and fusion module 25, the defect classification module 26, and the reporting result module 27 are shown.

[0113] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0114] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0115] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0116] In another embodiment of this application, a computer-readable storage medium is provided. The computer-readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the various method embodiments described above.

[0117] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, etc.

[0118] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0119] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for detecting crystal structure defects in a sapphire substrate, characterized in that, include: Obtain the process history information of the sapphire substrate to be tested, and query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information; The sapphire substrate is scanned in its entirety using a first detection technique to obtain a first distribution image; Based on the expected defect information, image analysis is performed on the first distribution image to identify one or more candidate defect regions; For the candidate defect region, a second detection technique with a spatial resolution or spectral resolution greater than the first detection technique is used for local scanning to obtain a second distribution image; wherein, the type of the second detection technique is selected based on the defect type in the expected defect information; the second detection technique is microscopic spectral imaging technology, and obtaining the second distribution image includes: scanning the candidate defect region using excitation light or an electron beam of a selected wavelength; acquiring the spectral signal generated at each scanning point; analyzing the spectral signal at each scanning point to extract one or more quantitative parameters; and generating the second distribution image based on the quantitative parameters of all scanning points; Spatial registration and data fusion are performed on the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region; The multimodal fusion feature data is input into the trained defect classification and quantization model to obtain the defect information of the candidate defect region; Based on the defect information of all the candidate defect regions, a defect detection report and quality control judgment result for the sapphire substrate are generated; based on the correlation analysis between the defect detection report and the process history information, the process step that caused the defect is located; based on the preset process optimization strategy library, adjustment parameter suggestions for the process step are generated.

2. The method for detecting crystal structure defects in a sapphire substrate according to claim 1, characterized in that, The step of querying typical defect types and distribution characteristics associated with the process history information from a preset defect knowledge base as expected defect information includes: The crystal growth method, thermal history, and surface processing parameters in the process history information are obtained and converted into a structured data format to obtain structured process history data. By using rule-based reasoning or case retrieval algorithms, the structured historical process data is matched with the defect features corresponding to different process conditions stored in the defect knowledge base to obtain matching results; The defect types in the matching results are sorted based on the correlation, and the defect types and their spatial distribution patterns in the crystal are selected based on the sorting results to determine the expected defect information.

3. The method for detecting crystal structure defects in a sapphire substrate according to claim 1, characterized in that, The step of performing image analysis on the first distribution image based on the expected defect information to identify one or more candidate defect regions includes: Calculate the global signal-to-noise ratio and local contrast distribution of the first distributed image, evaluate the image quality, and obtain the image quality evaluation result; Based on the image quality assessment results, the parameters of the feature extraction and image segmentation algorithms are adjusted; wherein, when the signal-to-noise ratio is less than the first threshold or the contrast non-uniformity is greater than the second threshold, an adaptive filtering algorithm is used to perform noise reduction preprocessing on the image, and the fusion weights of features at each scale are adjusted in the multi-scale decomposition. The similarity threshold of the adaptive optimization region growing algorithm and the connection distance of the clustering algorithm are optimized in order to make the algorithm stable in a noisy environment, resulting in the optimized region growing algorithm and clustering algorithm. Based on the defect spatial configuration in the expected defect information, a corresponding spatial feature enhancement filter is generated; The first distribution image is processed using the optimized region growing algorithm, clustering algorithm, and spatial feature enhancement filter to extract feature maps, and adaptive threshold segmentation and region clustering are performed to merge connected regions with similar features and identify them as candidate defect regions.

4. The method for detecting crystal structure defects in a sapphire substrate according to claim 3, characterized in that, The step of adjusting the parameters of the feature extraction and image segmentation algorithms based on the image quality assessment results also includes: Establish a lookup table that includes the optimal algorithm parameters under different noise levels and contrast conditions; Based on the calculated image quality assessment results, the corresponding optimal parameter set is retrieved from the lookup table and loaded, and applied to the analysis process of the current first distribution image.

5. The method for detecting crystal structure defects in a sapphire substrate according to claim 1, characterized in that, The step of analyzing the spectral signal at each scanning point to extract one or more quantitative parameters includes: Acquire the Raman scattering signal, photoluminescence signal, or cathodoluminescence signal from the spectral signal; For the Raman scattering signal, the peak position, full width at half maximum (FWHM) or intensity of the characteristic peaks are analyzed, and the indices of lattice stress, lattice disorder or defect density are calculated. For the photoluminescence or cathodoluminescence signal, the intensity, wavelength or lifetime of the characteristic peaks are analyzed, and the indices of luminescent center concentration, nonradiative recombination center density or carrier diffusion length are calculated. The quantitative parameter is one or more combinations of the above indicators.

6. The method for detecting crystal structure defects in a sapphire substrate according to claim 1, characterized in that, The step of spatially registering and fusing the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region includes: Using the encoder coordinates of the sample stage as a reference, a spatial coordinate correspondence between the first distribution image and the second distribution image is established, and the first registration is performed. Based on the first registration, a second registration is performed on the first distribution image and the second distribution image based on mutual information maximization or feature similarity measurement algorithms; Within the image space after the second registration, the feature values ​​of the first distribution image are correlated and combined with the quantitative parameters of the second distribution image to generate the multimodal fusion feature data.

7. The method for detecting crystal structure defects in a sapphire substrate according to claim 6, characterized in that, The algorithm based on maximizing mutual information or feature similarity measurement performs a second registration of the first distribution image and the second distribution image, including: In the first distributed image, stress concentration boundaries, high-contrast spots, or specific texture regions are extracted as the first feature set by edge detection, corner detection, or region segmentation methods. In the second distribution image, the boundaries where spectral parameters change abruptly, abnormal peak sites, or regions of uniformity mismatch are extracted as the second feature set; Calculate the similarity measure between the first feature set and the second feature set, and iteratively optimize the spatial transformation parameters to maximize the overall matching degree between the two feature sets.

8. The method for detecting crystal structure defects in a sapphire substrate according to claim 1, characterized in that, The trained defect classification and quantification model was obtained and optimized through the following methods: Obtain a historical dataset including multimodal fusion feature data samples and manually annotated defect ground truth labels; The historical dataset is used to train a deep neural network or ensemble learning model, enabling the defect classification and quantification model to predict the defect type and regress the quantitative attributes of the defect based on the input multimodal fusion feature data. Establish an online adaptive mechanism for the model: Calculate the statistical distribution of fusion feature data for the current batch of sapphire substrates in real time; When a deviation is detected between the current statistical distribution and the model training set distribution, the probability output by the defect classification and quantization model is recalibrated using temperature scaling or Platt scaling methods to correct the output confidence.

9. A device for detecting crystal structure defects in a sapphire substrate, characterized in that, include: The expected defect module is used to obtain the process history information of the sapphire substrate to be tested, and to query the associated typical defect types and distribution characteristics from the preset defect knowledge base based on the process history information as expected defect information. The first detection module is used to perform a full-field scan of the sapphire substrate using a first detection technology to obtain a first distribution image; The defect identification module is used to perform image analysis on the first distribution image based on the expected defect information to identify one or more candidate defect regions; The second detection module is used to locally scan the candidate defect region using a second detection technique with a spatial resolution or spectral resolution greater than that of the first detection technique to obtain a second distribution image. The type of the second detection technique is selected based on the defect type in the expected defect information. The second detection technique is microscopic spectral imaging technology. Obtaining the second distribution image includes: scanning the candidate defect region using excitation light or an electron beam of a selected wavelength; acquiring the spectral signal generated at each scanning point; analyzing the spectral signal at each scanning point to extract one or more quantitative parameters; and generating the second distribution image based on the quantitative parameters of all scanning points. The registration and fusion module is used to spatially register and fuse the first distribution image and the second distribution image of the same candidate defect region to generate multimodal fusion feature data of the candidate defect region. The defect classification module is used to input the multimodal fusion feature data into the trained defect classification and quantization model to obtain the defect information of the candidate defect region; The report results module is used to generate a defect detection report and quality control judgment results for the sapphire substrate based on the defect information of all the candidate defect regions; to locate the process steps that cause the defects based on the correlation analysis between the defect detection report and the process history information; and to generate adjustment parameter suggestions for the process steps based on a preset process optimization strategy library.

Images