Wafer defect detection method and device

Abstract

The invention provides a wafer defect detection method and device, and the method comprises the steps: obtaining a macroscopic image comprising the global or regional defect information of the surface of a wafer, and generating a rechecking task set based on the abnormal features in the macroscopic image; under the condition that the re-checking task set is not empty, driving the imaging unit and the wafer to generate relative displacement so as to obtain a microscopic feature image of the target area in the re-checking task set; determining the native defect attribute of the wafer according to the microscopic feature image; wherein the optical resolution of the microscopic feature image is higher than that of the macroscopic image.

Description

Technical Field

[0001] This disclosure relates to the field of semiconductor manufacturing inspection technology, and in particular to a method and apparatus for detecting defects in wafers. Background Technology

[0002] Wafers are the core material in semiconductor manufacturing, and their inherent surface defects (such as scratches, dislocations, and holes) directly affect chip yield and reliability. Existing inspection technologies mainly rely on manual visual inspection combined with microscopic operation, which suffers from low efficiency, high subjectivity, and high false negative and missed detection rates, and cannot meet the high-throughput requirements of large-scale production.

[0003] Some inspection equipment uses traditional visual algorithms, but the complex texture and high noise of wafer surfaces make it difficult for traditional algorithms to distinguish between real defects and artifacts. Furthermore, existing equipment mostly performs single macroscopic or microscopic inspections, lacking macro-microscopic coordination, resulting in low automation and requiring manual intervention. In addition, the application of deep learning in defect identification is limited by scarce samples, the difficulty of model deployment, and the inability to effectively feed detection results back to the crystal pulling process, making it difficult to form a closed-loop optimization.

[0004] In summary, existing technologies suffer from problems such as high reliance on manual labor, low efficiency, insufficient accuracy, poor macro-micro coordination, and lack of process feedback. There is an urgent need for a fully automated, high-precision detection solution with process feedback capabilities. Summary of the Invention

[0005] This disclosure provides a method and apparatus for wafer defect detection; it can solve the technical problems of high reliance on manual labor, low efficiency, inconsistent judgment standards, insufficient macro-micro detection coordination and lack of process feedback in existing wafer defect detection.

[0006] The technical solution disclosed herein is implemented as follows: In a first aspect, this disclosure provides a wafer defect detection method. The method first acquires a macroscopic image containing global or regional defect information on the wafer surface, and generates a verification task set based on the abnormal features in the macroscopic image. When the verification task set is not empty, the imaging unit is driven to move relative to the wafer to acquire a microscopic feature image of the target area in the verification task set, and the optical resolution of the microscopic feature image is higher than that of the macroscopic image. Finally, the original defect attributes of the wafer are determined based on the microscopic feature image.

[0007] Secondly, this disclosure provides a wafer defect detection device, which includes a macroscopic imaging module, a microscopic imaging module, a motion control module, and a processor. The macroscopic imaging module is used to acquire macroscopic images of the wafer surface; the microscopic imaging module has a higher optical resolution than the macroscopic imaging module and is used to acquire microscopic feature images of the wafer surface; the motion control module is used to support the wafer and drive it to move between the macroscopic imaging module and the microscopic imaging module; the processor is configured to execute the above-described detection method.

[0008] This disclosure provides a wafer defect detection method and apparatus that, through the coordinated use of macroscopic and microscopic imaging, combined with mechanical motion control and deep learning analysis, achieves full automation of the wafer native defect detection process, completely eliminating reliance on manual labor and significantly improving detection efficiency. By rapidly screening macroscopically to narrow the microscopic detection range, the workload of microscopic inspection is reduced. At the same time, the use of high-resolution microscopic imaging and deep learning models ensures the accuracy and repeatability of defect identification. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the modular architecture of a wafer defect detection system provided in this disclosure.

[0010] Figure 2 This is a schematic flowchart of a wafer defect detection method provided in this disclosure.

[0011] Figure 3 This is a schematic diagram illustrating the principle of line scan stitching for macroscopic image acquisition provided in this disclosure.

[0012] Figure 4 This is a flowchart illustrating the process of generating a set of review tasks provided in this disclosure.

[0013] Figure 5 This is a schematic diagram of the autofocus unit structure of a microscopic imaging module provided in this disclosure.

[0014] Figure 6 This is a schematic diagram of the network architecture of a defect analysis model provided in this disclosure.

[0015] Figure 7 This is a schematic diagram illustrating the connection between a testing device and upstream and downstream equipment on a production line, as provided in this disclosure.

[0016] Figure 8 This is a schematic diagram illustrating the actual working process of a detection device provided in this disclosure. Detailed Implementation

[0017] The technical solutions in this disclosure will now be clearly and completely described with reference to the accompanying drawings.

[0018] The automated method and apparatus for detecting native defects in wafers disclosed herein are mainly applied to wafer inspection stations in semiconductor production lines. They are applicable to wafers of different specifications and can detect native defect types including crystal origin grains, dislocations, stacking faults, scratches, holes, oxygen deposits, metal contamination, and surface particles.

[0019] like Figure 1 As shown, the detection system disclosed herein adopts a modular architecture design, divided into a hardware layer 101, a software layer 102, and an application layer 103. These three layers work collaboratively from top to bottom to form a complete detection system. The hardware layer 101 is the physical foundation of the system, including a macroscopic imaging module 1011, a microscopic imaging module 1012, a motion control module 1013, and a processor 1014. Each hardware module is connected through standardized interfaces to ensure stable signal transmission and scalability. The software layer 102 is the core control hub of the system, including an image preprocessing module, a defect identification module, motion control software, data management software, and model deployment software. The software modules, based on the physical resources provided by the hardware layer 101, realize core functions such as data processing, motion control, and intelligent identification. The application layer 103 is the system's external interaction interface, including interfaces for detection task management, result visualization, process feedback, and integration with the production line's MES (Manufacturing Execution System) / LIMS (Laboratory Information Management System), facilitating operator operation and production data integration.

[0020] This disclosure first provides a method for detecting defects in a wafer, referring to... Figure 2 Specifically, this may include steps S210 to S330.

[0021] In step S210, a macroscopic image including global or regional defect information of the wafer surface is acquired, and a set of verification tasks is generated based on the abnormal features in the macroscopic image.

[0022] In some exemplary embodiments of this disclosure, the core of the macroscopic image acquisition step 210 is to acquire a macroscopic image containing global or regional defect information of the wafer surface, providing a basis for subsequent screening of suspected defects. This step is mainly carried out by... Figure 1 The macroscopic imaging module 1011 and motion control module 1013 in hardware layer 101 work together to complete this task.

[0023] In this step, the macroscopic imaging module 1011 uses an industrial camera as the imaging device, along with a uniform light source, to image the wafer surface. Considering the large size of the wafer, a single image is unlikely to cover the entire wafer surface; therefore, a multi-line scan stitching method is used to obtain a complete macroscopic image.

[0024] Specifically, refer to Figure 3 The motion control module 1013 drives the wafer carrier stage 301 to move the wafer 302 at a constant speed along a set direction. The macroscopic camera follows a fixed line scan path in the opposite direction. During the wafer's movement, the macroscopic camera continuously acquires line scan images at a preset sampling frequency. Each line scan has a certain overlap rate to ensure the accuracy of subsequent stitching. After acquiring multiple line scan images through multiple line scans, the images are then processed by… Figure 1 The image preprocessing module in software layer 102 calls the image stitching algorithm to synthesize a complete macroscopic image of the wafer.

[0025] The optical resolution of the macroscopic image is set to a level that can effectively identify obvious defects while ensuring imaging speed to meet the requirements of high-throughput detection. This resolution parameter is determined through... Figure 1 Configure the inspection task management interface of application layer 103 to adapt to the inspection needs of wafers of different specifications.

[0026] After obtaining the aforementioned macroscopic images, the images are first preprocessed, including noise removal and contrast enhancement, to improve image quality and facilitate the extraction of suspected defect areas. Noise removal employs a Gaussian filtering algorithm, with the image preprocessing module adaptively adjusting the Gaussian kernel size based on the noise intensity of the macroscopic image. Contrast enhancement uses a histogram equalization algorithm to enhance the grayscale difference between defects and the background, making defect features more prominent.

[0027] Then, image processing algorithms are used to extract suspected defect regions from the macroscopic image. In this embodiment, a threshold segmentation algorithm combined with morphological processing is used for suspected defect extraction. The specific process is as follows: First, the defect recognition module determines the threshold using an automatic thresholding algorithm, converting the macroscopic image into a binary image. Regions with gray values ​​higher than the threshold are identified as suspected defect regions, while regions with gray values ​​lower than the threshold are identified as background regions. Then, morphological opening operations (erosion followed by dilation) are performed on the binary image to remove small noise points. Next, morphological closing operations (dilation followed by erosion) are performed to fill small holes inside the defect region. Finally, connected component analysis is used to extract connected components with areas greater than a preset threshold as suspected defect regions.

[0028] After extracting suspected defect areas, the defect recognition module calculates the first coordinates of each suspected defect area, which is the position coordinate of the suspected defect area in the macroscopic image coordinate system. The first coordinates are determined with respect to the set origin of the macroscopic image and the direction of the set coordinate axes, and the coordinate values ​​are in pixels.

[0029] Finally, the data management software encapsulates the first coordinates, area size, and corresponding acquisition parameters (such as lighting conditions and camera parameters during imaging) of all suspected defect areas into subtasks within the verification task set. Each subtask corresponds to one suspected defect area, and the verification task set is the collection of all subtasks. If no suspected defect area is detected in the macroscopic image, the verification task set is empty, and the motion control software directly triggers the material unloading process without further microscopic inspection.

[0030] In step S220, if the verification task set is not empty, the imaging unit is driven to move relative to the wafer to obtain a microscopic feature image of the target region in the verification task set.

[0031] When the aforementioned set of verification tasks is not empty, i.e., when defects are found in some areas, the driving imaging unit undergoes relative displacement with the wafer, moving the suspected defective area of ​​the wafer within the field of view of the microscopic imaging module 1012, preparing for the acquisition of microscopic feature images. This step is mainly performed by... Figure 1 The motion control module 1013 and processor 1014 in hardware layer 101 work together to complete this task.

[0032] In this embodiment, the motion control module 1013 can be a multi-axis linear motor slide. The wafer carrier stage 301 is mounted on the horizontal slide to drive the wafer 302 to move in the horizontal plane. The microscopic imaging module 1012 is mounted on the vertical slide to adjust the distance (focusing) between the microscopic imaging lens and the surface of the wafer 302. The motion control module 1013 has high-precision motion and feedback capabilities, which can meet the precise alignment requirements between macroscopic coordinates and microscopic field of view. Its motion parameters are determined by... Figure 1 Configure the motion control software in software layer 102.

[0033] When the set of review tasks is not empty, the processor 1014 extracts the first coordinates of the first subtask from the set of review tasks, converts them into mechanical coordinates of the motion control module 1013, and then sends a physical positioning command to the motion control module 1013. Based on the mechanical coordinates, the motion control module 1013 drives the horizontal slide to move the wafer 302, aligning the center of the suspected defect area with the center of the field of view of the microscopic imaging lens.

[0034] After completing the microscopic image acquisition of a subtask, the processor 1014 extracts the next subtask from the verification task set and repeats the above process until microscopic image acquisition has been completed for all suspected defect areas corresponding to all subtasks. In order to reduce the total movement stroke of the slide and improve the detection efficiency, the motion control software uses a path planning algorithm to plan the paths of the subtasks in the verification task set and optimize the movement sequence of the slide.

[0035] The microscopic imaging module 1012 has a higher optical resolution than macroscopic images, enabling it to clearly present the microscopic structure and features of defects. Its core components include an industrial microscope camera, a high-magnification objective lens, and a multimodal illumination unit. During imaging, the multimodal illumination unit automatically switches illumination modes based on the type of suspected defect, and the industrial microscope camera works in conjunction with the high-magnification objective lens for imaging. Before acquiring microscopic feature images, autofocus is required to ensure image sharpness. It should be noted that the multimodal illumination unit can include at least one of coaxial lighting, dark-field illumination, or back illumination. Coaxial lighting includes bright-field illumination and DIC illumination, where DIC involves adding a DIC prism for polarization under bright-field illumination.

[0036] After focusing, the microscopic imaging camera begins acquiring images of the microscopic features of the suspected defect area. Acquisition parameters (including image resolution, exposure time, gain, etc.) are determined by... Figure 1 The image preprocessing module in software layer 102 is dynamically adjusted according to the defect type and imaging requirements. The acquired microscopic feature images are transmitted to processor 1014 through the image transmission interface. Processor 1014 calls the image preprocessing module to perform preprocessing such as noise removal and contrast enhancement on the images, and then stores them in data storage module 1007.

[0037] In step S230, the native defect properties of the wafer are determined based on the microscopic feature image.

[0038] In this embodiment, the process of determining the native defect attributes of a wafer based on microscopic feature images can be implemented using a defect analysis model. The defect analysis model employs a combined architecture of a convolutional neural network and a defect analysis model, where the defect analysis model is used for precise localization of defect regions, and the convolutional neural network model is used for accurate classification of defect categories. The defect analysis model is deployed in the processor 1014 using model deployment software, and the model parameters are stored in the data storage module. Figure 1 The interface of application layer 103 is updated.

[0039] The defect analysis model is trained using a large number of wafer defect samples, including various defect types to ensure dataset diversity and coverage. Specifically, processor 1014 inputs preprocessed microscopic feature images into the defect analysis model. The model first extracts defect regions from the images, determining their locations and bounding boxes. Then, it crops the defect region image and inputs it into a convolutional neural network model for feature extraction and classification, outputting native defect attributes, which may include defect category labels and confidence levels. Category labels must cover at least one of the following: crystal origin grains, dislocations, and stacking faults. The confidence level represents the model's reliability of the classification results. A confidence threshold is set by the model deployment software. When the confidence level is greater than the threshold, the classification result is considered valid; when the confidence level is less than the threshold, it is marked as a suspected defect, and a manual review reminder is issued through the result visualization interface of application layer 103.

[0040] Once the native defect attributes are determined, the processor 1014 stores information such as native defect attributes (category label, confidence level), defect location (macroscopic coordinates, mechanical coordinates), defect size, and acquisition time to the data storage module, and the data management software generates an inspection report. The inspection report can be displayed through the result visualization interface of the application layer 103, presenting the distribution of defects on the wafer surface in the form of a heatmap, and also providing functions such as batch statistics and trend analysis to facilitate quality control and production management by staff.

[0041] This system achieves fully automated detection of native wafer defects through a two-stage detection model combining macroscopic screening and microscopic identification. Macroscopic imaging rapidly covers the entire wafer surface, identifying suspected defect areas and reducing the workload of microscopic inspection. Microscopic imaging provides high-resolution images, which, combined with a deep learning model, ensures the accuracy and repeatability of defect identification. Furthermore, the modular architecture design provides excellent scalability, allowing the system to adapt to different detection needs by adding or removing hardware modules or updating software algorithms.

[0042] In some exemplary embodiments of this disclosure, reference is made to Figure 4 The process of generating the review task set may include steps S410 to S430.

[0043] In step S410, an image processing algorithm is used to extract the suspected defect region in the macroscopic image, and the first coordinates of the suspected defect region are calculated.

[0044] In some exemplary embodiments of this disclosure, the extraction of suspected defect areas is fundamental to the generation of the review task set. A combination of various image processing algorithms can be employed to improve the accuracy of suspected defect area extraction. Edge detection algorithms and texture analysis algorithms can be used to optimize the extraction effect.

[0045] Specifically, the edge detection algorithm employs operators with noise resistance and edge localization accuracy. The process is as follows: First, Gaussian filtering is applied to the macroscopic image to remove noise interference; then, the gradient magnitude and direction of the image are calculated; next, non-maximum suppression is applied to the gradient magnitude to preserve local maxima at the edges; finally, edge pixels are determined using a double thresholding method and connected to form a complete defect edge. This edge detection algorithm effectively extracts the contour information of defects, aiding in the determination of the boundaries of suspected defect areas.

[0046] Texture analysis algorithms employ gray-level co-occurrence matrices to extract texture features (such as energy, entropy, contrast, and correlation) by calculating the spatial distribution of gray values ​​in an image. The texture features of defect areas on a wafer surface differ significantly from those of the background area. By calculating these texture features and setting appropriate thresholds, suspected defect areas can be further filtered out, reducing the false extraction of defects.

[0047] Specifically, the image preprocessing module divides the macroscopic image into several sub-blocks, calculates the gray-level co-occurrence matrix and corresponding texture feature values ​​for each sub-block, and then calls a classifier to classify each sub-block as either a defect sub-block or a background sub-block. Adjacent defect sub-blocks are merged to form suspected defect regions, which are then fused with the results of threshold segmentation and edge detection to finally determine the location and boundaries of the suspected defect regions. During the fusion process, a weighted voting mechanism is used, assigning different weights to the results of different algorithms to improve the reliability of the extraction.

[0048] After preprocessing is complete, the first coordinates can be calculated. The core of this calculation is to determine the precise location of the suspected defect area in the macroscopic image coordinate system, providing a foundation for subsequent coordinate transformations. This function is performed by... Figure 1 The defect identification module of software layer 102 is implemented. In this embodiment, the first coordinate is represented by a combination of center coordinates and boundary coordinates, which can clearly identify the center position of the suspected defect area and accurately describe its shape and size.

[0049] The center coordinates are calculated as the average coordinates of all pixels in the suspected defect area; the boundary coordinates are calculated as the extreme coordinates (minimum X coordinate, maximum X coordinate, minimum Y coordinate, maximum Y coordinate) of the suspected defect area. The width and height of the suspected defect area can be calculated using the boundary coordinates, and thus its area can be determined.

[0050] To improve the accuracy of coordinate calculations, a sub-pixel-level localization algorithm can be employed to perform sub-pixel-level fitting on the edges of suspected defect areas. Specifically, an edge detection algorithm extracts the pixels of the defect edges, and then a fitting algorithm is used to fit the edge pixels with straight lines or curves to obtain sub-pixel-level edge equations, thereby calculating the sub-pixel-level center and boundary coordinates. This algorithm can achieve an accuracy of 0.1 pixels, ensuring accurate conversion from macroscopic coordinates to mechanical coordinates in subsequent processes.

[0051] In step S420, the first coordinates are converted into physical positioning commands based on the spatial mapping relationship.

[0052] After obtaining the first coordinate, a coordinate transformation can be performed. The purpose of the coordinate transformation is to convert the first coordinate in the macroscopic image coordinate system into the mechanical coordinate of the motion control module 1013, thereby achieving the physical positioning of the suspected defect area. The macroscopic camera lens exhibits distortion (such as radial distortion and tangential distortion), which can cause a deviation between the pixel coordinates in the macroscopic image and the actual physical coordinates. If this is not corrected, it will affect the accuracy of subsequent coordinate mapping. This embodiment uses a polynomial distortion correction algorithm to correct the first coordinate, eliminating the influence of lens distortion. This algorithm is integrated into... Figure 1 In the motion control software of software layer 102.

[0053] First, the macroscopic camera needs to be calibrated to obtain distortion parameters. The calibration process adopts a standard calibration method. A checkerboard calibration board is prepared and placed on the wafer carrier stage 301. Images of the calibration board at different angles and positions are captured by the macroscopic camera. Then, the motion control software performs corner detection on each calibration board image and extracts the corner coordinates of the checkerboard. Finally, the camera's intrinsic parameters (focal length, principal point coordinates) and distortion parameters (radial distortion coefficient, tangential distortion coefficient) are solved by the calibration algorithm, and these parameters are stored in the data storage module.

[0054] Based on the obtained distortion parameters, the motion control software uses a preset distortion correction formula to correct the first coordinate, obtaining the corrected coordinate, effectively eliminating the influence of lens distortion. The accuracy of the corrected coordinate is improved by an order of magnitude compared to the uncorrected coordinate, providing precise input for subsequent affine transformations.

[0055] The affine transformation matrix is ​​used to establish the mapping relationship between the macroscopic image coordinate system and the machine coordinate system, achieving accurate coordinate transformation. The calibration and application of this matrix are determined by... Figure 1 Motion control software implementation in software layer 102.

[0056] The calibration process for the affine transformation matrix is ​​as follows: First, a calibration board with multiple marker points is prepared, and the marker points are evenly distributed on the calibration board. The actual physical coordinates of each marker point in the mechanical coordinate system are measured using a laser interferometer. Then, the calibration board is placed on the wafer carrier stage 301, and an image of the calibration board is captured by a macroscopic camera. The coordinates of each marker point in the macroscopic image coordinate system (coordinates after distortion correction) are extracted.

[0057] Based on the macroscopic and mechanical coordinates of the marked points, the motion control software uses the least squares method to solve for the affine transformation matrix. The affine transformation matrix contains scaling, rotation, and translation coefficients, which accurately map the macroscopic coordinates to the mechanical coordinates. The solved affine transformation matrix is ​​stored in the data storage module and can be periodically recalibrated and updated according to equipment maintenance needs.

[0058] During the coordinate transformation process, the motion control software substitutes the distortion-corrected first coordinates into the affine transformation formula to obtain the corresponding mechanical coordinates. This transformation process is highly real-time, with a transformation delay of less than 1ms, meeting the real-time requirements of motion control.

[0059] After the mechanical coordinates are generated, they need to be combined with adsorption error compensation and path planning to generate the final physical positioning command. This process is carried out by... Figure 1 Motion control software implementation in software layer 102.

[0060] When a wafer is adsorbed onto the wafer carrier stage 301, a slight displacement (adsorption error) may occur, causing a deviation between the actual position and the ideal position. Therefore, adsorption error compensation is required. The specific method for adsorption error compensation is as follows: After the wafer is adsorbed, the edge or notch of the wafer (such as the wafer positioning notch) is photographed by a macroscopic camera. The motion control software extracts the actual coordinates of the edge or notch and compares them with the ideal coordinates to calculate the adsorption error. Then, this adsorption error is superimposed on the mechanical coordinates obtained by affine transformation to obtain the compensated mechanical coordinates.

[0061] The path planning employs a path optimization algorithm to optimize the movement sequence of the slide table, reducing the total movement distance and time. Finally, the motion control software generates physical positioning commands based on the compensated mechanical coordinates and the planned path, including parameters such as the slide table's movement distance, speed, and acceleration, ensuring smooth and precise movement. These parameters are sent to [the relevant authority / system] via a standardized interface. Figure 1 The motion control module 1013 of the hardware layer 101 drives the slide table to perform motion.

[0062] The distortion correction eliminated the influence of lens distortion, the affine transformation matrix achieved a precise mapping between macroscopic coordinates and mechanical coordinates, and the positioning command was optimized through adsorption error compensation and path planning to ensure that the suspected defect area can accurately fall into the center of the field of view of the microscopic imaging lens with a positioning error of less than ±1μm, which meets the precise positioning requirements of microscopic detection.

[0063] In step S430, the physical positioning command and the corresponding acquisition parameters are encapsulated into subtasks in the verification task set.

[0064] Task encapsulation involves packaging relevant information about suspected defective areas into subtasks, forming a set of review tasks. This function is provided by... Figure 1 The data management software is implemented in software layer 102. In this embodiment, each subtask includes the following information: subtask number, wafer ID, first coordinates (center coordinates, boundary coordinates), mechanical coordinates, acquisition parameters (light source type, light intensity, camera exposure time, gain during macroscopic imaging), and preliminary defect judgment results (such as the defect type preliminarily judged through macroscopic images).

[0065] Recording the acquired parameters helps in subsequent analysis of the impact of different imaging conditions on defect extraction and optimizes the parameter settings for macroscopic imaging. The review task set is stored in a standardized data format, facilitating data transmission and parsing. This format supports [interface name missing - likely a specific technology or protocol]. Figure 1 The application layer 103 is compatible with the MES / LIMS interface, enabling seamless data flow.

[0066] By integrating multiple image processing algorithms, the accuracy of extracting suspected defect areas is improved, and the false extraction of defects is reduced. Subpixel-level first coordinate calculation ensures coordinate accuracy. Standardized task encapsulation provides a standardized data foundation for subsequent microscopic inspection and data management, ensuring the smoothness of the entire inspection process and data consistency.

[0067] In some exemplary embodiments of this disclosure, macroscopic image acquisition employs a multi-line scan stitching method, by... Figure 1 The macro camera, line scan driver module, and motion control module 1013 of the hardware layer 101 work together to complete the task.

[0068] The line scan drive module controls the sampling frequency and line scan path of the macro camera. The sampling frequency is determined based on the wafer movement speed and image resolution to ensure that the overlap rate between adjacent line scan images meets the stitching requirements. In this embodiment, the wafer movement speed is set to a constant speed, and the sampling frequency is matched with the movement speed to keep the overlap rate of adjacent line scan images within a reasonable range, ensuring stitching accuracy while avoiding excessive redundant data.

[0069] The motion control module 1013 drives the wafer carrier stage 301 to move the wafer 302 along a set direction. The macro camera is fixedly installed in the opposite direction. During the wafer movement, the macro camera continuously acquires line scan images at a set sampling frequency. The number of acquisitions is determined based on the wafer size and the camera's field of view to ensure complete coverage of the wafer surface.

[0070] During the acquisition process, a uniform white LED light source is used. The light intensity is adjusted by a light source controller to ensure uniform illumination of the wafer surface and reduce the impact of reflections and shadows on image quality. The light intensity is dynamically adjusted according to the roughness and reflectivity of the wafer surface to ensure clear imaging.

[0071] Image stitching by Figure 1 The image stitching module of software layer 102 is implemented using a stitching algorithm based on feature point matching. The specific process includes four steps: feature point extraction, feature point matching, image registration, and image fusion.

[0072] Feature point extraction employs an algorithm suitable for image stitching, capable of extracting stable feature points from images unaffected by image scaling, rotation, and lighting changes. This algorithm extracts a sufficient number of feature points from each line scan image, with each feature point containing location, scale, and orientation information.

[0073] Feature point matching employs a fast matching algorithm to quickly match feature points in adjacent line scan images and filter out feature point pairs with high matching scores. To improve matching accuracy, an algorithm is used to eliminate mismatched feature point pairs, retaining only the correct matching results.

[0074] Image registration calculates the transformation matrix between adjacent images based on matched feature point pairs, transforming the images to be stitched to the same coordinate system. The transformation matrix includes translation, rotation, and scaling parameters, which are obtained through a suitable algorithm.

[0075] Image fusion employs a linear fusion algorithm to smooth the transition of overlapping areas in the registered images, eliminating stitching gaps and generating a complete macroscopic image of the wafer. During the fusion process, a weighted average of pixel values ​​in the overlapping areas is applied to ensure a natural transition in the fused image, with no obvious stitching artifacts.

[0076] Image preprocessing by Figure 1 The background filtering module and illumination normalization module of software layer 102 are implemented to remove the effects of image noise and uneven illumination, and enhance the recognition of defect features.

[0077] The background filtering module employs adaptive background subtraction, which subtracts background components from the macroscopic image by establishing a background model of the wafer surface, highlighting defect areas. The background model is obtained through statistical analysis of local regions of the image and can adapt to variations in wafer surface texture and lighting differences, effectively removing background noise and artifacts.

[0078] The illumination normalization module employs an illumination compensation algorithm to adjust the brightness of unevenly illuminated areas in the image, making the illumination intensity of the entire image more uniform. Specifically, by calculating the illumination distribution of the image, it enhances the brightness of weakly illuminated areas and suppresses the brightness of strongly illuminated areas, eliminating the influence of illumination shadows and reflections on defect identification.

[0079] The preprocessed macroscopic image shows a significant improvement in the contrast between the defect area and the background, and effective suppression of noise and lighting interference, providing high-quality image data for subsequent extraction of suspected defect areas.

[0080] By employing multiple line scan stitching and a stitching algorithm based on feature point matching, complete acquisition of macroscopic images of large-size wafers was achieved. The stitched images have no obvious gaps and uniform resolution. Background filtering and illumination normalization preprocessing effectively improved image quality and enhanced the identification of defect features, laying a solid foundation for subsequent extraction of suspected defects.

[0081] In some exemplary embodiments of this disclosure, when acquiring a microscopic feature image of the target region in the set of review tasks, a laser displacement sensor can first be controlled to emit a probe beam toward the wafer surface and receive the reflected signal to generate real-time height measurement data. This height measurement data characterizes the vertical distance between the microscopic imaging unit and the wafer surface within the current field of view. Then, when the real-time height measurement data deviates from a preset focal length threshold, a driving signal is generated based on the deviation value to adjust the height of the microscopic imaging unit until the vertical distance conforms to the preset focal length range, and a microscopic feature image of the target region is acquired. (Mainly dependent on...) Figure 1 The microscopic imaging module 1012 and motion control module 1013 of the hardware layer 101 and the image preprocessing module of the software layer 102 work together.

[0082] Reference Figure 5The microscopic imaging module 1012 may include an autofocus unit, which is the core component of the microscopic imaging module 1012. The autofocus unit may include a laser displacement sensor 10121, a coarse focus adjustment module 10122, a sharpness evaluation module 10123, and a fine focus locking module 10124. The laser displacement sensor 10121 is mounted on one side of the microscopic imaging lens, maintaining a fixed distance from the lens, and is used to measure the distance between the wafer 302 surface and the laser displacement sensor 10121 in real time. The coarse focus adjustment module 10122 and the fine focus locking module 10124 are integrated into the microscopic imaging module 10122. Figure 1 In the motion control software of software layer 102, the movement of the vertical slide is controlled to adjust the distance between the microscopic imaging lens and the surface of wafer 302; the sharpness evaluation module 10123 is integrated into the image preprocessing module to calculate the sharpness of the microscopic image and provide a basis for fine focus locking.

[0083] The laser displacement sensor 10121 can quickly and accurately measure the height change of the wafer surface with a measurement accuracy of ±0.1μm and a sampling frequency of 1000Hz. Its measurement data is transmitted to the processor 1014 in real time through the data interface for coarse focus adjustment.

[0084] The purpose of coarse focus adjustment is to quickly adjust the focal plane of the microscopic imaging lens to the coarse focus range of the wafer surface, laying the foundation for fine focus locking. This process is carried out by... Figure 1 The motion control software in software layer 102 is dominant.

[0085] Specifically, the laser displacement sensor 10121 measures the height of the wafer surface (i.e., the distance between the laser displacement sensor 10121 and the wafer surface), and the processor 1014 calculates the current distance between the microscopic imaging lens and the wafer surface based on the fixed distance between the laser displacement sensor 10121 and the microscopic imaging lens. The ideal focusing distance of the microscopic imaging lens is determined based on the focal length and magnification of the objective lens. This parameter is stored in the data storage module and can be configured through the interface of the application layer 103.

[0086] The processor 1014 calculates the difference between the current distance and the ideal focusing distance, and then sends a motion command to the vertical slide, driving the vertical slide to move the microscopic imaging lens a corresponding distance, adjusting the focal plane to near the ideal focusing distance, i.e., the coarse focus range. The coarse focus adjustment speed is set to 1mm / s to ensure rapid adjustment while avoiding overshoot. After adjustment, the vertical slide stops moving, and the fine focus locking step begins.

[0087] The purpose of fine-focus locking is to precisely lock the fine-focus position within the coarse-focus range, ensuring optimal sharpness of the microscopic feature image. This process is performed by... Figure 1 The image preprocessing module and motion control software in software layer 102 work together to complete this task.

[0088] First, the sharpness evaluation module 10123 calculates the sharpness evaluation value of the microscopic image. In this embodiment, the variance method is used as the sharpness evaluation function; the larger the variance value, the sharper the image. Then, the motion control software uses a search algorithm to find the position with the largest sharpness evaluation value, i.e., the focus position.

[0089] The specific process is as follows: Initialize the position of the vertical slide to the position after coarse focus adjustment, acquire a microscopic image at this position, and calculate the sharpness evaluation value; drive the vertical slide to move at a set step size, acquire images and calculate the sharpness evaluation value; adjust the movement direction according to the changing trend of the sharpness evaluation value until the sharpness evaluation value reaches the maximum value, and the corresponding position is the fine focus position; drive the vertical slide to move to the fine focus position to complete the fine focus locking.

[0090] The step size for focusing is set according to the depth of field of the objective lens and the measurement accuracy, usually between 0.05μm and 0.1μm, to ensure that the accuracy of the focusing position meets the requirements of microscopic imaging.

[0091] After focusing, the microscopic imaging camera begins acquiring images of the microscopic features of the suspected defect area. Acquisition parameters (including image resolution, exposure time, gain, etc.) are determined by... Figure 1 The image preprocessing module of software layer 102 dynamically adjusts according to the defect type and imaging requirements, and the adjustment is based on the correspondence between historical imaging parameters and defect types stored in the data storage module.

[0092] The acquired microscopic feature images are transmitted to the processor 1014 via an image transmission interface (such as the Gig E Vision interface). The processor 1014 calls the image preprocessing module to perform preprocessing such as noise removal and contrast enhancement on the images, and then stores them in the data storage module. Noise removal uses a median filtering algorithm to remove salt-and-pepper noise; contrast enhancement uses an adaptive histogram equalization algorithm to enhance the detailed features of defects.

[0093] By employing coarse focus adjustment assisted by a laser displacement sensor and fine focus locking assisted by an image sharpness evaluation function, autofocus was achieved, solving the focusing problem for wafers of different thicknesses and roughnesses and ensuring the clarity of microscopic feature images. Testing showed that the autofocus time was less than 200ms, with a focusing accuracy of ±0.1μm. The acquired microscopic images clearly revealed the microscopic structure of defects, providing high-quality data support for subsequent defect attribute identification.

[0094] In some exemplary embodiments of this disclosure, reference is made to Figure 6 The defect analysis model adopts a combined architecture of convolutional neural network and defect analysis model, including input layer 601, convolutional layer 602, pooling layer 603, fully connected layer 604, and output layer 605.

[0095] The input layer 601 receives the preprocessed microscopic feature image. The image size is determined according to the model design requirements and can be a grayscale image or a color image. In this embodiment, a grayscale image of 640×640 pixels is used as input to reduce the amount of computation while ensuring the feature extraction effect.

[0096] Convolutional layer 602 uses multiple 3×3 kernels to extract features from the image, with the number of kernels gradually increasing from 64 to 256. Through multiple convolutional operations, it progressively extracts shallow features (such as edges and textures) and deep features (such as defect morphology and structure) from the image. The stride is set to 1, and edge padding is used to ensure that the image size remains unchanged after convolution. Specific convolution methods can also be customized according to user needs, which will not be elaborated here.

[0097] Pooling layer 603 uses a 2×2 max-pooling kernel with a stride of 2 to reduce the dimensionality of the feature map, thereby reducing computational cost while preserving key features and avoiding overfitting. In this embodiment, convolutional layers and pooling layers are alternated; for example, a total of 8 convolutional layers and 4 pooling layers are included.

[0098] The fully connected layer 604 includes two hidden layers and one output layer. The first hidden layer has 1024 neurons and the second hidden layer has 512 neurons. The fully connected layer maps the convolutional pooling features into a one-dimensional feature vector. The number of neurons in the output layer 605 is set according to the number of defect categories. In this embodiment, there are 8 defect categories, so the number of neurons in the output layer is 8, and the output layer outputs the probability value of each defect category.

[0099] The defect analysis model can adopt the YOLO series architecture for accurate localization of defect regions, and output the bounding box coordinates, confidence score, and class probability of the defect; the convolutional neural network model is used for defect classification, receiving the defect region image located by the defect analysis model, and outputting the defect category label and confidence score.

[0100] In terms of dataset preparation, the training dataset includes real defect samples and synthetic defect samples. Real defect samples are obtained by collecting wafer defect images from actual production, with no fewer than 1,000 real samples for each defect type. Synthetic defect samples are generated using simulation software, with no fewer than 5,000 synthetic samples for each defect type, ensuring the diversity and coverage of the dataset. In some cases, the training dataset can be expanded using synthetic defect images based on simulation or modeling; the specific expansion process will not be elaborated here.

[0101] In some example implementations of this disclosure, the dataset is divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set is used for training model parameters, the validation set is used for adjusting model hyperparameters (such as learning rate and regularization coefficient), and the test set is used for evaluating model performance. All samples are labeled, including the bounding box coordinates of defects and category labels. The labeling format adopts a standardized format for easy parsing during model training.

[0102] The loss function for training the defect analysis model can be a comprehensive loss function that includes bounding box regression loss, confidence loss, and class loss. The optimization algorithm is stochastic gradient descent, with the learning rate initially set to 0.001 and gradually decreasing as the number of training rounds increases.

[0103] The model inference process includes five steps: image preprocessing, defect localization, feature extraction, defect classification, and result output. In the image preprocessing stage, the acquired microscopic feature images are scaled to 640×640 pixels and normalized and noise-removing operations are performed. In the defect localization stage, the defect analysis model extracts defect regions from the image, determines the bounding box coordinates and confidence scores of the defects, and filters out defect regions with confidence scores greater than a threshold. In the feature extraction stage, multi-layer convolution and pooling operations are performed on the defect region images to extract deep features. In the defect classification stage, fully connected layers map the features to class probabilities, outputting the defect class label and confidence score. In the result output stage, the defect attribute information is transmitted to processor 1014 for storage and display.

[0104] In some exemplary embodiments of this disclosure, reference is made to Figure 7 The testing device 701 is connected to the loading device 702 and the unloading device 703 via the production line transfer track 704, forming an automated material flow channel. The loading device 702 uses a robotic arm to remove the wafer 302 to be tested from the wafer cassette and place it on the production line transfer track 704. The transfer track uses an adapted transfer method to transport the wafer 302 to the wafer carrier platform 301 of the testing device 701. After testing, the transfer track transports the wafer 302 to the unloading device 703, where the robotic arm places the wafer 302 into the corresponding wafer cassette, completing the testing process.

[0105] The inspection device 701 interacts with the crystal pulling equipment 705 and the MES / LIMS system via the data interaction module 706. The data interaction with the crystal pulling equipment 705 uses an adapted interface; the inspection device 701 sends process adjustment feedback signals to the crystal pulling equipment 705, and the crystal pulling equipment 705 sends crystal pulling process parameters to the inspection device 701. The data interaction with the MES / LIMS system uses a standardized interface; the inspection device 701 uploads inspection results, defect data, and equipment operating status to the system, and the system issues inspection tasks, process parameter thresholds, and other instructions to the inspection device 701.

[0106] Encrypted transmission protocols are used during data interaction to ensure data security and integrity. Inspection data is stored and transmitted in a standardized format, including fields such as wafer ID, inspection time, defect type, defect location, process parameters, and inspection results, facilitating data statistics and analysis by the system.

[0107] Reference Figure 8 In actual production, the working process of the testing device 1201 is as follows: First, step S801, task reception, is executed. The inspection device 701 receives the inspection task through the MES / LIMS system, including information such as wafer batch, inspection standard, and process parameter thresholds. Then, step S802, wafer loading, is executed. The loading device 702 places the wafer 302 to be inspected on the transport track, which transports the wafer to the wafer carrier stage 301 of the inspection device 701. The wafer carrier stage 301 fixes the wafer using vacuum adsorption. After loading, step S803, macroscopic inspection, is executed. The macroscopic imaging module 1011 acquires macroscopic images of the wafer, performs preprocessing and extracts suspected defects, generates a set of verification tasks, and then executes step S804, microscopic inspection, based on the verification task set. The motion control module 1013 drives the wafer carrier stage 301 to move, moving the suspected defect area into the field of view of the microscopic imaging module 1012, which performs automatic focusing and microscopic feature image acquisition. Then, step S805, defect identification, is executed. Processor 1014 calls the defect analysis model to identify defects in the microscopic feature image and determine the original defect attributes.

[0108] Further, after determining the original defect attributes, step S806, data storage and uploading, is executed. The detected data is stored in the data storage module and uploaded to the MES / LIMS system; then step S807, process feedback, is executed. A process adjustment feedback signal is generated based on the defect data and sent to the crystal pulling equipment 705; after completing the inspection, step S808, wafer unloading, is executed. After inspection, the transport track conveys the wafer 302 to the unloading equipment 703, which places the wafer into the corresponding wafer cassette; after completion, equipment maintenance can be performed. The inspection device 701 performs periodic self-diagnosis and maintenance, including lens cleaning, sensor calibration, and model updates, to ensure stable equipment operation.

[0109] In mass production scenarios, the 701 inspection device supports continuous inspection mode, enabling automatic inspection of multiple wafers without manual intervention. The equipment's inspection efficiency meets the high-throughput inspection requirements of large-scale production lines. Simultaneously, the equipment features a fault alarm function; when equipment malfunctions or wafer abnormalities occur during inspection, it promptly issues alarm signals and notifies personnel via the MES / LIMS system for handling.

[0110] The testing device can seamlessly connect with upstream and downstream equipment and management systems in semiconductor production lines, realizing the automation and intelligence of the testing process. Through encrypted data interaction protocols and standardized data formats, it ensures the security, integrity, and compatibility of data. The actual workflow is efficient and smooth, and the testing efficiency and accuracy can meet the needs of large-scale production, providing strong support for the quality control and process optimization of semiconductor companies.

[0111] It should be noted that the technical solutions described in this disclosure can be combined arbitrarily as long as they do not conflict.

[0112] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of detecting defects of a wafer, characterized by, The method comprises the following steps: acquiring a macro image including defect information of a global or regional wafer surface, and generating a review task set based on abnormal features in the macro image; in the case that the review task set is not empty, driving an imaging unit to relatively displace the wafer to acquire a micro feature image of a target region in the review task set; determining a native defect attribute of the wafer according to the micro feature image; wherein the optical resolution of the micro feature image is higher than that of the macro image.

2. The method of claim 1, wherein The method of generating the review task set based on the image comprises: extracting a suspected defect region in the macro image by using an image processing algorithm, and calculating a first coordinate of the suspected defect region; converting the first coordinate into a physical positioning instruction based on a spatial mapping relationship; encapsulating the physical positioning instruction and corresponding acquisition parameters as a subtask in the review task set.

3. The method of claim 2, wherein the step of detecting defects in the wafer is performed by a method comprising: The physical positioning instruction includes a target mechanical coordinate of a mechanical motion component; The method of converting the first coordinate into a physical positioning instruction based on a spatial mapping relationship comprises: inputting the first coordinate into a preset coordinate mapping model for processing to obtain a target mechanical coordinate for driving a mechanical motion component; wherein the coordinate mapping model is configured to perform at least one of the following processes: performing optical distortion correction on the first coordinate to eliminate nonlinear deformation in the image acquisition process; mapping the corrected image coordinate to a mechanical motion coordinate system based on an affine transformation matrix; and performing adsorption error compensation based on a pre-acquired wafer position offset.

4. The method of claim 1, wherein The method of acquiring a micro feature image of a target region in the review task set comprises: controlling a laser displacement sensor to emit a probe light beam to the wafer surface and receive a reflected signal to generate real-time height measurement data, which represents the vertical distance between the microscopic imaging unit and the wafer surface in the current field of view; when the real-time height measurement data deviates from a preset focal length threshold, generating a driving signal according to the deviation value to adjust the height of the microscopic imaging unit until the vertical distance meets the preset focal length range, and acquiring the micro feature image of the target region.

5. The method of claim 1, wherein The method of determining a native defect attribute of the wafer according to the micro feature image comprises: inputting the micro feature image into a defect analysis model to obtain the native defect attribute; wherein the defect analysis model is a classifier constructed based on a deep neural network; the native defect attribute includes a defect category label and a confidence level, and the category label covers at least one of a crystal origin particle, a dislocation, and a stacking fault.

6. The method of claim 5, wherein The method further comprises: statistically determining the spatial distribution density of the native defects on the wafer; generating a feedback signal for adjusting the thermodynamic process parameters of the upstream crystal growth equipment based on the spatial distribution density.

7. The method of claim 5, wherein the step of detecting defects in the wafer is performed by a method comprising: The construction or deployment of the defect analysis model comprises at least one of the following strategies: adopting model pruning or quantization techniques to perform lightweight processing on the deep neural network to adapt to the inference resources of the industrial computer; using synthetic defect images based on simulation or simulation to expand the training data set.

8. The method of claim 1, wherein The acquisition includes a macro image of global or regional defect information of the wafer surface, including: controlling the macro camera to perform multiple line scan imaging on the wafer; and splicing the images obtained by multiple line scans to generate the macro image; In the imaging process, the non-uniform texture interference of the wafer surface is eliminated by background filtering and illumination normalization algorithm.

9. A wafer defect detection apparatus characterized by comprising: Comprising: a macro imaging module for acquiring a macro image of the wafer surface; a micro imaging module having higher optical resolution than the macro imaging module, for acquiring a micro feature image of the wafer surface; a motion control module for carrying the wafer and driving it to move between the macro imaging module and the micro imaging module; a processor configured to perform the method of any one of claims 1 to 8.

10. The wafer defect inspection apparatus according to claim 9, wherein The micro imaging module further comprises an autofocus unit and a multi-modal illumination unit; The autofocus unit includes a laser displacement sensor for measuring the height variation of the wafer surface in real time to assist focusing; The multi-modal illumination unit is configured to provide at least one of coaxial light, dark field illumination or back illumination, and automatically switch the illumination mode according to the type of suspected defects.

Images