Artificial Intelligence-Based Wafer Surface Defect Detection Method and System

Abstract

The application provides a wafer surface defect detection method and system based on artificial intelligence. The method comprises the following steps: obtaining a first light source scheme, configuring a controllable light source unit, and combining an image acquisition unit to perform image acquisition on a target wafer to obtain a first image set; identifying a plurality of first defect points of the target wafer according to the first image set, and generating a first defect distribution; generating a second light source scheme based on the first defect distribution, reconfiguring the controllable light source unit, and combining the image acquisition unit to perform image acquisition on the target wafer again to obtain a second image set; identifying a plurality of second defect points of the target wafer according to the second image set, labeling the defect types of the second defect points, and combining the first defect distribution to generate a surface defect distribution of the target wafer. The technical problem of poor detection rate and precision of wafer surface defects in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of image inspection, and more particularly to a method and system for detecting wafer surface defects based on artificial intelligence. Background Technology

[0002] In current wafer surface defect detection processes, fixed illumination parameters are typically used to acquire images of the wafer surface, and then defect identification is performed based on the acquired images. However, different wafer surfaces produce different types of defects, and fixed illumination parameters are difficult to capture all types of defects. For example, some defects are not obvious under certain fixed illumination parameters, which leads to the acquired images not displaying the defects well, resulting in poor detection rate and accuracy of wafer surface defects. Summary of the Invention

[0003] This invention addresses the technical problem of poor detection rate and accuracy of wafer surface defects in existing technologies by providing an artificial intelligence-based method and system for wafer surface defect detection.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a wafer surface defect detection method based on artificial intelligence, comprising:

[0006] A first light source scheme is obtained, the controllable light source unit is configured, and the target wafer is image acquired in conjunction with the image acquisition unit to obtain a first image set;

[0007] Based on the first image set, multiple first defect points of the target wafer are identified, and a first defect distribution is generated;

[0008] Based on the first defect distribution, a second light source scheme is generated, the controllable light source unit is reconfigured, and the target wafer is image acquired again by the image acquisition unit to obtain a second image set;

[0009] Based on the second image set, multiple second defect points of the target wafer are identified, and the defect type of each second defect point is labeled. Combined with the first defect distribution, the surface defect distribution of the target wafer is generated.

[0010] Secondly, the present invention provides an artificial intelligence-based wafer surface defect detection system, comprising:

[0011] An initial light source module is used to acquire a first light source scheme, configure the controllable light source unit, and combine it with the image acquisition unit to acquire images of the target wafer and obtain a first image set.

[0012] The defect analysis module is used to identify multiple first defect points of the target wafer based on the first image set and generate a first defect distribution;

[0013] The light source module is optimized to generate a second light source scheme based on the first defect distribution, the controllable light source unit is reconfigured, and the image acquisition unit is combined to acquire images of the target wafer again to obtain a second image set;

[0014] An integrated output module is used to identify multiple second defect points of the target wafer based on the second image set, label the defect type of each second defect point, and generate the surface defect distribution of the target wafer by combining the first defect distribution.

[0015] The beneficial effects of this invention are:

[0016] Compared to existing technologies, this application first obtains a first light source scheme, configures a controllable light source unit, and combines it with an image acquisition unit to acquire images of the target wafer, obtaining a first image set. This enables intelligent reuse of historical data, improves the efficiency and accuracy of image acquisition, and provides necessary support for subsequent defect analysis. Secondly, based on the first image set, multiple first defect points on the target wafer are identified, generating a first defect distribution. This obtains the distribution of most defects on the wafer surface and provides a necessary foundation for obtaining a second defect distribution. Thirdly, based on the first defect distribution, a second light source scheme is generated, the controllable light source unit is configured again, and combined with the image acquisition unit, images of the target wafer are acquired again, obtaining a second image set. This effectively identifies potential defects, providing a reliable supplement to the first defect distribution and improving the accuracy and precision of defect detection. Finally, based on the second image set, multiple second defect points on the target wafer are identified, and the defect type of each second defect point is labeled. Combined with the first defect distribution, a surface defect distribution of the target wafer is generated, obtaining an accurate surface defect distribution of the target wafer.

[0017] Through the above technical solution, this application dynamically matches illumination configuration items from historical data based on the raw material characteristics and production process characteristics of the target wafer, and fully considers potential defects, significantly improving the detection rate and accuracy of wafer surface defect detection. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the artificial intelligence-based wafer surface defect detection method provided by the present invention.

[0019] Figure 2 This is a flowchart of the wafer defect identification network in the artificial intelligence-based wafer surface defect detection method provided by the present invention.

[0020] Figure 3This is a schematic diagram of the structure of the artificial intelligence-based wafer surface defect detection system provided by the present invention.

[0021] In the attached diagram, the components represented by each number are as follows:

[0022] Initial light source module 11, defect analysis module 12, optimized light source module 13, integrated output module 14. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0026] Example 1, as Figure 1 As shown, this embodiment of the invention provides a wafer surface defect detection method based on artificial intelligence, including:

[0027] S10: Obtain a first light source scheme, configure the controllable light source unit, and combine it with the image acquisition unit to acquire images of the target wafer to obtain a first image set;

[0028] In traditional wafer surface defect detection, fixed illumination parameters are typically used to acquire images of the wafer surface, and then defect identification is performed based on the acquired images. However, different wafer surfaces produce different types of defects, and fixed illumination parameters are difficult to capture all types of defects, resulting in low defect identification accuracy.

[0029] To address the aforementioned issues, this application retrieves multiple sample defect distributions based on the raw material characteristics and manufacturing process characteristics of the target wafer. Then, it obtains multiple historical light source schemes corresponding to the multiple sample defect distributions from the historical wafer production database. From these schemes, it selects illumination configuration items whose historical configuration frequency is greater than the preset configuration frequency, summarizes them to form a first light source scheme, configures a controllable light source unit, and combines it with an image acquisition unit to acquire images of the target wafer, thereby obtaining a first image set.

[0030] Specifically, step S10 in the method includes:

[0031] Obtain the raw material characteristics and manufacturing process characteristics of the target wafer;

[0032] Based on the characteristics of the raw materials and the characteristics of the production process, a historical wafer production database is retrieved, multiple sample wafers are selected, and the sample defect distribution of the multiple sample wafers is extracted to obtain multiple sample defect distributions.

[0033] Based on the defect distribution of the multiple samples, a first light source scheme for the target wafer is generated.

[0034] In this embodiment, the raw material characteristics and manufacturing process characteristics of the target wafer are first obtained. The obtained raw material characteristics include multi-dimensional feature vectors such as wafer substrate material (e.g., single-crystal silicon / polycrystalline silicon / compound semiconductor), doping type (e.g., N-type / P-type), and crystal structure (e.g., crystal orientation, defect density), which can reflect the inherent physical properties and optical characteristics of the wafer. The manufacturing process characteristics include multi-dimensional feature vectors such as lithography precision (e.g., 7nm process node), etching process (e.g., anisotropic / isotropic etching), thin film deposition (e.g., CVD / PVD process parameters), and mechanical polishing (e.g., pressure, time), which can reflect the potential defect types formed during wafer manufacturing. For example, the obtained material feature vector of the GaN wafer is [compound semiconductor, N-type doping, crystal orientation...]. <0001> The process feature vector for the 7nm FinFET wafer fabrication is [EUV lithography, copper interconnect, CMP pressure 5psi, low-k dielectric layer].

[0035] Secondly, based on raw material characteristics and manufacturing process characteristics, a historical wafer production database is retrieved to select multiple sample wafers. The defect distributions of these sample wafers are then extracted, resulting in multiple sample defect distributions. The historical wafer production database is built upon historical multi-dimensional data. Further, the retrieval process employs a hybrid metric method combining cosine similarity and Euclidean distance to match within the historical wafer production database: cosine similarity is used to calculate semantic relevance for raw material characteristics (nominal data) (e.g., the similarity between "silicon-based" and "GaN-based" is 0.3, and the similarity between "silicon-based" and "polycrystalline silicon" is 0.9); normalized Euclidean distance is used to measure differences for manufacturing process characteristics (numerical data) (e.g., similarity decreases by 20% when CMP pressure difference > 2 psi). This process selects 30-50 sample wafers with a similarity ≥ 0.85 (this threshold is determined based on a large amount of sample data). For example, the feature vector of the raw material is [Si-based material, N-type doping, 7nm EUV lithography, copper interconnect, CMP pressure 5psi]. A hybrid metric method that combines cosine similarity and Euclidean distance is used to select 42 samples with similarity ≥0.85 from the historical wafer production database.

[0036] Furthermore, the defect distribution of multiple sample wafers is extracted, including spatial distribution, type distribution, and depth distribution. Specifically, for each sample wafer, its three-dimensional defect distribution is extracted: 1. Spatial distribution: A polar coordinate system (e.g., accuracy ±1μm) is established with the wafer center as the origin, and the wafer surface is divided into 500×500 grids (e.g., resolution 100μm), and the defect density of each grid is calculated. 2. Type distribution: Based on morphological characteristics (e.g., aspect ratio) and spectral response, defect types are classified into particulate contamination (environmental dust, equipment wear), scratches (mechanical contact, polishing process abnormalities), voids (chemical vapor deposition (CVD) process defects), etc., and the proportion of each type of defect is then statistically analyzed. For example, for a certain wafer, particulate contamination accounts for 45%, scratches account for 30%, and voids account for 25%. 3. Depth Distribution: Utilizing the penetration characteristics of multi-wavelength light sources (e.g., 940nm infrared light penetrating a 50μm silicon layer), combined with optical slicing technology, a defect depth probability distribution function is constructed (e.g., P(d≤10μm)=0.7 indicates that 70% of defects are located within the surface 10μm). Finally, the DBSCAN clustering algorithm can be used to identify defect-dense areas, generating a multi-dimensional defect distribution containing spatial density, type proportion, and depth features. For example, defect distribution data from multiple sample wafers are extracted and integrated using the DBSCAN clustering algorithm as follows: spatial density (40% defect rate in edge regions), defect type (45% particulate contamination, 30% scratches, 25% voids), and depth distribution (25% subsurface defects), forming a three-dimensional defect distribution that reflects the potential defect distribution on the wafer surface under current raw material and manufacturing process characteristics.

[0037] In this way, based on the raw material characteristics and production process characteristics of the target wafer, the historical wafer production database is searched to select multiple similar sample wafers. Then, by extracting the defect distribution of multiple similar sample wafers, a bridge is built between historical data and real-time detection requirements, providing the necessary data foundation for subsequent determination of the light source scheme.

[0038] Finally, based on the defect distributions of multiple samples, a first light source scheme for the target wafer is generated. Specifically, multiple historical light source schemes corresponding to the defect distributions of multiple samples are obtained from the historical wafer production database. Each historical scheme sequence contains multiple illumination configuration items. Finally, based on the multiple historical light source schemes, a first light source scheme for the target wafer is generated.

[0039] Specifically, the step of "generating a first light source scheme for the target wafer based on the defect distribution of the plurality of samples" includes:

[0040] Multiple historical light source schemes corresponding to the defect distributions of the multiple samples are obtained from the historical wafer production database. Each historical light source scheme includes multiple illumination configuration items. Each illumination configuration item includes the wavelength parameter, angle parameter, polarization parameter and illumination mode parameter used by the controllable light source unit in a single image acquisition.

[0041] The first light source scheme is formed based on the multiple historical light source schemes.

[0042] In this embodiment of the application, multiple historical light source schemes corresponding to the defect distribution of multiple samples are first obtained from the historical wafer production database. The light source scheme is a sequence, and each historical light source scheme contains multiple illumination configuration items. Each illumination configuration item includes wavelength parameters (e.g., 250-1100nm), angle parameters (e.g., incident angle 0-90°, azimuth angle 0-360°), polarization parameters (e.g., linear polarization, circular polarization), and illumination mode parameters (e.g., dark field / bright field / combined mode) used by the controllable light source unit in a single image acquisition. For example, the defect distribution of multiple samples is categorized by spatial density (40% of defects are in the edge region), defect type (45% for particulate contamination, 30% for scratches, and 25% for voids), and depth distribution (25% for subsurface defects). Based on this, multiple historical light source schemes are obtained from the historical wafer production database. Each historical light source scheme contains multiple illumination configuration items. One of these illumination configuration items is: 532nm wavelength (wavelength parameter), 45° incident angle (angle parameter), horizontal polarization (polarization parameter), and annular dark field illumination (illumination mode parameter). In this way, a mapping relationship between defect distribution and light source scheme is established.

[0043] Secondly, based on the multiple historical light source schemes, the first light source scheme is formed. Specifically, the union of the illumination configuration items in the multiple historical light source schemes is taken to obtain an illumination configuration set. Then, the illumination configuration set is traversed, and the historical configuration frequency of each illumination configuration item in the illumination configuration set is counted. Finally, illumination configuration items with historical configuration frequencies greater than a preset configuration frequency are selected and summarized to form the first light source scheme. In this way, by selecting illumination configuration items with frequencies greater than a preset configuration frequency from multiple historical light source schemes, the first light source scheme is obtained, providing necessary support for subsequent light source scheme optimization.

[0044] Specifically, the phrase "forming the first light source scheme based on the plurality of historical light source schemes" includes:

[0045] The illumination configuration set is obtained by taking the union of the illumination configuration items in multiple historical light source schemes;

[0046] Traverse the set of lighting configurations, extract each lighting configuration item in turn, and count the historical configuration frequency of each lighting configuration item in the multiple historical light source schemes;

[0047] Lighting configuration items with historical configuration frequencies greater than preset configuration frequencies are selected and summarized to form the first light source scheme.

[0048] In this embodiment, the illumination configuration items in multiple historical light source schemes are first combined to obtain an illumination configuration set. For example, all illumination configuration items in multiple historical light source schemes are combined to obtain 100 illumination configuration sets, which reflect all historical illumination configurations corresponding to the sample defect distribution.

[0049] Secondly, the lighting configuration set is traversed, and each lighting configuration item is extracted sequentially. The historical configuration frequency of each lighting configuration item in multiple historical light source schemes is calculated. The historical configuration frequency of each lighting configuration item is calculated as: the number of times the lighting configuration item appears / the total number of lighting configurations in the set. For example, if the lighting configuration set is 100, lighting configuration item 1 (532nm, incident angle 45°, horizontal polarization, annular dark field) appears 55 times in the set, and lighting configuration item 2 (940nm, incident angle 15°, circular polarization, backscattered illumination) appears 42 times. Therefore, the historical configuration frequency of lighting configuration item 1 is 55 / 100 = 0.55, and the historical configuration frequency of lighting configuration item 2 is 42 / 100 = 0.42. The historical configuration frequency reflects the proportion of the lighting configuration item in the historical lighting schemes; a higher historical configuration frequency indicates a better lighting configuration item.

[0050] Finally, illumination configuration items with historical configuration frequencies greater than the preset configuration frequency are selected and summarized to form the first light source scheme. The preset configuration frequency recommended in this application is 0.4, which can be dynamically adjusted by those skilled in the art based on the actual sample size. Further, illumination configuration items with historical configuration frequencies greater than the preset configuration frequency (e.g., 0.4) are selected and summarized to form the first light source scheme. The summarization process can be based on descending order of the angle parameters of the illumination configuration items, thus generating the first light source scheme. For example, with a preset configuration frequency of 0.4, the following are selected: Illumination configuration item 1 (532nm, incident angle 45°, horizontal polarization, annular dark field) has a historical configuration frequency of 0.55, and illumination configuration item 2 (940nm, incident angle 15°, circular polarization, backscattered illumination) has a historical configuration frequency of 0.42. These are then sorted in descending order of angle parameters to obtain the first light source scheme: illumination configuration item 1 and illumination configuration item 2. This eliminates inefficient configurations and obtains a more optimal illumination configuration.

[0051] Furthermore, based on the acquired first light source scheme, a controllable light source unit is configured, and combined with an image acquisition unit, images of the target wafer are acquired to obtain a first image set. Specifically, according to the first light source scheme, multiple illumination configurations (multiple wavelength parameters, angle parameters, polarization parameters, and illumination mode parameters) in the first light source scheme are sequentially converted into specific control commands for the controllable light source unit. Combined with an image acquisition unit (such as a 12K resolution line scan camera), wafer images under different illumination configurations are sequentially acquired. Finally, multiple wafer images are summarized to obtain the first image set, thus providing reliable feature input for the subsequent defect recognition network.

[0052] In summary, compared with existing technologies, this application retrieves multiple sample defect distributions based on the raw material characteristics and manufacturing process characteristics of the target wafer. Then, it obtains multiple historical light source schemes corresponding to the multiple sample defect distributions from the historical wafer production database. Next, it selects illumination configuration items with historical configuration frequencies greater than preset configuration frequencies from the multiple historical light source schemes and summarizes them to form a first light source scheme. Finally, based on the obtained first light source scheme, it configures a controllable light source unit and combines it with an image acquisition unit to acquire images of the target wafer and obtain a first image set. In this way, by filtering multiple illumination configurations in similar historical data and acquiring images accordingly, it realizes intelligent reuse of historical data, improves the efficiency and accuracy of image acquisition, and provides necessary support for subsequent defect analysis.

[0053] S20: Based on the first image set, identify multiple first defect points of the target wafer and generate a first defect distribution;

[0054] The aforementioned steps have obtained a first image set. In order to more intuitively reflect the distribution of defects on the surface of the target wafer, multiple defect points on the target wafer and their corresponding spatial coordinates and defect types can be identified based on the first image set.

[0055] To address the aforementioned issues, this application, based on a first image set, uses a wafer defect recognition network to identify multiple first defect points in multiple target wafer images, along with their corresponding coordinates and defect types. Finally, these points are annotated on the three-dimensional structure of the target wafer to generate a first defect distribution.

[0056] Specifically, step S20 in the method includes:

[0057] The wafer defect identification network is retrieved, and multiple first images from the first image set are sequentially input into the wafer defect identification network to obtain multiple first defect points, the coordinate positions of each first defect point in the target wafer, and the defect type corresponding to each first defect point.

[0058] Obtain the three-dimensional structure of the target wafer;

[0059] Based on the coordinate positions of each first defect point in the target wafer, the plurality of first defect points and the defect types corresponding to each first defect point are labeled on the three-dimensional structure to generate the first defect distribution.

[0060] In this embodiment, a pre-trained wafer defect recognition network is first retrieved. Multiple first images from the first image set (one image is acquired for each illumination configuration in the first light source scheme) are sequentially input into the wafer defect recognition network to obtain multiple first defect points, the coordinates (three-dimensional coordinates) of each first defect point in the target wafer, and the defect type (such as particle contamination, scratches, voids, etc.) corresponding to each first defect point. For example, a first image from the first image set is input into the wafer defect recognition network to obtain multiple first defect points and their corresponding coordinates and defect types. The coordinates of a certain defect point are (5nm, -12nm, 3nm), and the defect type is void. This achieves accurate localization and type identification of the target wafer defect.

[0061] Secondly, the three-dimensional structure of the target wafer is obtained. Specifically, the three-dimensional structure of the target wafer is obtained by constructing a three-dimensional model of the target wafer. For example, the three-dimensional structure of the wafer can be constructed based on design files (such as GDS files) or process simulation data (such as TCAD) with an accuracy of ±5nm. Then, a three-dimensional coordinate system (x, y, z) is established, thus obtaining the three-dimensional structure of the target wafer.

[0062] Finally, based on the coordinate positions of each first defect point in the target wafer, multiple first defect points and their corresponding defect types are labeled in the three-dimensional structure to generate a first defect distribution. Specifically, based on the coordinate positions of each first defect point in the target wafer, all first defect points (x, y, z defect types) are labeled in the three-dimensional structure of the target wafer to form an intuitive three-dimensional first defect distribution.

[0063] Furthermore, the construction steps of the "wafer defect identification network" include:

[0064] Historical wafer images are collected from the historical wafer production database to form a sample wafer image set;

[0065] Image defect annotation is performed on the sample wafer image set to generate a sample wafer defect set and a sample defect type set;

[0066] Machine learning is used to generate a defect presence identification branch based on the sample wafer image set and the sample wafer defect set, and a defect type judgment branch is generated based on the sample wafer image set and the sample defect type set.

[0067] The defect presence identification branch and the defect type determination branch are connected and configured to obtain the wafer defect identification network.

[0068] In this embodiment, historical wafer images are first collected from a historical wafer production database to form a sample wafer image set. The image collection process is random sampling, meaning the collected images include both defective and defect-free images to improve the accuracy and robustness of the model detection. For example, 5000 historical wafer images (e.g., covering 28nm to 3nm processes) are first randomly collected from the historical database to form the sample wafer image set.

[0069] Secondly, image defect annotation is performed on the sample wafer image set to generate a sample wafer defect set and a sample defect type set. For example, each wafer image is manually labeled with its corresponding defect type to obtain the sample wafer defect set and sample defect type set. Further, an automatic annotation module can be used to reduce labor costs. For example, common defect types can be initially labeled based on threshold segmentation, and then complex defects (such as minor scratches) can be corrected by those skilled in the art. Simultaneously, weakly supervised annotation (only labeling the image level with "defect exists") is introduced to train the detection branch.

[0070] Again, such as Figure 2As shown, machine learning is used to generate a defect presence recognition branch based on a set of sample wafer images and a set of sample wafer defects, and a defect type judgment branch based on the set of sample wafer images and a set of sample defect types. Specifically, two branches are constructed based on machine learning: one is a defect presence recognition branch, used to identify whether there is a defect in the image; the other is a defect type judgment branch, used to determine the specific defect type of the image with a defect. The image will only enter the defect type judgment branch for type judgment if the defect presence recognition branch determines that the image has a defect; if the image has no defects, it will directly output "none".

[0071] For example, the defect detection branch can be based on the improved YOLOv7 framework, using ResNet50 as the backbone network. The first 10 convolutional layers extract multi-scale features, where the shallow layers (Conv1-Conv2_x) capture the edge details of defects at the 0.3μm level (such as the pixel-level contours of scratches), and the deep layers (Conv3_x) expand the receptive field through dilated convolutions to identify clumped defects larger than 50μm (such as large-area contamination). The embedded CBAM attention module suppresses interference from normal textures on the wafer surface through a dual channel and spatial attention mechanism, focusing on the defect region (such as enhancing the feature response of particulate contamination by 2 times). The feature fusion part adopts the FPN+PANet structure to generate three-level feature maps (P3-P5), corresponding to the detection of small, medium and large defects, respectively. The detection head outputs the coordinates of the defect bounding box (accuracy ±0.5μm) and confidence score through Focal Loss (to solve the imbalance of positive and negative samples) and GIoU Loss (to improve the regression accuracy of extreme-sized boxes), and locates the defect location after non-maximum suppression.

[0072] For example, the defect type judgment branch is based on the low-level features extracted by the first 5 convolutional layers of ResNet50. It compresses the spatial dimension through global average pooling and retains semantic information such as texture roughness and edge directionality of defects (such as the difference between the linear texture of metal bridging and the circular contour of particulate contamination). The fully connected layer performs two nonlinear transformations (such as 2048-dimensional → 16-dimensional) and combines Softmax to output the probability distribution of 16 types of defects. Cross-entropy loss is used to optimize the classification accuracy. In particular, differentiated feature weights are designed for visually similar defects (such as subsurface holes and surface pits) (such as enhancing the penetration depth features of infrared images).

[0073] Furthermore, the two branches share the parameters of the first 5 convolutional layers. The defect presence identification branch then enhances spatial localization through FPN+PANet, while the defect type judgment branch focuses on semantic abstraction through global pooling, forming a position-semantic complementarity. The total loss function balances the detection and classification tasks with a weight of 0.6:0.4. In this way, by sharing the parameters of the first 5 convolutional layers, the model size is reduced, and by first determining whether a defect exists and then identifying the defect type, the efficiency of defect recognition is improved.

[0074] Finally, the defect presence recognition branch and the defect type determination branch are connected and configured to obtain the wafer defect recognition network. Specifically, the wafer image is input into the wafer defect recognition network. First, the defect presence recognition branch determines whether there is a defect in the image. If there is no defect, "none" is output directly. If there is a defect, the defect type determination branch is entered to determine the specific defect type. In this way, defect recognition is performed on the wafer image.

[0075] In summary, compared to existing technologies, this application, based on a first image set, uses a wafer defect recognition network to identify multiple first defect points and their corresponding coordinates and defect types in multiple target wafer images. These are then annotated on the three-dimensional structure of the target wafer to generate a first defect distribution. This obtains the distribution of most defects on the wafer surface and provides the necessary foundation for obtaining the subsequent second defect distribution.

[0076] S30: Based on the first defect distribution, generate a second light source scheme, reconfigure the controllable light source unit, and combine it with the image acquisition unit to acquire images of the target wafer again to obtain a second image set;

[0077] The aforementioned steps acquired a first image set, which can reflect most of the defects in the target wafer. However, the target wafer may also have other potential defect distributions.

[0078] To address the aforementioned issues, this application, based on a first defect distribution, searches a historical wafer production database, filters out multiple historical wafer samples, removes samples from the first defect distribution, obtains a potential defect distribution, extracts first potential defect points and their corresponding coordinates and defect types from the potential defect distribution, then, based on the first potential defect types, retrieves multiple sets of candidate illumination configuration items with detection rate indicators, selects the candidate illumination configuration item with the highest detection rate as the first preferred illumination configuration item and adds it to the second light source scheme, finally configures a controllable light source unit, and combines it with an image acquisition unit to perform image acquisition on the target wafer again to obtain a second image set.

[0079] Specifically, step S30 in the method includes:

[0080] Based on the first defect distribution, the historical wafer production database is searched to select multiple historical wafer samples;

[0081] Extract the complete defect distribution of the multiple historical wafer samples, and take the union of the complete defect distributions of the multiple historical wafer samples to form the overall predicted defect distribution;

[0082] The first defect distribution is removed from the overall predicted defect distribution to obtain the potential defect distribution;

[0083] The second light source scheme is generated based on the potential defect distribution.

[0084] In this embodiment of the application, firstly, based on the first defect distribution, a historical wafer production database is retrieved to select multiple historical wafer samples. Specifically, based on the first defect distribution of the target wafer (the coordinate position of the first defect point in the target wafer and the corresponding defect type), multiple historical wafer samples are selected from the historical database, wherein the selected multiple historical wafer samples contain most of the defect distributions in the first defect distribution.

[0085] Secondly, the complete defect distributions of the multiple historical wafer samples are extracted, and the union of the complete defect distributions of the multiple historical wafer samples is taken to form the overall predicted defect distribution. Specifically, the union of the complete defect distributions of the multiple selected historical wafer samples (including potential defects not detected in the first detection) is taken to form the overall predicted defect distribution, which includes the first defect distribution and other potential defect distributions.

[0086] Next, the first defect distribution is removed from the overall predicted defect distribution to obtain the potential defect distribution. Specifically, the first defect distribution of detected defects is removed from the overall predicted defect distribution, and potential defect types that exist in historical samples but were not identified during the first detection of the target wafer are retained.

[0087] Finally, based on the potential defect distribution, the second light source scheme is generated. Specifically, a first potential defect point and its corresponding coordinates and defect type are extracted from the potential defect distribution. Then, based on the first potential defect type, multiple sets of candidate lighting configuration items with detection rate identifiers corresponding to the first potential defect type are retrieved. The candidate lighting configuration item with the highest detection rate is selected as the first preferred lighting configuration item. Finally, the first preferred lighting configuration item is added to the second light source scheme.

[0088] Specifically, the step of "generating the second light source scheme based on the potential defect distribution" includes:

[0089] Extract the first potential defect point from the potential defect distribution, and extract the first potential coordinates and the first potential defect type of the first potential defect point;

[0090] Based on the first potential defect type, the defect-lighting mapping database is retrieved to obtain multiple sets of candidate lighting configuration items corresponding to the first potential defect type. Each candidate lighting configuration item has a detection rate identifier.

[0091] The candidate illumination configurations are arranged in descending order according to the detection rate identifier, and the candidate illumination configuration with the highest detection rate is selected as the first preferred illumination configuration.

[0092] Add the first preferred lighting configuration to the second light source scheme.

[0093] In this embodiment of the application, firstly, a first potential defect point is extracted from the potential defect distribution, and then the first potential coordinates and the first potential defect type of the first potential defect point are extracted. For example, the potential defect distribution is traversed to extract all first potential defect points and their corresponding first potential coordinates (e.g., (4nm, -10nm, 5nm)) and first potential defect types (e.g., interlayer bridging).

[0094] Secondly, based on the first potential defect type, a defect-lighting mapping database is retrieved to obtain multiple sets of candidate lighting configurations corresponding to the first potential defect type. Each candidate lighting configuration has a detection rate identifier. The defect-lighting mapping database contains the mapping relationship between different defect types and lighting detection parameters. The detection rate = the number of detected real defects / the total number of actual defects, which reflects the accuracy of defect detection under that lighting configuration. A higher detection rate indicates a better defect detection effect, and the lighting configuration is more superior. For example, based on the first potential defect type, the defect-lighting mapping database is retrieved to obtain multiple sets of candidate lighting configurations corresponding to the first potential defect type, and each candidate lighting configuration has a detection rate identifier. Examples include candidate lighting configuration 1 (500nm, incident angle 35°, horizontal polarization, annular dark field, detection rate 80%) and candidate lighting configuration 2 (900nm, incident angle 25°, circular polarization, backscattered illumination, detection rate 50%).

[0095] Next, the multiple candidate illumination configurations are arranged in descending order according to the detection rate identifier, and the candidate illumination configuration with the highest detection rate is selected as the first preferred illumination configuration. For example, the retrieved multiple candidate illumination configurations include: candidate illumination configuration 1 (500nm, incident angle 35°, horizontal polarization, annular dark field, detection rate 80%) and candidate illumination configuration 2 (900nm, incident angle 25°, circular polarization, backscattered illumination, detection rate 50%), arranged in descending order according to the detection rate identifier as: candidate illumination configuration 1, candidate illumination configuration 2. The higher the detection rate, the better the defect detection effect, and the better the illumination configuration. Thus, the candidate illumination configuration with the highest detection rate (candidate illumination configuration 1) is selected as the first preferred illumination configuration.

[0096] Finally, the first preferred illumination configuration is added to the second light source scheme. For example, the first preferred illumination configuration—candidate illumination configuration 1 (500nm, incident angle 35°, horizontal polarization, annular dark field)—is added to the second light source scheme. In this way, the second light source scheme is the illumination configuration with the highest detection rate for potential defect distribution. By acquiring images based on this, potential defects can be effectively identified, improving the accuracy and precision of defect detection.

[0097] Furthermore, according to the second light source scheme, the controllable light source unit is reconfigured, and combined with the image acquisition unit, images of the target wafer are acquired again to obtain a second image set. Specifically, according to the second light source scheme, the illumination configuration (including wavelength parameters, angle parameters, polarization parameters, and illumination mode parameters) is converted into specific control commands for the controllable light source unit, and combined with the image acquisition unit (such as a 12K resolution line scan camera), corresponding wafer images are acquired to obtain the second image set. Thus, the second image set is a reliable supplement to the first image set.

[0098] In summary, compared to existing technologies, this application, based on a first defect distribution, searches a historical wafer production database, selects multiple historical wafer samples, and removes those from the first defect distribution to obtain a potential defect distribution. From this potential defect distribution, it extracts first potential defect points and their corresponding coordinates and defect types. Then, based on the first potential defect types, it retrieves multiple sets of candidate illumination configurations with detection rate indicators, selects the candidate illumination configuration with the highest detection rate as the first preferred illumination configuration, and adds it to the second light source scheme. Finally, it configures a controllable light source unit and, combined with an image acquisition unit, performs image acquisition on the target wafer again to obtain a second image set. Thus, the second light source scheme effectively identifies potential defects, providing a reliable supplement to the first defect distribution and improving the accuracy and precision of defect detection.

[0099] S40: Based on the second image set, identify multiple second defect points of the target wafer, label the defect type of each second defect point, and generate the surface defect distribution of the target wafer by combining the first defect distribution.

[0100] In this embodiment of the application, firstly, following the same method as in S20, "based on the first image set, identify multiple first defect points of the target wafer and generate a first defect distribution", based on the second image set, identify multiple second defect points of the target wafer, label the defect type of each second defect point, and generate a second defect distribution.

[0101] Furthermore, by combining the first and second defect distributions, the surface defect distribution of the target wafer is generated. The combination of the first and second defect distributions allows for global coordinate system calibration (accuracy ±0.5μm) using wafer edge Mark points, unifying the first and second defect distributions onto the three-dimensional structural model. This yields an accurate surface defect distribution for the target wafer.

[0102] In summary, the embodiments of this application have at least the following technical effects:

[0103] Compared to existing technologies, this application retrieves multiple sample defect distributions based on the raw material characteristics and manufacturing process characteristics of the target wafer. Then, it obtains multiple historical light source schemes corresponding to the multiple sample defect distributions from the historical wafer production database. Next, it selects illumination configuration items with historical configuration frequencies greater than preset configuration frequencies from the multiple historical light source schemes and summarizes them to form a first light source scheme. Finally, based on the obtained first light source scheme, it configures a controllable light source unit and combines it with an image acquisition unit to acquire images of the target wafer and obtain a first image set. In this way, by filtering multiple illumination configurations in similar historical data and acquiring images accordingly, it realizes intelligent reuse of historical data, improves the efficiency and accuracy of image acquisition, and provides necessary support for subsequent defect analysis.

[0104] Secondly, based on the first image set, this application uses a wafer defect recognition network to identify multiple first defect points and their corresponding coordinates and defect types in multiple target wafer images. These are then annotated on the three-dimensional structure of the target wafer to generate a first defect distribution. In this way, the distribution of most defects on the wafer surface is obtained, providing the necessary foundation for obtaining the subsequent second defect distribution.

[0105] Furthermore, based on the first defect distribution, this application searches a historical wafer production database, filters out multiple historical wafer samples, and removes those from the first defect distribution to obtain a potential defect distribution. From this distribution, it extracts first potential defect points and their corresponding coordinates and defect types. Then, based on the first potential defect types, it retrieves multiple sets of candidate illumination configurations with detection rate indicators. The candidate illumination configuration with the highest detection rate is selected as the first preferred illumination configuration and added to the second light source scheme. Finally, a controllable light source unit is configured, and combined with an image acquisition unit, images are acquired again from the target wafer to obtain a second image set. Thus, the second light source scheme effectively identifies potential defects, providing a reliable supplement to the first defect distribution and improving the accuracy and precision of defect detection.

[0106] Finally, based on the second image set, this application identifies multiple second defect points on the target wafer and labels the defect type of each second defect point. Combined with the first defect distribution, it generates the surface defect distribution of the target wafer, thus obtaining an accurate surface defect distribution of the target wafer.

[0107] Through the above technical solution, this application dynamically matches illumination configuration items from historical data based on the raw material characteristics and production process characteristics of the target wafer, and fully considers potential defects, significantly improving the detection rate and accuracy of wafer surface defect detection.

[0108] Example 2, as Figure 3 As shown, based on the same inventive concept as the AI-based wafer surface defect detection method provided in Embodiment 1, this embodiment of the invention also provides an AI-based wafer surface defect detection system, including:

[0109] The initial light source module 11 is used to acquire a first light source scheme, configure the controllable light source unit, and combine it with the image acquisition unit to acquire images of the target wafer to obtain a first image set;

[0110] The defect analysis module 12 is used to identify multiple first defect points of the target wafer based on the first image set and generate a first defect distribution;

[0111] The optimized light source module 13 is used to generate a second light source scheme based on the first defect distribution, reconfigure the controllable light source unit, and combine it with the image acquisition unit to perform image acquisition on the target wafer again to obtain a second image set;

[0112] The integrated output module 14 is used to identify multiple second defect points of the target wafer based on the second image set, and label the defect type of each second defect point, and generate the surface defect distribution of the target wafer in combination with the first defect distribution.

[0113] The initial light source module 11 is specifically used for:

[0114] Obtain the raw material characteristics and manufacturing process characteristics of the target wafer;

[0115] Based on the characteristics of the raw materials and the characteristics of the production process, a historical wafer production database is retrieved, multiple sample wafers are selected, and the sample defect distribution of the multiple sample wafers is extracted to obtain multiple sample defect distributions.

[0116] Based on the defect distribution of the multiple samples, a first light source scheme for the target wafer is generated.

[0117] Specifically, the step of "generating a first light source scheme for the target wafer based on the defect distribution of the plurality of samples" includes:

[0118] Multiple historical light source schemes corresponding to the defect distributions of the multiple samples are obtained from the historical wafer production database. Each historical light source scheme includes multiple illumination configuration items. Each illumination configuration item includes the wavelength parameter, angle parameter, polarization parameter and illumination mode parameter used by the controllable light source unit in a single image acquisition.

[0119] The first light source scheme is formed based on the multiple historical light source schemes.

[0120] Furthermore, the phrase "forming the first light source scheme based on the plurality of historical light source schemes" includes:

[0121] The illumination configuration set is obtained by taking the union of the illumination configuration items in multiple historical light source schemes;

[0122] Traverse the set of lighting configurations, extract each lighting configuration item in turn, and count the historical configuration frequency of each lighting configuration item in the multiple historical light source schemes;

[0123] Lighting configuration items with historical configuration frequencies greater than preset configuration frequencies are selected and summarized to form the first light source scheme.

[0124] Specifically, the defect analysis module 12 is used for:

[0125] The wafer defect identification network is retrieved, and multiple first images from the first image set are sequentially input into the wafer defect identification network to obtain multiple first defect points, the coordinate positions of each first defect point in the target wafer, and the defect type corresponding to each first defect point.

[0126] Obtain the three-dimensional structure of the target wafer;

[0127] Based on the coordinate positions of each first defect point in the target wafer, the plurality of first defect points and the defect types corresponding to each first defect point are labeled on the three-dimensional structure to generate the first defect distribution.

[0128] Furthermore, the construction steps of the "wafer defect identification network" include:

[0129] Historical wafer images are collected from the historical wafer production database to form a sample wafer image set;

[0130] Image defect annotation is performed on the sample wafer image set to generate a sample wafer defect set and a sample defect type set;

[0131] Machine learning is used to generate a defect presence identification branch based on the sample wafer image set and the sample wafer defect set, and a defect type judgment branch is generated based on the sample wafer image set and the sample defect type set.

[0132] The defect presence identification branch and the defect type determination branch are connected and configured to obtain the wafer defect identification network.

[0133] Specifically, the optimized light source module 13 is used for:

[0134] Based on the first defect distribution, the historical wafer production database is searched to select multiple historical wafer samples;

[0135] Extract the complete defect distribution of the multiple historical wafer samples, and take the union of the complete defect distributions of the multiple historical wafer samples to form the overall predicted defect distribution;

[0136] The first defect distribution is removed from the overall predicted defect distribution to obtain the potential defect distribution;

[0137] The second light source scheme is generated based on the potential defect distribution.

[0138] Further, the step of "generating the second light source scheme based on the potential defect distribution" includes:

[0139] Extract the first potential defect point from the potential defect distribution, and extract the first potential coordinates and the first potential defect type of the first potential defect point;

[0140] Based on the first potential defect type, the defect-lighting mapping database is retrieved to obtain multiple sets of candidate lighting configuration items corresponding to the first potential defect type. Each candidate lighting configuration item has a detection rate identifier.

[0141] The candidate illumination configurations are arranged in descending order according to the detection rate identifier, and the candidate illumination configuration with the highest detection rate is selected as the first preferred illumination configuration.

[0142] Add the first preferred lighting configuration to the second light source scheme.

[0143] The integrated output module 14 is specifically used for:

[0144] Based on the second image set, multiple second defect points of the target wafer are identified, and the defect type of each second defect point is labeled. Combined with the first defect distribution, the surface defect distribution of the target wafer is generated.

[0145] In summary, the embodiments of this application have at least the following technical effects:

[0146] Compared to existing technologies, this application first obtains a first light source scheme through an initial light source module, configures a controllable light source unit, and combines it with an image acquisition unit to acquire images of the target wafer, obtaining a first image set. This enables intelligent reuse of historical data, improves the efficiency and accuracy of image acquisition, and provides necessary support for subsequent defect analysis. Secondly, through a defect analysis module, multiple first defect points on the target wafer are identified based on the first image set, generating a first defect distribution. This obtains the distribution of most defects on the wafer surface and provides a necessary foundation for obtaining a second defect distribution. Thirdly, by optimizing the light source module to generate a second light source scheme based on the first defect distribution, the controllable light source unit is reconfigured, and combined with the image acquisition unit, images of the target wafer are acquired again to obtain a second image set. This effectively identifies potential defects, providing a reliable supplement to the first defect distribution and improving the accuracy and precision of defect detection. Finally, through an integrated output module, multiple second defect points on the target wafer are identified based on the second image set, and the defect type of each second defect point is labeled. Combined with the first defect distribution, a surface defect distribution of the target wafer is generated, obtaining an accurate surface defect distribution of the target wafer. This improves the detection rate and accuracy of wafer surface defect detection.

[0147] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0148] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0153] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wafer surface defect detection method based on artificial intelligence, characterized by, The application is applied to a wafer surface defect detection device, the wafer surface defect detection device includes a controllable light source unit and an image acquisition unit; the method comprises: Obtaining a first light source scheme, configuring the controllable light source unit, and combining the image acquisition unit to acquire a first image set by image acquisition of a target wafer; According to the first image set, a plurality of first defect points of the target wafer are identified, and a first defect distribution is generated; Based on the first defect distribution, a second light source scheme is generated, the controllable light source unit is configured again, and the target wafer is image acquired again by combining the image acquisition unit, and a second image set is obtained; According to the second image set, a plurality of second defect points of the target wafer are identified, and the defect types of each second defect point are labeled, and the surface defect distribution of the target wafer is generated in combination with the first defect distribution; According to the first image set, a plurality of first defect points of the target wafer are identified, and a first defect distribution is generated, comprising: Call the wafer defect recognition network, input a plurality of first images in the first image set into the wafer defect recognition network in turn, obtain a plurality of first defect points, the coordinate position of each first defect point in the target wafer, and the defect type corresponding to each first defect point; Obtaining the three-dimensional structure of the target wafer; Based on the coordinate position of each first defect point in the target wafer, the plurality of first defect points and the defect type corresponding to each first defect point are labeled on the three-dimensional structure to generate the first defect distribution; Based on the first defect distribution, a second light source scheme is generated, comprising: According to the first defect distribution, the historical wafer production database is searched, and a plurality of historical wafer samples are screened out; Extract the sample complete defect distribution of the plurality of historical wafer samples, and take the union of the sample complete defect distribution of the plurality of historical wafer samples to form an overall predicted defect distribution; Eliminate the first defect distribution in the overall predicted defect distribution to obtain a potential defect distribution; According to the potential defect distribution, the second light source scheme is generated; According to the potential defect distribution, the second light source scheme is generated, comprising: Extracting a first potential defect point in the potential defect distribution, and extracting a first potential coordinate and a first potential defect type of the first potential defect point; According to the first potential defect type, the defect-illumination mapping database is searched to obtain a plurality of candidate illumination configuration items corresponding to the first potential defect type, each candidate illumination configuration item has a detection rate identifier; The plurality of candidate illumination configuration items are arranged in descending order according to the detection rate identifier, the candidate illumination configuration item with the highest detection rate is selected as the first preferred illumination configuration item; The first preferred illumination configuration item is added to the second light source scheme.

2. The method of claim 1, wherein, Obtaining a first light source scheme, comprising: Obtaining the material characteristics and production process characteristics of the target wafer; Based on the material characteristics and the production process characteristics, the historical wafer production database is searched, a plurality of sample wafers are screened out, and the sample defect distribution of the plurality of sample wafers is extracted to obtain a plurality of sample defect distributions; According to the plurality of sample defect distributions, a first light source scheme of the target wafer is generated.

3. The method of claim 2, wherein, According to the plurality of sample defect distributions, a first light source scheme of the target wafer is generated, comprising: From the historical wafer production database, a plurality of historical light source schemes corresponding to the plurality of sample defect distributions are obtained, each of the historical light source schemes includes a plurality of light configuration items, and each of the light configuration items includes wavelength parameters, angle parameters, polarization parameters and illumination mode parameters used by the controllable light source unit in single image acquisition; According to the plurality of historical light source schemes, the first light source scheme is formed.

4. The method of claim 3, wherein, According to the plurality of historical light source schemes, the first light source scheme is formed, comprising: Taking the union of the light configuration items in the plurality of historical light source schemes, a light configuration set is obtained; Traverse the light configuration set, and sequentially extract each light configuration item, and count the historical configuration frequency of each light configuration item in the plurality of historical light source schemes; Selecting the light configuration item with a historical configuration frequency greater than a preset configuration frequency, and collecting to form the first light source scheme.

5. The method of claim 1, wherein, The construction steps of the wafer defect recognition network include: Collecting historical wafer images in the historical wafer production database to form a sample wafer image set; Performing image defect annotation on the sample wafer image set to generate a sample wafer defect set and a sample defect type set; Using machine learning, generating a defect existence recognition branch based on the sample wafer image set and the sample wafer defect set, and generating a defect type judgment branch based on the sample wafer image set and the sample defect type set; Connecting and configuring the defect existence recognition branch and the defect type judgment branch to obtain the wafer defect recognition network.

6. A wafer surface defect detection system based on artificial intelligence, characterized by, For performing the method of any one of claims 1-5, comprising: An initial light source module is configured to obtain a first light source scheme, configure the controllable light source unit, and combine the image acquisition unit to perform image acquisition on a target wafer to obtain a first image set; A defect analysis module is configured to identify a plurality of first defect points of the target wafer according to the first image set, and generate a first defect distribution; An optimized light source module is configured to generate a second light source scheme based on the first defect distribution, reconfigure the controllable light source unit, and combine the image acquisition unit to perform image acquisition on the target wafer again to obtain a second image set; An integrated output module is configured to identify a plurality of second defect points of the target wafer according to the second image set, label the defect types of each of the second defect points, and combine the first defect distribution to generate a surface defect distribution of the target wafer.

Images