Wafer defect identification method, device and system
By acquiring the texture and chromaticity distribution characteristics of the chip image, combined with the deep learning model, the chip defects are automatically identified, which solves the problems of low manual recognition efficiency and large errors, and achieves efficient and accurate chip defect detection and quantitative analysis.
Patent Information
- Application Number
- CN202510507681.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-07-25
AI Technical Summary
In the prior art, the detection of semiconductor chip defects relies on manual visual recognition, which has low efficiency, large errors, and cannot be quantified, so as to meet mass production needs.
By acquiring wafer images, extracting texture features and global chromaticity distribution features, and fusing them with deep learning models, we automatically identify wafer defects, including defect locations and types.
It realizes automated, high-precision identification and quantitative analysis of chip defects, improves detection efficiency and accuracy, and at the same time, no damage detection, and supports rapid analysis and process optimization.
Smart Images
Figure CN120374585A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of semiconductor detection technology, and in particular, to a method, device, and system for identifying wafer defects. Background Art
[0002] Various defects may occur during the manufacturing process of semiconductor wafers. For example, silicon carbide crystals exist in multiple polytypes, such as 3C-SiC, 4H-SiC, 6H-SiC, etc., which differ in the stacking order of atomic layers. Polytype defects may occur in silicon carbide wafers during the growth process due to process fluctuations. For example, a 6H-SiC region may appear in 4H-SiC. The traditional detection of polytype defects in silicon carbide wafers is to place the silicon carbide wafer under a strong light, and the polytype defects show local color differences. For example, a color difference appears in the mixed region of 4H-SiC and 6H-SiC, and it relies on manual visual observation under a strong light to identify the polytype defects in the silicon carbide wafer.
[0003] However, the traditional method of relying on manual visual inspection to identify defects in wafers has the following problems: First, the efficiency is low. Manual operation takes a long time and cannot meet the mass production requirements. Second, the subjective error is large. The determination of the color difference of defects is easily affected by the experience of personnel and is prone to misjudgment. Third, there is a lack of quantification, that is, the defect area and distribution density in the wafer cannot be accurately measured.
[0004] Therefore, how to automatically and accurately identify, classify, and quantitatively analyze the defects in wafers is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention
[0005] To solve the above technical problems, embodiments of this application provide a method, device, and system for identifying wafer defects to achieve automatic and accurate identification, classification, and quantitative analysis of defects in wafers.
[0006] To achieve the above object, embodiments of this application provide the following technical solutions:
[0007] In a first aspect, embodiments of this application provide a method for identifying wafer defects, including:
[0008] Obtain an image of the wafer to be identified;
[0009] Extract the texture features of the image, and extract the global chromaticity distribution features of the image;
[0010] Fuse the texture features and the global chromaticity distribution features to obtain fused features;
[0011] Based on the fused features, determine the defect identification result of the wafer to be identified, where the defect identification result includes: defect location and defect type.
[0012] Optionally, the wafer to be recognized is a silicon carbide wafer. Obtaining an image of the wafer to be recognized includes:
[0013] Obtaining an image of the wafer to be recognized under blue light or ultraviolet light irradiation.
[0014] Optionally, the wafer defect recognition method is implemented by a configured wafer defect recognition model, and the wafer defect recognition model includes:
[0015] A first encoding module for extracting texture features from the input image;
[0016] A second encoding module for extracting global chromaticity distribution features from the input image;
[0017] A feature fusion module for fusing the texture features and the global chromaticity distribution features to obtain fused features;
[0018] An output module for determining a defect recognition result of the crystal to be recognized based on the fused features.
[0019] Optionally, the first encoding module is a ResNet-34 network structure.
[0020] Optionally, the second encoding module is a ViT network structure.
[0021] Optionally, the feature fusion module is specifically configured to fuse the texture features and the global chromaticity distribution features by using a cross-attention mechanism to obtain fused features.
[0022] Optionally, the output module includes:
[0023] A detection head for outputting a defect position of the wafer to be recognized based on the fused features;
[0024] A classification head for outputting a defect type of the wafer to be recognized based on the fused features.
[0025] Optionally, the defect recognition result further includes a pixel-level defect mask, and the output module further includes:
[0026] A segmentation head for outputting a pixel-level defect mask of the wafer to be recognized based on the fused features.
[0027] In a second aspect, an embodiment of the present application provides a wafer defect recognition device, including:
[0028] An image acquisition unit for acquiring an image of a wafer to be recognized;
[0029] A feature extraction unit, configured to extract the texture features of the image and the global chromaticity distribution features of the image;
[0030] A feature fusion unit, configured to fuse the texture features and the global chromaticity distribution features to obtain fused features;
[0031] A defect prediction unit, configured to determine the defect recognition result of the wafer to be recognized based on the fused features, where the defect recognition result includes: defect location and defect type.
[0032] In a third aspect, an embodiment of the present application provides a wafer defect recognition system, including: a light source module, a camera module, and an image processing module;
[0033] The light source module is configured to emit light to irradiate the wafer to be recognized;
[0034] The camera module is configured to capture an image of the wafer to be recognized under the irradiation of the light emitted by the light source module;
[0035] The image processing module is configured to execute each step of any of the above wafer defect recognition methods.
[0036] Optionally, the wafer to be recognized is a silicon carbide wafer, the light source module includes an LED array, and the LED array includes a red LED, a green LED, and a blue LED;
[0037] The light source module further includes a control module, and the control module is configured to modulate the proportions of red, green, and blue light in the light emitted by the LED array, so that the proportion of blue light in the light emitted by the LED array is in the range of 60% - 90%, including the end values, the spectral peak of the light emitted by the LED array is in the blue light band, and the light emitted by the LED array appears blue.
[0038] Optionally, the wafer defect recognition system further includes a polarization module;
[0039] The polarization module includes a polarizer and an analyzer. The polarizer is disposed between the light source module and the wafer to be recognized, and the analyzer is disposed between the wafer to be recognized and the camera module. The light source module, the polarizer, the wafer to be recognized, and the analyzer are coaxially disposed, and the polarization direction of the polarizer is perpendicular to the polarization direction of the analyzer.
[0040] Compared with the prior art, the above technical solution has the following advantages:
[0041] The wafer defect recognition method provided by the embodiments of the present application first obtains an image of the wafer to be recognized, then extracts the texture features of the image and the global chromaticity distribution features of the image, and then fuses the extracted texture features and global chromaticity distribution features to obtain fused features. Finally, based on the fused features, the defect recognition result of the wafer to be recognized is determined, including the defect position and defect type. It can be seen that the wafer defect recognition method provided by the embodiments of the present application does not require manual recognition of defects in the wafer, but can execute each step in the wafer defect recognition method through a configured deep learning model, realizing automatic recognition of defects in the wafer, significantly improving the detection efficiency of wafer defects; moreover, the wafer defect recognition method provided by the embodiments of the present application fuses the texture features and global chromaticity distribution features of the wafer image to determine the position and type of defects in the wafer, enabling high-precision recognition of defects in the wafer, significantly improving the accuracy of wafer defect detection, and facilitating quantitative analysis of defects in the wafer; at the same time, it will not cause damage to the wafer surface and can realize non-destructive detection of the wafer surface; this method is conducive to the rapid analysis of wafer quality and process optimization in the semiconductor manufacturing process. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0043] Figure 1 It is a result block diagram of a wafer defect recognition model for executing the wafer recognition method provided by the embodiments of the present application;
[0044] Figure 2 It is a schematic structural diagram of a wafer defect recognition device provided by the embodiments of the present application;
[0045] Figure 3 It is a schematic framework diagram of a wafer defect recognition system provided by the embodiments of the present application.
[0046] Reference Signs:
[0047] 10 - First Encoding Module; 20 - Second Encoding Module; 30 - Feature Fusion Module; 40 - Output Module; 41 - Detection Head; 42 - Classification Head; 43 - Segmentation Head; 100 - Image Acquisition Unit; 110 - Feature Extraction Unit; 120 - Feature Fusion Unit; 130 - Defect Prediction Unit; 200 - Wafer Stage; 201 - Quartz Support Sheet; 210 - Light Source Module; 220 - Camera Module; 230 - Image Processing Module; 240 - Polarization Module; 241 - Polarizer; 242 - Analyzer. Detailed implementation manners
[0048] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.
[0049] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned accompanying drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such terms can be interchanged under appropriate circumstances, which is only a way of distinguishing when describing objects with the same attributes in the embodiments of the present application. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device including a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these processes, methods, products or devices.
[0050] As described in the background art section, how to automatically and accurately identify, classify, and quantitatively analyze defects in wafers is a technical problem that those skilled in the art urgently need to solve.
[0051] In view of this, the embodiments of the present application provide a wafer defect identification method, and the wafer defect identification method includes:
[0052] S1: Obtain an image of the wafer to be identified.
[0053] Optionally, the wafer to be identified may be a silicon carbide wafer. For example, to identify polytype defects in a silicon carbide wafer. However, the present application does not limit this. The wafer to be identified may also be other semiconductor wafers, depending on the specific situation.
[0054] Optionally, an image of the wafer to be identified taken under a certain special light (such as blue light) can be obtained, so that the defects in the wafer to be identified show differences in color, etc. under the special light, which is convenient for subsequent identification of the defects in the wafer. However, the present application does not limit this. An image of the wafer to be identified taken under ordinary light (such as natural light or white light) can also be obtained, depending on the specific situation.
[0055] S2: Extract the texture features of the image of the wafer to be identified, and extract the global chromaticity distribution features of the image of the wafer to be identified.
[0056] It can be understood that there may be structural differences between the defective regions and non-defective regions in the wafer to be recognized. For example, the lattice structure in the defective region is incomplete, with defects such as lattice distortion, dislocation, stacking fault, vacancy, interstitial atom, etc. These defects disrupt the periodic arrangement of the crystal, resulting in disordered atomic arrangement or deviation from the normal position in the local region. In contrast, the non-defective region has a complete lattice structure, and atoms are arranged according to strict periodic rules, forming an ordered crystal lattice.
[0057] There may also be characteristic differences between the defective regions and non-defective regions in the wafer to be recognized. For example, the absorption, scattering, and reflection characteristics of light in the defective region are different from those in the non-defective region, showing different colors under special illumination.
[0058] Thus, in the image of the wafer to be recognized, the defective regions and non-defective regions of the wafer to be recognized will show differences in gray level and color, etc.
[0059] It can also be understood that the texture feature of an image is an important visual attribute that describes the spatial distribution pattern among pixels in a local or global region of the image. The texture feature reflects structural information such as repeatability, directionality, roughness, and contrast in the image. Therefore, the texture features of the defective regions and non-defective regions in the image of the wafer to be recognized are different. By extracting the texture features of the image of the wafer to be recognized, the defects in the wafer to be recognized can be identified based on the extracted texture features of the image.
[0060] Furthermore, it can be understood that the global chromaticity distribution feature of an image refers to the statistical characteristics of the color distribution of the entire image in the color space, which can reflect the overall color composition and distribution law of the image. The color space can be the RGB color space or the HSV color space. For example, in the HSV color space, color is decomposed into hue, saturation, and value, which is more suitable for chromaticity analysis. Therefore, the chromaticity features of the defective regions and non-defective regions in the image of the wafer to be recognized are different. By extracting the global chromaticity distribution feature of the image of the wafer to be recognized, the defects in the wafer to be recognized can be identified based on the extracted global chromaticity distribution feature of the image.
[0061] S3: Fuse the extracted texture features and global chromaticity distribution features to obtain fused features.
[0062] S4: Based on the fused features, determine the defect recognition result of the wafer to be recognized. The defect recognition result includes: defect location and defect type.
[0063] Considering that there may be more than one type of defect in the wafer, but multiple types exist. Therefore, not only the defects in the wafer need to be identified, but also the types of defects in the wafer. However, some different types of defects may be very easy to confuse. Therefore, in this application, after extracting the texture features and global chromaticity distribution features of the image of the wafer to be identified, the extracted texture features and global chromaticity distribution features are then fused to obtain a fused feature, that is, a texture-color fused feature. Thus, based on the fused feature, the defect recognition result of the wafer to be identified is determined. In this way, the recognition accuracy of various types of defects in the wafer to be identified can be improved.
[0064] In this application, the defect recognition result of the wafer to be identified includes the defect position and the defect type. For example, a defect bounding box can be output to characterize the position of the defect. Another example, if the wafer to be identified is a silicon carbide wafer and the identified defect is a polytype defect, then the types of polytype defects such as 3C-SiC, 4H-SiC, and 6H-SiC and the category rate (indicating the proportion of a certain type of defect in the total defects) can be output.
[0065] It can be seen that the wafer defect recognition method provided by the embodiments of this application first obtains an image of the wafer to be identified, then extracts the texture features of the image and extracts the global chromaticity distribution features of the image, and then fuses the extracted texture features and global chromaticity distribution features to obtain a fused feature. Finally, based on the fused feature, the defect recognition result of the wafer to be identified is determined, including the defect position and the defect type. It can be seen that the wafer defect recognition method provided by the embodiments of this application does not require manual identification of defects in the wafer, but can execute each step in the wafer defect recognition method through a configured deep learning model to achieve automatic recognition of defects in the wafer, significantly improving the detection efficiency of wafer defects. Moreover, the wafer defect recognition method provided by the embodiments of this application fuses the texture features and global chromaticity distribution features of the wafer image to determine the position and type of defects in the wafer, can achieve high-precision recognition of defects in the wafer, significantly improve the accuracy of wafer defect detection, and is convenient for quantitative analysis of defects in the wafer. At the same time, it will not cause damage to the wafer surface and can achieve non-destructive detection of the wafer surface. This method is beneficial to the rapid analysis of wafer quality and process optimization in the semiconductor manufacturing process.
[0066] Optionally, in some embodiments of the present application, the wafer to be recognized is a silicon carbide wafer. For silicon carbide wafers, polytype defects and microtube dense defects are two common types of key defects in the silicon carbide crystal growth process. Polytype defects are caused by incorrect stacking order of atomic layers, and microtube dense defects are caused by the aggregation of hollow dislocations or screw dislocations. When identifying polytype defects in silicon carbide wafers in the prior art, white light emitted by a high color rendering white light LED light source is usually used to irradiate the silicon carbide wafer. However, the inventor found that polytype defects and microtube dense defects are very easy to be confused under the white light emitted by the white light LED light source, resulting in easy misjudgment in the identification of polytype defects and microtube dense defects.
[0067] Based on this, the inventor's research found that if the silicon carbide wafer is irradiated with blue light or ultraviolet light, the polytype defects and microtube dense defects in the silicon carbide wafer are relatively easier to distinguish. For example, polytype defects appear green under blue light irradiation. Therefore, when the wafer to be recognized is a silicon carbide wafer, step S1 of obtaining an image of the wafer to be recognized includes:
[0068] S11: Obtain an image of the wafer to be recognized taken under blue light or ultraviolet light irradiation.
[0069] It can be understood that when the wafer to be recognized is a silicon carbide wafer and polytype defects in the silicon carbide wafer need to be recognized, an image of the silicon carbide wafer taken under blue light or ultraviolet light irradiation is obtained. Thus, in the obtained image of the silicon carbide wafer, polytype defects and microtube dense defects are relatively easier to distinguish. Furthermore, when extracting and fusing the texture features and global chromaticity distribution features of the image of the silicon carbide wafer and identifying polytype defects in the silicon carbide wafer based on the fused features, the accuracy of identifying polytype defects in the silicon carbide wafer can be further improved.
[0070] It should be noted that when irradiating a silicon carbide wafer with blue light, optionally, the blue light emitted by a blue light LED light source or the ultraviolet light emitted by an ultraviolet light LED light source can be used to irradiate the silicon carbide wafer. However, the costs of both the blue light LED light source and the ultraviolet light LED light source are relatively high. Another option is to use a high-color rendering white light LED light source. The high-color rendering white light LED light source includes an LED array, and the LED array includes a red LED, a green LED, and a blue LED. By Pulse Width Modulation (PWM), the proportions of red, green, and blue light in the light emitted by the high-color rendering white light LED light source are controlled, such that the proportion of blue light in the light emitted by the high-color rendering white light LED light source is in the range of 60% - 90%, including the end values. The spectral peak of the light emitted by the high-color rendering white light LED light source is in the blue light band (450nm - 470nm), presenting blue light. In this way, instead of using a relatively high-cost blue light LED light source, blue light enhancement is achieved through a relatively low-cost white light LED light source, greatly reducing the hardware transformation cost. Moreover, the white light LED can quickly switch the spectral peak of the output light through PWM control. It can not only output blue light to detect polytype defects in silicon carbide wafers, but also flexibly switch to output other colors of light, such as green light, to detect other material defects, with strong scalability.
[0071] The wafer defect recognition method provided by the embodiments of the present application can be implemented through a configured wafer defect recognition model. Optionally, Figure 1 The result block diagram of the wafer defect recognition model for implementing the wafer recognition method provided by the embodiments of the present application is shown. As Figure 1 shown, the wafer defect recognition model includes:
[0072] A first encoding module 10 for extracting texture features from the input image of the wafer to be recognized;
[0073] A second encoding module 20 for extracting global chromaticity distribution features from the input image of the wafer to be recognized;
[0074] A feature fusion module 30 for fusing the texture features and the global chromaticity distribution features to obtain fused features;
[0075] An output module 40 for determining the defect recognition result of the crystal to be recognized based on the fused features.
[0076] Optionally, the first encoding module 10 can be a ResNet-34 network structure.
[0077] Specifically, an image of a wafer to be recognized with a resolution of 2048×2048 and single-channel grayscale can be input to the first encoding module 10. The first encoding module 10 can adopt a pre-trained ResNet-34 network structure, remove the original fully connected layer to avoid directly giving the result, and retain the feature extraction part - the convolutional layer and the residual block.
[0078] As we know, the ResNet-34 network structure is composed of multiple stacked residual blocks. After every few residual blocks, the feature map is downsampled (halved in size and doubled in the number of channels) through a convolution or pooling operation with a stride of 2. Thus, the feature map can be extracted from the second residual block of the ResNet network. The feature map output by the second residual block has a relatively large size (512×256×256) and can retain the high-definition local information of the image, that is, retain the high-resolution texture information of the image.
[0079] Moreover, a CBAM attention module (Convolutional Block Attention Module) is inserted after the residual block to focus on the defect area through channel attention and spatial attention.
[0080] Finally, the first encoding module 10 can output a 512-dimensional high-resolution texture feature map (256×256), which characterizes local details such as defect edges and lattice distortion.
[0081] Optionally, the second encoding module 20 can be a ViT (Vision Transformer) network structure.
[0082] Specifically, the RGB image of the wafer to be recognized taken can be converted to the HSV space, and the saturation (S) and value (V) channels are extracted to enhance the saturation and reduce the value of the defect area, forming a sharp contrast with the background to form an HSV enhanced image (with three channels of H, S, and V, and a resolution of 2048×2048). The HSV enhanced image is input to the second encoding module 20, and the second encoding module 20 adopts a ViT network structure.
[0083] The ViT network structure mainly consists of three parts: image chunking and embedding, the Transformer encoder, and the output layer.
[0084] Among them, image chunking and embedding include image chunking, linear projection, classification token, and positional encoding. Image chunking divides the input image into 16×16 pixel patches, a total of 128×128 patches. Linear projection linearly maps each patch into a 256-dimensional vector. Since the Transformer itself has no position information, positional encoding needs to be added to the embedding vector of each patch. Therefore, the positional encoding adds learnable positional embeddings (PositionalEmbedding) to retain spatial information.
[0085] The Transformer encoder consists of multiple stacked Transformer Blocks. Each Block contains multi-head self-attention, a feed-forward neural network, layer normalization, and residual connection. Multi-Head Self-Attention (MHSA) is used to model the global dependencies between patches. The self-attention mechanism dynamically assigns attention weights by calculating the similarities between queries, keys, and values. Multi-head self-attention calculates in parallel through multiple attention heads to capture information in different subspaces. The Feed-Forward Network (FFN) is used to further process the features of each patch, performing a non-linear transformation on the features of each patch. The feed-forward neural network (FNN) consists of two linear layers with an activation function (such as GELU) in the middle. Layer Normalization (LN) is performed before each sub-layer (MHSA and FFN). The Residual Connection adds the input to the output of the sub-layer to alleviate the problem of vanishing gradients.
[0086] The output layer extracts the output of the 6th Transformer block (global context features) and the output of the 3rd layer (mid-level semantic features), corresponding to the class tendency and chromaticity distribution of the defect respectively.
[0087] Finally, the second encoding module 20 outputs a fused 512-dimensional global chromaticity distribution feature map (128×128), representing macroscopic information such as defect color distribution and regional consistency.
[0088] Optionally, the feature fusion module 30 is specifically used to fuse the extracted texture features and global chromaticity distribution features by using the cross-attention mechanism (Cross-Attention) to obtain fused features, focusing on the local texture and global color associations related to the defect.
[0089] Specifically, first, feature alignment and preprocessing are performed to align the texture feature Fpol and the global chromaticity distribution feature Fcolor in terms of spatial dimensions and number of channels, providing a consistent feature representation for subsequent cross-attention calculation. Bilinear interpolation is performed on the texture feature Fpol to match its spatial dimensions with those of the global chromaticity distribution feature Fcolor (128×128), ensuring the consistency of the two feature maps in spatial positions and facilitating subsequent pixel-by-pixel attention calculation; the number of channels of the texture feature Fpol and the global chromaticity distribution feature Fcolor is unified to 512 dimensions through 1×1 convolution, reducing the impact of channel dimension differences on attention calculation, compressing redundant information, and improving computational efficiency.
[0090] Then, the cross-attention mechanism is calculated to fuse the texture feature Fpol and the global chromaticity distribution feature Fcolor by dynamically allocating attention weights, enhancing the model's attention to key texture details. Among them, the definitions of query, key, and value are first carried out; the texture feature Fpol is used as the query vector to learn "which texture features need to be focused on" to capture the part in the global chromaticity distribution feature Fcolor that needs to be aligned with the texture feature; the texture feature Fpol is used as the key and value to provide texture detail information, where the key is used to calculate the attention weights and the value is used for weighted fusion. Then, the attention weights are calculated, the similarity between the texture feature Fpol and the global chromaticity distribution feature Fcolor is calculated to generate the attention weights, and the attention weights are multiplied by the corresponding feature values to generate the weighted texture-color fusion feature.
[0091] Next, the fused feature is added to the original global chromaticity distribution feature to retain the original color information and avoid color distortion caused by the attention mechanism.
[0092] Optionally, as Figure 1 shown, the output module 40 includes a detection head 41, and the detection head 41 is used to output the defect position of the wafer to be recognized based on the fused feature. Specifically, a defect bounding box can be output to represent the defect position; the output module 40 includes a classification head 42, and the classification head 42 is used to output the defect type of the wafer to be recognized based on the fused feature. Specifically, the type of the defect and the category rate can be output.
[0093] Further optionally, the defect recognition result also includes a pixel-level defect mask. As Figure 1 shown, the output module 40 can also include a segmentation head 43, and the segmentation head 43 is used to output the pixel-level defect mask of the wafer to be recognized based on the fused feature. It can be understood that the pixel-level defect mask is used to identify whether each pixel on the wafer surface belongs to the defect area.
[0094] It can be understood that the wafer defect recognition model can finally generate a wafer surface defect distribution map, on which the defect type, coordinates and area ratio are marked, so as to facilitate the rapid analysis of wafer quality and process optimization.
[0095] It can also be understood that the wafer defect recognition model can be trained using wafer sample images marked with defect recognition result labels. For example, if the wafer defect recognition model is to identify polytype defects in silicon carbide wafers, synthetic images with different blue light ratios and different defect sizes can be generated based on the silicon carbide birefringence model to expand the training sample images. A defect simulation engine can also be introduced to simulate interference items such as scratches and dust on the wafer surface to improve the robustness of the model. After verification, the trained wafer defect recognition model can stably identify defects of 0.05 mm 2 in wafers, while the traditional method can only identify defects with a threshold of 0.2 mm 2 in wafers at most.
[0096] The above introduces a wafer defect recognition method provided by an embodiment of the present application. The following will introduce the device for executing the above wafer defect recognition method.
[0097] Figure 2 FIG. shows a schematic structural diagram of a wafer defect recognition device provided by an embodiment of the present application. As Figure 2 shown, the wafer defect recognition device includes:
[0098] An image acquisition unit 100, configured to acquire an image of the wafer to be recognized;
[0099] A feature extraction unit 110, configured to extract the texture feature of the image of the wafer to be recognized, and extract the global chromaticity distribution feature of the image of the wafer to be recognized;
[0100] A feature fusion unit 120, configured to fuse the extracted texture feature and global chromaticity distribution feature to obtain a fusion feature;
[0101] A defect prediction unit 130, configured to determine the defect recognition result of the wafer to be recognized based on the fusion feature, where the defect recognition result includes: defect position and defect type.
[0102] Optionally, the defect recognition result may further include a pixel-level defect mask.
[0103] Since the wafer defect recognition device provided by the embodiment of the present application corresponds to the wafer defect recognition method provided by the foregoing embodiment of the present application, and the wafer defect recognition method provided by the embodiment of the present application has been described in detail in the foregoing embodiments, it will not be elaborated herein.
[0104] The embodiment of the present application further provides a wafer defect recognition system.Figure 3 The figure shows a schematic framework diagram of a wafer defect recognition system provided by an embodiment of the present application. As Figure 3 shown, the wafer defect recognition system includes a wafer stage 200. The wafer stage 200 can be a circular aluminum alloy frame with a hollow center and four highly transparent quartz support sheets 201 (such as with a thickness of 0.5 mm) symmetrically distributed at the edge, which are used to fix the edge of the wafer to be recognized. The contact area between the quartz support sheet 201 and the wafer can be less than 1 mm 2 , so as to reduce the occlusion of the wafer. Specifically, the robot can grasp the wafer from the loading port, correct the position through the edge finder, and place it on the quartz support sheet 201 of the wafer stage 200. Moreover, a force feedback sensor is equipped at the end of the robot to adjust the grasping force in real time to prevent the wafer from being damaged.
[0105] As Figure 3 shown, the wafer defect recognition system includes a light source module 210, a camera module 220, and an image processing module 230. Among them, the light source module 210 is used to emit light to irradiate the wafer to be recognized; the camera module 220 is used to capture an image of the wafer to be recognized under the irradiation of the light emitted by the light source module 210; the image processing module 230 is used to execute each step of the wafer defect recognition method provided by the foregoing embodiment of the present application. Specifically:
[0106] S1: Obtain the image captured of the wafer to be recognized.
[0107] S2: Extract the texture features of the image of the wafer to be recognized, and extract the global chromaticity distribution features of the image of the wafer to be recognized.
[0108] S3: Fuse the extracted texture features and global chromaticity distribution features to obtain fused features.
[0109] S4: Based on the fused features, determine the defect recognition result of the wafer to be recognized. The defect recognition result includes: the defect position and the defect type.
[0110] Optionally, the wafer to be recognized is a silicon carbide wafer. As known from the foregoing, the polytype defects and microtube dense defects in the silicon carbide wafer are very easy to be confused under the white light emitted by the white light LED light source, resulting in easy misjudgment of the recognition of polytype defects and microtube dense defects. If the silicon carbide wafer is irradiated with blue light or ultraviolet light, the polytype defects and microtube dense defects in the silicon carbide wafer are relatively easier to distinguish. For example, the polytype defects show green under blue light irradiation. However, the costs of both the blue light LED light source and the ultraviolet light LED light source are relatively high.
[0111] Based on this, optionally, in some embodiments of the present application, the light source module 210 includes an LED array. The LED array includes red LEDs, green LEDs, and blue LEDs, that is, the LED array can be a white LED light source. The light source module 210 further includes a control module, and the control module is used to modulate (i.e., PWM modulation) the proportions of red, green, and blue light in the light emitted by the LED array, so that the proportion of blue light in the light emitted by the LED array is within the range of 60%-90%, including the endpoint values. The spectral peak of the light emitted by the LED array is in the blue light band (450nm - 470nm), and the light emitted by the LED array presents blue light. In this way, instead of using a relatively expensive blue LED light source, blue light enhancement is achieved through a relatively low-cost white LED light source, greatly reducing the hardware transformation cost. Moreover, the white LED light source can quickly switch the spectral peak of the output light through PWM control. It can not only output blue light to detect polytype defects in silicon carbide wafers, but also flexibly switch to output other colors of light, such as green light, to detect other material defects, with strong scalability.
[0112] For example, if you want to identify 4H-SiC polytype defects in a silicon carbide wafer, you can control the proportion of blue light in the light emitted by the LED array to be 85%.
[0113] Based on any of the above embodiments, optionally, in some embodiments of the present application, as Figure 3 shown, the wafer defect identification system may further include a polarization module 240. The polarization module 240 includes a polarizer 241 and an analyzer 242. The polarizer 241 is disposed between the light source module 210 and the wafer to be identified, and the analyzer 242 is disposed between the wafer to be identified and the camera module 220. The light source module 210, the polarizer 241, the wafer to be identified, and the analyzer 242 are coaxially arranged, and the polarization direction of the polarizer 241 is perpendicular to the polarization direction of the analyzer 242.
[0114] It can be understood that the light source module 210, the polarizer 241, the wafer to be identified, and the analyzer 242 are coaxially arranged, and the polarization direction of the polarizer 241 is perpendicular to the polarization direction of the analyzer 242, that is, the polarization direction of the polarizer 241 and the polarization direction of the analyzer 242 are orthogonally fixed. For example, the polarization direction of the polarizer 241 is 0°, and the polarization direction of the analyzer 242 is 90°. In this way, the light emitted by the light source module 210 becomes the first linearly polarized light after passing through the polarizer 241, and the polarization direction of the first linearly polarized light is the same as the polarization direction of the polarizer 241. After the first linearly polarized light passes through the wafer to be identified, due to the birefringence effect of the wafer to be identified, it becomes elliptically polarized light. After the elliptically polarized light passes through the analyzer 242, it becomes the second linearly polarized light, and the second linearly polarized light is then received by the camera module 220, and the camera module 220 captures an image of the wafer to be identified.
[0115] The polarization direction of the polarizer 241 is orthogonally fixed with the polarization direction of the analyzer 242, which can suppress the reflection noise on the surface of the wafer to be recognized, reducing the intensity of the reflected light on the surface of the wafer to be recognized to less than 8%. Moreover, the coaxial optical path design of the light source module 210, the polarizer 241, the wafer to be recognized, and the analyzer 242 can ensure that the transmission direction of the light emitted by the light source module 210 is consistent with the normal direction of the wafer surface, that is, the light emitted by the light source module 210 is perpendicularly incident on the wafer surface, thus reducing the birefringence effect error of the wafer.
[0116] It should be noted that when the wafer to be recognized is a silicon carbide wafer and a white light LED light source is used to irradiate the silicon carbide wafer with blue light through PWM control for image capture, a blue light filter with a bandwidth of 10 nm (center wavelength 460 nm) can be added at the camera end of the camera module 220 to suppress the non-blue light components in the light emitted by the white light LED light source, increasing the chromaticity contrast of the polytype defect region in the silicon carbide wafer to ΔEab = 38.2, while the chromaticity contrast of the polytype defect region in the silicon carbide wafer is only ΔEab = 15.5 under traditional white light. In this way, the accuracy of identifying polytype defects in the silicon carbide wafer can be further improved.
[0117] Moreover, the camera module 220 can convert the captured RGB image of the wafer to be recognized into the HSV space, extract the saturation (S) and value (V) channels, enhance the polytype defect features in the silicon carbide wafer under blue light, increasing the saturation and decreasing the value of the polytype defects (high blue light response), forming a sharp contrast with the background, which can further improve the accuracy of identifying polytype defects in the silicon carbide wafer.
[0118] The orthogonal polarization of the above-mentioned polarizer 241 and analyzer 242 plus the blue light defect enhancement can increase the signal-to-noise ratio (SNR) of the polytype defects in the silicon carbide wafer to 28 dB, which is 2.5 times higher than that of the traditional white light polarization scheme.
[0119] In addition, the polarization direction of the polarizer 241 is orthogonally fixed with the polarization direction of the analyzer 242, which can also eliminate the mechanical rotation error and is suitable for industrial sites in a vibrating environment.
[0120] Optionally, the image processing module 230 can be an industrial computer (equipped with a GPU acceleration card). Optionally, a programmable logic controller (PLC controller) coordinates the actions of the manipulator, the light source module 210, and the camera module 220. Thus, fully automated detection of the defects of the wafer to be recognized can be achieved.
[0121] In this specification, each part is described in a way that combines parallelism and progression. The key points of each part are the differences from other parts. For the same or similar parts among the parts, reference can be made to each other.
[0122] With respect to the above description of the disclosed embodiments, the features described in the various embodiments in this specification may be replaced or combined with each other, enabling those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but rather will be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for identifying wafer defects, characterized in that, Including: Obtain an image of the wafer to be recognized; Extract the texture features of the image, and extract the global chromaticity distribution features of the image; Fuse the texture features and the global chromaticity distribution features to obtain a fused feature; Based on the fused feature, determine the defect recognition result of the wafer to be recognized, where the defect recognition result includes: defect location and defect type.
2. The wafer defect recognition method according to claim 1, characterized in that, The wafer to be recognized is a silicon carbide wafer, and obtaining an image of the wafer to be recognized includes: Obtain an image of the wafer to be recognized taken under blue light or ultraviolet light irradiation.
3. The wafer defect identification method according to claim 1, wherein The wafer defect recognition method is implemented by a configured wafer defect recognition model, and the wafer defect recognition model includes: A first encoding module for extracting texture features from the input image; A second encoding module for extracting global chromaticity distribution features from the input image; A feature fusion module for fusing the texture features and the global chromaticity distribution features to obtain a fused feature; An output module for determining the defect recognition result of the wafer to be recognized based on the fused feature.
4. The wafer defect identification method according to claim 3, wherein The first encoding module is a ResNet-34 network structure.
5. The wafer defect recognition method according to claim 3, wherein The second encoding module is a ViT network structure.
6. The wafer defect identification method according to claim 3, wherein The feature fusion module is specifically configured to use a cross-attention mechanism to fuse the texture features and the global chromaticity distribution features to obtain a fused feature.
7. The wafer defect identification method according to claim 3, wherein The output module includes: A detection head for outputting the defect location of the wafer to be recognized based on the fused feature; A classification head for outputting the defect type of the wafer to be recognized based on the fused feature.
8. The wafer defect identification method according to claim 7, wherein The defect recognition result further includes a pixel-level defect mask, and the output module further includes: A segmentation head for outputting the pixel-level defect mask of the wafer to be recognized based on the fused feature.
9. A wafer defect identification device, characterized in that, Including: An image acquisition unit for obtaining an image of the wafer to be recognized; A feature extraction unit for extracting the texture features of the image and extracting the global chromaticity distribution features of the image; A feature fusion unit for fusing the texture features and the global chromaticity distribution features to obtain a fused feature; A defect prediction unit for determining the defect recognition result of the wafer to be recognized based on the fused feature, where the defect recognition result includes: defect location and defect type.
10. A wafer defect identification system, characterized in that, Including: A light source module, a camera module, and an image processing module; The light source module is used to emit light to irradiate the wafer to be recognized; The camera module is used to take an image of the wafer to be recognized under the irradiation of the light emitted by the light source module; The image processing module is used to execute each step of the wafer defect recognition method according to any one of claims 1-9.
11. The wafer defect identification system according to claim 10, characterized in that, The wafer to be recognized is a silicon carbide wafer, the light source module includes an LED array, and the LED array includes a red LED, a green LED, and a blue LED; The light source module further includes a control module, which is used to modulate the proportion of red light, green light, and blue light in the light emitted by the LED array, so that the proportion of blue light in the light emitted by the LED array is in the range of 60% - 90%, including the end values. The spectral peak of the light emitted by the LED array is in the blue light band, and the light emitted by the LED array presents blue light.
12. The wafer defect identification system according to claim 10 or 11, characterized in that, The wafer defect recognition system further includes a polarization module; The polarization module includes a polarizer and an analyzer. The polarizer is disposed between the light source module and the wafer to be recognized, and the analyzer is disposed between the wafer to be recognized and the camera module. The light source module, the polarizer, the wafer to be recognized, and the analyzer are coaxially arranged, and the polarization direction of the polarizer is perpendicular to the polarization direction of the analyzer.