Defect detection model and method based on DIC point scanning wafer signals

Abstract

This invention discloses a defect detection model and method based on DIC point scanning wafer signals. The method includes: S1, acquiring DIC point scanning detection results from wafer defect test pieces under different scanning spot spacing settings, converting them into Polar images, and segmenting and labeling the defect targets in the view; using image scale transformation technology to balance the total amount of data under different scanning spot spacings; cropping the entire image to the input size set by the model and generating diverse training samples using data augmentation methods; S2, selecting and optimizing the detection base model, including small-scale defect optimization and automated hyperparameter selection optimization; S3, feeding the data-augmented image dataset into the model for network training to obtain the DIC wafer defect detection model. Before model inference, the wide scanning spot spacing data in the data to be detected is scaled. This invention ensures the stability and reliability of the detection results under different equipment or process conditions.

Description

Technical Field

[0001] This invention relates to wafer defect detection technology, and more particularly to a defect detection model and method based on DIC point scanning wafer signals. Background Technology

[0002] Differential Interference Contrast (DIC) microscopy is a key method in wafer defect detection, used to enhance height difference contrast. It reveals defect morphologies with extremely subtle height or optical path differences by splitting polarized light into two orthogonal beams and re-interfering them upon sample return. DIC can reveal minute structural changes in wafer patterns caused by minute height differences more clearly than traditional bright-field or dark-field microscopy, aiding in the development of hard-to-detect defects.

[0003] Currently, defect detection methods for DIC point scanning wafer signals in industrial scenarios can be divided into machine learning-based detection methods and deep learning-based detection methods.

[0004] Machine learning-based detection methods rely on traditional feature engineering to design features and then use algorithms to learn patterns. For example, Chinese patent CN119600002A discloses an automatic detection method for semiconductor wafer defects. This method extracts defect features and configures a detection wheel through clustering, and then performs iterative matching to detect defects.

[0005] Deep learning-based detection methods automatically extract and learn features through multi-layer neural networks. For example, Chinese patent CN120385689A proposes a method, device, and equipment for detecting wafer appearance defects, using a difference model and an AI model for defect detection. However, this method also suffers from long training times and difficulty in model convergence during training.

[0006] Meanwhile, semiconductor testing equipment employs various testing modes to address the diverse testing needs of manufacturers. Different testing modes, with varying scanning spot spacing settings, produce image signals of different scales. Defect morphologies on images of different scales will exhibit different appearances. Significant differences in image quality exist between different modes, leading to a higher false negative rate in imaging with wider spot spacing.

[0007] In summary, current defect detection methods for DIC point scanning wafer signals have the following problems:

[0008] (1) Existing machine learning-based detection methods rely on manual feature design. When faced with atypical or novel defect patterns, the feature expression ability is insufficient, resulting in a decrease in detection accuracy and limiting the detection stability and applicability of the method.

[0009] (2) Existing deep learning-based detection methods rely heavily on high-quality training data. However, wafer defect data in industrial production is sensitive and scarce, making it difficult to obtain sufficient and high-quality pre-training samples, thus limiting the detection performance of the model.

[0010] (3) The scanning spot spacing has a significant impact on the image representation of defects in DIC point scanning wafers. Especially with wide spot spacing, small defects often exhibit characteristics very similar to the surrounding background, making it difficult to effectively detect such defects in images with wide spot spacing.

[0011] (4) In actual data collection, there is a large amount of data under the narrow spot spacing condition, while there is very little data under the wide spot spacing condition. This will lead to a serious imbalance in the distribution of training data. The model learns well for the samples with a large proportion, but misses a lot of samples with a small proportion, showing a typical long-tail problem. It is difficult to ensure the consistency of detection results for the same defect under different spot spacing conditions. Summary of the Invention

[0012] The purpose of this invention is to provide a defect detection model and method for DIC point scanning wafer signals, so as to alleviate the problem of low model generalization and long-tail problem caused by insufficient high-quality data sources in industrial scenarios, while adapting to the scale and intensity differences brought about by different scanning sources, and improving the automation level of DIC defect detection.

[0013] To this end, the present invention provides a DIC wafer defect detection model, the acquisition method of which includes: S1, collecting DIC point scanning detection results from wafer defect test pieces under different scanning spot spacing settings, converting them into Polar images and performing defect target segmentation and annotation on the view; using image scale transformation technology to balance the total amount of data under different scanning spot spacings; cropping the entire Polar image to the input size set by the model and using data augmentation methods to generate diverse training samples; S2, selecting and optimizing a basic model for DIC wafer defect detection, including small-scale defect optimization and automated hyperparameter selection optimization; S3, feeding the data-augmented image dataset into the model for network training to obtain a deployable DIC wafer defect detection model.

[0014] According to another aspect of the present invention, a defect detection method for DIC point scan wafer signals is provided, comprising: S4, using a super-resolution reconstruction method to refine the defect features of a Polar image with a wide scanning spot spacing in the data to be detected, and cropping the refined Polar image to the input size set by the model; S5, inputting the cropped image into the DIC wafer defect detection model to obtain the defect coordinates detected in each cropped image; S6, performing defect coordinate mapping based on the position of the cropped image in the whole image for the defect coordinates detected in each cropped image to obtain the coordinates of the defect detection result in the complete wafer Polar image, which is used as the final defect detection result.

[0015] The present invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the above-described defect detection method.

[0016] Compared to traditional detection methods that rely solely on training with limited samples, this invention improves the effectiveness of model training from a data perspective through the synergistic effect of data balancing, amplification, and refinement, significantly enhancing the robustness and universality of the detection system against complex wafer defects.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0019] Figure 1 This is a flowchart of the method for obtaining the DIC wafer defect detection model of the present invention;

[0020] Figure 2 This is a schematic diagram of defect patching in the data augmentation method of the present invention;

[0021] Figure 3 This is a schematic diagram of image blending in the data enhancement method of the present invention;

[0022] Figure 4 This is a flowchart of the automated hyperparameter selection process of the present invention;

[0023] Figure 5 This is a flowchart of the defect detection method of the present invention;

[0024] Figure 6 This is a schematic diagram of the detection results of wide scanning spot spacing data plus long tail suppression according to the present invention;

[0025] Figure 7 This is a schematic diagram of the detection results of low-resolution data scaling and refinement according to the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Combined with reference Figures 1 to 3 The method for obtaining the DIC wafer defect detection model of the present invention includes the following steps S1 to S3.

[0028] S1. Construct a training dataset. In this stage, long-tail suppression and data augmentation are performed on the DIC wafer defect data.

[0029] S2. Select and optimize the base model for DIC wafer defect detection, including small-scale defect optimization and automated hyperparameter selection optimization;

[0030] S3. Feed the enhanced image dataset into the model for network training to obtain a deployable DIC wafer defect detection model.

[0031] Specifically, step S1 above includes steps S11 to S14.

[0032] S11. Collect DIC point scanning detection results from professional wafer defect test pieces under different scanning spot spacing settings, convert the Mercator view of the original detection results into a Polar view with analyzable spatial location, and complete the defect target segmentation and annotation based on the Polar view.

[0033] S12. Image scale transformation techniques using super-resolution reconstruction and degradation enhancement are employed to balance the total data volume under different scanning spot spacings, thereby reducing the impact of uneven data distribution caused by scale differences on detection accuracy. This step includes:

[0034] S121. For images with narrow scanning spot spacing and high resolution, a degradation enhancement method based on downsampling is used to generate corresponding images with wide scanning spot spacing and low resolution, balancing the total amount of data under different scanning spot spacing settings.

[0035] S122. For images with wide scanning spot spacing and low resolution, a super-resolution reconstruction method based on upsampling is used to generate corresponding images with narrow scanning spot spacing and high resolution, enriching the feature distribution of defects at this scale. Considering the impact of the super-resolution method on the image feature distribution, the scale change range should not be too large at this time, so as to preserve the defect distribution features at the current scale to the greatest extent.

[0036] S13. Image Cropping. The entire Polar view wafer image is cropped to the input size required for network training according to preset values. Due to the sparsity of defect distribution, images that do not contain defect targets are filtered out from the dataset during this process to improve network training efficiency.

[0037] S14. Data Augmentation. A series of data augmentation methods are employed to generate diverse training samples based on the cropped image data, effectively expanding the number of samples and improving the model's ability to learn defect features under conditions of limited data sources.

[0038] The data augmentation methods used mainly include:

[0039] S141, Defect Pasting. A substrate image far from the wafer edge is randomly selected from the image set. Defect region images corresponding to the segmentation labels are randomly cropped from the currently retained dataset. These defect images are pasted onto the substrate image at random positions and angles, generating an image with new defect distribution characteristics and a corresponding annotation file. Unlike the "hard pasting" in traditional enhancement methods, to mitigate the impact of image differences between defect edges and the substrate image on model training, when cropping the defect region image, a certain pixel value expansion and edge smoothing should be performed on the segmentation label region to ensure smooth transition of the pasted edges. In implementation, the current dataset is first traversed and read. After a certain degree of expansion based on the segmentation labels, defect targets are cropped to establish a defect target repository for the current dataset. During pasting, defect targets are randomly selected from the defect target repository, and pixel weights near the edges are generated using a Gaussian distance decay function. These weighted targets are then pasted onto the current image, such as... Figure 2 As shown.

[0040] S142. Image Mixing. Randomly select two to four images from the current dataset containing defects. Perform random cropping, scaling, and other transformations on each image. Then, stitch these processed images together onto a new canvas to form a composite defect image containing multiple targets and their different contextual backgrounds. During the stitching process, the corresponding ground truth annotations for each image should also be adjusted and integrated into the new annotation file, such as... Figure 3 As shown.

[0041] In step S2 above, the selected base model is YOLOv11, and it is optimized. The optimization mainly includes two aspects:

[0042] (1) Model optimization for small-scale defects. To expand the receptive field of the feature map and improve the model detection efficiency, classic target detection models are often based on C3 and deeper feature maps for detection. However, in DIC point scan wafer defect data, the defect targets to be detected mostly show small-scale features in the Polar view. This results in the defect target being represented as only a pixel in the deep feature map after multiple downsampling, which makes it difficult to effectively model the structure and detailed features of the defect. To address this problem, an additional detection branch is designed on the basis of the baseline model, incorporating the C2 layer feature map with richer detailed features into the detection. At the same time, the C2 feature layer is added as the shallowest feature layer to the feature pyramid feature fusion process, so that the deep feature map can also learn the detailed distribution of small-scale defects, enhancing the network's ability to model target defects;

[0043] (2) Automated Hyperparameter Selection and Optimization. To reduce the randomness and trial-and-error costs of intuitive hyperparameter tuning and meet the model training requirements for defect detection in industrial scenarios, the manual hyperparameter tuning process is optimized into automated hyperparameter selection. A series of continuous hyperparameters to be tuned are combined into a high-dimensional vector. Based on Bayesian optimization theory, the probability distribution relationship between the hyperparameter vector and model performance is established through iterative sampling and fitting, transforming the hyperparameter optimization problem into finding the extrema of a function in a high-dimensional space. The automated hyperparameter selection process is as follows: Figure 4 As shown:

[0044] To address the diversity of target probability distribution functions, Gaussian processes are preferentially chosen as surrogate models for fitting. Gaussian processes are a class of stochastic processes. Its arbitrary n-dimensional distribution All of them follow a multivariate normal distribution. That is, for each observation point of this function... All satisfy a Gaussian distribution:

[0045]

[0046] in, Represents the mean function, This represents the covariance function. A commonly used covariance function is the radial basis function (RBF), which can be expressed as:

[0047]

[0048] in, It represents the signal variance and is used to control the fluctuation range of the covariance function; This indicates the length scale and is used to control the decay distance of variable correlation.

[0049] After the previous round of surrogate model updates, the sampling location for the next hyperparameter space is selected using an acquisition function. To comprehensively consider both the estimated sampling point value and accuracy, the EI function is used as the acquisition function.

[0050]

[0051]

[0052] in, This represents the target value corresponding to the best hyperparameter set in the current proxy model; This represents the exploration parameters, used to control the degree of relaxation in accuracy during the selection of sampling points; Represents the cumulative distribution function. This represents the probability density function.

[0053] The model is trained by iteratively selecting hyperparameter sets, continuously increasing the number of sampling points for the target probability distribution function, and utilizing the maximum a posteriori estimation method. and By iteratively fitting the parameters, the surrogate model can continuously approximate the target probability distribution function, fit the probability distribution relationship between the hyperparameter vector and the model performance, and then calculate the optimal hyperparameter set for the current task.

[0054] like Figure 2 As shown, the defect detection method for DIC point scanning wafer signals of the present invention includes the following steps S4 to S6.

[0055] S4. The Polar image with wide scanning spot spacing in the data to be detected is scaled up using the super-resolution reconstruction method to refine its defect features, and the refined Polar image is cropped to the input size set by the model.

[0056] S5. Input the cropped image into the DIC wafer defect detection model to obtain the defect coordinates detected in each cropped image.

[0057] S6. For the defect coordinates detected in each cropped image, perform defect coordinate mapping based on the position of the cropped image in the full image to obtain the coordinates of the defect detection result in the complete wafer Polar image, which is used as the final defect detection result.

[0058] Step S4 is used to preprocess the data to be detected before model inference. For Polar views with wide scanning spot spacing in the data to be detected, a super-resolution reconstruction method is used to refine the defect features, giving more easily discernible details to the originally blurry small defects, so as to achieve effective detection of defects in wide scanning spot spacing images; for images with narrow scanning spot spacing, no additional processing is required. Also considering the impact of the super-resolution method on the distribution of image features, the scale change range should not be too large at this time to avoid changes in defect features caused by the super-resolution reconstruction itself.

[0059] For scale refinement, sampling interpolation algorithms are mainly used. The tangential signal is downsampled slightly, while the radial signal is interpolated. Interpolation algorithms can include nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation. Unlike traditional interpolation algorithms, this involves coordinate transformation from the original signal to the image. The bilinear interpolation algorithm can be represented as:

[0060]

[0061] Among them To scale up the image The grayscale value. To refine the image The original signal obtained by coordinate transformation The signal strength. These are the weights. The weights are as follows:

[0062]

[0063] Among them They are respectively The value obtained by performing coordinate transformation. They are respectively The result obtained by rounding down is... .

[0064] For bicubic interpolation (the bicubic difference is taken from the surrounding 4), (4 pixels), the corresponding interpolation algorithm can be expressed as:

[0065]

[0066] Among them It can be represented as

[0067]

[0068] Step S5 is used for model inference, which involves cropping the scaled Polar image to the size required by the model, and then inputting the cropped images into the trained deployable DIC wafer defect detection model in sequence to obtain the corresponding defect detection results for each image.

[0069] Step S6 is used for post-processing of the detection results. It mainly includes the following steps:

[0070] S61. Edge False Detection Removal. Based on a pre-set defect radius threshold, false detections of defects at wafer edges are removed to address false detections in the Polar image caused by wafer edge pixel filling.

[0071]

[0072] in, This is the edge detection function. A value of 0 indicates that the current defect is judged as a non-edge false detection, and a value of 1 indicates that it is judged as an edge false detection. This indicates the defect currently being assessed and its coordinates. This represents the defect radius threshold, with a value range of (0, 1). This indicates the radius of the scanned wafer.

[0073] S62. Defect Coordinate Mapping. For the defect coordinates detected in each cropped image, defect coordinate mapping is performed based on the position of the cropped image in the full image to obtain the coordinates of the defect detection result in the full wafer Polar view, which is used as the final defect detection result.

[0074] This invention has the following technical advantages / effects:

[0075] (1) By introducing a long-tail suppression mechanism in the training data preparation stage and adopting a scale transformation method that combines super-resolution reconstruction and downsampling, the scale distribution of defect images can be balanced under different scanning spot spacings, effectively balancing the amount of data and reducing the problem of data distribution imbalance caused by scale differences. This improves the detection performance of the detection model on wide scanning spot spacing data with a small number of samples. Specifically, as follows: Figure 6 As shown in the image, after adding the long-tail suppression mechanism, the detection model becomes more sensitive to minute defects visible to the naked eye in the image, achieving more comprehensive defect detection.

[0076] (2) By using diverse data augmentation methods such as image blending and defect patchwork, the number of training samples can be significantly increased under the condition of limited data sources, enriching the morphology and distribution characteristics of defects, improving the model's learning ability for scarce or complex defects, avoiding the problem of low generalization of the model due to insufficient samples, and improving the model's generalization ability, as shown in the table below:

[0077] Table 1. Comparison of Data Augmentation Effectiveness

[0078] Data set A contains 209 images; dataset B, obtained by augmenting dataset A, contains 522 images; and dataset C, obtained by augmenting dataset A in a different way than B, contains 522 images. All datasets are randomly divided into training, validation, and test sets in a 7:1:2 ratio. As shown in the table, both models A and B exhibit performance degradation when applied to dataset C; however, model B shows a smaller performance degradation than model A, indicating that model B has stronger generalization ability.

[0079] (3) During the inference stage, a scale refinement operation based on super-resolution reconstruction is introduced into the low-resolution image with a wide spacing between scanned spots. This allows for the detailed restoration of blurred defect features, thereby enhancing the discriminability of minute defects and improving the detection rate of subtle defects in low-resolution data. Specifically, as follows: Figure 7 As shown, after scale refinement, the features of minute defects in the image to be detected are enhanced, enabling the model to detect minute defects that are easily missed in the original image.

[0080] The method of this invention maintains the consistency of data sources while adapting and optimizing defect images under various scanning modes, ensuring the stability and reliability of detection results under different equipment or process conditions.

[0081] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A defect detection model for DIC wafers, characterized in that, The methods for obtaining it include: S1. Collect DIC point scanning detection results from wafer defect test pieces under different scanning spot spacing settings, convert them into Polar images, and perform defect target segmentation and annotation on the view; use image scale transformation technology to balance the total amount of data under different scanning spot spacing; crop the entire Polar image to the input size set by the model and use data augmentation methods to generate diverse training samples; S2. Select and optimize the basic model for DIC wafer defect detection, including small-scale defect optimization and automated hyperparameter selection optimization; S3. Input the augmented image dataset into the model for network training to obtain a deployable DIC wafer defect detection model.

2. The DIC wafer defect detection model according to claim 1, characterized in that, Step S12 includes: S121. For images with narrow scanning spot spacing and high resolution, use a downsampling-based degradation enhancement method to generate corresponding images with wide scanning spot spacing and low resolution. S122. For images with wide scanning spot spacing and low resolution, use a super-resolution reconstruction method based on upsampling to generate corresponding images with narrow scanning spot spacing and high resolution.

3. The DIC wafer defect detection model according to claim 1, characterized in that, The entire Polar image is cropped to the input size set by the model. This process includes filtering out images from the dataset that do not contain defective targets.

4. The DIC wafer defect detection model according to claim 1, characterized in that, Step S13 includes defect stitching, which involves randomly selecting a substrate image far from the wafer edge from the image set, randomly cropping the defect region image corresponding to the segmentation label from the currently retained dataset, and pasting the defect image into the substrate image at random positions and angles to generate an image with new defect distribution characteristics, while generating a corresponding annotation file.

5. The DIC wafer defect detection model according to claim 4, characterized in that, In defect collage, the current dataset is first traversed and read. Based on the segmentation labels, the defect targets are partially expanded and then cropped to establish a defect target warehouse for the current dataset. During collage, defect targets are randomly selected from the defect target warehouse, and the Gaussian distance decay function is used to generate pixel weights near the edges. These weighted targets are then pasted into the current image.

6. The DIC wafer defect detection model according to claim 4, characterized in that, Step S13 also includes image blending, which involves randomly selecting two to four images from the current dataset containing defects, performing random transformations on each image, including cropping and scaling, and then stitching these processed images onto a new canvas to form a composite defect image containing multiple targets and their different contextual backgrounds. During the stitching process, the real annotations corresponding to each image are also adjusted and integrated into the new annotation file.

7. The DIC wafer defect detection model according to claim 1, characterized in that, In step S2, the automated hyperparameter selection optimization is used to combine the continuous hyperparameters to be tuned into a high-dimensional vector. Based on Bayesian optimization theory, the probability distribution relationship between the hyperparameter vector and the model performance is established through iterative sampling fitting. Then, the extreme values ​​of the function in the high-dimensional space are found to obtain the hyperparameters.

8. A defect detection method for DIC dot-scan wafer signals, characterized in that, include: S4. Use super-resolution reconstruction to refine the defect features of the Polar image with wide scanning spot spacing in the data to be detected, and then crop the refined Polar image to the input size set by the model. S5. Input the cropped image into the DIC wafer defect detection model according to any one of claims 1 to 7 to obtain the defect coordinates detected in each cropped image; S6. For the defect coordinates detected in each cropped image, perform defect coordinate mapping based on the position of the cropped image in the full image to obtain the coordinates of the defect detection result in the complete wafer Polar image, which is used as the final defect detection result.

9. The defect detection method for DIC dot-scan wafer signals according to claim 8, characterized in that, The defect features are scaled up using a super-resolution reconstruction method, including: since the tangential signal resolution of the point scan wafer signal is higher than the radial signal resolution, a small amount of downsampling is performed on the tangential signal, and interpolation is performed on the radial signal.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the defect detection method for DIC point scanning wafer signals according to claim 8.

Images