Wafer defect detection method and system

Abstract

The application discloses a wafer defect detection method and system, and relates to the technical field of wafer defect detection, and comprises the following steps: acquiring a multi-focus section image of a wafer to be detected; extracting edge features of the multi-focus section image under different focal planes to generate a depth mask; fusing the multi-focus section image into a full-depth wafer image and a full-depth edge condition image by using the depth mask; inputting the full-depth wafer image into a condition diffusion model, and taking the full-depth edge condition image as a structure guide condition to constrain the inverse denoising process of the condition diffusion model, and outputting a defect-free reconstructed image; comparing the full-depth wafer image with the defect-free reconstructed image and extracting residual features; dividing the residual features into local difference zones corresponding to different depths by using the depth mask; and extracting defect regions based on local statistical threshold values by using the features of the local difference zones. By using the above scheme, the detection accuracy and precision of a wafer with a high aspect ratio structure during defect detection are improved.

Description

Technical Field

[0001] This invention relates to the field of wafer defect detection technology, and in particular to a wafer defect detection method and system. Background Technology

[0002] Currently, wafer defect detection and microstructure measurement technologies are widely used in semiconductor manufacturing, advanced integrated circuit process control, and wafer yield analysis. As semiconductor manufacturing processes continue to evolve towards advanced nodes, device structures are gradually shifting from traditional two-dimensional planar structures to complex three-dimensional structures, such as three-dimensional NAND flash memory, FinFET transistors, and high aspect ratio through-silicon vias (TSVs). During the fabrication of these three-dimensional devices, the wafer surface typically forms numerous high aspect ratio step structures, deep trench structures, and complex cavity microstructures. In critical process steps such as etching, thin film deposition, and cleaning, nanoscale particle contamination, etching residues, bridging defects, and localized morphological distortions are easily generated. Therefore, high-precision, non-destructive defect detection and morphological characterization are required.

[0003] In existing technologies, defect detection for wafers with high aspect ratio structures typically relies on a single two-dimensional imaging method. For example, an optical microscope is used to acquire images of the wafer surface at a fixed focal plane, and then image processing algorithms are used to identify defects. However, due to the significant height difference in the vertical direction of high aspect ratio stepped structures, images from a single focal plane can only clearly present the local morphology of the top or bottom surface of the stepped structure, and cannot simultaneously achieve high-fidelity imaging of both the top and bottom surfaces. When the acquisition device focuses on the top surface, the bottom area is out of focus, resulting in a blurred image and loss of defect information; the reverse is also true. This defocusing blind spot caused by the depth of focus limitation makes it difficult for conventional two-dimensional imaging methods to obtain complete wafer surface information, thus affecting the comprehensiveness and accuracy of defect detection.

[0004] Furthermore, even when attempting to capture and stitch together images from multiple focal lengths, existing technologies are prone to producing ghosting or fusion artifacts at the edges of steps when fusing images from different focal planes. Moreover, they lack effective means to distinguish the differences in defects at different depths, resulting in insufficient accuracy and reliability in defect extraction. Summary of the Invention

[0005] The purpose of this invention is to provide a wafer defect detection method and system, which aims to solve the technical problem in the prior art that it is difficult to balance detection precision and accuracy when performing defect detection on wafers with high aspect ratio structures.

[0006] To achieve the above objectives, in a first aspect, this application provides a wafer defect detection method, applied to a wafer to be inspected having a high aspect ratio structure, comprising the following steps: Acquire multi-focal-length images of the wafer to be inspected. These images are obtained by capturing images at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer, traversing the defocus blind zone. Edge features of multi-focal length images under different focal planes are extracted to generate depth masks. The depth masks are then used to fuse the multi-focal length images into a panoramic deep wafer image and a panoramic deep edge condition map. The panoramic depth wafer image is input into the conditional diffusion model, and the panoramic depth edge condition map is used as the inverse denoising process of the structure-guided conditional constraint diffusion model to output a defect-free reconstructed image. By comparing the panoramic depth wafer image with the defect-free reconstructed image and extracting residual features, the residual features are divided into local difference regions corresponding to different depths using a depth mask. The defect region is then extracted based on the features of each local difference region and a local statistical threshold.

[0007] This invention provides a wafer defect detection method. First, by acquiring multi-focal-length images spanning the defocus blind zone, the acquisition process can cover the top and bottom surfaces of the stepped structure at two different focal planes, thus overcoming the limitation of single-focal-plane imaging by high aspect ratio structures and fully acquiring image information from different depth planes. Based on this, edge features under different focal planes are extracted to generate a depth mask. This depth mask essentially marks and distinguishes the depth level to which each pixel in the image belongs. Using the depth mask to guide image fusion avoids the information confusion at the step edges caused by traditional direct stitching or superposition, fusing the multi-focal-length images into a clear full-depth wafer image and a corresponding full-depth edge condition map. In the defect detection stage, this application reconstructs the full-depth wafer image using a conditional diffusion model and uses the full-depth edge condition map as a structurally guided condition constraint for inverse denoising. This ensures that the reconstructed defect-free image maintains the original stepped structure morphology of the wafer, while defects, as abnormal features deviating from the normal structure, are removed during the reconstruction process. Finally, residual features are extracted by comparing the full-depth wafer image with the defect-free reconstructed image. A depth mask is then used to divide the residual features into different depth regions. Defects are extracted separately based on local statistical thresholds for each local difference region. This allows defects in the top and bottom surfaces to be independently judged using thresholds adapted to the noise characteristics of that region, avoiding false positives or false negatives caused by using a single global threshold. Compared with existing technologies, the technical solution of this application enables end-to-end detection of wafers with high aspect ratio stepped structures, from image acquisition, fusion, and reconstruction to layered defect extraction, improving the accuracy and precision of defect detection.

[0008] Further, the comparison of the full-depth wafer image and the defect-free reconstructed image, and the extraction of residual features, followed by the use of a depth mask to divide the residual features into local difference regions corresponding to different depths, includes: The absolute pixel error and structural similarity deviation at corresponding pixel positions are extracted from the panoramic depth wafer image and the defect-free reconstructed image. A residual matrix representing the degree of reconstruction deviation is constructed using the absolute pixel error and structural similarity deviation. By using depth masking to decouple the residual matrix, the residual matrix is ​​decoupled into multiple independent difference submatrices corresponding to the top and bottom surfaces of the stepped structure, respectively.

[0009] Further, the step of extracting defect regions based on local statistical thresholds using the features of each of the local difference regions includes: Extract the residual values ​​within each independent difference submatrix, and calculate the local mean and local standard deviation of each independent difference submatrix; The local mean and local standard deviation are assigned preset weight coefficients and then summed in a weighted manner to generate a hierarchical filtering threshold corresponding to the depth of each independent difference submatrix. In the independent difference submatrix, pixels with residual values ​​lower than the hierarchical filtering threshold are screened out as background noise, and pixels with residual values ​​higher than or equal to the hierarchical filtering threshold are identified as defect areas.

[0010] Furthermore, acquiring the multi-focal-length image of the wafer to be inspected includes: Z-axis scanning is controlled by the acquisition equipment based on the high aspect ratio depth prior parameters of the wafer to be inspected. After determining the first focusing height of the top surface of the corresponding stepped structure, based on the high aspect ratio depth prior parameter, the second focusing height of the bottom surface of the corresponding stepped structure is obtained by crossing the defocus blind zone downwards, so as to acquire focal length images of the corresponding top and bottom surfaces respectively.

[0011] Furthermore, the fusion of multi-focal length images using depth masks includes: Edge feature matrices of multi-focal length images under different focal planes are extracted respectively, and depth masks are generated by comparing the pixel-level magnitude of the edge feature matrices point by point. Based on preset coordinate transformation relationships and depth masks, multi-focal length images are recombined pixel by pixel in a unified coordinate system to obtain a panoramic depth wafer image.

[0012] Furthermore, the extraction of edge feature matrices under different focal planes includes: Using camera extrinsic parameters including rotation matrix R and translation vector t, the pixel coordinates of the acquired top and bottom images are obtained. Mapped to world coordinates The mapping relationship is A spatially aligned top focal length image is obtained through pixel interpolation. and the bottom focal length image ; The top focal length image was processed using edge detection operators. and the bottom focal length image Calculate the gradient component gx in the horizontal direction and the gradient component gy in the vertical direction, and calculate the total gradient magnitude to generate the top edge feature matrix G1 and the bottom edge feature matrix G2.

[0013] In the above embodiments, the top and bottom images are mapped to a unified world coordinate system through camera extrinsic parameters, and spatial alignment is achieved by using pixel interpolation. This eliminates the spatial offset that may occur when the acquisition device scans at different focal planes, and provides a spatially consistent basis for subsequent point-by-point comparison of edge feature matrices and accurate generation of depth masks.

[0014] Furthermore, the output defect-free reconstructed image includes: Reduce the dimensionality of the panoramic deep wafer image to latent space variables; In the process of performing inverse denoising on latent space variables in the conditional diffusion model, the panoramic depth edge condition map is input into the cross-attention mechanism to constrain edge generation in the inverse denoising process.

[0015] In the above implementation, by inputting the panoramic depth edge condition map into the cross-attention mechanism, constraints are imposed on the generation of edge regions during the inverse denoising process of the conditional diffusion model, so that the reconstructed image can maintain the clarity and positional accuracy of the step structure edges while removing defects, and prevent edge information from being blurred or shifted during the denoising process.

[0016] Further, the step of generating the depth mask by comparing the pixel-level magnitudes of the edge feature matrix point by point includes: A generalized logic function is introduced as a nonlinear mapping function to map the pixel difference between the top edge feature matrix G1 and the bottom edge feature matrix G2 into a depth mask matrix M. The formula for calculating the depth mask matrix M at pixel coordinates (x, y) is as follows: ; in, It is a natural constant. The sharpening coefficient k is a preset scalar value, which is greater than zero and is used to control the sharpness of the mask weight conversion to prevent image fusion ghosting. when (x,y) is much larger than When (x,y), the weights of the depth mask matrix M tend to be 1; when (x,y) is much smaller than When (x,y), the weights of the depth mask matrix M tend to be 0.

[0017] Secondly, this application provides a wafer defect detection system for performing a wafer defect detection method as described in the first aspect or any possible implementation thereof, the system comprising: The multi-focal-length image acquisition module is used to acquire multi-focal-length images of the wafer to be inspected. The multi-focal-length images are acquired by crossing the defocus blind zone at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer. The feature extraction and fusion module is used to extract edge features of multi-focal length images under different focal planes to generate a depth mask, and use the depth mask to fuse multi-focal length images into a panoramic deep wafer image and a panoramic deep edge condition map. The conditional diffusion module is used to input the full-depth wafer image into the conditional diffusion model and use the full-depth edge condition map as the inverse denoising process of the structure-guided conditional constraint conditional diffusion model to output a defect-free reconstructed image. The defect analysis module is used to compare the full-depth wafer image with the defect-free reconstructed image and extract residual features. It uses a depth mask to divide the residual features into local difference regions corresponding to different depths, and extracts the defect region based on the features of each local difference region and the local statistical threshold.

[0018] In conjunction with the second aspect, the multi-focal length image acquisition module is further configured to: The acquisition device performs Z-axis scanning based on the high aspect ratio depth prior parameters of the wafer under test; after determining the first focusing height of the top surface of the corresponding step structure, it crosses the defocus blind zone downwards based on the high aspect ratio depth prior parameters to optimize the acquisition of the second focusing height of the bottom surface of the corresponding step structure, so as to acquire the focal length images of the corresponding top and bottom surfaces respectively.

[0019] The beneficial effects of the second aspect described above can be referred to in the first aspect or any possible implementation of the first aspect, and will not be elaborated here. Based on the implementations provided in the above aspects, this application can also be further combined to provide more implementations. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a wafer defect detection method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the defocused blind zone of the wafer to be tested in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the logic of depth mask generation and pixel recombination and fusion in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the reverse denoising process of the conditional diffusion model in an embodiment of the present invention. Figure 5 This is a schematic diagram of a wafer defect detection system according to an embodiment of the present invention.

[0021] Reference numerals: 1. Optical acquisition equipment; 2. First focusing height; 3. Second focusing height; 4. Wafer to be inspected; 5. Defocused blind zone. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0023] Combined with appendix Figure 1 - Appendix Figure 3 The specific embodiments of the present invention will be further described in detail below.

[0024] In the semiconductor manufacturing field, wafer defect detection and microstructure measurement technologies are widely used in advanced integrated circuit process control and wafer yield analysis. As device structures evolve from traditional two-dimensional planes to complex three-dimensional structures, wafer surfaces often exhibit step structures with high aspect ratios, deep trench structures, and complex cavity microstructures. These micro-defects, such as nanoscale particle contamination, etching residues, bridging defects, and local morphological distortions, are easily generated during etching, thin film deposition, and cleaning processes. Therefore, high-precision, non-destructive defect detection and morphological characterization methods are required.

[0025] In related technologies, to address the depth-of-field limitation caused by high aspect ratio structures, some solutions employ multi-focal plane image acquisition, capturing images at different heights and then using image registration and fusion algorithms to obtain images with extended depth of field. For example, a multi-focus image fusion method based on wavelet transform or Laplacian pyramid is used to fuse images from different focal planes after frequency domain decomposition.

[0026] However, such schemes are relatively rough in the transition of information from different focal planes at the edge of the step structure during image fusion, which is prone to fusion ghosting or edge artifacts. Although the fused image expands the depth of field, the subsequent defect detection stage usually adopts a globally uniform judgment threshold or template comparison method, which does not fully consider the differences in imaging conditions and noise characteristics in different depth areas. As a result, it is difficult to ensure the accuracy of defect detection in the top and bottom areas at the same time, which is prone to missed detection or false detection.

[0027] Reference Figure 1 and Figure 2 To address the aforementioned technical problems, in a first aspect, this application provides a wafer defect detection method, applied to a wafer 4 to be inspected with a high aspect ratio structure, comprising the following steps: Step S100: Acquire multi-focal images of the wafer 4 to be inspected. The multi-focal images are acquired by capturing images at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer, across the defocus blind zone 5. Step S200: Extract edge features of multi-focal length images under different focal planes to generate a depth mask, and use the depth mask to fuse the multi-focal length images into a panoramic deep wafer image and a panoramic deep edge condition map. Step S300: Input the panoramic depth wafer image into the conditional diffusion model, and use the panoramic depth edge condition map as the reverse denoising process of the structure-guided conditional constraint diffusion model to output a defect-free reconstructed image; Step S400: Compare the full-depth wafer image with the defect-free reconstructed image and extract residual features. Use a depth mask to divide the residual features into local difference regions corresponding to different depths, and use the features of each local difference region to extract the defect region based on the local statistical threshold.

[0028] In the above embodiments, firstly, multi-focal-length images spanning the defocus blind zone 5 are acquired using optical acquisition device 1. This allows the acquisition process to cover the top and bottom surfaces of the stepped structure, two different focal planes, thereby overcoming the limitation of high aspect ratio structures on single-focal-plane imaging and fully acquiring image information from different depth planes. Based on this, edge features under different focal planes are extracted to generate depth masks. A depth mask essentially marks and distinguishes the depth level to which each pixel in the image belongs. Using depth masks to guide image fusion avoids the information confusion at the steps' edges caused by traditional direct stitching or superposition, fusing the multi-focal-length images into a clear panoramic depth wafer image and a corresponding panoramic depth edge condition map.

[0029] In the defect detection stage, this application reconstructs the full-depth wafer image using a conditional diffusion model and uses the full-depth edge condition map as a structure-guided constraint for inverse denoising. This ensures that the reconstructed defect-free image maintains the original stepped structure of the wafer, while defects, as abnormal features deviating from the normal structure, are removed during reconstruction. Finally, residual features are extracted by comparing the full-depth wafer image with the defect-free reconstructed image. A depth mask is used to divide the residual features into different depth regions, and defects are extracted based on local statistical thresholds for each local difference region. This allows defects in the top and bottom surfaces to be independently judged using thresholds adapted to the noise characteristics of that region, avoiding false positives or false negatives caused by using a single global threshold.

[0030] Reference Figure 3and Figure 4 In conjunction with the first aspect, according to one embodiment of this application, comparing a full-depth wafer image with a defect-free reconstructed image and extracting residual features, and using a depth mask to divide the residual features into local difference regions corresponding to different depths, includes: The absolute pixel error and structural similarity deviation at corresponding pixel positions are extracted from the panoramic depth wafer image and the defect-free reconstructed image. A residual matrix representing the degree of reconstruction deviation is constructed using the absolute pixel error and structural similarity deviation. By using depth masking to decouple the residual matrix, the residual matrix is ​​decoupled into multiple independent difference submatrices corresponding to the top and bottom surfaces of the stepped structure, respectively.

[0031] In the above embodiments, the residual matrix is ​​constructed by simultaneously extracting difference information from two dimensions: absolute pixel error and structural similarity deviation. Absolute pixel error reflects the degree of grayscale deviation of a single pixel, while structural similarity deviation reflects the degree of change in local texture structure. Combining the two can more comprehensively capture the reconstruction deviation caused by defects. Based on this, a depth mask is used to decouple the residual matrix, dividing the original residual matrix, which mixed the difference information from different depth regions, into independent difference sub-matrices corresponding to the top and bottom surfaces. This allows subsequent defect analysis to be performed independently within their respective depth regions, avoiding mutual interference between different depth regions due to different imaging conditions and noise characteristics, and improving the targeting of defect feature extraction.

[0032] In conjunction with the first aspect, according to one embodiment of this application, defect regions are adaptively extracted based on the features of each local difference region, including: Extract the residual values ​​within each independent difference submatrix, and calculate the local mean and local standard deviation of each independent difference submatrix; The local mean and local standard deviation are assigned preset weight coefficients and then summed in a weighted manner to generate a hierarchical filtering threshold corresponding to the depth of each independent difference submatrix. In the independent difference submatrix, pixels with residual values ​​lower than the hierarchical filtering threshold are screened out as background noise, and pixels with residual values ​​higher than or equal to the hierarchical filtering threshold are identified as defect areas.

[0033] In the above embodiments, the local mean and local standard deviation are calculated for different depth regions, and a layered filtering threshold adapted to the noise level of the region is generated based on the local statistics. This allows different thresholds to be used for defect judgment on the top and bottom surfaces. Since the noise distribution and contrast of images at different focal planes differ, using layered thresholds can better adapt to the actual image characteristics of each depth region, improving the adaptability of defect region identification compared to a globally fixed threshold.

[0034] In conjunction with the first aspect, according to one embodiment of this application, obtaining a multi-focal-length image of the wafer 4 to be inspected includes: Z-axis scanning is controlled by the acquisition device based on the high aspect ratio depth prior parameters of the wafer 4 to be inspected. After determining the first focusing height 2 of the top surface of the corresponding step structure, based on the high aspect ratio depth prior parameter, the second focusing height 3 of the bottom surface of the corresponding step structure is obtained by crossing the defocus blind zone 5 downwards, so as to acquire the focal length images of the corresponding top and bottom surfaces respectively.

[0035] In the above embodiments, by using the high aspect ratio depth prior parameter of the wafer 4 to be inspected to guide the Z-axis scanning process, the focus search range can be determined in a targeted manner, avoiding blind large-scale scanning. After determining the top surface focus height, the bottom surface focus height is optimized directly based on the depth prior parameter across the defocus blind zone 5, improving the efficiency and accuracy of bottom surface focus optimization and ensuring that both the acquired top and bottom surface images are in sharp focus.

[0036] In conjunction with the first aspect, according to one embodiment of this application, fusing multi-focal length images using a depth mask includes: Edge feature matrices of multi-focal length images under different focal planes are extracted respectively, and depth masks are generated by comparing the pixel-level magnitude of the edge feature matrices point by point. Based on preset coordinate transformation relationships and depth masks, multi-focal length images are recombined pixel by pixel in a unified coordinate system to obtain a panoramic depth wafer image.

[0037] In the above embodiments, a depth mask is generated by comparing the pixel-level amplitudes of edge feature matrices under different focal planes. This allows for depth-level classification of pixels based on the image's own information, eliminating the need for an additional depth sensor. Pixel reassembly using the depth mask in a unified coordinate system enables selective fusion at the depth level, building upon spatial alignment. This effectively suppresses ghosting and yields a clear panoramic depth wafer image.

[0038] In conjunction with the first aspect, according to one embodiment of this application, the extraction of edge feature matrices under different focal planes includes: Using camera extrinsic parameters including rotation matrix R and translation vector t, the pixel coordinates of the acquired top and bottom images are obtained. Mapped to world coordinates The mapping relationship is as follows: A spatially aligned top focal length image is obtained through pixel interpolation. and the bottom focal length image ; The top focal length image was processed using edge detection operators. and the bottom focal length image Calculate the gradient component gx in the horizontal direction and the gradient component gy in the vertical direction, and calculate the total gradient magnitude to generate the top edge feature matrix G1 and the bottom edge feature matrix G2.

[0039] In the above embodiments, the top and bottom images are mapped to a unified world coordinate system through camera extrinsic parameters, and spatial alignment is achieved by using pixel interpolation. This eliminates the spatial offset that may occur when the acquisition device scans at different focal planes, and provides a spatially consistent basis for subsequent point-by-point comparison of edge feature matrices and accurate generation of depth masks.

[0040] In conjunction with the first aspect, according to one embodiment of this application, outputting a defect-free reconstructed image includes: Reduce the dimensionality of the panoramic deep wafer image to latent space variables; In the process of performing inverse denoising on latent space variables in the conditional diffusion model, the panoramic depth edge condition map is input into the cross-attention mechanism to constrain edge generation in the inverse denoising process.

[0041] In the above embodiments, by inputting the panoramic depth edge condition map into the cross-attention mechanism, constraints are imposed on the generation of edge regions during the inverse denoising process of the conditional diffusion model, so that the reconstructed image can maintain the clarity and positional accuracy of the step structure edges while removing defects, and prevent edge information from being blurred or shifted during the denoising process.

[0042] In conjunction with the first aspect, according to one embodiment of this application, the step of generating a depth mask by comparing pixel-level magnitudes of edge feature matrices point-by-point includes: Introducing a generalized logic function as a nonlinear mapping function, the top surface edge feature matrix is... With bottom edge feature matrix The pixel differences are mapped to the depth mask matrix M; The formula for calculating the depth mask matrix M at pixel coordinates (x, y) is: ; in, It is a natural constant. The sharpening coefficient k is a preset scalar value, which is greater than zero and is used to control the sharpness of the mask weight conversion to prevent image fusion ghosting. when (x,y) is much larger than When (x,y), the weights of the depth mask matrix M tend to be 1; when (x,y) is much smaller than When (x,y), the weights of the depth mask matrix M tend to be 0.

[0043] Specifically, the top surface edge feature matrix The formula for calculating at pixel coordinates (x, y) is: The bottom edge feature matrix In cell coordinates The calculation formula at the location is: . In the above embodiments, a generalized logic function is used as the mapping function, which can continuously map discrete pixel differences into mask weights between 0 and 1, achieving a smooth transition from the top surface to the bottom surface and avoiding sharp segmentation jumps at the step edges. By adjusting the steepness of the mapping through a preset sharpening coefficient k, the sharpness of this transition can be flexibly controlled, clearly distinguishing the top and bottom surface regions while effectively suppressing ghosting artifacts generated at the step edges during the fusion process. When a certain pixel's (x,y) is much larger than When (x,y), it indicates that the pixel has a stronger edge response in the top image and belongs to the top region, and the mask weight tends to 1; conversely, the mask weight tends to 0 and belongs to the bottom region.

[0044] It should be noted that in conventional processing logic, image fusion algorithms often result in high-frequency artifacts at the physical edges of stepped structures. If a fused image with artifacts is directly input into a conventional deep learning model, the model will misjudge the fusion artifacts as structural defects such as etching residues. Simultaneously, conventional diffusion models are prone to irreversibly disrupting the smoothness of complex stepped edges with high aspect ratios during denoising and reconstruction. To address these contradictions, this invention uses a depth mask extracted from multiple focal lengths as the core control constraint throughout the process. The depth mask not only eliminates physical spatial ghosting in image fusion at the front end, extracting a clean full-depth edge condition map; more importantly, the edge condition map, as strong prior information, is directly used within the cross-attention mechanism of the conditional diffusion model, forcibly constraining the locking of physical edge position features when reconstructing a defect-free image. Because the front-end depth mask provides precise structural anchor points, the subsequent residual matrix can be decoupled without interference based on depth information, thereby adaptively stripping away real local defects.

[0045] Reference Figure 5 Secondly, this application provides a wafer defect detection system for performing a wafer defect detection method as described in the first aspect or any possible embodiment of the first aspect, the system comprising: The multi-focal-length image acquisition module is used to acquire multi-focal-length images of the wafer 4 to be inspected. The multi-focal-length images are acquired at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer, spanning the defocus blind zone 5. The feature extraction and fusion module is used to extract edge features of multi-focal length images under different focal planes to generate a depth mask, and use the depth mask to fuse multi-focal length images into a panoramic deep wafer image and a panoramic deep edge condition map. The conditional diffusion module is used to input the full-depth wafer image into the conditional diffusion model and use the full-depth edge condition map as the inverse denoising process of the structure-guided conditional constraint conditional diffusion model to output a defect-free reconstructed image. The defect analysis module is used to compare the full-depth wafer image with the defect-free reconstructed image and extract residual features. It uses a depth mask to divide the residual features into local difference regions corresponding to different depths, and extracts the defect region based on the features of each local difference region and the local statistical threshold.

[0046] The beneficial effects of the above embodiments can be referred to the description of the corresponding embodiments in the first aspect, and will not be repeated here.

[0047] In conjunction with the second aspect, according to one embodiment of this application, the multi-focal length image acquisition module is further configured to: Based on the high aspect ratio depth prior parameters of the wafer 4 to be inspected, the acquisition device is controlled to perform Z-axis scanning; after determining the first focusing height 2 of the top surface of the corresponding step structure, the device is used to cross the defocus blind zone 5 downwards based on the high aspect ratio depth prior parameters to optimize and obtain the second focusing height 3 of the bottom surface of the corresponding step structure, so as to acquire the focal length images of the corresponding top and bottom surfaces respectively.

[0048] The beneficial effects of the above embodiments can be referred to the description of the corresponding embodiments for multi-focal length image acquisition in the first aspect, and will not be repeated here.

[0049] It should be noted that the order of description of the embodiments in this application is not intended to limit the priority of the embodiments, and the execution of the wafer defect detection method described in this invention strongly depends on the physical action of the front-end multi-focal length image acquisition module, that is, based on the real depth prior parameters of high aspect ratio devices such as 3D NAND, the optical focusing system of the hardware acquisition device is controlled in real time to perform Z-axis physical scanning to cross the defocus blind zone in physical space; on the other hand, the depth mask matrix generated in this invention is essentially a precise mapping of the three-dimensional microscopic physical morphology (step top surface and bottom surface) of the wafer surface in two-dimensional digital space.

[0050] Furthermore, the data processed by the conditional diffusion model's inverse denoising process, as well as the final extracted local difference regions, directly correspond to real physical defects generated in wafer manufacturing processes, such as nanoscale particle contamination and etching residues. The entire data processing process constitutes feedback and optimization of the real semiconductor physical measurement process, solving the technical problem that existing optical microscopy inspection equipment is limited by physical depth of focus and cannot achieve high-fidelity imaging. It has clear industrial applicability and physical and technical effects.

[0051] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0052] Furthermore, the signal processing modules in the various embodiments of this application can be integrated into one processing unit, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0053] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute some steps of the transmission and reception methods described in the various embodiments of this application. The aforementioned storage medium can be a volatile or non-volatile computer-readable storage medium, including: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0054] It is understood that the signal processing modules in the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, modules, units, submodules, subunits, etc., can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0055] The above descriptions are some embodiments of this application. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

[0056] For details of this embodiment of the application, please refer to the description of the foregoing method embodiments. To avoid repetition, the description will not be repeated here.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0058] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the present invention. Furthermore, the present invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A wafer defect detection method, applied to a wafer with a high aspect ratio structure, characterized in that, Includes the following steps: Acquire multi-focal-length images of the wafer to be inspected. These multi-focal-length images are acquired by capturing images at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer, spanning the defocus blind zone. Edge features of the multi-focal-length images under different focal planes are extracted to generate a depth mask. The depth mask is then used to fuse the multi-focal-length images into a panoramic deep wafer image and a panoramic deep edge condition map. The panoramic depth wafer image is input into the conditional diffusion model, and the panoramic depth edge condition map is used as a structure-guided condition to constrain the conditional diffusion model in the reverse denoising process, outputting a defect-free reconstructed image. The panoramic depth wafer image is compared with the defect-free reconstructed image and residual features are extracted. The residual features are divided into local difference regions corresponding to different depths using the depth mask. The defect region is extracted based on the features of each local difference region and a local statistical threshold.

2. The wafer defect detection method according to claim 1, characterized in that, The step of comparing the panoramic depth wafer image with the defect-free reconstructed image and extracting residual features, and using the depth mask to divide the residual features into local difference regions corresponding to different depths, includes: Extract the absolute pixel error and structural similarity deviation of the panoramic depth wafer image and the defect-free reconstructed image at corresponding pixel positions, and construct a residual matrix to characterize the degree of reconstruction deviation using the absolute pixel error and the structural similarity deviation; The residual matrix is ​​decoupled using the depth mask, resulting in multiple independent difference sub-matrices corresponding to the top and bottom surfaces of the stepped structure.

3. The wafer defect detection method according to claim 2, characterized in that, The step of extracting defect regions based on local statistical thresholds using the features of each of the local difference regions includes: Extract the residual values ​​within each independent difference submatrix, and calculate the local mean and local standard deviation of each independent difference submatrix; The local mean and the local standard deviation are assigned preset weight coefficients and then summed in a weighted manner to generate a hierarchical filtering threshold corresponding to the depth of each independent difference submatrix. Pixels in the independent difference submatrix with residual values ​​lower than the hierarchical filtering threshold are filtered out as background noise, and pixels with residual values ​​higher than or equal to the hierarchical filtering threshold are identified as defect areas.

4. The wafer defect detection method according to claim 1, characterized in that, The process of acquiring multi-focal-length images of the wafer to be inspected includes: The acquisition device is controlled to perform Z-axis scanning based on the high aspect ratio depth prior parameters of the wafer to be inspected; After determining the first focusing height corresponding to the top surface of the stepped structure, based on the high aspect ratio depth prior parameter, the second focusing height corresponding to the bottom surface of the stepped structure is obtained by crossing the defocus blind zone downwards, so as to acquire focal length images corresponding to the top and bottom surfaces respectively.

5. The wafer defect detection method according to claim 1, characterized in that, The step of fusing the multi-focal length image using the depth mask includes: The edge feature matrices of the multi-focal length images under different focal planes are extracted respectively, and the depth mask is generated by comparing the pixel-level magnitude of the edge feature matrices point by point; Based on the preset coordinate transformation relationship and the depth mask, the multi-focal length image is pixel-wise recombined in a unified coordinate system to obtain the panoramic depth wafer image.

6. The wafer defect detection method according to claim 5, characterized in that, The extraction of edge feature matrices under different focal planes includes: Using camera extrinsic parameters including rotation matrix R and translation vector t, the pixel coordinates of the acquired top and bottom images are obtained. Mapped to world coordinates The mapping relationship is as follows: A spatially aligned top focal length image is obtained through pixel interpolation. and the bottom focal length image ; The top focal length image was processed using edge detection operators. and the bottom focal length image Calculate the gradient component gx in the horizontal direction and the gradient component gy in the vertical direction, and calculate the total gradient magnitude to generate the top edge feature matrix G1 and the bottom edge feature matrix G2.

7. The wafer defect detection method according to claim 1, characterized in that, The output defect-free reconstructed image includes: The panoramic deep wafer image is reduced to a latent space variable; During the inverse denoising process of the latent space variables by the conditional diffusion model, the panoramic depth edge condition map is input to the cross-attention mechanism to constrain the edge generation of the inverse denoising process.

8. The wafer defect detection method according to claim 5, characterized in that, The step of generating the depth mask by comparing the pixel-level magnitudes of the edge feature matrix point by point includes: Introducing a generalized logic function as a nonlinear mapping function, the top surface edge feature matrix is... With the bottom edge feature matrix The pixel differences are mapped to the depth mask matrix M; The formula for calculating the depth mask matrix M at pixel coordinates (x, y) is as follows: ； in, It is a natural constant. The sharpening coefficient k is a preset scalar value, which is greater than zero and is used to control the sharpness of the mask weight conversion to prevent image fusion ghosting. when (x,y) is much larger than When (x,y), the weights of the depth mask matrix M tend to be 1; when (x,y) is much smaller than When (x,y), the weights of the depth mask matrix M tend to be 0.

9. A wafer defect detection system, used to perform a wafer defect detection method as described in any one of claims 1-8, characterized in that, The system includes: A multi-focal-length image acquisition module is used to acquire multi-focal-length images of the wafer to be inspected. The multi-focal-length images are acquired by crossing the defocus blind zone at different focal planes on the top and bottom surfaces of the stepped structure of the corresponding wafer. The feature extraction and fusion module is used to extract the edge features of the multi-focal length image under different focal planes to generate a depth mask, and use the depth mask to fuse the multi-focal length image into a panoramic depth wafer image and a panoramic depth edge condition map. The conditional diffusion module is used to input the panoramic depth wafer image into the conditional diffusion model, and use the panoramic depth edge condition map as a structure-guided condition to constrain the inverse denoising process of the conditional diffusion model, and output a defect-free reconstructed image. The defect analysis module is used to compare the panoramic depth wafer image with the defect-free reconstructed image and extract residual features. The residual features are divided into local difference regions corresponding to different depths using the depth mask, and the defect region is extracted based on the features of each local difference region and the local statistical threshold.

10. A wafer defect detection system according to claim 9, characterized in that, The multi-focal length image acquisition module is also used for: Based on the high aspect ratio depth prior parameters of the wafer to be inspected, the acquisition device is controlled to perform Z-axis scanning; after determining the first focusing height corresponding to the top surface of the stepped structure, the device is then guided downward across the defocus blind zone based on the high aspect ratio depth prior parameters to optimize and obtain the second focusing height corresponding to the bottom surface of the stepped structure, so as to acquire focal length images corresponding to the top and bottom surfaces respectively.

Images