Method for detecting wafer defects by singular value decomposition

The wafer image is preprocessed and defect processing through the singular value decomposition method, which successfully solved the problem of insufficient accuracy and reliability of semiconductor wafer defect detection in the prior art, and achieved efficient early identification of defects and avoid wasted production cycles.

CN117788364BActive Publication Date: 2024-12-03UNITED SEMICONDUCTOR (XIAMEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310352417.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-04-04
Publication Date
2024-12-03
Estimated Expiration
2043-04-04

AI Technical Summary

Technical Problem

In the prior art, when detecting semiconductor wafer defects, especially low-rank semiconductor patterns and sparse defect patterns, there are difficulties in separating and identifying them, which leads to insufficient detection accuracy and reliability. It is often necessary to wait until the final test to find defects, wasting production cycles. and resources.

Method used

The wafer image is preprocessed by the singular value decomposition method, and the low-rank semiconductor pattern is separated from the sparse defect pattern through histogram equalization and singular value decomposition, and defect amplification, reconstruction and denoising are performed, and defect recognition and Detection degree.

Benefits of technology

It significantly improves the recognition and detectability of wafer defects, can detect defects on the wafer surface in the early stage, avoid waste of production cycles, and improves the accuracy and reliability of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117788364B_ABST
    Figure CN117788364B_ABST
Patent Text Reader

Abstract

The present invention proposes a method for detecting wafer defects using singular value decomposition, the steps of which include inputting a wafer image, preprocessing the wafer image using a histogram equalization method to obtain a preprocessed image, performing singular value decomposition on the preprocessed image, amplifying defects in the wafer image after the singular value decomposition, reconstructing the defect-amplified wafer image to obtain a defect-significant image, denoising the defect-significant image, and detecting defects in the denoised wafer image.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention generally relates to a method for detecting wafer defects, and more particularly, to a method for detecting wafer defects by singular value decomposition. Background Art

[0002] Semiconductor wafers contain multiple layers. Each layer undergoes complex and precise material deposition, photoresist coating, etching and / or ion implantation process steps to produce IC patterns. Before proceeding to the next layer, it is necessary to detect whether the layer contains defects before continuing to the next layer. Common wafer defects may include scratches, abnormal exposure, particle contamination, hot spots, wafer edge defects, and other various defects that affect the electrical properties and performance of the final chip. If they are not detected immediately after the wafer layer is deposited, such defects can only be discovered during the final test, wasting valuable production cycles and resources.

[0003] Currently, the most intuitive way to detect wafer defects is to visually inspect the wafer surface, such as manual visual inspection such as after etching inspection (AEI), or machine vision inspection such as KLA scanning. Wafer defects may be formed at any position on the wafer surface, and are usually mixed in the complex IC background of the wafer layer formed by the previous process. In this case, even with the use of highly accurate machine vision, it is still very difficult to detect and identify these defects. Those skilled in the art still need to improve the existing wafer defect detection methods in order to improve the accuracy and reliability of wafer defect detection. Summary of the invention

[0004] In view of the above situation, the present invention proposes a method for detecting wafer defects, which is characterized in that low rank semiconductor patterns and sparse defect patterns in wafer images are separated by a singular value decomposition method, thereby greatly improving the recognition and detectability of defects.

[0005] The purpose of the present invention is to propose a method for detecting wafer defects using singular value decomposition, the steps of which include inputting a wafer image, preprocessing the wafer image using a histogram equalization method to obtain a preprocessed image, performing singular value decomposition on the preprocessed image, amplifying the defects of the wafer image after the singular value decomposition, reconstructing the defect-amplified wafer image to obtain a defect-significant image, denoising the defect-significant image, and detecting defects in the denoised wafer image.

[0006] These and other objects of the present invention will become more apparent to those skilled in the art after reading the following detailed description of the preferred embodiment which is described in various diagrams and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A flow chart of a method for detecting wafer defects according to a preferred embodiment of the present invention;

[0008] Figure 2 A schematic diagram of decomposing a wafer image by a singular value decomposition method according to a preferred embodiment of the present invention;

[0009] Figure 3 is a schematic diagram of the result of singular value decomposition of an m×n matrix; and

[0010] Figure 4 FIG. 4 is a diagram showing the evolution of a wafer image during defect detection according to a preferred embodiment of the present invention.

[0011] Please note that all illustrations in this manual are of a legend nature. For the sake of clarity and convenience of illustration, the components in the illustrations may be exaggerated or reduced in size and proportion. Generally speaking, the same reference symbols in the figures will be used to indicate corresponding or similar component features after modification or in different embodiments.

[0012]

Main component symbol description

[0013] 10-12 images

[0014] 20-24 images

[0015] S r×r order diagonal matrix

[0016] Steps S1 to S6

[0017] U is a unitary matrix of order m×r

[0018] V T r×n-order transposed unitary matrix

[0019] X is an m×n-order pixel value matrix DETAILED DESCRIPTION

[0020] Now the exemplary embodiments of the present invention will be described in detail below, and it will illustrate the described features with reference to the accompanying drawings so that those skilled in the art can understand and achieve the technical effect. Those skilled in the art will appreciate that the description in the text is only carried out by way of example, and is not intended to limit the application. Various embodiments of the present application and various features that do not conflict with each other in the embodiments can be combined or reset in various ways. Without departing from the spirit and scope of the present invention, modifications, equivalents or improvements to the present application are understandable to those skilled in the art, and are intended to be included in the scope of the present application.

[0021] Preferably, one or more of the embodiments of the present invention are implemented on a computer-implemented program or control system. The computer executing such a program or control system may be a general-purpose computer architecture, wherein a user may input instructions for executing the computer-implemented method of the present invention through a user interface including a display, keyboard, mouse, etc., or the instructions may be automatically implemented by an automatic control system. The processor in the computer may read computer-readable code and data from the memory and perform operations and processing on the computer-implemented method. A high-capacity storage device, such as a disk drive, may provide program code or data related to the computer-implemented method of the present invention to be loaded into the memory. An input / output device may provide a data connection to transmit data to other devices, such as a network, a modem, a printer, a testing machine, etc.

[0022] First, please refer to Figure 1 , which depicts a flow chart of a method for detecting wafer defects according to a preferred embodiment of the present invention. First, at the beginning of the process, a wafer image is provided as a target object of the wafer defect detection method of the present invention, such as Figure 4 The image 20 in the figure. The wafer image may be a wafer surface image of the wafer after any semiconductor process, such as a wafer image after an etching process, which may be a color analog or digital image format in the embodiment of the present invention. The image 20 may include normal semiconductor patterns of the current layer and the previous layer of the wafer, such as IC patterns or photoresist patterns produced in the previous process, and may also contain abnormal patterns such as wafer defects, such as the inconspicuous scratches in the upper left corner of the image 20, which are the defect features to be detected by the wafer defect detection method of the present invention.

[0023] After providing the wafer image 20 to be inspected, in step S1, a preprocessing is performed on the image 20, which may include inputting the image 20 into a computer system and converting it into a digital, grayscale image that is convenient for computer processing. Specifically, the digital grayscale image 20 can be regarded as a matrix composed of multiple pixels, each pixel has its corresponding pixel value, such as 256 grayscale values ​​ranging from 0 to 255, where pure black is defined as 0 and pure white is defined as 255. Then, the number of pixels with each grayscale value in the image matrix can be counted by histogram equalization to establish a histogram of the image grayscale value. By adjusting the cumulative distribution function (CDF) in the histogram, the area with high pixel values ​​can be dispersed to the surrounding area, thereby stretching the pixel intensity distribution range, so that the image information is more balanced in the grayscale value distribution of 0 to 255, which can significantly improve the global contrast of the image, achieve the effect of improving the subjective visual effect of the image, and facilitate the subsequent image processing.

[0024] Furthermore, in this step, the image after histogram equalization can be normalized to normalize the pixel values ​​after equalization to a value between 0 and 1, that is, the percentage of the pixel value to the sum of all pixel values ​​in the image matrix, so as to facilitate subsequent image processing. Figure 4 From image 21, we can see that the global contrast of the wafer image is significantly improved after processing.

[0025] After completing the preprocessing of the wafer image, in step S2, singular value decomposition is performed on the image 21. Figure 2 , which is a schematic diagram of decomposing a wafer image by a singular value decomposition method according to a preferred embodiment of the present invention. For a wafer image 10, the pattern therein may include a normal low-rank part 11, such as a regular semiconductor pattern formed by the current layer and the previous layer, and an abnormal sparse part 12, such as a wafer defect such as a scratch to be detected, wherein the rank represents the richness, redundancy, and noise entropy of the image, and the lower the rank, the higher the redundancy of the pattern, and the higher the dependence between the columns of its image matrix, such as a regular semiconductor pattern part. On the other hand, the higher the rank, the higher the richness of the pattern, and the lower the correlation between the columns of its image matrix, such as an abnormal wafer defect part, which is not related to the normal semiconductor pattern, and is an independent pattern part on the whole wafer, and is also the target part to be detected by the present invention. Since the wafer image has structural information, it is suitable to use a low-rank matrix to express the pixel value matrix of the image 21 after the aforementioned histogram equalization normalization, and use a large amount of redundant information therein to decompose and restore the abnormal information therein.

[0026] In step S2, the pixel value matrix of the image 21 after the histogram equalization normalization is subjected to singular value decomposition, and the formula is as follows:

[0027] X=USV T

[0028] like Figure 3 As shown, X is the m×n pixel value matrix of the aforementioned image 21, U is the m×r unitary matrix after its singular value decomposition, S is the r×r diagonal matrix after its singular value decomposition, V T is the r×n order transposed unitary matrix after singular value decomposition, where the resulting matrix U and matrix V T The value in is the corresponding eigenvector, and the value λ on the diagonal of the obtained diagonal matrix S i (i is between 1 and r) is the eigenvalue of the pixel value matrix X, r is the rank of the pixel value matrix X, λr The larger the value, the more important the content information in the image.

[0029] After completing the singular value decomposition of the image 21, the defect amplification is then performed in step S3. The specific method of this step is to multiply the singular value decomposition result by a coefficient K, and the formula is as follows:

[0030] K.USV T =X′

[0031] The coefficient K is the number of eigenvalues ​​between 0 and 1 multiplied by the number of non-zero eigenvalues ​​in the diagonal matrix S of the singular value decomposition result.

[0032] After the above-mentioned defect magnification action, the pixel values ​​of the low-rank normal semiconductor pattern parts in the pixel value matrix of the original image 21 will be greater than a predetermined value, for example, greater than the maximum grayscale value 255 in the 256-level grayscale setting, which will appear pure white in the grayscale image, while the wafer defect parts representing sparse abnormalities in the pixel value matrix will be less than the predetermined value, which will appear black. Through such an action, the defective parts in the wafer image can be highlighted. Figure 4 The image 22 in the figure is the wafer image obtained after the above defect magnification operation of the image 21. It can be seen that the low-rank normal semiconductor pattern parts in the original image 21 are visually filtered out, leaving only the sparse abnormal wafer defect parts. Figure 2 The separation action of a low-rank semiconductor pattern and a sparse defect pattern is shown.

[0033] After the above defect magnification is completed, the defect saliency image is reconstructed in step S4. The specific method of this step is to convert the above defect magnified grayscale image into a binary image through a binarization method, for example, the grayscale of pixels with grayscale values ​​greater than 255 is set to the grayscale maximum value 1, and the grayscale of pixels with grayscale values ​​less than 255 is set to the grayscale minimum value 0, thereby achieving binarization. Figure 4 Image 23 is the wafer image obtained after the above-mentioned defect-significant image reconstruction operation of image 22. It can be seen that the blurred and dim defect pattern in the original image 22 becomes visually very obvious against the normal semiconductor pattern background, which facilitates the subsequent defect detection operation.

[0034] After completing the above-mentioned reconstruction of the defective significant image, the image denoising operation is then performed in step S5. The specific method of this step is to process the aforementioned valued wafer image by a closing operation, which may include two operations: dilation and erosion. The dilation process is to set the value of the input pixel to 1 when one or more of the pixel values ​​of the input pixel and its surrounding pixels relative to the structural element of 0 is 1. The result of the operation will make the image look enlarged, and the gap between images can be filled with appropriate structural elements. On the other hand, the erosion process is to set the value of the input pixel to 255 when the pixel values ​​of the input pixel and its surrounding pixels relative to the structural element of 0 are both 255. The result of the operation will make the image look shrunk, and appropriate structural elements can be used to remove unnecessary elements in the image, such as the redundant semiconductor pattern background in image 23. Figure 4 The image 24 in the figure is the wafer image obtained after the aforementioned defective image 23 is processed by the above-mentioned closing operation. It can be seen that the background noise entropy patterns such as the normal semiconductor patterns in the original image 23 are visually completely removed, leaving only the defective patterns desired by the invention.

[0035] After obtaining the final wafer image 24 with only significant defect patterns, the defect detection action of step S6 can be performed. The specific method of this step is to first remove the interference factors outside the wafer image, such as the black background in image 24. Then obtain the coordinate axis of each black point in the image, and count the number of points around each of these points, as shown in the box in image 24, which is the counting area range. When the maximum value counted is greater than a preset value, it can be determined that there is a defect in the wafer image.

[0036] The above descriptions are only preferred embodiments of the present invention. All equivalent changes and modifications made according to the claims of the present invention should fall within the scope of the present invention.

Claims

1. A method for detecting wafer defects using singular value decomposition, comprising: Input wafer image; The wafer image is preprocessed using a histogram equalization method to obtain a preprocessed image; Performing singular value decomposition on the preprocessed image; Performing defect magnification on the wafer image after singular value decomposition, wherein the defect magnification operation includes determining a coefficient by multiplying the number of eigenvalues ​​between (0, 1) and the number of non-zero eigenvalues ​​in the diagonal matrix after singular value decomposition, and the wafer image magnified by the coefficient can distinguish defects from background parts; Reconstructing the wafer image after the defect is enlarged to obtain a defect-significant image; Denoising the image with significant defects; as well as Detect defects in the denoised wafer image.

2. The method for detecting wafer defects using singular value decomposition as described in claim 1, wherein the wafer image is a matrix composed of multiple pixels, each of the pixels has a corresponding pixel value, and the histogram equalization method includes converting the wafer image into a grayscale image and normalizing it.

3. The method for detecting wafer defects using singular value decomposition as described in claim 2, wherein the pixel value before the histogram equalization step is between 0 and 255, and the pixel value after the histogram equalization step is between 0 and 1, which is the percentage of the pixel value to the sum of all the pixel values ​​in the matrix.

4. The method for detecting wafer defects using singular value decomposition as described in claim 2, wherein the singular value decomposition step includes dividing the wafer image into a low-rank part and a sparse part, the low-rank part represents the background image of the wafer, and the sparse part represents the defect image on the wafer.

5. The method for detecting wafer defects using singular value decomposition as described in claim 4, wherein the defect magnification step includes multiplying the pixel value by the coefficient so that the pixel values ​​of the background image are all greater than 255 and the pixel values ​​of the defect image are all less than 255, thereby magnifying the defect image.

6. A method for detecting wafer defects using singular value decomposition as described in claim 5, wherein the singular value decomposition divides the matrix of m×n order into three multiplied matrices, namely, an m×r unitary matrix, an r×r diagonal matrix, and an r×n transposed unitary matrix, wherein the coefficient is the number of eigenvalues ​​between 0 and 1 in the diagonal matrix S multiplied by the number of non-zero eigenvalues.

7. The method for detecting wafer defects using singular value decomposition as described in claim 1, wherein reconstructing the wafer image after the defect is magnified includes using a binarization method to set pixel values ​​greater than a set value in the image after the defect is magnified as maximum pixel values ​​and pixel values ​​less than the set value as minimum pixel values, thereby obtaining the defect-significant image.

8. The method for detecting wafer defects using singular value decomposition as claimed in claim 1, wherein the denoising step comprises removing noise points using a closing operation, including sequentially performing a dilation process and an erosion process.

9. The method for detecting wafer defects using singular value decomposition as described in claim 1, wherein the defect detection action further includes obtaining the coordinate axis of each black dot in the image and counting the number of points in the area around each of the points. If the maximum value of the number of points exceeds a preset value, the wafer image is determined to have defects.