A semiconductor defect detection method and system

Abstract

The application discloses a semiconductor defect detection method and system, and the semiconductor defect detection method comprises the following steps: step S1, a light source is irradiated to a wafer surface, and then an original image of the wafer surface is collected by a detector; step S2, an image processing unit receives the original image of the wafer surface collected by the detector, and performs image decomposition on the original image of the wafer surface to calculate an AGC image and a GAIN image; step S3, a unified threshold is used for defect detection on the AGC image; and step S4, after the defect detection on the AGC image is completed, AGC defect local images and GAIN defect local images are intercepted respectively, information of the AGC defect local images and the GAIN defect local images is merged, and original light intensity information of a defect image is obtained. The application provides a semiconductor defect detection method and system, the detection method is faster and more accurate, and has better universality.

Description

Technical Field

[0001] This invention relates to a semiconductor defect detection method and system, belonging to the field of semiconductor manufacturing technology. Background Technology

[0002] Currently, in the semiconductor manufacturing field, wafer fabs use defect inspection equipment to inspect each process step of semiconductor manufacturing, thereby improving manufacturing yield. Defect inspection equipment typically scans and images samples, then analyzes the images to extract defects.

[0003] In advanced manufacturing processes, as linewidths shrink, even small defects can impact yield. To detect these smaller defects, equipment uses highly sensitive sensors, such as PMTs (photomultiplier tubes), to receive signals. These sensors have a high dynamic range and can simultaneously capture both strong and weak signals.

[0004] The conventional approach is to obtain an image from a sensor that reflects the true light intensity and then perform defect detection. However, such images often contain multiple regions with completely different content, resulting in significant differences in the signal-to-noise ratio (SNR) of defects in these regions. For example, in patterned wafer inspection using dark-field equipment, the scattered signal is very strong in areas containing rich circuit patterns, but weaker in areas without patterns, leading to variations in background noise intensity. The same applies to defect signals; some types of defects produce very strong signals after imaging, while others are relatively weak. If the same image data contains several defects with significantly different SNRs due to different regions, defect detection becomes challenging. This challenge primarily lies in optimizing the sensitivity across the entire detection area.

[0005] Current mainstream defect detection algorithms compare adjacent dies, quantify the differences, and extract defects by setting a threshold. If the same algorithm threshold is used for the entire detection area, the sensitivity of defect detection is often poor. One solution is to configure appropriate Regions of Interest (CAs) in the recipe to segment regions with different signal-to-noise ratios. Different algorithm thresholds can be configured in each CA. This way, if each algorithm threshold is suitable for its respective CA, the sensitivity of defect detection can be significantly improved. However, using segmented CAs has several drawbacks. First, it requires high accuracy in the location of the segmented regions; otherwise, a large number of false defects are easily generated at the boundaries. Second, when there are multiple regions that need to be segmented, recipe creation becomes very inconvenient, especially since multiple sets of suitable algorithm thresholds need to be set. In fact, many logic chip designs contain a large number of regions that need to be segmented, to the point that manual segmentation is impossible. Finally, conventional CAs are rectangular and cannot accurately segment regions.

[0006] Therefore, there is an urgent need for a defect detection method to solve the above problems. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a semiconductor defect detection method and system that is faster, more accurate and has better universality.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a semiconductor defect detection method, comprising the following steps: Step S1: Illuminate the wafer surface with a light source, and then the detector acquires the original image of the wafer surface; Step S2: The image processing unit receives the original image of the wafer surface acquired by the detector, and performs image decomposition on the original image of the wafer surface to calculate the AGC image and the GAIN image. Step S3: Perform defect detection on the AGC image using a uniform threshold; Step S4: After completing the defect detection of the AGC image, extract the local images of the AGC defect and the GAIN defect respectively, and merge the information of the local images of the AGC defect and the GAIN defect to obtain the original light intensity information of the defect image.

[0009] Furthermore, the calculation steps for the GAIN image in step S2 are as follows: Step S21: Calculate the gain that needs to be added to all pixels in the original image column by column, then obtain the brightness of the current pixel or the local brightness of the area where the current pixel is located in the original image, and then obtain the gain value that needs to be added to pixel (i, j) by looking up the table. Step S22: Use the gain value obtained from the lookup table as the value of pixel (i,j) in the GAIN image, GAIN(i,j); Step S23: Perform low-pass filtering on the GAIN image; Step S24: After determining the gain of pixel (i, j), calculate the gray value of pixel (i, j) after gain based on the current gray value of pixel (i, j) and the gain value.

[0010] Furthermore, the calculation steps for the AGC image in step S2 are as follows: Step S25: Obtain the gray value of pixel (i,j) from the original image, denoted as ORIGINAL(i,j); obtain the value of pixel (i,j) from the GAIN image, denoted as GAIN(i,j); apply GAIN(i,j) to ORIGINAL(i,j) to obtain the AGC image of pixel (i,j); Step S26: Use ORIGINAL(i,j) and GAIN(i,j) as keys to perform a lookup table retrieval to obtain the value AGC(i,j) after gain of pixel (i,j); Step S27: Repeat steps S25 to S26 to traverse all pixels and obtain the complete AGC image.

[0011] Furthermore, step S3 specifically includes the following steps: Step S31: Perform image preprocessing on the AGC image; Step S32: Extract defective pixels; Step S33: Merge and filter the extracted defective pixels.

[0012] Furthermore, step S32 specifically includes the following steps: Subtract the pixel value corresponding to the REF image from the pixel value corresponding to the TARGET image, and then compare the absolute value of the difference with a threshold. If the absolute value of the difference is less than or equal to the threshold, it is determined that there is no defect; if the absolute value of the difference is greater than the threshold, the current pixel is determined to be a defective pixel.

[0013] Furthermore, step S33 specifically includes the following steps: Set a distance threshold D. If the distance between any two defective pixels is less than the distance threshold D, it means that the two defective pixels come from the same defective image block. Using any machine learning or neural network model, select defective image blocks that meet the detection requirements and discard the remaining defective image blocks.

[0014] Furthermore, step S4 specifically includes the following steps: After completing the defect detection of the AGC image, the corresponding positions of the defects in the AGC image and the GAIN image are obtained. The corresponding positions of the defects in the AGC image and the GAIN image are the same. Based on the location of the defect in the AGC image and the GAIN image, the local image of the defect is extracted from the AGC image and the GAIN image respectively to obtain the AGC defect local image and the GAIN defect local image. The extracted defect local image is denoted as Patch Image. The information from the AGC image and the GAIN image is merged to obtain the original light intensity information of the defect.

[0015] Another aspect of the present invention provides a system that applies the semiconductor defect detection method, comprising a wafer, a stage, an image processing unit, a vertical incident energy source, a semi-reflective mirror, a vertical incident illumination objective, a vertical incident acquisition device, an oblique incident energy source, an oblique incident illumination objective, and an oblique incident acquisition device. The wafer is placed on a display platform; The light emitted by the vertically incident energy source passes through a semi-reflective lens and then shines vertically onto the vertically incident illumination objective. The vertically incident illumination objective focuses the light emitted by the vertically incident energy source onto the wafer surface. The vertically incident acquisition device is used to acquire images of the wafer surface in the vertical direction. The oblique incident energy source illuminates the wafer surface from the oblique angle, the oblique incident illumination objective lens focuses the light emitted by the oblique incident energy source onto the wafer surface, and then the oblique incident acquisition device acquires an image of the wafer surface from the oblique angle. The image processing unit is used to perform defect detection on the raw image data acquired by the vertical incident acquisition device and the oblique incident acquisition device.

[0016] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention can normalize the defect signal-to-noise ratio, simplify the setting of scanning formulas, and use a uniform threshold for defect detection. It reduces false defects at CA edges caused by insufficient CA position accuracy in complex formulas.

[0017] 2. In areas with rich patterns, where it is impossible to accurately segment the area by setting the recipe, this method can be used to normalize the signal-to-noise ratio to improve the defect detection rate.

[0018] 3. This invention is applicable to various implementation methods. If computational efficiency is emphasized, the image decomposition algorithm can be implemented using an FPGA; if algorithmic flexibility is emphasized, it can be implemented using an image processor.

[0019] 4. The core of this invention is to use an image with normalized signal-to-noise ratio (SNR) in defect detection, rather than an image reflecting the original light intensity. By adjusting the local grayscale values ​​of the image in real-time low-frequency gain, the SNR is normalized. When detecting different types of regions within a die, compared to configuring complex formulas, this invention directly processes the image, resulting in better versatility. Attached Figure Description

[0020] Figure 1 This is a flowchart of the semiconductor defect detection method of the present invention; Figure 2 This is a schematic diagram of traversing pixels column by column in the original image coordinate system of this invention; Figure 3 This is the original image before image decomposition according to the present invention; Figure 4 This is the AGC image obtained after image decomposition according to the present invention; Figure 5 This is the GAIN image obtained after image decomposition according to the present invention; Figure 6 This is a schematic diagram of a defect die on a wafer according to the present invention; Figure 7 This is a schematic diagram of the semiconductor defect detection system of the present invention. Detailed Implementation

[0021] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0022] Example 1

[0023] like Figure 1 As shown, this embodiment provides a semiconductor defect detection method, including the following steps: Step S1: Irradiate the wafer surface with an energy source. In this embodiment, the energy source can be a deep ultraviolet laser, electron beam, plasma beam, etc. Then, a detector acquires the original image of the wafer surface. In this embodiment, the detector can be a PMT photomultiplier tube or a CCD camera. One energy source can be used for irradiation and one detector can be used for acquisition, or multiple energy sources can be used to irradiate from multiple angles and multiple detectors can be used for acquisition.

[0024] Step S2: The image processing unit receives the raw image of the wafer surface acquired by the detector and performs image decomposition on the raw image (ORIGINAL image) to obtain the AGC image and the GAIN image. Specifically, this includes the following steps: The AGC image is the image after gain processing; the GAIN image is a gain information image used to describe the added gain; both the AGC and GAIN images are the same size as the original image.

[0025] This embodiment employs a scanning mode with vertical spot scanning and horizontal wafer movement. Other scanning modes, such as horizontal spot scanning and vertical wafer movement scanning, can achieve the same effect using similar calculation methods and by adjusting relevant parameters.

[0026] The above-mentioned methods for generating AGC and GAIN images employ a method of adding gain to the image based on the brightness of the region where each pixel is located, specifically as follows: First, pre-set a user-defined or default target grayscale value, such as 2000. After setting the target grayscale value, you can add the corresponding gain according to the target grayscale value.

[0027] The following description of image pixel positions will be based on a Cartesian coordinate system, where pixels are integer grid points in the coordinate system, with both x and y starting at 0.

[0028] Calculate the GAIN image. Specifically: Step S21, as follows Figure 2 As shown, the gain that needs to be added to all pixels in the original image is calculated column by column. The column traversal order is from left to right, and the traversal in each column is from top to bottom. The order is consistent with the pixel order in the scanning process. Then, the brightness of the current pixel or the local brightness of the area where the current pixel is located in the original image is obtained. Then, the gain value that needs to be added to pixel (i, j) is obtained by looking up a table. In this embodiment, a one-dimensional lookup table 1 is used for the lookup.

[0029] For pixel (i, j), that is, the j-th row in the i-th column.

[0030] This embodiment provides a gain lookup method: obtain the gray value ORIGINAL(i,j) of pixel (i,j) in the original image (ORIGINAL image), and use this gray value as the key for searching in the one-dimensional lookup table 1.

[0031] This embodiment also provides another gain lookup method: using the grayscale value of pixel (i, j) in the mean-filtered image MEAN as the key for searching in the one-dimensional lookup table 1. The advantage of this method is that it reduces gain instability caused by rapid changes in image content. In this embodiment, a rectangular neighborhood with width w=9 and height h=9 is used for mean filtering. The specific mean filtering parameters can be chosen using neighborhoods of arbitrary shape and size to adapt to different imaging methods.

[0032] One-dimensional lookup table 1

[0033] One-dimensional lookup table 1 is used to find the gain corresponding to the grayscale value: Assume the pixel count of the original image is log (light intensity), which is the logarithm of the light intensity. The gain value in the one-dimensional lookup table can be calculated as: Gain = Target gray value - Current gray value. Furthermore, upper and lower boundaries can be set for the gain, for example, controlling the gain between -1000 and +1000.

[0034] Step S22: Obtain the gain value corresponding to ORIGINAL(i,j) or MEAN(i,j) from the one-dimensional lookup table 1, and use the gain value obtained from the lookup table as the value of pixel (i,j) in the GAIN image, GAIN(i,j).

[0035] Step S23: Perform a low-pass filter on the GAIN image to remove rapid changes in gain and high-frequency noise in the gain. For example, a moving average with a length of 20 pixels can be used.

[0036] Step S24: After determining the gain of pixel (i, j), it is necessary to calculate the grayscale value of pixel (i, j) after gain based on the current grayscale value of pixel (i, j) and the gain magnitude. In this embodiment, a two-dimensional lookup table 2 can be used, where the key is the grayscale value and the gain, and the value is the grayscale value after gain. The two-dimensional lookup table 2 can be obtained by applying different gains to the detector under different light intensities. It can also be obtained through calculation, such as directly adding the gain to the grayscale value. Alternatively, a more complex two-dimensional linear function can be used, where the grayscale value after gain has a linear relationship with the current grayscale value and the gain.

[0037] Two-dimensional lookup table 2

[0038] Two-dimensional lookup table 2 is used to find the grayscale value and the corresponding gain-adjusted grayscale value.

[0039] By calculation: Gray value after gain = Current gray value + Gain.

[0040] When the post-gain grayscale value exceeds the range of 0 to 4095, the lookup table needs to be processed. One way to do this is to perform the following operation on all calculated post-gain grayscale values ​​in the table: gl=MIN(MAX(gl,0),4095); Where gl is the grayscale value.

[0041] Then, the AGC image is calculated. Specifically: Step S25: Obtain the grayscale value of pixel (i,j) from the original image, denoted as ORIGINAL(i,j); obtain the value of pixel (i,j) from the GAIN image, denoted as GAIN(i,j); apply GAIN(i,j) to ORIGINAL(i,j) to obtain the AGC image of pixel (i,j).

[0042] Step S26: Use ORIGINAL(i,j) and GAIN(i,j) as keys to search in the two-dimensional lookup table 2 to obtain the value AGC(i,j) after gain of pixel (i,j).

[0043] Step S27: Repeat steps S25 to S26 to traverse all pixels in the same order as in the GAIN image calculation steps to obtain a complete AGC image.

[0044] In this embodiment, Figure 3 This is the original image before image decomposition according to the present invention. Figure 4 This is the AGC image obtained after image decomposition according to the present invention. Figure 5 This is the GAIN image obtained after image decomposition according to the present invention.

[0045] Step S3: Perform defect detection on the AGC image using a uniform threshold. This includes the following steps: Because dies on a wafer have a repetitive characteristic, conventional defect detection methods (die-to-die inspection) compare multiple dies, such as... Figure 6 As shown, the parts that differ are considered defects.

[0046] Step S31: Perform image preprocessing on the AGC image. Specifically: The AGC image of the target die on the wafer is called the TARGET image, and the image of the die adjacent to the target die (which can be the die to the left or the die to the right of the target die) is called the REF image (refer to the image).

[0047] Align the REF image with the target die image by calculating the NCC (normalized cross correlation) values ​​of the REF image and the target die image under multiple offsets in the x and y directions, and then selecting the offset corresponding to the best NCC value for movement.

[0048] After alignment, noise will be present in the image. Noise can include several types, such as salt-and-pepper noise and shot noise. The noise is then reduced through filtering.

[0049] Step S32: Extract defective pixels. Specifically: Subtract the pixel value corresponding to the REF image from the pixel value corresponding to the TARGET image, and then compare the absolute value of the difference with a threshold. If the absolute value of the difference is less than or equal to the threshold, it is determined that there is no defect; if the absolute value of the difference is greater than the threshold, the current pixel is determined to be a defective pixel.

[0050] Step S33: Merge and filter the extracted defective pixels. Specifically: Merging defective pixels includes setting a distance threshold D. If the distance between any two defective pixels is less than the distance threshold D, it means that the two defective pixels come from the same defective image block.

[0051] Defect filtering includes: using any machine learning or neural network model to select defective image blocks that meet the detection requirements, and discarding the remaining defective image blocks. Detection requirements are customized according to customer specifications, capturing only the defect types required by the customer.

[0052] Step S4: After completing the defect detection of the AGC image, extract the local images of the AGC defect and the GAIN defect respectively, and merge the information from the local images of the AGC defect and the GAIN defect to obtain the original light intensity information of the defect image. Specifically, this includes the following steps: After defect detection is completed on the AGC image, the corresponding positions of the defects in the AGC and GAIN images are obtained. Since the corresponding positions are the same in both images, local images of the defects are extracted from both images based on these positions. These extracted local defect images are denoted as Patch Images. The defect local image region can be of any size, such as 64*64 pixels. Extraction is performed on both the AGC and GAIN images. The pixels on the Patch Image of the AGC image correspond one-to-one with the pixel positions on the Patch Image of the GAIN image.

[0053] The information from the AGC defect local image and the GAIN defect local image is merged. After merging, the gray value of the defect image is obtained by looking up the two-dimensional lookup table 3. This is the original light intensity information of the defect image, thus obtaining a patch image that reflects the original intensity of the defect.

[0054] Two-dimensional lookup table 3

[0055] Two-dimensional lookup table 3 is used to find the original grayscale value corresponding to the grayscale value and gain.

[0056] By calculation: Original grayscale value = Grayscale value - Gain.

[0057] When the original grayscale value exceeds the range of 0 to 4095, the lookup table needs to be processed. One way to do this is to perform the following operation on all calculated original grayscale values ​​in the table: gl = MIN(MAX(gl,0),4095); Where gl is the grayscale value.

[0058] Example 2

[0059] like Figure 7 As shown, this embodiment provides a semiconductor defect detection system, which includes: a wafer 108, a stage 109, an image processing unit 110, a vertical incident energy source 101, a semi-reflective mirror 102, a vertical incident illumination objective lens 103, a vertical incident acquisition device 104, an oblique incident energy source 105, an oblique incident illumination objective lens 106, and an oblique incident acquisition device 107.

[0060] During testing, wafer 108 is first placed on stage 109.

[0061] The light emitted by the vertically incident energy source 101 passes through the semi-reflective mirror 102 and then shines vertically onto the vertically incident illumination objective 103. The vertically incident illumination objective 103 focuses the light emitted by the vertically incident energy source 101 onto the surface of the wafer 108. Then, the vertically incident acquisition device 104 acquires an image of the surface of the wafer 108 in the vertical direction.

[0062] The oblique incident energy source 105 illuminates the surface of the wafer 108 from the oblique direction. The oblique incident illumination objective lens 106 focuses the light emitted by the oblique incident energy source 105 onto the surface of the wafer 108. Then, the oblique incident acquisition device 107 acquires an image of the surface of the wafer 108 from the oblique direction.

[0063] The image processing unit 110 performs defect detection on the raw image data acquired by the vertical incident acquisition device 104 and the oblique incident acquisition device 107.

[0064] The image processing unit 110 in this embodiment can be an image processor or a combination of an FPGA chip and an image processor. The image decomposition is performed in the FPGA chip, and the decomposed AGC image and GAIN image are then transmitted to the image processor for defect detection.

[0065] Both the vertical incident acquisition device 104 and the oblique incident acquisition device 107 in this embodiment can be PMT photomultiplier tubes or CCD cameras.

[0066] The specific embodiments described above further illustrate the technical problems, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semiconductor defect detection method, characterized in that, Includes the following steps: Step S1: Illuminate the wafer surface with a light source, and then the detector acquires the original image of the wafer surface; Step S2: The image processing unit receives the original image of the wafer surface acquired by the detector, and performs image decomposition on the original image of the wafer surface to calculate the AGC image and the GAIN image; the AGC image is the image after gain processing, and the GAIN image is the gain information image. Step S3: Perform defect detection on the AGC image using a uniform threshold; Step S4: After completing the defect detection of the AGC image, the corresponding positions of the defects in the AGC image and the GAIN image are obtained. The corresponding positions of the defects in the AGC image and the GAIN image are the same. Based on the corresponding positions of the defects in the AGC image and the GAIN image, the local images of the defects are extracted from the AGC image and the GAIN image respectively to obtain the AGC defect local image and the GAIN defect local image. The extracted defect local images are denoted as Patch Images. The pixel positions on the Patch Image of the AGC image correspond one-to-one with the pixel positions on the Patch Image of the GAIN image. The information of the AGC defect local image and the GAIN defect local image are merged. After merging, the gray value of the defect image is obtained by looking up a table. The gray value of the defect image is the original light intensity information of the defect image.

2. The semiconductor defect detection method according to claim 1, characterized in that, The calculation steps for the GAIN image in step S2 are as follows: Step S21: Calculate the gain that needs to be added to all pixels in the original image column by column, then obtain the brightness of the current pixel or the local brightness of the area where the current pixel is located in the original image, and then obtain the gain value that needs to be added to pixel (i, j) by looking up the table. Step S22: Use the gain value obtained from the lookup table as the value of pixel (i,j) in the GAIN image, GAIN(i,j); Step S23: Perform low-pass filtering on the GAIN image; Step S24: After determining the gain of pixel (i, j), calculate the gray value of pixel (i, j) after gain based on the current gray value of pixel (i, j) and the gain value.

3. The semiconductor defect detection method according to claim 1, characterized in that, The calculation steps for the AGC image in step S2 are as follows: Step S25: Obtain the gray value of pixel (i,j) from the original image, denoted as ORIGINAL(i,j); obtain the value of pixel (i,j) from the GAIN image, denoted as GAIN(i,j); apply GAIN(i,j) to ORIGINAL(i,j) to obtain the AGC image of pixel (i,j); Step S26: Use ORIGINAL(i,j) and GAIN(i,j) as keys to perform a lookup table retrieval to obtain the value AGC(i,j) after gain of pixel (i,j); Step S27: Repeat steps S25 to S26 to traverse all pixels and obtain the complete AGC image.

4. The semiconductor defect detection method according to claim 1, characterized in that, Step S3 specifically includes the following steps: Step S31: Perform image preprocessing on the AGC image; Step S32: Extract defective pixels; Step S33: Merge and filter the extracted defective pixels.

5. The semiconductor defect detection method according to claim 4, characterized in that, Step S32 specifically includes the following steps: The pixel values ​​corresponding to the REF image are subtracted from the pixel values ​​corresponding to the TARGET image. The absolute value of the difference is then compared with a threshold. If the absolute value of the difference is less than or equal to the threshold, the pixel is determined to be without defects. If the absolute value of the difference is greater than the threshold, the current pixel is determined to be a defective pixel. The TARGET image is the AGC image of the target die on the wafer, and the REF image is the image of a die adjacent to the target die.

6. The semiconductor defect detection method according to claim 5, characterized in that, Step S33 specifically includes the following steps: Set a distance threshold D. If the distance between any two defective pixels is less than the distance threshold D, it means that the two defective pixels come from the same defective image block. Using machine learning or neural network models, defective image blocks that meet the detection requirements are selected, and the remaining defective image blocks are discarded.

7. A system employing the semiconductor defect detection method as described in any one of claims 1 to 6, characterized in that: It includes a wafer, a stage, an image processing unit, a vertical incident energy source, a semi-reflective mirror, a vertical incident illumination objective, a vertical incident acquisition device, an oblique incident energy source, an oblique incident illumination objective, and an oblique incident acquisition device. The wafer is placed on a display platform; The light emitted by the vertically incident energy source passes through a semi-reflective lens and then shines vertically onto the vertically incident illumination objective. The vertically incident illumination objective focuses the light emitted by the vertically incident energy source onto the wafer surface. The vertically incident acquisition device is used to acquire images of the wafer surface in the vertical direction. The oblique incident energy source illuminates the wafer surface from the oblique angle, the oblique incident illumination objective lens focuses the light emitted by the oblique incident energy source onto the wafer surface, and then the oblique incident acquisition device acquires an image of the wafer surface from the oblique angle. The image processing unit is used to perform defect detection on the raw image data acquired by the vertical incident acquisition device and the oblique incident acquisition device.

Images