Evaluation method, system and electronic device for wafer defect detection

Abstract

The application provides a wafer defect detection evaluation method, system and electronic equipment. The evaluation method comprises: acquiring an image of a wafer surface; performing a primary evaluation based on the image to obtain a primary evaluation index, the primary evaluation index being used to represent the quality of the image; performing algorithm optimization processing on the image satisfying a first threshold, the first threshold being used to represent that the quality of the image meets the requirement of being used for optimization; performing a secondary evaluation on the image after the optimization processing to obtain a secondary evaluation index, the secondary evaluation index being used to represent the quality of the image after the optimization processing; and outputting the image satisfying a second threshold of the secondary evaluation index, which is used for subsequent defect detection, the second threshold being used to represent that the quality of the image meets the requirement of being used for defect detection. The second threshold is higher than the first threshold. Through the step-by-step screening of the primary evaluation and the secondary evaluation, and the introduction of the algorithm optimization processing between the two layers of evaluation, the hierarchical management and improvement of the image quality can be achieved.

Description

Technical Field

[0001] This application relates to the field of wafer defect detection technology, and in particular to an evaluation method, system and electronic device for wafer defect detection. Background Technology

[0002] As semiconductor process nodes shrink, higher demands are placed on inspection technologies. Optical inspection equipment can quickly and accurately detect and identify various defects and anomalies that occur during the manufacturing process without contacting or damaging the wafer, thus meeting the challenges of advanced processes.

[0003] In high-precision optical inspection equipment, it is usually necessary to rely on optical systems to acquire images of the wafer surface and identify defects based on these images. As the patterns on wafers become increasingly complex and the precision requirements become increasingly stringent, the quality of the acquired images directly affects the accuracy and stability of defect detection.

[0004] Currently, wafer images are typically corrected and optimized before defect detection. For example, image quality is improved by optimizing metrics such as sharpness and contrast.

[0005] However, during the wafer image capture process, it is easily affected by the internal factors of the detection system, such as laser intensity jitter, uneven wafer surface coating, random noise, and motion platform errors. These combined interferences can cause large fluctuations in the image quality of some wafers, which seriously affect the accuracy of subsequent defect detection. Summary of the Invention

[0006] In view of this, this application provides an evaluation method, system, and electronic device for wafer defect detection, aiming to at least solve the problem that current wafer defect detection is easily affected by the internal factors of the detection system, resulting in low detection accuracy.

[0007] The first aspect of this application provides an evaluation method for wafer defect detection, the evaluation method comprising: acquiring an image of the wafer surface; performing an initial evaluation based on the image to obtain an initial evaluation index, the initial evaluation index being used to characterize the quality of the image; performing algorithm optimization processing on an image whose initial evaluation index meets a first threshold, the first threshold being used to characterize that the quality of the image meets the requirements for optimization; performing a secondary evaluation on the optimized image to obtain a secondary evaluation index, the secondary evaluation index being used to characterize the quality of the optimized image; and outputting an image whose secondary evaluation index meets a second threshold for subsequent defect detection, the second threshold being used to characterize that the quality of the image meets the requirements for defect detection, wherein the second threshold is higher than the first threshold.

[0008] In one specific implementation of the first aspect of this application, outputting an image whose secondary evaluation index meets a second threshold for subsequent defect detection includes: outputting an image whose secondary evaluation index meets the second threshold and whose secondary evaluation index is not inferior to the initial evaluation index before optimization processing for subsequent defect detection, wherein the secondary evaluation index is not inferior to the initial evaluation index before optimization processing means that the secondary evaluation index is greater than the initial evaluation index.

[0009] In one specific implementation of the first aspect of this application, obtaining an image of the wafer surface includes: obtaining a first image of a first unit among a plurality of repeating units on the wafer surface, and a second image of a second unit adjacent to the first unit.

[0010] In one specific implementation of the first aspect of this application, the initial evaluation index and the secondary evaluation index each include a first type of index, which includes information entropy and signal-to-noise ratio. The information entropy is used to characterize the degree of grayscale disorder in the first image, and the signal-to-noise ratio is used to characterize the ratio of effective signal to noise intensity in the first image.

[0011] In one specific implementation of the first aspect of this application, the initial evaluation index and the secondary evaluation index each include a second type of index, which includes a mutual information value. The mutual information value is used to characterize the consistency between the first image and the second image.

[0012] In one specific implementation of the first aspect of this application, the second type of index further includes mean absolute error and mean square error, used to characterize the degree of alignment between the first image and the second image.

[0013] In one specific implementation of the first aspect of this application, the algorithm optimization process includes translation correction or rotation correction.

[0014] In one specific implementation of the first aspect of this application, the algorithm optimization process includes brightness correction or contrast correction.

[0015] In one specific implementation of the first aspect of this application, the algorithm optimization process includes denoising or sharpening.

[0016] A second aspect of this application provides a wafer defect detection system, comprising an image acquisition module, a primary evaluation module, an algorithm optimization module, a secondary evaluation module, and an output module. The image acquisition module acquires images of the wafer surface. The primary evaluation module performs an initial evaluation on the images to obtain primary evaluation indicators, which characterize the image quality. The algorithm optimization module performs algorithm optimization processing on images whose primary evaluation indicators meet a first threshold. The first threshold characterizes the image quality as meeting optimization requirements. The secondary evaluation module performs a secondary evaluation on the optimized images to obtain secondary evaluation indicators. The secondary evaluation indicators characterize the quality of the optimized images. The output module outputs images whose secondary evaluation indicators meet a second threshold for defect detection, where the second threshold characterizes the image quality as meeting defect detection requirements. The second threshold is higher than the first threshold.

[0017] A third aspect of this application provides an electronic device, which includes a processor and a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by the processor, it causes the processor to perform the control method mentioned above.

[0018] The wafer defect detection evaluation method provided in this application employs a step-by-step screening process involving initial and secondary evaluations, and introduces algorithm optimization between the two evaluation layers. Furthermore, images that meet the criteria after the secondary evaluation are used for subsequent defect detection, facilitating hierarchical control and improvement of image quality. Moreover, without significantly sacrificing the original workflow, it helps suppress false positives and false negatives caused by low-quality images entering the detection process. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts for an evaluation method for wafer defect detection provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of the structure of a wafer surface provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of the structure of a wafer defect detection system provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0024] Figure 5 This is a second flowchart of an evaluation method for wafer defect detection provided in an embodiment of this application.

[0025] Explanation of reference numerals in the attached figures: A - Repeating cell; A1~A6 - First to sixth cells; W - Wafer; 110 - Image acquisition module; 120 - Initial evaluation module; 140 - Algorithm optimization module; 150 - Secondary evaluation module; 170 - Output module.

[0026] 200 - Electronic device; 210 - Processor; 220 - Memory. Detailed Implementation

[0027] The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that there are various implementations of this application, and it should not be construed as being limited to the embodiments described herein. The embodiments described herein are merely for a more thorough and clear understanding of this application. Those skilled in the art can make various equivalent substitutions or modifications to the specific structures, parameters, and connection relationships described in this application without departing from the core concept of this application, and all such substitutions or modifications should be considered to fall within the protection scope of this application.

[0028] In the field of semiconductor patterned wafer defect detection, optical systems are typically used to acquire images of the wafer surface, and defects are identified based on these images. To ensure the reliability of defect detection, related technologies usually optimize the wafer image before defect detection. For example, image quality is improved by optimizing indicators such as image sharpness and contrast.

[0029] However, wafer W images are easily affected by the internal factors of the detection system during the acquisition process. For example, the combined interference of multiple factors such as laser intensity jitter, uneven coating on the wafer W surface, random noise, and motion platform error can cause significant differences in the quality of the acquired images. If images with poor quality are directly used for defect detection, the accuracy of defect detection will decrease.

[0030] Because the current defect detection process lacks a systematic, tiered evaluation and optimization of image quality, images of poor quality are directly introduced into the defect detection stage, impacting detection accuracy. Research has revealed that image quality degradation is not entirely irreversible. Some images, while not meeting the direct standards for defect detection, can still satisfy detection requirements after targeted algorithmic correction. Conversely, other images are of such poor quality that they lack optimization potential; in other words, optimizing these images would not only fail to improve their quality but also waste computational resources.

[0031] Based on this understanding, this application proposes a wafer defect detection and evaluation method that includes a two-layer evaluation mechanism and intermediate algorithm optimization processing. Through a step-by-step screening process involving initial and secondary evaluations, and introducing algorithm optimization processing between the two evaluation layers, images that meet the conditions after the secondary evaluation are used for subsequent defect detection. This helps to achieve hierarchical control and improvement of image quality. Furthermore, without significantly sacrificing the original workflow, it helps to suppress false positives and false negatives caused by low-quality images entering the detection process.

[0032] refer to Figure 1 The evaluation method steps S100 to S500 provided in this application embodiment are as follows.

[0033] Step S100: Obtain an image of the surface of wafer W. This image may be a photolithographic pattern (e.g., a grayscale image) on wafer W captured by a sensor (e.g., a position sensor, an image sensor, etc.).

[0034] Specifically, the image may include a grayscale image with multiple channels. In other embodiments, the image may also be acquired by means of electron beam scanning or the like. The image of the wafer W surface provides raw data for subsequent evaluation and defect detection. The image may contain information about the lithography pattern and background information on the wafer W surface. Based on the lithography pattern and background information, a preliminary assessment of image quality (such as grayscale analysis) can be performed.

[0035] In at least one embodiment of this application, reference is made to Figure 2 The surface of wafer W typically contains a large number of repeating chip units formed by photolithography. Each repeating unit A corresponds to a patterned area of ​​a chip and contains the same pattern structure. The photolithographic patterns on the surface of wafer W have a high degree of repeatability (see reference). Figure 2 The first unit (A1) to the sixth unit (A6) in the image (ideally, adjacent repeating units should have a highly consistent pattern structure, including highly consistent graphic and grayscale features). Based on this inherent repeatability, deviations introduced by the system (such as motion platform errors, light intensity fluctuations, etc.) can be detected by comparing the images of adjacent units. Therefore, acquiring an image of the surface of wafer W can include the following steps: Acquire a first image of the first unit A1 among multiple repeating units on the surface of wafer W, and a second image of the second unit A2 adjacent to the first unit A1.

[0036] By selecting adjacent first unit A1 and second unit A2 (or even the fourth unit A4) for comparison, the interference of systematic differences (such as coating thickness gradient) at different locations on the wafer W surface on the index analysis can be minimized, making the evaluation results more accurately reflect the image deviation caused by the introduction of factors inside the detection system.

[0037] Of course, you can also choose a repeating image that is not adjacent to the first unit A1 (such as the third unit A3, the fifth unit A5, or the sixth unit A6), without making any specific restrictions here.

[0038] Furthermore, after acquiring the first and second images, a data foundation can be provided for subsequent consistency and alignment evaluation based on multi-image metrics (such as mutual information value, mean absolute error, mean square error, etc.), enabling the detection system to fully utilize the consistency between repeating pattern units to evaluate image quality.

[0039] Step S200: Perform an initial evaluation based on the image to obtain an initial evaluation index, which is used to characterize the quality of the image.

[0040] Specifically, after acquiring the image, an initial evaluation is performed to obtain initial evaluation metrics. These metrics characterize the image quality, which can be quantified across multiple dimensions, such as grayscale uniformity, noise level, and consistency between images. By performing an initial evaluation and quantifying the quality of the original image, the image's quality status can be identified early in the process, providing a basis for deciding whether to perform subsequent algorithm optimization and preventing images with poor quality from entering the optimization process and wasting computational resources.

[0041] Step S300: For images whose initial evaluation metrics meet the first threshold, perform algorithm optimization processing. Specifically, after the initial evaluation, the initial evaluation metrics are compared with the preset first threshold. The first threshold is used to characterize that the image quality meets the requirements for optimization; that is, only images whose initial evaluation metrics meet the first threshold will enter the algorithm optimization processing stage.

[0042] The first threshold serves as the initial evaluation and screening threshold, excluding images of poor quality or lacking optimization potential from the algorithm's optimization process. This ensures that the images received by the algorithm have a basic basis for optimization, thereby improving optimization efficiency and reducing unnecessary computational overhead.

[0043] For images that meet the first threshold in the initial evaluation, algorithm optimization is performed. This optimization process is used to specifically correct images that pass the initial evaluation, thereby improving image quality. Specifically, this optimization helps to eliminate or reduce image quality degradation caused by factors within the detection system (such as light intensity fluctuations, motion errors, noise, etc.), further enhancing image quality.

[0044] It should be noted that the first threshold cannot be met under poor conditions within the detection system or if the pattern on the surface of wafer W is poor. In other words, the images intercepted or discarded by the evaluation method provided in this application may be due to unclear wafer patterns or poor internal environment of the detection system resulting in images of too poor quality to be used for subsequent defect detection.

[0045] Step S400: Perform a secondary evaluation on the optimized image to obtain secondary evaluation indicators, which are used to characterize the quality of the optimized image.

[0046] Specifically, after the algorithm optimization process is completed, a secondary evaluation is performed on the optimized image to obtain secondary evaluation metrics. These secondary evaluation metrics characterize the quality of the optimized image. By evaluating the quality of the optimized image again, the secondary evaluation metrics can also include multiple quantitative indicators such as the uniformity of grayscale, the noise level of the image, and the consistency between images.

[0047] The secondary evaluation after optimization can not only verify whether the algorithm optimization process effectively improves the image quality, but also ensure that the optimization process does not introduce new quality degradation, thereby helping to ensure the stability and reliability of the system.

[0048] Step S500: Output an image that meets the second threshold for the secondary evaluation index, for subsequent defect detection. The second threshold is used to characterize the quality of the image that is suitable for defect detection, wherein the second threshold is higher than the first threshold.

[0049] Specifically, by comparing the secondary evaluation index with a preset second threshold, an image whose secondary evaluation index meets the second threshold is output for subsequent defect detection. The second threshold is used to characterize the quality of the image to meet the requirements for defect detection.

[0050] The second threshold of the secondary evaluation index is higher than the first threshold of the primary evaluation index. This means that the second threshold and the first threshold form a progressive quality screening mechanism. That is, the first threshold ensures that the image has basic quality that can be optimized, and the second threshold ensures that the image that can be used for defect detection reaches a higher quality standard.

[0051] This layered threshold design helps avoid setting a single threshold too high, which would cause optimized images to be discarded prematurely, or setting it too low, which would cause inferior images to enter subsequent detection stages, resulting in detection results that do not accurately reflect the defects of wafer W.

[0052] It's understandable that if the secondary evaluation metrics improve, it means the image degradation can be mitigated by eliminating or reducing the combined interference from multiple sources, such as laser intensity fluctuations, random noise, and motion platform errors, within the detection system. Furthermore, a higher second threshold helps improve image quality, thereby enhancing the reliability and accuracy of subsequent defect detection. If the secondary evaluation metrics do not improve, it indicates that the image degradation is due to poor wafer quality or a poor detection system, rendering the acquired image unusable. In such cases, reminders or alarms can be used to alert technicians to promptly inspect the detection equipment.

[0053] To further improve the reliability of the evaluation method, outputting an image that meets the second threshold for the secondary evaluation index, which can be used for subsequent defect detection, may include the following steps: The output image, in which the secondary evaluation index meets the second threshold and is not inferior to the initial evaluation index before optimization, is used for subsequent defect detection. Here, "the secondary evaluation index is not inferior to the initial evaluation index before optimization" means that the secondary evaluation index is greater than the initial evaluation index (i.e., the non-deterioration principle).

[0054] Specifically, before outputting an image whose secondary evaluation metric meets the second threshold, it is necessary to compare the initial and secondary evaluation metrics of the images before and after optimization, and obtain images whose secondary evaluation metrics are not worse than the initial evaluation metrics before optimization. In other words, the initial evaluation metrics obtained before algorithm optimization are compared with the secondary evaluation metrics obtained after algorithm optimization to determine whether the algorithm optimization has a positive effect on image quality. If the secondary evaluation metrics improve, the corresponding image is output; if the secondary evaluation metrics deteriorate, the image is intercepted and an alarm is triggered. Of course, the order of comparing the secondary evaluation metrics and the second threshold, as well as comparing the metrics before and after optimization, is not critical; they can be compared simultaneously or separately. No specific limitation is made here.

[0055] By directly quantifying and comparing the evaluation metrics of the same image before and after optimization, we can objectively judge the effect of the algorithm optimization process and avoid relying solely on the absolute value of the secondary evaluation metrics while ignoring the quality degradation that may be introduced during the optimization process.

[0056] It should be noted that "the secondary evaluation index is not inferior to the initial evaluation index before optimization" means that for the same image, the values ​​of each index in the secondary evaluation index obtained after algorithm optimization are not lower than (or not worse than) the corresponding initial evaluation index values ​​obtained before optimization. This constraint helps ensure that the algorithm optimization process provides a non-decreasing improvement to image quality; that is, the optimized image is at least not inferior to its state before optimization in each evaluation dimension. For example, if an optimization process causes a degradation in a certain index (e.g., denoising introduces blur, leading to an increase in signal-to-noise ratio but a decrease in information entropy), that image will not be output for defect detection, thereby preventing degraded images from entering subsequent detection processes and ensuring the reliability of defect detection results.

[0057] In at least one embodiment of this application, the initial evaluation metrics and the secondary evaluation metrics can cover the metric analysis of a single image and / or the metric analysis of multiple images. Among them, single-image metrics (such as information entropy and signal-to-noise ratio) can be used to evaluate the gray-level uniformity and noise level of a single image, while multi-image metrics (such as mutual information value) can be used to evaluate the consistency and alignment between multiple images.

[0058] Evaluating single-image and / or multi-image metrics of the acquired images is equivalent to comprehensively assessing multiple influencing factors in the detection system, which helps improve the accuracy of the evaluation. Furthermore, secondary evaluation can also include comparative analysis of the metrics of the same image before and after correction to ensure that the metrics of the corrected image do not deteriorate. By comparing the metrics before and after correction, it is possible to verify whether the algorithm optimization process has introduced negative effects (e.g., excessive denoising leading to the loss of effective information), ensuring the effectiveness of the optimization process and the correctness of the system.

[0059] In other embodiments, multi-image index analysis is not limited to the comparison of two images (such as the first and second images mentioned above), but can be extended to the comparison of indices of multiple images, obtaining statistical information from more repeated units, thereby improving the robustness and reliability of the evaluation results.

[0060] It is understandable that the constraint of non-degradation of metrics mentioned above applies to both metrics for a single image and comparative metrics between multiple images. That is, for metrics of a single image (such as information entropy and signal-to-noise ratio), the corrected value cannot be worse than the original value; for comparative metrics between multiple images (such as mutual information value), the corrected value cannot be worse than the original value. This helps ensure that algorithm optimization improves one dimension of metrics without causing degradation in other dimensions, thus achieving comprehensive guarantee of image quality.

[0061] The initial evaluation indicators and secondary evaluation indicators mentioned above can cover single-map indicators and multi-map indicators. The categories and calculation methods of these indicators will be explained below.

[0062] The initial evaluation metrics and the secondary evaluation metrics can each include the first category of metrics (i.e., single-image metrics), which include information entropy and signal-to-noise ratio (SNR). Information entropy and SNR are quantitative assessments of the quality of a single image based on two dimensions: gray-level uniformity and noise level. They reflect the effects of different types of influencing factors in the system on image quality, and when used together, they provide a comprehensive, multi-dimensional evaluation of the quality of a single image.

[0063] Information entropy can be used to characterize the degree of disorder in the grayscale of a first image. For example, factors such as laser intensity fluctuations and uneven coating on the surface of a wafer can cause uneven grayscale distribution in an image. Information entropy can quantify this unevenness; that is, the stronger the grayscale unevenness caused by various interferences, the higher the information entropy and the worse the image quality.

[0064] By quantifying the degree of disorder in image grayscale using the information entropy index, grayscale anomalies caused by factors such as light intensity fluctuations and uneven coating can be identified in the initial stage of evaluation, thus preventing images with uneven grayscale from entering the defect detection stage and causing a decrease in defect detection accuracy.

[0065] For example, the information entropy H can be calculated using the following formula: In Formula 1 above, p(i) represents the probability of a pixel with a gray value of i in the image, where the gray value can range from 0 to 255. Calculating the information entropy based on the pixel's gray-level probability distribution directly reflects the dispersion of the image's gray-level distribution. The more concentrated the gray-level distribution (i.e., the more uniform the image), the lower the information entropy; the more dispersed the gray-level distribution (i.e., the more chaotic the image's gray-level distribution), the higher the information entropy.

[0066] Furthermore, in some embodiments, the grayscale range can be divided into multiple segments for calculating information entropy separately, and the final information entropy can be calculated according to weights. Segmented weighted calculation of information entropy helps to take into account the complexity of the pattern in the patterned wafer W. That is, different grayscale range segments may correspond to different pattern areas or background areas. Segmented weighting can more accurately reflect the grayscale uniformity of each area and improve the adaptability of the information entropy index to complex patterns.

[0067] In one example, the information entropy of the first threshold can be 6.5, and the information entropy of the second threshold can be 6.0. If the information entropy calculated in the initial evaluation of the image is equal to or less than 6.5, then the image has the potential to recover its quality through algorithmic correction. If the information entropy calculated in the secondary evaluation of the optimized image is equal to or less than 6.0, then after optimization by algorithms such as geometric correction, brightness correction, or denoising, the optimization is truly effective, and the gray-level distribution of the image becomes more concentrated and uniform, and the gray-level uniformity of the corrected image meets the standard for defect detection. If the corrected information entropy increases instead (e.g., from 6.2 to 6.7), it indicates that the algorithmic optimization (such as excessive brightness correction or contrast stretching) has led to abnormal dispersion in the gray-level distribution, introducing new non-uniformity. For example, non-uniform brightness compensation may cause blocky gray-level differences in an originally uniform background area. In this case, even if the corrected information entropy still meets the second threshold, the system should determine that the optimization is invalid, and the parameters need to be readjusted or a warning issued to avoid sending the degraded image to the defect detection stage.

[0068] Signal-to-noise ratio (SNR) characterizes the ratio of effective signal intensity to noise intensity in a first image. A higher SNR indicates a more uniform background and lower noise, which is beneficial for subsequent defect detection. SNR directly reflects the relative strength of the effective signal to the noise in an image. A high SNR means that the effective lithographic pattern signal is much stronger than the background noise, making the grayscale difference between defective pixels and normal pixels easier to identify, thus helping to improve the accuracy of defect detection.

[0069] For example, the signal-to-noise ratio (SNR) can be calculated using the following formula: In formula two above, This indicates that the average value of the pixel grayscale values ​​in the image is selected (i.e., the effective signal). This represents the standard deviation of pixel grayscale values ​​in the selected image (i.e., noise intensity), calculated in decibels as the signal-to-noise ratio (SNR). A logarithmic transformation is applied to the ratio of the mean grayscale value of the effective signal to the standard deviation of the background noise. This provides a clear reflection of the SNR quality of the initially acquired image over a large dynamic range, facilitating comparison with a first or second threshold. For example, molecules... and denominator This means that the calculation is performed on the same small, flat gray area selected from the image.

[0070] Furthermore, in some embodiments, multiple sets of signal-to-noise ratio (SNR) values ​​can be calculated, and the final SNR can be calculated by weighting them. Calculating the SNR using multiple weighted sets can cover the SNR characteristics of different regions in the image, avoiding overall evaluation distortion caused by SNR deviations in local areas, thereby helping to improve the overall integrity and accuracy of the SNR metric.

[0071] In one example, the signal-to-noise ratio (SNR) for the first threshold can be 20 dB, and the SNR for the second threshold can be 30 dB. If the SNR of the initial evaluation metric is equal to or greater than 20 dB, the first threshold is met, and the acquired image will be optimized by the algorithm. If the SNR of the initial evaluation metric is less than 20 dB, then the background noise of the image is too high, indicating that the effective signal is severely overwhelmed, and it does not meet the basic conditions for subsequent algorithm optimization, and may be intercepted. If the SNR of the optimized image reaches or exceeds 30 dB, then the SNR level of the optimized image meets the requirements for high-quality defect detection and can be used for subsequent defect detection.

[0072] The first type of indicators has been explained above; the second type will now be explained. Both the initial evaluation indicators and the secondary evaluation indicators can include the second type of indicators, which may include mutual information values. Mutual information values ​​are used to characterize the consistency between the first and second images; that is, image quality is evaluated through the consistency dimension of multiple images.

[0073] Mutual information value can specifically quantify the statistical dependency between the first image and the second image. The larger the mutual information value, the better the consistency between the first image and the second image, that is, the smaller the alignment deviation and optical difference introduced by the detection system.

[0074] It is understandable that, since the first unit A1 and the second unit A2 are adjacent repeating units, their images should ideally be highly consistent. The mutual information value can quantify the degree to which the actual acquired first and second images deviate from the ideal consistency. By evaluating the consistency between images of adjacent repeating units through the mutual information value, alignment deviations and grayscale differences introduced by detection system factors such as motion platform errors and lithography errors can be detected. This allows for the identification of images with insufficient alignment quality before defect detection, avoiding false defects or missed defects caused by image inconsistencies, thereby improving the accuracy and reliability of subsequent detection.

[0075] For example, the mutual information value can be calculated using the following formula three: In Formula 3 above, p_A(a) represents the probability that the gray value of the first image is 'a', p_B(b) represents the probability that the gray value of the second image is 'b', and p_AB(a,b) represents the joint gray value distribution probability of the first and second images. It can be understood that mutual information, based on the relationship between the joint probability distribution and the marginal probability distribution in information theory, can capture the non-linear statistical dependencies between images. Compared to simple pixel difference metrics, it has stronger robustness to non-uniform gray value mappings (such as overall gray value shifts caused by light intensity fluctuations).

[0076] In one example, the mutual information value for the first threshold can be 0.65, and the mutual information value for the second threshold can be 0.85. If the mutual information value between the first and second images is less than 0.65, then the statistical dependency and spatial alignment between the first and second images are poor, and they cannot be effectively recovered through translation or rotation correction. This set of images will be judged as having no optimization potential and trigger a system alarm. If the mutual information value between adjacent images is greater than or equal to 0.65, the adjacent images are optimized by the algorithm. If the mutual information value between the two optimized images increases to above 0.85, then the consistency of the image pair is extremely high, and it is accurately aligned by the algorithm, making it suitable for precise pixel-level defect differential comparison.

[0077] In addition to mutual information values, the second type of metric can also include Mean Absolute Error (MAE) and Mean Squared Error (MSE) to supplement the quantification of the difference between the first and second images from the perspective of alignment. The smaller the values ​​of MAE and MSE, the better the alignment between the first and second images.

[0078] Mean absolute error can be used to characterize the average absolute value of the grayscale difference between the first and second images (i.e., the method of performing operations one by one with a single pixel as the basic unit). It treats the differences in the position of each pixel equally and can intuitively quantify the overall alignment.

[0079] Mean squared error (MSE) is used to characterize the mean of the squared differences in gray levels between the first and second images. It gives higher weight to larger pixel gray level differences (i.e., it amplifies them after squaring). Compared with MAE, it is more sensitive to local misalignment areas and helps to discover significant local deviations in the image.

[0080] Mutual information values, along with MAE and MSE, can be computed in parallel to jointly characterize the consistency and alignment between the first and second images. Since mutual information values ​​focus on the global consistency evaluation of statistical dependencies, while MAE and MSE focus on the pixel-level alignment accuracy evaluation, using mutual information values, MAE, and MSE in parallel computation allows for a comprehensive quantification of the alignment quality between multiple images from different dimensions, helping to improve the completeness and reliability of multi-image metric evaluation.

[0081] In some optional embodiments, the initial or secondary evaluation can be used to assess image quality without intercepting subsequent processes. Alternatively, the evaluation results can be output as reference information for system status monitoring, rather than as a mandatory interception condition for subsequent processes. This optional implementation of the non-interception mode allows the system to acquire image quality evaluation data even in scenarios requiring continuous operation, for post-analysis and system optimization, thereby increasing the flexibility of system deployment.

[0082] It's understandable that the "no degradation" principle for single-image and multi-image metrics can be specifically interpreted as requiring not only that the first and second thresholds be met before and after optimization and correction, but also that the corrected single-image and multi-image metrics not degrade. In other words, when evaluating corrected single-image and / or multi-image metrics, it's not only required that the corrected metrics be superior to the original metrics—meaning that the single-image and / or multi-image metrics for the same image must not degrade—but also that the original and corrected single-image and / or multi-image metrics meet the first and / or second thresholds. In other words, after correction, the detection system compares the original and corrected single-image and multi-image metrics. If degradation is detected, a protection mechanism is triggered, preventing the image from entering subsequent processes. Of course, whether to block or allow entry into subsequent processes is up to the user.

[0083] By constraining the non-deterioration of single-image metrics and / or multi-image metrics, a safety locking mechanism is further set up for algorithm optimization based on multiple threshold progressive evaluations. This ensures that each correction operation makes a positive contribution to image quality, thereby helping to guarantee the stability of the defect detection system under multi-source interference and the correctness of detection results, and ensuring that the corrected image is truly improved in different dimensions.

[0084] Specifically, if the corrected information entropy increases (e.g., from 6.2 to 6.7), it indicates that the algorithm optimization (such as excessive brightness correction or contrast stretching) has led to abnormal dispersion in the grayscale distribution, introducing new non-uniformity. For example, non-uniform brightness compensation may cause patchy grayscale differences in an originally uniform background area. In this case, even if the corrected information entropy still meets the second threshold, the system should determine that the optimization is ineffective, and the parameters should be readjusted or a warning should be issued to avoid using the degraded image for subsequent defect detection.

[0085] If the corrected signal-to-noise ratio (SNR) decreases (e.g., from 25dB to 22dB), it indicates that the denoising or sharpening process may have overly smoothed the edges of the effective lithographic pattern, resulting in signal component damage, or that the denoising algorithm introduced new artifacts in flat areas. For example, overly aggressive denoising may mistakenly eliminate subtle pattern textures as noise, thus reducing the energy of the useful signal. By introducing a non-degradation constraint, the detection system compares the SNR before and after correction. If degradation is detected, a protection mechanism is triggered to prevent the image from proceeding to subsequent processes.

[0086] For multi-image metrics, if the corrected mutual information value decreases (e.g., from 0.70 to 0.68), it indicates a potential deviation in translation, rotation, or brightness correction, leading to new spatial misalignment between the first and second images, or an incorrect stretching of the grayscale mapping relationship. For example, an incorrect rotation angle estimate can cause partially aligned edge regions to become completely misaligned, thus reducing statistical dependence. By using non-degradation constraints, the detection system can promptly identify and eliminate such improper corrections.

[0087] If the corrected MAE or MSE increases (e.g., MAE increases from 10.8 to 12.3, MSE increases from 185 to 230), it indicates that the correction algorithm may have introduced new pixel-level misalignments or grayscale distortions in local areas. For example, the interpolation algorithm during translation correction may produce a ringing effect in edge regions, causing the differences between edge pixels to be artificially amplified. Compared to the global statistical evaluation of mutual information values, MAE and MSE can more sensitively capture such local degradations, effectively complementing mutual information values. Through non-degradation constraints, the detection system can comprehensively guarantee the effectiveness of multi-image correction from both global consistency and local alignment accuracy perspectives.

[0088] The categories and calculation methods of evaluation indicators have been explained above. The following section will further explain the types of algorithm optimization processing.

[0089] In at least one embodiment of this application, the algorithm optimization process may include translation correction or rotation correction. Translation correction and rotation correction belong to geometric algorithm optimization processes and are used to eliminate or reduce image spatial position shifts and angular deflections caused by factors such as motion platform errors and lithography errors, thereby improving the alignment accuracy of the image.

[0090] Translation correction can be used to eliminate positional offsets of images in the horizontal and / or vertical directions. Such offsets are usually caused by positioning errors of the motion platform. Translation correction can make the images of adjacent repeating units accurately aligned in spatial position, thereby helping to improve the evaluation results of subsequent multi-image metrics (such as mutual information value, MAE, MSE) and the accuracy of defect detection.

[0091] Rotation correction can be used to eliminate angular deflection in an image, which may be caused by rotational errors of the motion platform or angular deviations during wafer loading. Rotation correction can correct rotational offsets in the image so that the lithographic pattern in the image is aligned with the reference orientation.

[0092] In at least one embodiment of this application, the algorithm optimization process may further include brightness correction or contrast correction. Brightness correction and contrast correction belong to grayscale algorithm optimization processes and are used to eliminate or reduce image grayscale deviations caused by factors such as laser light intensity jitter and uneven coating on the wafer W surface, thereby improving the grayscale uniformity and consistency of the image.

[0093] Brightness correction can be used to eliminate or reduce overall or local brightness deviations in an image. These deviations are usually caused by fluctuations in laser light intensity or uneven coating thickness on the wafer surface. Brightness correction can make the grayscale level of the image more consistent, reduce the information entropy index, and improve the quality assessment results of a single image.

[0094] Contrast correction can be used to adjust the dynamic range of gray values ​​in an image, making the gray difference between the effective lithographic pattern and the background more distinct. Contrast correction helps to improve the signal-to-noise ratio, making the gray distinction between defective pixels and normal pixels higher, which is beneficial for subsequent defect detection.

[0095] Furthermore, in another embodiment of this application, the algorithm optimization process may also include denoising or sharpening. Denoising and sharpening belong to signal enhancement algorithm optimization processes, and are used to reduce the noise level in the image and enhance the clarity of edge details in the image, respectively.

[0096] Denoising can be used to reduce the intensity of random noise in an image and improve the signal-to-noise ratio. The reduction of noise makes the effective signal (grayscale information of the lithographic pattern) more prominent, and the difference between defective pixels and normal pixels is easier for detection algorithms to identify, thereby helping to improve the accuracy of defect detection.

[0097] Sharpening can be used to enhance the clarity of edges and details in an image, making the outline of the lithographic pattern more distinct. In defect detection, defects often manifest as abnormal changes in the edges of the pattern; sharpening helps to enhance the discernibility of these abnormal features, thereby improving the sensitivity of defect detection.

[0098] It should be noted that the above-mentioned geometric optimization algorithms, grayscale optimization algorithms, and signal enhancement optimization algorithms can be used individually or in combination according to the actual image quality. This application does not limit this.

[0099] It should also be noted that some information based on the two evaluation metrics (such as the results of the initial evaluation metrics, the results of the secondary evaluation metrics after algorithm optimization, etc.), process information such as logs or alarms can also be reported.

[0100] The evaluation method has been explained in detail above. The following section will combine... Figure 5 A complete and specific embodiment of the wafer defect detection evaluation method provided in this application is given.

[0101] First, the motion platform carries the wafer to be inspected and moves it along a preset scanning trajectory. The optical imaging unit images the surface of the wafer W under oblique laser illumination, and the camera acquires images containing multiple repeating pattern units (see reference). Figure 2 The original full image of the first unit A1 to the sixth unit A6 in the image.

[0102] In this embodiment, the entire original image can be automatically split into multiple sub-images corresponding to repeating pattern units based on preset chip layout information, unit size information, or cutting coordinate templates. For example, a first image (corresponding to the first unit) and a second image (corresponding to the second unit adjacent to the first unit) are obtained, along with images corresponding to other adjacent or similar repeating units. The first and second images can be images of two adjacent repeating units on the wafer surface, or they can be further extended to multiple images such as the first image, second image, third image, fourth image, fifth image, and sixth image, which participate in subsequent evaluation.

[0103] The first image can be used as one of the images to be evaluated, and the second image can be used as an adjacent reference image for consistency comparison with the first image. Since adjacent repeating units should maintain consistency in theoretical pattern structure, the differences between their images can more sensitively reflect the impact of factors such as motion platform error, light intensity fluctuation, brightness drift, noise interference, or local process disturbance on image quality.

[0104] After obtaining the first and second images, the system first performs an initial evaluation on the original first image. The initial evaluation includes single-image index analysis and multi-image index analysis.

[0105] For the first image, the system calculates single-image metrics and multi-image metrics respectively. In this embodiment, single-image metrics include information entropy and signal-to-noise ratio. Multi-image metrics include at least mutual information values, and optionally mean absolute error (MAE) and mean squared error (MSE). Specific calculation methods can be found above and will not be described in detail here.

[0106] The system compares the initial set of evaluation indicators with a preset first threshold. This first threshold is not a single value, but can be a set of threshold conditions set for different indicators.

[0107] When the initial evaluation metrics corresponding to the first image and the second image both meet the first threshold, the system considers that the first image has met the basic requirements for optimization and can be used for subsequent algorithm optimization.

[0108] If at least one of the single-image metrics and multi-image metrics (such as information entropy, signal-to-noise ratio, or mutual information value) does not meet the first threshold, it indicates that the image quality is too poor, the noise is too high, the offset is too heavy, or the grayscale is abnormal. The system can directly determine that the image or image group has no optimization value and terminate its subsequent processing, or mark it as an abnormal sample and output it to the log system for device status analysis.

[0109] For the first and second images that pass the first threshold screening, the system performs algorithm optimization processing. The optimization processing can be adaptively selected based on the problem type reflected by the initial evaluation indicators, or it can be executed sequentially according to a preset processing chain. In this embodiment, the optimization processing may include one or more of the following: translation correction, rotation correction, brightness correction, contrast correction, noise reduction processing, and sharpening processing.

[0110] After algorithm optimization, a second evaluation is performed on the optimized first and second images. The index system used in the second evaluation can be the same as that used in the initial evaluation, which facilitates the comparative analysis of the indices before and after the evaluation.

[0111] After optimization, the set of secondary evaluation metrics is compared with a second threshold. The second threshold is set higher than the first threshold to ensure that the quality of images entering the defect detection stage is superior to that of the initially acquired images.

[0112] To avoid introducing new quality degradation during the optimization process itself, a non-deterioration principle judgment of the indicators before and after optimization is further performed before the output image is used for defect detection.

[0113] Specifically, the initial evaluation index set is compared with the secondary evaluation index set, and it is determined whether the optimized index is not worse than the corresponding index before optimization. That is, each index (such as information entropy, signal-to-noise ratio, and mutual information value) is not degraded when compared before and after optimization.

[0114] If an optimization step brings some metrics to the second threshold but causes significant degradation in other key metrics (such as information entropy, signal-to-noise ratio, and mutual information), the image can still be deemed not to meet the final output conditions. By using non-degradation constraints, we can ensure that algorithm optimization has an overall non-negative effect, rather than sacrificing other quality dimensions for local metric improvements.

[0115] In other words, if the corresponding conditions are not met at any stage, the image will not proceed to the final defect detection stage. Images that fail to meet the first threshold are considered to have insufficient original acquisition quality; images that meet the first threshold but fail the second threshold are considered to have failed to meet defect detection requirements even after optimization; images that meet the second threshold but not the non-deterioration constraint are considered to have potential side effects from optimization. For each of these situations, corresponding logs can be recorded, and the anomaly information can be fed back to the equipment maintenance module, process monitoring module, or production line management system for subsequent tracking and analysis. Of course, the order of comparison between the non-deterioration principle and the second threshold is not specifically limited; as long as all requirements are met, the image can be output for subsequent defect detection.

[0116] The evaluation method has been described in detail above. The following section describes the wafer defect detection system from a system architecture perspective. In a wafer defect detection system, if the system lacks a systematic evaluation and optimization capability for image quality, poor-quality images will be directly used for defect detection, affecting detection accuracy. To address this issue, this application provides a wafer defect detection system (see reference...) Figure 3 The system may include an image acquisition module 110, an initial evaluation module 120, an algorithm optimization module 140, a secondary evaluation module 150, and an output module 170.

[0117] The image acquisition module 110 can be used to acquire images of the surface of wafer W. Exemplarily, this module may include a camera and related hardware components such as a laser optical path, an oblique incidence imaging system, and a high-precision motion platform control system. The image acquisition module 110 can provide the inspection system with raw image data of the surface of wafer W, and its acquisition quality directly affects the effectiveness of subsequent evaluation and inspection.

[0118] The initial evaluation module 120 can be used to perform an initial evaluation of the images acquired by the image acquisition module 110 to obtain initial evaluation metrics. These initial evaluation metrics can include different metrics, such as single-image metrics (e.g., information entropy, signal-to-noise ratio) and / or multi-image metrics (e.g., mutual information value). The initial evaluation module can quantify the quality of the original images and generate initial evaluation metrics, providing data for subsequent screening decisions.

[0119] The algorithm optimization module 140 can be used to perform algorithm optimization processing on images that meet a first threshold. Algorithm optimization processing may include one or more of the following: translation correction, rotation correction, brightness correction, contrast correction, denoising, or sharpening. The algorithm optimization module 140 performs targeted corrections on images that pass the initial screening, eliminating or reducing image quality degradation caused by system factors, and improving image quality to meet the requirements of subsequent, higher-standard secondary evaluations.

[0120] The secondary evaluation module 150 can be used to perform a secondary evaluation on the optimized image to obtain secondary evaluation indicators. After the algorithm optimization process, the secondary evaluation module 150 performs a second quantitative evaluation of the image quality to verify the optimization effect and ensure that the optimization process does not introduce new quality degradation.

[0121] The output module 170 can be used to output images that meet the secondary evaluation criteria for subsequent defect detection. By outputting images that meet the quality requirements after double-layer evaluation and optimization processing to subsequent defect detection, the output module 170 helps ensure that defect detection is based on high-quality images, thereby helping to improve the accuracy of defect detection.

[0122] The image acquisition module 110, the initial evaluation module 120, the algorithm optimization module 140, the secondary evaluation module 150, and the output module 170 work together to realize a complete closed-loop process from image acquisition to quality evaluation, optimization, re-evaluation, and final output. This helps to ensure the quality of images used for defect detection, thereby improving the accuracy and reliability of defect detection.

[0123] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements any of the aforementioned evaluation methods. The computer-readable storage medium may be, for example, a non-volatile storage medium such as a hard disk, solid-state drive (SSD), optical disk, USB flash drive, or memory card. By storing the computer program on the computer-readable storage medium, the aforementioned evaluation methods can be distributed and deployed in the form of a program product, and the processor can implement the aforementioned evaluation methods by reading and executing the program.

[0124] This application also provides an electronic device 200, with reference to... Figure 4The electronic device 200 may include a processor 210 and a memory 220. The memory 220 stores a computer program, which, when executed by the processor 210, causes the processor 210 to perform any of the aforementioned evaluation methods. The electronic device 200 may be, for example, an industrial control computer, a server, an embedded computing device, or other computing device with data processing capabilities. The processor 210 may be, for example, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), or a Field Programmable Gate Array (FPGA). The memory 220 may be, for example, Random Access Memory (RAM), Read Only Memory (ROM), or Flash memory. By executing the computer program stored in the memory through the processor of the electronic device to implement the aforementioned evaluation method, the evaluation method can be deployed in software on general-purpose or special-purpose electronic devices, facilitating integration into existing wafer inspection systems.

[0125] It is understood that in this application, directional descriptions such as "upper," "lower," "inner," and "outer" are relative rather than absolute. These directional terms may be applicable when the wafer defect detection system provided in this application is placed in the posture and position shown in the accompanying drawings.

[0126] It should be understood that although terms such as "first" or "second" may be used in this application to describe various elements (such as the first image and the second image), these elements are not limited by these terms, which are only used to distinguish one element from another.

[0127] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0128] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

[0129] The components and devices described in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the accompanying drawings. As those skilled in the art will recognize, these components and devices can be connected, arranged, and configured in any manner.

[0130] The above are merely specific embodiments 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.

Claims

1. An evaluation method for wafer defect detection, comprising: Acquire an image of the wafer surface; An initial evaluation is performed based on the image to obtain an initial evaluation index, wherein the initial evaluation index is used to characterize the quality of the image. For images whose initial evaluation metrics meet the first threshold, an algorithm optimization process is performed, wherein the first threshold is used to characterize that the quality of the image meets the requirements for optimization. The optimized image is then subjected to a secondary evaluation to obtain a secondary evaluation index, wherein the secondary evaluation index is used to characterize the quality of the optimized image; and The image that meets the second threshold of the secondary evaluation index is output for subsequent defect detection, wherein the second threshold is used to characterize the quality of the image that is suitable for defect detection, and the second threshold is higher than the first threshold.

2. The evaluation method according to claim 1, wherein, The output image that satisfies the second threshold of the secondary evaluation index is used for subsequent defect detection, including: The image of the secondary evaluation index that satisfies the second threshold and is not inferior to the initial evaluation index before optimization is output for subsequent defect detection. The secondary evaluation index being not inferior to the initial evaluation index before optimization means that the secondary evaluation index is greater than the initial evaluation index.

3. The evaluation method according to claim 1, wherein, The process of acquiring an image of the wafer surface includes: A first image of a first unit among multiple repeating units on the wafer surface, and a second image of a second unit adjacent to the first unit are obtained.

4. The evaluation method according to any one of claims 1 to 3, wherein, The initial evaluation indicators and the secondary evaluation indicators each include a first category of indicators, which include information entropy and signal-to-noise ratio. The information entropy is used to characterize the degree of grayscale disorder in the first image, and the signal-to-noise ratio is used to characterize the ratio of effective signal to noise intensity in the first image.

5. The evaluation method according to any one of claims 1 to 3, wherein, The initial evaluation index and the secondary evaluation index each include a second type of index, which includes a mutual information value. The mutual information value is used to characterize the consistency between the first image and the second image.

6. The evaluation method according to claim 5, wherein, The second category of metrics also includes mean absolute error and mean square error, used to characterize the alignment between the first image and the second image.

7. The evaluation method according to any one of claims 1 to 3, wherein, The algorithm optimization process includes translation correction or rotation correction.

8. The evaluation method according to any one of claims 1 to 3, wherein, The algorithm optimization process includes brightness correction or contrast correction.

9. The evaluation method according to any one of claims 1 to 3, wherein, The algorithm optimization process includes noise reduction or sharpening.

10. A wafer defect detection system, wherein, include: An image acquisition module is used to acquire images of the wafer surface; The initial evaluation module is used to perform an initial evaluation on the image and obtain initial evaluation indicators, wherein the initial evaluation indicators are used to characterize the quality of the image. The algorithm optimization module is used to perform algorithm optimization processing on the image whose initial evaluation index meets the first threshold, wherein the first threshold is used to characterize that the quality of the image has reached the requirements for optimization. The secondary evaluation module is used to perform a secondary evaluation on the optimized image and obtain secondary evaluation indicators, wherein the secondary evaluation indicators are used to characterize the quality of the optimized image. The output module is used to output an image that meets the second threshold of the secondary evaluation index for defect detection, wherein the second threshold is used to characterize the quality of the image that is suitable for defect detection, and the second threshold is higher than the first threshold.

11. An electronic device comprising a processor and a memory, wherein, The memory stores a computer program that, when executed by the processor, causes the processor to perform the evaluation method as described in any one of claims 1 to 9.

Images