Automatic detection method of semiconductor wafer defects based on deep learning

Abstract

The present invention provides a method for automatic detection of semiconductor wafer defects based on deep learning, which relates to the field of semiconductor detection technology. Through the quality monitoring unit of the interactive target wafer production line, multiple wafer defect detection data are extracted, defect features are identified for the multiple wafer defect detection data, K defect feature clusters are obtained, the K defect feature edge intervals are compared with the target wafer quality feature tolerance intervals, and an automatic detection wheel is configured based on K defect factors to obtain an automatic detection plan. According to the automatic detection plan, automatic detection of product defects of the target wafer production line is performed to obtain defect detection results. The method solves the technical problems in the prior art of low efficiency, easy missed detection and false detection of semiconductor defects, resulting in low wafer detection accuracy and insufficient real-time performance. The method achieves the technical effect of automatic real-time dynamic detection of wafer defects and improving the efficiency, accuracy and reliability of wafer defect detection.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor detection technology, and in particular to a method for automatic detection of semiconductor wafer defects based on deep learning. Background Art

[0002] In the semiconductor manufacturing industry, wafer quality control is a critical step in ensuring the quality of the final product. With the continuous advancement of integrated circuit technology and the increasing size of wafers, defect detection on wafers has become increasingly complex. Traditional wafer defect detection methods often rely on manual visual inspection or post-manufacturing inspection using optical and electron beam methods. While simple and easy to implement, these methods suffer from low efficiency, poor reliability, and the tendency to miss and misdetect.

[0003] The existing technology has technical problems such as low efficiency of semiconductor defect detection, easy missed detection and false detection, resulting in low wafer detection accuracy and insufficient real-time performance. Summary of the Invention

[0004] This application provides a method for automatic detection of semiconductor wafer defects based on deep learning, which is used to solve the technical problems in the existing technology that semiconductor defect detection is inefficient, prone to missed detection and false detection, resulting in low wafer detection accuracy and insufficient real-time performance.

[0005] In view of the above problems, the present application provides a method for automatic detection of semiconductor wafer defects based on deep learning, the method comprising: interacting with the quality monitoring unit of the target wafer production line, extracting multiple wafer defect detection data within a preset detection window, wherein each wafer defect detection data includes a production time mark; performing defect feature identification on the multiple wafer defect detection data to obtain multiple defect feature sets; clustering the multiple defect feature sets with the defect feature type as an index to obtain K defect feature clusters, each defect feature cluster including multiple defect feature values; traversing the K defect feature clusters to perform defect feature value edge identification to generate K defect feature edge intervals; comparing the K defect feature edge intervals with the target wafer quality feature tolerance interval to determine K defect factors; configuring an automatic detection wheel based on the K defect factors, and turning the automatic detection wheel multiple times according to the preset number of detection samples to obtain an automatic detection plan, wherein the automatic detection plan includes K detection sample quantities of K defect features; performing automatic detection of product defects of the target wafer production line based on the automatic detection plan to obtain defect detection results.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] The method provided in the embodiment of the present application extracts multiple wafer defect detection data within a preset detection window through the quality monitoring unit of the interactive target wafer production line, wherein each wafer defect detection data includes a production time mark; performs defect feature recognition on the multiple wafer defect detection data to obtain multiple defect feature sets; clusters the multiple defect feature sets with the defect feature type as an index to obtain K defect feature clusters, each defect feature cluster including multiple defect feature values; traverses the K defect feature clusters to perform defect feature value edge recognition to generate K defect feature edge intervals; compares the K defect feature edge intervals with the target wafer quality feature tolerance interval to determine K defect factors; configures an automatic detection wheel based on the K defect factors, and dials the automatic detection wheel multiple times according to a preset number of detection samples to obtain an automatic detection scheme, wherein the automatic detection scheme includes K detection sample sizes of K defect features; and performs automatic detection of product defects of the target wafer production line based on the automatic detection scheme to obtain defect detection results. The method achieves the technical effect of automatic real-time dynamic detection of wafer defects and improving the efficiency, accuracy and reliability of wafer defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 A flowchart of the deep learning-based automatic detection method for semiconductor wafer defects provided in this application;

[0010] Figure 2 A schematic diagram of the process of generating K defect feature edge intervals in the deep learning-based semiconductor wafer defect automatic detection method provided in this application;

[0011] Figure 3 A schematic diagram of the process of configuring an automatic detection wheel in the deep learning-based automatic detection method for semiconductor wafer defects provided in this application. DETAILED DESCRIPTION

[0012] This application provides a deep learning-based automatic semiconductor wafer defect detection method to address the existing technical issues of low semiconductor defect detection efficiency, easy missed detection and false detection, resulting in low wafer detection accuracy and insufficient real-time performance. Through intelligent data analysis using deep learning, the method achieves the technical effect of automatic real-time dynamic detection of wafer defects, improving the efficiency, accuracy, and reliability of wafer defect detection.

[0013] Below, the technical solutions of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the accompanying drawings.

[0014] like Figure 1 As shown, the present application provides a method for automatic detection of semiconductor wafer defects based on deep learning, the method comprising:

[0015] The quality monitoring unit of the interactive target wafer production line extracts a plurality of wafer defect detection data within a preset detection window, wherein each wafer defect detection data includes a production time mark.

[0016] Specifically, the quality monitoring unit is a key component of the wafer production line, which is used to monitor the wafer production process in real time and collect relevant quality data. The monitoring data includes images of the wafer surface during the generation process, dimensional measurements, electrical performance test results, etc. The preset detection window is a time range or condition range pre-set based on production batch, production time, product type or other relevant factors, which is used to limit the scope of wafer defect detection data extracted from the quality monitoring unit. The quality monitoring unit of the interactive target wafer production line extracts multiple wafer defect detection data within the preset detection window, which means extracting multiple wafer defect detection data within the preset detection window range from the quality monitoring unit of the target wafer production line. The wafer defect detection data includes wafer image, defect location, defect type, defect size, and each wafer defect detection data includes a production time stamp to record when the data was generated, providing strong support for subsequent quality analysis and automatic detection.

[0017] Defect feature recognition is performed on the plurality of wafer defect detection data to obtain a plurality of defect feature sets.

[0018] Specifically, a neural network model is used to identify defect features from multiple extracted wafer defect inspection data, obtaining multiple defect feature sets. The neural network model used for defect feature identification is trained using historical wafer inspection data, including multiple normal wafers and multiple wafers containing various types of defects. The trained neural network model enables efficient and accurate defect feature identification. Each defect feature set includes defect feature type, geometric features, texture features, and more, providing data support for subsequent quality assessment and automated inspection.

[0019] The plurality of defect feature sets are clustered using the defect feature type as an index to obtain K defect feature clusters, each defect feature cluster including a plurality of defect feature values.

[0020] Specifically, clustering algorithms, such as K-means clustering and hierarchical clustering, are used to index defect feature types and cluster multiple defect feature sets. This clustering involves grouping defect feature values with similar feature values and the same feature type based on their similarity. Different defect feature types are then assigned to different clusters, forming K defect feature clusters, where K is a positive integer ≥ 1. Each cluster contains multiple similar defect feature values belonging to the same defect type. Cluster analysis can help better understand defect type characteristics and provide guidance for subsequent automatic defect detection.

[0021] The K defect feature clusters are traversed to perform defect feature value edge recognition, and K defect feature edge intervals are generated.

[0022] Specifically, the eigenvalues within each of the K defect feature clusters are analyzed to determine the boundaries of their eigenvalues. These boundaries can be natural boundaries of the eigenvalues within the cluster, such as minimum and maximum values, or manually set thresholds to distinguish between samples within and outside the cluster. Based on the identified boundaries, K defect feature edge intervals are generated for each defect feature cluster based on the edge recognition method. These defect feature edge intervals are numerical intervals or probability distributions. Obtaining these defect feature edge intervals through identification facilitates more accurate defect detection and classification during wafer production.

[0023] The K defect feature edge intervals are compared with the target wafer quality feature tolerance interval to determine K defect factors.

[0024] Specifically, a target wafer quality characteristic tolerance interval is set based on the target wafer's application requirements, historical data, and industry standards. The target wafer quality characteristic tolerance interval refers to the acceptable range for various characteristics of the target wafer, and the tolerance interval can be represented by a numerical range. Each defect characteristic cluster edge interval of the K defect characteristic edge intervals is then compared with the target wafer quality characteristic tolerance interval to check whether the defect characteristic edge interval is completely within the target wafer quality characteristic tolerance interval. Based on the comparison results, K defect factors are obtained. The defect factors are used to reflect the severity of the defect or its potential impact on wafer quality. For example, if the defect characteristic edge interval is completely within the target wafer quality characteristic tolerance interval, the defect factor can be set to 0, indicating that the defect characteristic of the cluster is completely within the acceptable range. If the defect characteristic edge interval overlaps with the tolerance interval but partially exceeds the tolerance interval, the defect factor is calculated based on the degree of excess. For example, the proportion of the excess portion to the entire interval can be used as the defect factor, indicating that the defect has a certain impact on wafer quality. If the defect characteristic edge interval is completely outside the tolerance interval, the defect factor should be set to a larger value, indicating that the defect characteristic of the cluster has a serious impact on wafer quality.

[0025] An automatic detection wheel is configured based on the K defect factors, and the automatic detection wheel is dialed multiple times according to a preset number of detection samples to obtain an automatic detection plan, wherein the automatic detection plan includes K detection sample quantities of K defect features.

[0026] Specifically, the size or severity of each of the K defect factors is analyzed, and a detection priority is assigned to each defect feature. The higher the defect factor, the higher the detection priority. Each defect factor is then mapped to a corresponding area of the automatic detection wheel based on the detection priority. The larger the defect factor, the larger the area allocated to it in the wheel. Accordingly, when the subsequent wave wheel is run, the greater the probability of being selected for that test, thereby achieving the configuration of the automatic detection wheel. Each area of the automatic detection wheel represents a specific defect feature, which is used to guide the detection order and frequency of multiple defect features in the automated detection of wafer defects. According to the preset number of detection samples, that is, the preset number of samples that need to be detected during the automatic detection process for each defect feature determined based on defect history data, defect severity, and wafer detection goals, the automatic detection wheel is run multiple times to obtain an automatic detection solution with K detection samples containing K defect features, which helps to achieve efficient and accurate detection of key defects in the wafer production process.

[0027] Based on the automatic detection solution, automatic detection of product defects in the target wafer production line is performed to obtain defect detection results.

[0028] Specifically, the target wafer production line's products are automatically inspected according to the defect characteristics and corresponding inspection sample size in the automatic inspection plan, identifying the defect characteristics of the target wafer production line's products. Based on the K number of inspection samples in the automatic inspection plan, wafer products are continuously and highly accurately inspected, including steps such as image capture, feature extraction, and defect identification. The inspection results of each defect characteristic are recorded in real time to obtain defect detection results, including information such as defect type, number, location, and severity. This not only improves inspection efficiency and ensures the accuracy and reliability of inspection results, but also enables dynamic and real-time detection of wafer defects, thereby improving the wafer quality of the target wafer production line.

[0029] In one embodiment, Figure 2 As shown, traversing the K defect feature clusters to perform defect feature value edge recognition and generate K defect feature edge intervals, including: extracting a first defect feature cluster from the K defect feature clusters, wherein the first defect feature cluster includes multiple first defect feature values; randomly extracting M first defect feature values from the multiple first defect feature values as M transition particles, wherein the M transition particles have M transition densities; the M transition particles move in the first defect feature cluster according to a preset step size to obtain M stage transition particles, wherein the M stage transition particles have M stage transition densities; iteratively updating the M transition particles according to the M stage transition particles to obtain M updated transition particles; determining a first defect feature edge interval of the first defect feature cluster based on the M updated transition particles; performing defect feature value edge recognition on the remaining K-1 defect feature clusters to generate K-1 defect feature edge intervals; generating the K defect feature edge intervals according to the first defect feature edge interval and the K-1 defect feature edge intervals.

[0030] Specifically, first, a defect feature cluster is extracted from the K defect feature clusters as the first defect feature cluster, and the first defect feature cluster contains multiple first defect feature values. Then, M values are randomly selected from the multiple first defect feature values as M transition particles, each transition particle represents a defect feature value, M is a positive integer ≥ 1, and the transition density is the ratio of the number of feature values in the area constructed with the transition particle as the center and the preset step length as the radius to the interval length. The interval length is two preset step lengths. The transition density reflects the distribution density of feature values gathered around the transition particle. M transition particles have M transition densities. Then, the M transition particles are moved within the range of the first defect feature cluster according to the preset step length. The preset step length determines the distance the particle moves each time in the feature space and can be set according to actual needs. During the movement process, each transition particle is adjusted based on the eigenvalue distribution at its location. Based on the eigenvalue distribution at its adjusted location, M stage-transition particles are formed. Based on the new positions of the stage-transition particles, their stage-transition densities are calculated. The stage-transition density reflects the degree of particle aggregation in the feature space, and each stage-transition particle has a corresponding stage-transition density. Subsequently, the initial M transition particles are iteratively updated based on the density information of the stage-transition particles. This iterative update includes adjusting the particle's position, velocity, or density. This iterative update enables the transition particles to more accurately reflect the distribution of eigenvalues within a specific interval, thereby generating M updated transition particles. Furthermore, by analyzing the density changes, cluster boundaries, or other statistical methods of the M updated transition particles, the defect feature edge interval of the first defect feature cluster is identified. The defect feature edge interval defines the abnormal range of the defect feature values in this cluster, representing the left and right intervals of the defect feature values corresponding to the defective wafer. The above steps are repeated to perform defect feature value edge identification on the remaining K-1 defect feature clusters, generating corresponding K-1 defect feature edge intervals. Finally, the defect feature edge interval of the first defect feature cluster and the remaining K-1 defect feature edge intervals are merged to form a complete K defect feature edge intervals. This can realize comprehensive defect feature value edge recognition of the K defect feature clusters of the wafer and generate accurate defect feature edge intervals, providing an important reference basis for subsequent product defect detection of the target wafer production line, ensuring that the detection process can accurately identify various defects and improve the accuracy and reliability of detection.

[0031] Furthermore, the M transition particles are iteratively updated according to the M stage transition particles to obtain M updated transition particles, including: comparing the M transition densities with the M stage transition densities, and when the M stage transition densities are greater than the M transition densities, using the M stage transition particles to update the M transition particles to obtain M stage updated transition particles; and taking the M stage updated transition particles as the starting point, moving in the first defect feature cluster according to the M first transition directions according to a preset step size, and iteratively updating according to the movement results until a preset number of iterations is met to obtain the M updated transition particles.

[0032] Specifically, after obtaining M phase transition particles and their corresponding M phase transition densities, the transition densities of the M phase transition particles are compared with the phase transition densities of the M phase transition particles formed after they are moved. If the phase transition density of a phase transition particle is greater than the transition density of its corresponding initial transition particle, it indicates that the position of the phase transition particle may be closer to or located in a high-density area of the data. In this case, the M phase transition particles are used to update the M original transition particles, forming M phase-updated transition particles, that is, the positions of the particles are moved from their original positions to the new positions of the phase transition particles. Then, using the M phase-updated transition particles as the new starting point, the particles are again moved within the first defect feature cluster according to a preset step size. Based on the results of this movement, the particles are iteratively updated again, so that the transition particles are closer to the true eigenvalue distribution. This step is repeated until the preset number of iterations is met. When the preset maximum number of iterations is reached, M updated transition particles are obtained after multiple iterative updates, providing an important basis for the subsequent determination of the defect feature edge interval, thereby improving the accuracy and reliability of subsequent wafer defect detection.

[0033] Furthermore, the M transition particles are iteratively updated according to the M stage transition particles to obtain M updated transition particles, and the method further includes: when the M stage transition densities are less than or equal to the M transition densities, M first transition directions are added to a taboo table, wherein the taboo table has a number of taboo iterations; and again taking the M transition particles as the starting point, moving in the first defect feature cluster in a direction other than the transition direction stored in the taboo table according to a preset step size, and iteratively updating according to the movement results until the preset iteration parameters are met to obtain the M updated transition particles.

[0034] Specifically, when the M transition densities are compared with the M stage transition densities and the result is that the M stage transition densities are less than or equal to the M transition densities, it indicates that the position of the stage transition particle is not better than the position of the original transition particle, or that no better solution has been found. At this time, the M first transition directions are added to the taboo table. The taboo table is one of the core components of the taboo search algorithm, which is used to record solutions or search directions that should be avoided during the search process. The taboo table can be a list or array that stores transition directions that have been marked as taboo. The taboo table also has a taboo iteration count. For example, if the taboo iteration count is set to 3, the corresponding direction cannot be used within 3 iterations to avoid falling into an inferior solution.

[0035] Next, using these M transition particles as the starting point again, they move within the first defect feature cluster according to the preset step size. However, this time, the movement excludes the transition directions stored in the taboo table. That is, the movement only occurs in non-taboo directions according to the preset step size, avoiding repeated selection of known bad directions during the search process, thereby improving search efficiency and solution quality. In each iteration, the eigenvalue distribution information of the current transition particles is updated based on the movement results. When the transition density at a certain stage is greater than the corresponding transition density, the transition particles at that stage are used to update the corresponding transition particles, and iterative updates are continued. Each iteration checks whether the directions in the taboo table are used until the preset iteration parameters are met. Ultimately, at the end of the iterative process, the M updated transition particles obtained represent the optimized exploration results after considering the taboo search strategy. This avoids re-entry into the search area that previously led to unsatisfactory results, thereby increasing the possibility of finding the global optimal solution, helping to improve the accuracy of defect feature value edge recognition, and thus improving the accuracy and reliability of wafer defect detection.

[0036] Furthermore, the first defect characteristic edge interval of the first defect characteristic cluster is determined based on the M updated transition particles, including: taking the updated transition particle corresponding to the maximum transition density among the M updated transition particles as the target transition particle; constructing a first partitioning interval with the target transition particle as the midpoint of the interval and the preset step size as the interval radius; taking the maximum value of multiple first defect characteristic values in the first defect characteristic cluster located in the first partitioning interval as the first right endpoint; taking the minimum value of multiple first defect characteristic values in the first defect characteristic cluster located in the first partitioning interval as the first left endpoint; and constructing the first defect characteristic edge interval based on the first left endpoint and the first right endpoint.

[0037] Specifically, from the M updated transition particles, the updated transition particle corresponding to the maximum transition density is selected as the target transition particle. The target transition particle refers to the location most likely to contain the defect feature edge during the current search process. Then, with the target transition particle's location as the midpoint of the interval and a preset step size as the interval radius, a first partition interval is constructed around the midpoint. The first partition interval is a neighborhood centered on the target transition particle, and its size is determined by the step size. Within the first defect feature cluster, all first defect feature values within the first partition interval are screened. The maximum value from the screened feature values is selected as the first right endpoint, and the minimum value from the screened feature values is selected as the first left endpoint. Finally, using the determined first left endpoint and first right endpoint, a closed interval is constructed, namely the first defect feature edge interval. The first defect feature edge interval encompasses the range of feature values that may represent the defect feature edge within the current exploration range. By effectively utilizing the information provided by the updated transition particles, a relatively accurate defect feature edge interval is determined within the first defect feature cluster, enabling better identification of defect features and improving defect detection efficiency and accuracy.

[0038] In one embodiment, Figure 3 As shown, configuring an automatic detection wheel based on the K defect factors includes: collecting a preset wheel area; calculating the ratio of the K defect factors to the sum of the K defect factors respectively, and using the calculation results as K wheel coefficients; using the K wheel coefficients as weight ratios, dividing the preset wheel area according to the weight ratios, and marking K defect features in corresponding divided areas of the wheel to generate the automatic detection wheel.

[0039] Specifically, the total area of the preset roulette wheel is determined based on actual needs. Then, for K known defect factors, the ratio of each defect factor to the sum of the K defect factors is calculated. This ratio reflects the importance or influence weight of the defect factor among all factors. These ratios are then used as K roulette wheel coefficients. The size of the roulette wheel coefficient represents the relative importance of the corresponding defect factor, and each roulette wheel coefficient corresponds to a defect factor. Using the calculated K roulette wheel coefficients as weight ratios, the preset roulette wheel area is divided. Each defect factor is assigned to a specific area on the roulette wheel based on its roulette wheel coefficient. The area of this area is proportional to the weight of the factor. K defect features are then identified in the corresponding divided areas of the roulette wheel. The defect feature identifiers are used to indicate the corresponding defect factor in each area. After all defect features are identified, an automatic detection roulette wheel is generated. By configuring the automatic detection roulette wheel, the weight and importance of each defect factor in overall defect detection can be intuitively displayed, thereby improving the efficiency and accuracy of defect detection.

[0040] Furthermore, the automatic detection wheel is dialed multiple times according to the preset number of detection samples to obtain an automatic detection plan, including: dialing the automatic detection wheel multiple times according to the preset number of detection samples to obtain an initial automatic detection plan; extracting K initial detection sample quantities of K defect features in the initial automatic detection plan, and calculating K random inspection coefficients based on the K initial detection sample quantities; calculating the random inspection similarity between the K wheel coefficients and the K random inspection coefficients, and when the random inspection similarity meets the preset similarity threshold, using the initial automatic detection plan as the automatic detection plan.

[0041] Specifically, the automatic inspection wheel is randomly rotated multiple times according to the preset number of inspection samples. After the wheel is rotated multiple times, the number of samples of defect features included in each inspection is recorded based on the defect features marked on the wheel and the corresponding divided areas, forming an initial automatic inspection plan.

[0042] From the initial automatic detection plan, K initial detection sample sizes for K defect features are extracted. These sample sizes represent the detection frequency or number of each defect feature in the initial detection plan. Based on these K initial detection sample sizes, a sampling coefficient is calculated for each defect feature. The sampling coefficient is the ratio of the detection sample size to the total sample size, or another indicator reflecting the detection frequency. The K roulette wheel coefficients (i.e., weight ratios) are then compared with the K sampling coefficients to calculate the sampling similarity between them. The sampling similarity can be determined through correlation analysis or by calculating a similarity metric between two coefficient vectors, such as cosine similarity. A preset similarity threshold is used to determine whether the sampling similarity is sufficiently high. If the sampling similarity meets or exceeds this threshold, it indicates that the sample size distribution in the initial automatic detection plan is relatively consistent with the weight distribution on the roulette wheel, indicating that the initial automatic detection plan is reasonable. This plan is then used as the final automatic detection plan for subsequent automatic detection processes. This approach ensures that the automatic detection plan matches the roulette wheel coefficients, improving the accuracy and effectiveness of automatic detection.

[0043] Furthermore, the automatic detection wheel is dialed multiple times according to a preset number of detection samples to obtain an automatic detection scheme, and it also includes: when the random inspection similarity does not meet the preset similarity threshold, the automatic detection wheel is dialed multiple times again according to the preset number of detection samples to obtain an iterative automatic detection scheme; calculating the iterative random inspection similarity of K iterative random inspection coefficients and K wheel coefficients in the iterative automatic detection scheme, and if the iterative random inspection similarity meets the preset similarity threshold, the iterative automatic detection scheme is used as the automatic detection scheme.

[0044] Specifically, when the sampling similarity does not meet the preset similarity threshold, it means that the sample size distribution in the initial automatic detection scheme is quite different from the weight distribution on the roulette wheel, and the automatic detection roulette wheel needs to be dialed multiple times according to the preset number of detection samples. Based on the results of dialing the roulette wheel multiple times, the detection data is collected and sorted to form an iterative automatic detection scheme. The iterative scheme contains new detection sample sizes for K defect features. Extract K iterative detection sample sizes of K defect features in the iterative automatic detection scheme, and calculate the iterative sampling coefficient of each defect feature based on these iterative detection sample sizes. Compare the K roulette coefficients, that is, the weight ratio, with the K iterative sampling coefficients to calculate the iterative sampling similarity between them. It can also be obtained by correlation analysis or calculating the similarity measure between two coefficient vectors, such as cosine similarity and other methods to measure similarity. Check whether the iterative sampling similarity meets the preset similarity threshold. If it does, it means that the iterative automatic detection scheme matches the roulette coefficient, and the iterative automatic detection scheme is used as the automatic detection scheme. If the iterative sampling similarity still does not meet the preset similarity threshold, consider further iterations or adjusting parameters such as the number of inspection samples and the roulette wheel coefficient, then repeating the above steps until a satisfactory automatic inspection solution is found. This iterative approach allows the automatic inspection solution to be gradually optimized to ensure that it matches the roulette wheel coefficient, thereby improving the accuracy and reliability of automatic semiconductor wafer defect detection.

[0045] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0046] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A method for automatic detection of semiconductor wafer defects based on deep learning, characterized in that: The method comprises: Interacting with a quality monitoring unit of a target wafer production line to extract multiple wafer defect detection data within a preset detection window, wherein each wafer defect detection data includes a production time stamp; Performing defect feature recognition on the plurality of wafer defect detection data to obtain a plurality of defect feature sets; Clustering the multiple defect feature sets using the defect feature type as an index to obtain K defect feature clusters, each defect feature cluster including multiple defect feature values; Traversing the K defect feature clusters to perform defect feature value edge recognition and generate K defect feature edge intervals; Comparing the K defect feature edge intervals with the target wafer quality feature tolerance interval to determine K defect factors; configuring an automatic detection wheel based on the K defect factors, and turning the automatic detection wheel multiple times according to a preset number of detection samples to obtain an automatic detection plan, wherein the automatic detection plan includes K detection samples of K defect characteristics; Performing automatic detection of product defects on a target wafer production line based on the automatic detection solution to obtain defect detection results; Turn the automatic detection wheel multiple times according to the preset number of test samples to obtain the automatic detection plan, including: Turn the automatic detection wheel multiple times according to the preset number of test samples to obtain the initial automatic detection plan; Extracting K initial detection sample sizes of K defect features in the initial automatic detection scheme, and calculating K sampling coefficients based on the K initial detection sample sizes; Calculating the sampling similarity between the K roulette coefficients and the K sampling coefficients, and when the sampling similarity meets a preset similarity threshold, using the initial automatic detection scheme as the automatic detection scheme; When the random inspection similarity does not meet the preset similarity threshold, the automatic detection wheel is dialed multiple times according to the preset number of detection samples to obtain an iterative automatic detection solution; The iterative sampling similarity of the K iterative sampling coefficients and the K roulette coefficients in the iterative automatic detection scheme is calculated. If the iterative sampling similarity meets a preset similarity threshold, the iterative automatic detection scheme is used as the automatic detection scheme.

2. The method for automatic semiconductor wafer defect detection based on deep learning according to claim 1, wherein: Traversing the K defect feature clusters to perform defect feature value edge identification, generating K defect feature edge intervals, including: Extracting a first defect feature cluster from the K defect feature clusters, wherein the first defect feature cluster includes a plurality of first defect feature values; Randomly extracting M first defect characteristic values from the plurality of first defect characteristic values as M transition particles, wherein the M transition particles have M transition densities; The M transition particles move in the first defect characteristic cluster according to a preset step length to obtain M stage transition particles, wherein the M stage transition particles have M stage transition densities; Iteratively updating the M transition particles according to the M stage transition particles to obtain M updated transition particles; Determining a first defect feature edge interval of the first defect feature cluster based on the M updated transition particles; Perform defect feature value edge recognition on the remaining K-1 defect feature clusters to generate K-1 defect feature edge intervals; The K defect feature edge intervals are generated according to the first defect feature edge interval and the K-1 defect feature edge intervals.

3. The method for automatic semiconductor wafer defect detection based on deep learning according to claim 2, wherein: Iteratively updating the M transition particles according to the M stage transition particles to obtain M updated transition particles, including: comparing the M transition densities with the M stage transition densities, and when the M stage transition densities are greater than the M transition densities, updating the M transition particles using the M stage transition particles to obtain M stage-updated transition particles; And taking the M stage updated transition particles as the starting point, moving in the first defect feature cluster according to the M first transition directions according to the preset step size, iteratively updating according to the movement results until the preset number of iterations is met, and obtaining the M updated transition particles.

4. The method for automatic semiconductor wafer defect detection based on deep learning according to claim 3, wherein: include: When the M stage transition densities are less than or equal to the M transition densities, adding the M first transition directions into a taboo table, wherein the taboo table has a taboo iteration number; Taking the M transition particles as the starting point again, move in the first defect feature cluster in a direction other than the transition direction stored in the taboo table according to a preset step size, and iterate and update according to the movement result until the preset iteration parameters are met to obtain the M updated transition particles.

5. The method for automatic semiconductor wafer defect detection based on deep learning according to claim 3, wherein: Determining a first defect feature edge interval of the first defect feature cluster based on the M updated transition particles includes: The updated transition particle corresponding to the maximum transition density among the M updated transition particles is used as the target transition particle; Constructing a first partition interval with the target transition particle as the midpoint of the interval and the preset step length as the interval radius; taking the maximum value of a plurality of first defect feature values in the first defect feature cluster and located in the first divided interval as a first right endpoint; taking the minimum value of a plurality of first defect feature values in the first defect feature cluster located in the first divided interval as a first left endpoint; The first defect feature edge interval is constructed according to the first left endpoint and the first right endpoint.

6. The method for automatic semiconductor wafer defect detection based on deep learning according to claim 1, wherein: Configuring an automatic inspection wheel based on the K defect factors includes: Collect the preset wheel area; Calculating the ratios of the K defect factors to the sum of the K defect factors respectively, and using the calculation results as K roulette coefficients; The K wheel coefficients are used as weight ratios, the preset wheel area is divided according to the weight ratios, and K defect features are marked in corresponding divided areas of the wheel to generate the automatic detection wheel.

Images