Unsupervised learning for repeater defect detection

Abstract

To find repeater defects, optical inspection results are obtained for one or more semiconductor wafers. Based on the optical inspection results, a plurality of defects on the one or more semiconductor wafers are identified. Among the plurality of defects, defects having identical die locations on multiple dies of the one or more semiconductor wafers are classified as repeater defects. Based on the optical inspection results, unsupervised machine learning is used to cluster the repeater defects into a plurality of clusters. The repeater defects are scored. Scoring the repeater defects includes assigning a respective score to each repeater defect within the plurality of clusters based on the degree to which the cluster contains multiple instances of the respective repeater defect. The repeater defects are ranked based on their respective scores.

Description

[Technical field]

[0001] The present disclosure relates to semiconductor defect detection, and more particularly, to detecting repeater defects using unsupervised learning. [Background technology]

[0002] Related Applications This application claims priority to U.S. Provisional Patent Application No. 63 / 068,380 (August 21, 2020), which is incorporated by reference in its entirety for all purposes.

[0003] Repeater defects on a semiconductor wafer are defects that occur at the same die location on multiple dies on the wafer. Repeater defects can be caused by defects on a reticle (i.e., photomask) that cause a layer to be patterned incorrectly at the same location on a particular die on the wafer (i.e., at the same reticle level location on multiple dies on the wafer). Finding and removing repeater defects is an important part of yield improvement in semiconductor manufacturing.

[0004] Defects on semiconductor wafers, including repeater defects, can be identified by optical inspection of the wafer. However, optical inspection identifies nuisance defects as well as defects of interest. Defects of interest are true defects that impair the functionality of a semiconductor device, while nuisance defects arise from minor process variations and do not impair the functionality of the device. The presence of defects of interest on a semiconductor wafer reduces the wafer yield (i.e., the number or percentage of functional dies on a wafer). The presence of nuisance defects on a semiconductor wafer does not affect the yield. Thus, engineers are interested in finding defects of interest but not in finding nuisance defects. However, with modern optical inspection, nuisance defects can outnumber the defects of interest by several orders of magnitude.

[0005] Repeater defects can be identified by performing optical inspection and identifying defects that have the same die-level or reticle-level location (e.g., by stacking i images of multiple dies). However, this approach suffers from a low signal-to-noise ratio for finding repeater defects that are defects of interest. Due to the large number of nuisance defects identified by optical inspection, many of the repeater defects identified using this technique will be nuisance defects. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] US Patent Application Publication No. 2018 / 0130199 Summary of the Invention [Problem to be solved by the invention]

[0007] Therefore, there is a need for a more accurate method and system for identifying repeater defects which are defects of interest. [Means for solving the problem]

[0008] In some embodiments, a method for finding repeater defects includes obtaining optical inspection results of one or more semiconductor wafers. Based on the optical inspection results, a plurality of defects on the one or more semiconductor wafers are identified. Of the plurality of defects, defects having identical die locations on a plurality of dies of the one or more semiconductor wafers are classified as repeater defects. Based on the optical inspection results, unsupervised machine learning is used to cluster the repeater defects into a plurality of clusters. Scoring the repeater defects. Scoring the repeater defects includes assigning a respective score to each repeater defect based on a degree to which a cluster in the plurality of clusters includes a plurality of instances of the respective repeater defect. The repeater defects are ranked based on the respective scores.

[0009] In some embodiments, a non-transitory computer-readable storage medium stores one or more programs for execution by one or more processors. The one or more programs include instructions for performing the methods described above.

[0010] In some embodiments, a system includes an optical inspection tool, one or more processors, and a memory storing one or more programs for execution by the one or more processors, the one or more programs including instructions for performing the methods described above.

[0011] The method, system, and non-transitory computer-readable storage medium enable repeater defects that are defects of interest (i.e., true repeater defects) to be identified with a high signal-to-noise ratio. [Brief description of the drawings]

[0012] For a better understanding of the various described embodiments, reference should be made to the following detailed description taken in conjunction with the following drawings, in which like reference numerals refer to corresponding parts throughout the drawings and specification, in which: [Figure 1] 1 is a wafer map showing defect locations on a semiconductor wafer. [Diagram 2] 1 is a flowchart illustrating a method for locating repeater defects according to some embodiments. [Figure 3A] 3 shows an example of clustering results performed using unsupervised machine learning in the method of FIG. 2 according to some embodiments. [Figure 3B] 3B illustrates a graph of repeater defect scores based on the clustering of FIG. 3A in accordance with some embodiments. [Figure 4] 1 illustrates a graph used to select hyper-parameter values, according to some embodiments. [Diagram 5] FIG. 1 is a block diagram of a semiconductor defect identification system according to some embodiments. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Reference will now be made in detail to various embodiments, examples of which are illustrated in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the various embodiments being described. However, it will be apparent to those skilled in the art that the various embodiments being described may be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.

[0014] FIG. 1 is a wafer map showing defect locations on a semiconductor wafer 100. The semiconductor wafer 100 has multiple dies 104 that are fabricated using photolithography by stepping a reticle (i.e., a photomask) across the wafer 100 and successively exposing respective regions 102 on the wafer to pattern respective layers of the multiple dies 104. Each region 102 includes an array of semiconductor dies 104. Each semiconductor die 104 in the array (i.e., within a region 102) corresponds to a respective patterned region on a reticle. Each reticle has an array of nominally identical patterned regions, each of which corresponds to a respective semiconductor die 104 in the array of each region 102. In the example of FIG. 1, each region 102 has a 2×3 array of semiconductor dies 104-1 through 104-6, and each reticle has a corresponding 2×3 array of nominally identical patterned regions. Other examples of the number of semiconductor dies 104 on the wafer 100 are possible, as are other examples of the area 102 and reticle array sizes.

[0015] The defects shown on the wafer map of semiconductor wafer 100 include repeater defects 106 and non-repeater defects 108. In the example of FIG. 1, the repeater defects include repeater defects 106-1, 106-2, and 106-3. Repeater defect 106-1 occurs at the same die level location in six instances of semiconductor die 104-1, and therefore occurs at the same reticle level location in six different dies. Repeater defect 106-2 occurs at the same die level location in two instances of semiconductor die 104-4, and therefore occurs at the same reticle level location in two different dies. Repeater defect 106-3 occurs at the same die level location in three instances of semiconductor die 104-6, and therefore occurs at the same reticle level location in three different dies. Non-repeater defects 108 do not occur at the same die level location in multiple semiconductor dies 104, and therefore do not occur at the same reticle level location in multiple semiconductor dies 104. In general, defects have the same reticle level location if they are at the same location in multiple dies and if multiple dies are at the same location in their respective arrays in region 102.

[0016] 2 is a flow chart illustrating a method 200 for locating repeater defects (e.g., repeater defects 106-1, 106-2, and / or 106-3 of FIG. 1) in accordance with some embodiments. Method 200 may be performed by a computer system communicatively coupled to an optical inspection tool (e.g., optical inspection tool 530 of FIG. 5). For example, method 200 may be performed by a computer system within semiconductor defect identification system 500 of FIG. 5.

[0017] In method 200, optical inspection results are obtained (202) for one or more semiconductor wafers (e.g., see FIG. 1 for one or more instances of semiconductor wafer 100). The optical inspection results include differential images of semiconductor dies on the one or more semiconductor wafers. Each differential image is the difference between a respective target image and a reference image: the differential image is generated by comparing the target image of the die to the reference image (e.g., by subtracting the reference image from the target image or vice versa). The target images are generated by scanning the one or more wafers using an optical inspection tool (e.g., optical inspection tool 530 of FIG. 5). In some embodiments, the optical inspection results also include the target images. The optical inspection results may be obtained directly from the optical inspection tool and / or by being retrieved from a previously stored memory of a computer system (e.g., memory 510 of FIG. 5).

[0018] Based on the optical inspection results, a plurality of defects are identified (204) on the one or more semiconductor wafers. For example, the plurality of defects are identified in the difference image. The plurality of defects include both repeater defects and non-repeater defects, but it has not yet been determined which are which at this point in the method 200. Similarly, the plurality of defects include both defects of interest and nuisance defects, but it is not known which are which at this point in the method 200.

[0019] Defects among the plurality of defects having the same die location on multiple dies of one or more semiconductor wafers are classified as repeater defects (206). In some embodiments, defects among the plurality of defects having the same reticle level location on one or more semiconductor wafers are classified as repeater defects (208). Defects that do not occur at the same location on multiple dies or do not occur at the same reticle level location are not classified as repeater defects (e.g., classified as non-repeater defects). For example, defects 106-1, 106-2, and 106-3 (FIG. 1) are classified as repeater defects, but defect 108 (FIG. 1) is not classified as a repeater defect (e.g., classified as a non-repeater defect).

[0020] In some embodiments, defect attributes of the repeater defects are extracted from the optical inspection results 210. The defect attributes include, for example, size, die location, wafer location, and / or light intensity (e.g., as indicated by gray levels in a difference image) 212. Once extracted, the defect attributes can be included in the optical inspection results.

[0021] Based on the optical inspection results, unsupervised machine learning (also referred to as unsupervised learning) is used to cluster the repeater defects into a plurality of clusters (214). According to some embodiments, non-repeater defects are excluded from the clustering. The unsupervised machine learning can be performed using an unsupervised machine learning algorithm that does not specify a predetermined number of clusters into which the plurality of defects should be clustered (216). For example, a density-based spatial clustering for applications with noise (DBSCAN) algorithm may be used that does not specify a predetermined number of clusters into which the items (in this case, the defects) should be clustered. The unsupervised machine learning can determine that some of the repeater defects are outliers that are not included in any of the plurality of clusters. The outliers may be discarded and therefore may not be considered in subsequent steps of the method 200.

[0022] In some embodiments, the unsupervised machine learning uses (218) the defect attributes for the repeater defects extracted in step 210. The unsupervised machine learning algorithm takes the defect attributes as input and clusters the repeater defects into multiple clusters using the defect attributes. The defect attributes can be used as dimensions of the space in which the repeater defects are clustered.

[0023] In some embodiments (e.g., without defect attribute extraction 210), a target image from the optical inspection is provided 220 to an unsupervised machine learning algorithm. The target image may also be referred to as a patch image. The unsupervised machine learning algorithm receives the target image as input and uses the target image to cluster defects into clusters. Thus, the unsupervised machine learning algorithm may perform clustering without receiving the extracted defect attributes. Such an embodiment is possible because information about the defect attributes is inherently embedded in the target image.

[0024] The repeater defects are scored (222). A respective score is assigned to each repeater defect (i.e., for each group of defects having the same die location (e.g., the same reticle level location)) based on the degree to which a cluster in the plurality of clusters contains multiple instances of the respective repeater defect. Thus, the score is based on the occurrence of repeater defects having the same die location (e.g., the same reticle level location) in each cluster.

[0025] For example, the proportion (eg, traction or percentage) of each repeater defect in each cluster is calculated. P ij =C ij / T j (1) where Tj is the number of instances (number of defects) of the i-th repeater defect in the i-th cluster, Tj is the total number of defects in the j-th cluster (i.e., the total number of repeater defects regardless of location), and Pij is the proportion of the i-th repeater defect in the i-th cluster. A score may be determined based on these ratios. In some embodiments, each score corresponds to the maximum percentage (or proportion) of instances of the respective repeater defect in a cluster of multiple clusters. Score i =max j (P ij ) (2) where i refers to the repeater defect and i refers to the cluster. In some other embodiments, each score corresponds to the average of the non-zero percentage of instances of the respective repeater defect in the multiple clusters. Score i =mean ∀Pij≠0 (P ij ) (3) where i is the index of the repeater defect and j is the index of the cluster.

[0026] The repeater defects are ranked based on their respective scores (224). In some embodiments, repeater defects having respective scores that meet a threshold are selected (226). For example, repeater defects having scores that exceed, or equal to, or exceed the threshold are selected. In another example, a specified number or percentage of repeater defects having the highest scores are selected (i.e., the threshold is that the score of a repeater defect places it among a specified number or percentage of the top scoring repeater defects).

[0027] 3A shows an example of a hypothetical result 300 of clustering performed using unsupervised machine learning in step 214, according to some embodiments. The result 300 includes four clusters 302-1 to 302-4 and an outlier 304 that is not included in any of the clusters 302. The outlier 304 is discarded and therefore not considered when scoring and ranking the repeater defects in steps 222 and 224. The repeater defects classified in step 206 include five different repeater defects 306-1 to 306-5. Cluster 302-1 includes four instances of repeater defect 306-1 and one instance of repeater defect 306-2. Cluster 302-2 includes three instances of repeater defect 306-2, one instance of repeater defect 306-3, and one instance of repeater defect 306-4. Cluster 302-3 includes two instances of repeater defect 306-3, one instance of repeater defect 306-2, and one instance of repeater defect 306-4. Cluster 302-4 includes three instances of repeater defect 306-4 and one instance of repeater defect 306-1. None of clusters 302-1 through 302-4 includes any instance of repeater defect 306-5.

[0028] 3B illustrates a graph 310 of scores 312 for repeater defects 306-1 through 306-5, according to some embodiments. As shown, the highest percentage of repeater defects 306-1 in cluster 302 is 80% (4 out of 5) in cluster 302-1. The highest percentage of repeater defects 306-2 in cluster 302 is 60% (3 out of 5) in cluster 302-2. The highest percentage of repeater defects 306-3 in cluster 302 is 50% (2 out of 4) in cluster 302-3. The highest percentage of repeater defects 306-4 in cluster 302 is 75% (3 out of 4) in cluster 302-4. Additionally, the highest percentage of repeater defects 306-5 in cluster 302 is zero, since no instances of repeater defects 306-5 occur in any of clusters 302-1 through 302-4. Thus, repeater defects 306-1 through 306-5 have scores of 0.8, 0.6, 0.5, 0.75, and 0, respectively (in some embodiments, repeater defect 306-5 is not assigned a score because it occurs only as an outlier). As ranked based on their scores, the ordering of repeater defects 306 is (from highest to lowest) 306-1, 306-4, 306-2, 306-3, and 306-5. The scores of repeater defects 306-1, 306-2, and 306-4 meet threshold 314, but the scores of repeater defects 306-3 and 306-5 do not. Thus, repeater defects 306-1, 306-2, and 306-4 are selected, but repeater defects 306-3 and 306-5 are not selected.

[0029] In some embodiments, a report is generated (228) that identifies the selected repeater defects. For example, the report may identify only the selected repeater defects. In another example, the report includes both selected and unselected repeater defects and indicates which repeater defects were selected. The report may include a list of the repeater defects and / or corresponding clusters, and / or may include a wafer map. The ratio of defects of interest to nuisance defects is significantly higher for the selected repeater defects than for the set of defects classified as repeater defects in step 206.

[0030] The unsupervised machine learning algorithm used in step 214 of method 200 (FIG. 2) may have one or more hyperparameters. A hyperparameter is a parameter that has a value selected before performing machine learning. Thus, the selected value of the hyperparameter is not based on the data processed during machine learning. An example of such a hyperparameter is "eps" (also called epsilon) of DBSCAN. During the unsupervised machine learning algorithm development process, method 200 may be repeatedly run with different hyperparameter values ​​(e.g., with different values ​​of eps) for the same optical inspection results (and thus for the same one or more semiconductor wafers). The value that produces the highest repeater count (e.g., which produces a selected cluster with the largest total number of defects) may be selected as the hyperparameter value. This hyperparameter value is then used when method 200 is implemented in a production environment.

[0031] 4 illustrates a hypothetical graph 400 used to select hyperparameter values, according to some embodiments. The y-axis of the graph 400 is the repeater count 404, and the x-axis of the graph 400 is the hyperparameter value 402. Different hyperparameter values ​​402 result in different repeater counts 404 for different iterations of the method 200. The hyperparameter value 406 that produces the highest repeater count 404 is selected and used for unsupervised machine learning when the method 200 is implemented in a production environment.

[0032] 5 is a block diagram of a semiconductor defect identification system 500 according to some embodiments. The semiconductor defect identification system 500 includes an optical inspection tool 530 and a computer system with one or more processors 502 (e.g., CPU), a user interface 506, a memory 510, and a communication bus 504 interconnecting these components. In some embodiments, the optical inspection tool 530 is communicatively coupled to the computer system via one or more wired and / or wireless networks. In some embodiments, the semiconductor defect identification system 500 includes a plurality of optical inspection tools 530 communicatively coupled to the computer system. The computer system may further include one or more wired and / or wireless network interfaces for communicating with the optical inspection tool 530 and / or remote computer systems.

[0033] User interface 506 may include a display 507 and one or more input devices 508 (e.g., a keyboard, a mouse, a touch-sensitive surface of display 507, etc.). Display 507 may display results including reports identifying repeater defects (e.g., a report identifying clusters selected according to step 228 of method 200 of FIG. 2).

[0034] The memory 510 includes volatile and / or non-volatile memory. The memory 510 (e.g., the non-volatile memory in the memory 510) includes a non-transitory computer-readable storage medium. The memory 510 optionally includes one or more storage devices located remotely from the processor 502 and / or a non-transitory computer-readable storage medium removably inserted into the system 500. The memory 510 (e.g., the non-transitory computer-readable storage medium of the memory 510) includes instructions for performing all or a portion of the method 200 (FIG. 2).

[0035] In some embodiments, memory 510 (eg, the non-transitory computer-readable storage medium of memory 510) stores the following modules and data, or a subset or superset thereof: An operating system 512 including procedures for handling various basic system services and performing hardware-dependent tasks; a defect identification module 514 (e.g., for performing step 204 of the method 200 of FIG. 2); a repeater classification module 516 (e.g., for performing step 206 of the method 200 of FIG. 2); an attribute extraction module 518 (e.g., for performing step 210 of the method 200 of FIG. 2); a clustering module 520 (e.g., for performing step 214 of the method 200 of FIG. 2); a cluster scoring, ranking, and selection module 524 (e.g., for performing steps 222, 224, and / or 226 of the method 200 of FIG. 2); a reporting module 526 (e.g., for performing step 228 of the method 200 of FIG. 2); and optical inspection results 528 (e.g., including difference images, target images, and / or defect attributes). The clustering module 520 can perform unsupervised machine learning 522 (e.g., can include instructions for implementing an unsupervised machine learning algorithm). In some embodiments, the unsupervised machine learning 522 does not specify a predetermined number of clusters into which the defects are clustered.

[0036] Each of the modules stored in memory 510 corresponds to a set of instructions for performing one or more functions described herein. Separate modules need not be implemented as separate software programs. Modules and various subsets of modules may be combined or otherwise reconfigured. In some embodiments, memory 510 stores a subset or superset of the modules and / or data structures identified above.

[0037] 5 is intended as a functional description of various features that may be present in a semiconductor defect identification system, rather than as a structural schematic. For example, the functionality of the computer systems in semiconductor defect identification system 500 may be divided among multiple devices. Some of the modules stored in memory 510 may alternatively be stored in one or more other computer systems communicatively coupled to the computer systems of semiconductor defect identification system 500 via one or more networks.

[0038] The foregoing description has been described with reference to specific embodiments for purposes of illustration. However, the above exemplary description is not intended to be exhaustive or to limit the scope of the claims to the precise form disclosed. Many modifications and variations are possible in light of the above teachings. The embodiments have been selected to best explain the principles underlying the claims and their practical application, thereby enabling those skilled in the art to best utilize the embodiments with various modifications suited to the particular use contemplated.

Claims

1. 1. A method comprising: obtaining optical inspection results for one or more semiconductor wafers; identifying a plurality of defects on the one or more semiconductor wafers based on results of the optical inspection; classifying, among the plurality of defects, defects having identical die locations on multiple dies of the one or more semiconductor wafers as repeater defects; clustering the repeater defects into a plurality of clusters based on the optical inspection results using unsupervised machine learning; scoring the repeater defects by assigning a respective score to each of the repeater defects based on the degree to which a cluster in the plurality of clusters contains multiple instances of the respective repeater defect; ranking the repeater defects based on their respective scores; A method for providing the above.

2. selecting the repeater defects having respective scores that satisfy a threshold; generating a report identifying the selected repeater defects; The method of claim 1 further comprising:

3. 2. The method of claim 1, wherein the classifying step comprises classifying defects among the plurality of defects that have identical reticle level locations on the one or more semiconductor wafers as the repeater defects.

4. The method of claim 1 , wherein each of the scores corresponds to a maximum percentage of instances of the respective repeater defect in a cluster of the plurality of clusters.

5. The method of claim 1 , wherein each of the scores corresponds to an average of a non-zero percentage of instances of the respective repeater defect in the plurality of clusters.

6. 10. The method of claim 1, further comprising extracting defect attributes of the repeater defects from the optical inspection results, wherein the unsupervised machine learning uses the defect attributes.

7. The method of claim 6 , wherein the defect attributes include defect attributes selected from the group consisting of size, die location, wafer location, and light intensity.

8. the optical inspection results include target images of dies on the one or more semiconductor wafers and difference images of dies on the one or more semiconductor wafers, each difference image being a difference between a respective target image and a reference image; wherein the identifying step includes identifying a plurality of defects in the difference image; The method of claim 1 , wherein the clustering the repeater defects using unsupervised machine learning comprises providing a target image to the unsupervised machine learning algorithm.

9. 2. The method of claim 1, wherein the clustering of the repeater defects using unsupervised machine learning is performed using an unsupervised machine learning algorithm that does not specify a predetermined number of clusters.

10. 1. A computer-readable storage medium storing one or more programs for execution by one or more processors, the one or more programs comprising: obtaining optical inspection results for one or more semiconductor wafers; identifying a plurality of defects on the one or more semiconductor wafers based on results of the optical inspection; classifying, among the plurality of defects, defects having identical die locations on multiple dies of the one or more semiconductor wafers as repeater defects; clustering the repeater defects into a plurality of clusters based on the optical inspection results using unsupervised machine learning; scoring the repeater defects by assigning a respective score to each repeater defect based on the degree to which a cluster in the plurality of clusters contains multiple instances of the respective repeater defect; ranking the repeater defects based on their respective scores; 23. A computer readable storage medium comprising instructions for:

11. The one or more programs include selecting the repeater defects having respective scores that satisfy a threshold; generating a report identifying the selected repeater defects; The computer-readable storage medium of claim 10 further comprising the instructions:

12. 11. The computer-readable storage medium of claim 10, wherein the instructions for classifying comprise instructions for classifying defects of the plurality of defects that have a same reticle level location on the one or more semiconductor wafers as the repeater defects.

13. The computer-readable storage medium of claim 10 , wherein each of the scores corresponds to a maximum percentage of instances of the respective repeater defects in a cluster of the plurality of clusters.

14. 11. The computer-readable storage medium of claim 10, wherein each of the scores corresponds to an average of a non-zero percentage of instances of the respective repeater defect in the plurality of clusters.

15. 11. The computer-readable storage medium of claim 10, wherein the one or more programs further comprise instructions for extracting defect attributes of the repeater defects from the optical inspection results, and the instructions for using unsupervised machine learning comprise instructions for clustering the repeater defects using the defect attributes.

16. 16. The computer-readable storage medium of claim 15, wherein the defect attributes comprise defect attributes selected from the group consisting of size, die location, wafer location, and light intensity.

17. the optical inspection results include target images of dies on the one or more semiconductor wafers and difference images of dies on the one or more semiconductor wafers, each difference image being a difference between a respective target image and a reference image; the instructions for identifying include instructions for identifying a plurality of defects in the difference image; 11. The computer-readable storage medium of claim 10, wherein the instructions for clustering the repeater defects using unsupervised machine learning include instructions for providing a target image to the unsupervised machine learning algorithm.

18. 11. The computer-readable storage medium of claim 10, wherein the instructions for clustering the repeater defects using unsupervised machine learning comprise instructions for using an unsupervised machine learning algorithm that does not specify a predetermined number of clusters.

19. 1. A system comprising: An optical inspection tool; one or more processors; A memory storing one or more programs for execution by one or more processors, the one or more programs comprising: acquiring optical inspection results of the one or more semiconductor wafers from the optical inspection tool; identifying a plurality of defects on the one or more semiconductor wafers based on results of the optical inspection; classifying, among the plurality of defects, defects having identical die locations on multiple dies of the one or more semiconductor wafers as repeater defects; clustering the repeater defects into a plurality of clusters based on the optical inspection results using unsupervised machine learning; scoring the repeater defects by assigning a respective score to each repeater defect based on the degree to which a cluster in the plurality of clusters contains multiple instances of the respective repeater defect; ranking the repeater defects based on their respective scores; a memory containing instructions for: A system comprising:

20. The one or more programs include selecting the repeater defects having respective scores that satisfy a threshold; generating a report identifying the selected repeater defects; 20. The system of claim 19, comprising the instructions:

21. 20. The system of claim 19, wherein the instructions for classifying comprise instructions for classifying defects of the plurality of defects that have a same reticle level location on the one or more semiconductor wafers as the repeater defects.

22. 20. The system of claim 19, wherein each of the scores corresponds to a maximum percentage of instances of the respective repeater defect in a cluster of the plurality of clusters.

23. 20. The system of claim 19, wherein each of the scores corresponds to an average of a non-zero percentage of instances of the respective repeater defect in the plurality of clusters.

24. The one or more programs include further comprising instructions for extracting defect attributes of the repeater defects from the optical inspection results; 20. The system of claim 19, wherein the instructions for using unsupervised machine learning include instructions for clustering the repeater defects using defect attributes.

25. 25. The system of claim 24, wherein the defect attributes include defect attributes selected from the group consisting of size, die location, wafer location, and light intensity.

26. the optical inspection results include target images of dies on the one or more semiconductor wafers and difference images of dies on the one or more semiconductor wafers, each difference image being a difference between a respective target image and a reference image; the instructions for identifying include instructions for identifying the defects in the difference image; 20. The system of claim 19, wherein the instructions for clustering the repeater defects using unsupervised machine learning include instructions for providing a target image to the unsupervised machine learning algorithm.

27. 20. The system of claim 19, wherein the instructions for clustering the repeater defects using unsupervised machine learning include instructions for using an unsupervised machine learning algorithm that does not specify a predetermined number of clusters.

Images