Wafer defect analysis method and device, equipment and storage medium

Abstract

The application relates to the field of semiconductor manufacturing, and particularly relates to a wafer defect analysis method and device, equipment and a storage medium. The method comprises the following steps: acquiring a plurality of defects on a wafer surface and position information of each defect; clustering and grouping the plurality of defects based on the position information of each defect, extracting a plurality of special defect clusters with special morphologies; including the remaining defects not belonging to any special defect cluster into a sporadic defect cluster; for each defect cluster, determining a corresponding image sampling number according to the type of the defect cluster, and individually sampling and collecting images of the defects in the defect cluster based on the image sampling number; performing defect type identification based on the collected images, and individually performing generalization classification on the defect cluster according to the identification result. The hierarchical, directional and interpretable intelligent sampling mechanism with the defect cluster as a basic sampling unit improves the accuracy and representativeness of sampling under limited detection resources, and further improves the efficiency of generalizing the defect types in the subsequent wafer.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing, and more particularly to a method, apparatus, device, and storage medium for analyzing wafer defects. Background Technology

[0002] In semiconductor manufacturing, the detection and analysis of wafer surface defects is a crucial step in ensuring chip yield. As process nodes continue to shrink to the nanometer or even angstrom level, the impact of tiny defects on device electrical performance, reliability, and yield is becoming increasingly significant. This places unprecedented demands on the accuracy of defect identification, the precision of classification, and the ability to diagnose root causes.

[0003] Currently, after a wafer completes one or more key process steps (such as photolithography, etching, thin film deposition, or chemical mechanical polishing), a full-wafer scan of the wafer surface is typically performed using high-precision optical inspection equipment to identify potential abnormal areas or suspected defects. Because the number of suspected defects marked by the inspection system is enormous, and subsequent manual verification using electron microscopy or high-resolution imaging is costly and time-consuming, the industry generally adopts a sampling strategy—that is, selecting a small number of representative samples from a massive number of suspected defects for photographing and manual confirmation to infer the overall defect distribution characteristics and the root cause of process problems.

[0004] For example, patent application CN117197617A discloses a method and system for classifying repetitive defects. The method includes: acquiring at least one set of location information for repetitive defects on a wafer; acquiring the total number of samples and the sampling ratio corresponding to sampling locations on the wafer, so as to determine the number of samples in different sampling regions of the wafer from multiple pre-divided sampling regions according to the total number of samples and the sampling ratio; selecting corresponding sampling locations from the multiple sampling regions according to the number of samples in each region; taking pictures at the selected sampling locations to obtain a set of defect photos corresponding to at least one set of repetitive defects; and classifying the repetitive defects using the set of defect photos.

[0005] For example, patent application CN 119601483A discloses a wafer defect detection method and apparatus, belonging to the field of semiconductor manufacturing technology. The wafer defect detection method includes: acquiring defect data on the surface of each wafer, the defect data including defect type and defect location coordinates; performing cluster analysis on the defect location coordinates of wafers in the same batch using a spatial clustering algorithm to determine the defect distribution pattern and defect distribution density of the batch of wafers, wherein the wafers in the same batch use the same production process parameters; calculating a risk score for the batch of wafers based on the defect distribution pattern, defect distribution density, and the proportion of abnormal wafers in the batch; comparing the risk score of the batch of wafers with a preset risk threshold; if the risk score of the batch of wafers is greater than the preset risk threshold, the batch of wafers is determined to be abnormal wafers.

[0006] However, existing sampling methods often struggle to balance accuracy and representativeness with limited review costs, which can easily lead to distorted generalization results or abnormal underreporting. Summary of the Invention

[0007] The main objective of this application is to provide a method, apparatus, device, and storage medium for analyzing wafer defects. To solve the aforementioned technical problems, this application specifically adopts the following technical solution: A first aspect of this application is to provide a method for analyzing wafer defects, the method comprising: S21, acquire multiple defects on the wafer surface and the location information of each defect; S22, based on the location information of each defect, the multiple defects are clustered and extracted to extract several special defect clusters with special shapes; S23, Incorporate the remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; S24. For each defect cluster, determine the corresponding image sampling quantity according to the type of the defect cluster, and perform individual sampling and image acquisition of defects within the defect cluster based on the image sampling quantity. S25, based on the acquired images, defect type identification is performed, and the defect clusters are individually generalized and classified according to the identification results.

[0008] In some embodiments, the special shape includes at least one of the following: straight line, arc, ring, fan shape, block shape, wavy shape, and a combination of at least two identical or different shapes.

[0009] In some embodiments, the special defect clusters include single-type defect clusters and composite defect clusters; wherein, the morphology of the single-type defect cluster corresponds to a unique process root cause, and the morphology of the composite defect cluster corresponds to multiple process root causes; S24 includes: allocating a first sampling quantity to the composite defect clusters, allocating a second sampling quantity to the single-type defect clusters, and the first sampling quantity is greater than the second sampling quantity; allocating a third sampling quantity to the sporadic defect clusters, and the third sampling quantity is greater than the first sampling quantity.

[0010] In some embodiments, S24 includes: S241, calculating the remaining sampling quantity based on the required sampling quantity for each special defect cluster and a preset upper limit; S242, using the remaining sampling quantity as the image sampling quantity for the sporadic defect cluster.

[0011] In some embodiments, the method further includes: when the morphology of any single-type defect cluster matches a preset morphology standard with a higher degree of matching than a preset matching threshold, and / or when the remaining sampling quantity for sporadic defect clusters is less than a preset minimum sampling quantity, setting the second sampling quantity corresponding to the single-type defect cluster to zero; updating the sampling quantity required for each special defect cluster, and recalculating the remaining sampling quantity.

[0012] In some embodiments, the method further includes: if the identification result of a single type defect cluster is inconsistent with the preset unique process root cause, then the single type defect cluster is reclassified as a composite defect cluster, and steps S24 to S25 are re-executed.

[0013] In some embodiments, the method further includes: acquiring a plurality of special defect cluster samples, each of the special defect cluster samples including a wafer defect distribution map and a defect cluster type label; training an image segmentation model based on the special defect cluster samples, the image segmentation model being used to identify and output special defect clusters from the input wafer defect distribution map.

[0014] A second aspect of this application is to provide an apparatus for analyzing wafer defects, the apparatus comprising: The defect acquisition module is used to acquire multiple defects on the wafer surface and the location information of each defect; The clustering module is used to cluster the multiple defects based on the location information of each defect, and extract several special defect clusters with special shapes. The sporadic defect module is used to include remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; The sampling acquisition module is used to determine the corresponding number of image samples for each defect cluster according to the type of the defect cluster, and to perform individual sampling and image acquisition of defects within the defect cluster based on the number of image samples. The defect identification module is used to identify the type of defect based on the acquired images and to perform generalization classification on the defect clusters individually according to the identification results.

[0015] A third aspect of this application is to provide a computer device, the device comprising: Memory, used to store computer programs; A processor is configured to execute the computer program and, in executing the computer program, implement the steps of the wafer defect analysis method provided in any embodiment of this application.

[0016] A fourth aspect of this application is that a computer-readable storage medium is also provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the wafer defect analysis method provided in any embodiment of this application.

[0017] Beneficial technical effects: This application provides a method, apparatus, device, and storage medium for analyzing wafer defects. Specifically, it provides a hierarchical, directional, and interpretable intelligent sampling mechanism with defect clusters as the basic sampling unit. Under limited review resources, it significantly improves the accuracy and representativeness of sampling, thereby improving the efficiency of subsequent generalization of defect types in wafers.

[0018] First, clustering is used to accurately extract special morphological defect clusters with geometric shapes, directions, or symmetries, separating systematic anomalies from massive random defects. At the same time, unmatched remaining defects are uniformly classified into sporadic defect clusters.

[0019] Based on this, each defect cluster serves as an independent sampling unit. The sampling strategy has been changed from the traditional "random selection on the entire wafer" to "balanced coverage by cluster". Even rare special-shaped defects can be included in the sampling range as a whole, fundamentally avoiding the risk of missing key anomalies due to the randomness of global sampling, and also avoiding the risk of re-examination.

[0020] Furthermore, a dynamic resource scheduling mechanism is introduced to allocate sampling quantities differently based on the diagnostic needs of various defect clusters, thereby achieving refined allocation of review resources: a higher sampling quantity is allocated to complex defect clusters with multiple causes to support multi-root cause discrimination; a lower sampling quantity is allocated to single-type defect clusters with clear causes, and even skipping the review directly under high confidence; while the highest sampling quantity is allocated to sporadic defect clusters that require statistical modeling to ensure the comprehensive representativeness of spatial distribution.

[0021] In addition, to maximize resource utilization, resources are first allocated to special defect clusters, and then the remaining quota is used for sporadic defect clusters. When resources are extremely scarce, single-type defect clusters with relatively certain characteristics can be skipped intelligently, and the remaining quota is updated in real time. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of this application; for those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0023] Figure 1 This is a schematic flowchart illustrating a wafer defect identification method provided in an embodiment of this application; Figure 2 This is a schematic diagram of a defect cluster identification result and a traditional defect sampling result provided in an embodiment of this application; Figure 3 This is a schematic flowchart of a wafer defect analysis method provided in an embodiment of this application; Figure 4 This is a schematic diagram of another defect sampling result provided in an embodiment of this application; Figure 5 This is a schematic diagram of another defect sampling result provided in an embodiment of this application; Figure 6 This is a schematic block diagram of a wafer defect identification device provided in an embodiment of this application; Figure 7 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation. Furthermore, although functional modules are divided in the device diagram, in some cases, a different module division may be used.

[0026] In this document, suffixes such as “module,” “part,” or “unit” used to denote elements are used only for illustrative purposes and have no specific meaning in themselves. Therefore, “module,” “part,” or “unit” may be used interchangeably.

[0027] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0028] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," and "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0029] In this document, the term “and / or” includes any and all combinations of one or more of the listed related items.

[0030] In this article, the term "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0031] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0032] In this article, a wafer refers to a circular silicon substrate (or other semiconductor material substrate, such as germanium, silicon carbide, etc.) used for semiconductor manufacturing. After one or more process steps (such as photolithography, etching, thin film deposition, ion implantation, etc.), various physical or electrical defects may be generated on its surface.

[0033] In some embodiments, multiple defects on the wafer surface can be identified as a type of defect cluster with a special morphology and / or a type of defect cluster with an aggregated state by different clustering analysis rules.

[0034] One type of defect cluster is often closely related to the component structure, operating mechanism, or aging state of manufacturing equipment, and its spatial distribution often exhibits an interpretable geometric shape associated with these factors. For example, the surface of the jigs (or carriers) used for wafer transfer or support is usually provided with an anti-wear layer (such as a silicone-based elastic material). With increased use, this material gradually wears down, hardens, or loses its elasticity, causing periodic physical scratches on the front and / or back sides of the wafer during the wafer clamping process, which appears as two parallel straight-line defect clusters on the defect distribution map. As another example, in the chemical mechanical polishing (CMP) process, if micron-sized particles or dust are mixed into the polishing pad or slurry, these foreign objects may scratch the surface along the trajectory during wafer rotation, forming arc-shaped defect clusters with a fixed curvature.

[0035] In some embodiments, based on historical wafer inspection data, defect cluster samples related to one or more specific morphologies are collected periodically. By performing statistical analysis and pattern summarization on the defect cluster samples, a corresponding preset morphology standard can be established for each specific morphology. The defect cluster samples include wafer defect distribution maps of the specific morphology defect clusters and their corresponding high-resolution inspection images, labeled with their cause tags (such as "CMP particle contamination," "wafer fixture aging scratches," etc.) or process context information (such as equipment number, process step, batch number, chamber status, or equipment sensor logs, etc.).

[0036] In some embodiments, the special morphology includes at least one of the following: straight line, arc, ring, fan shape, block shape, wavy shape, and a combination of at least two identical or different morphologies. Each morphology corresponds to a set of preset morphology criteria, used to identify candidate defect clusters during the primary clustering process.

[0037] In some embodiments, the preset morphological standard includes at least one of geometry, orientation, or symmetry.

[0038] In this context, geometric shape refers to the mathematical or graphical characteristics of the overall outline or arrangement of defects within a defect cluster. For example, the graphic of a ring-shaped defect cluster is approximately circular or elliptical, with its defects distributed along a curve centered on a certain point; the graphic of a block-shaped defect cluster is a locally dense area with blurred boundaries; and the graphic of a wavy defect cluster is a periodically changing curved shape.

[0039] Directionality refers to the characteristic that defects within a defect cluster are arranged in an orderly manner along a certain dominant axis. For example, in a linear defect cluster, all defects are arranged collinearly and have a clear directionality; in an arc-shaped defect cluster, the defects are arranged along an arc with a specific radius of curvature.

[0040] Symmetry refers to the characteristic that defects within a defect cluster are arranged in a mirror or rotationally symmetrical manner about a point, line, or surface in space. For example, the defects in a fan-shaped defect cluster are arranged radially symmetrically with a certain point as the vertex.

[0041] In semiconductor manufacturing, a type of defect cluster is relatively few in number and exhibits a regular spatial arrangement. In some cases, specific morphologies are strongly correlated with specific process steps or equipment conditions, making them highly identifiable and diagnostically certain. They can serve as early warning signals for early equipment degradation or process drift. It should be understood that some defects caused by equipment aging have fixed patterns. For example, after the anti-wear layer of a wafer jig hardens, two parallel straight-line defects may be generated on the front or back of the wafer during clamping. Therefore, the root cause can be inferred from the straight-line defects, which may be "jig aging".

[0042] In contrast, type II defect clusters are often caused by sudden or localized process anomalies, manifesting as localized areas of high defect density. For example, in photolithography, uneven photoresist coating may lead to dense, spot-like defects in localized areas; during etching or cleaning steps, particles or residual contaminants detached from the chamber may randomly adhere to the wafer surface, forming irregular clusters.

[0043] The high density of type II defect clusters indicates a quality risk on the production line that requires immediate intervention. Root cause determination relies on manual visual inspection, making it difficult to directly deduce the specific cause or location of the fault. Furthermore, current technologies mostly focus on identifying type II defect clusters, relying on density or quantity thresholds for clustering and alarms. For "scattered but orderly" defects occurring in the early stages of equipment aging, they are often unrecognizable or misjudged as random noise and ignored, leading to missed opportunities for preventative maintenance as process anomalies cannot be addressed within the optimal window.

[0044] Based on this, embodiments of this application provide a method, device, and storage medium for identifying wafer defects. Specifically, it provides an intelligent defect diagnosis scheme based on a multi-level clustering architecture and a differentiated response strategy, which significantly improves the accuracy of identifying early, weak, and systematic defects while effectively suppressing false alarms or missed alarms caused by a single criterion, thus balancing identification sensitivity and response accuracy.

[0045] Furthermore, embodiments of this application also provide a method, apparatus, device, and storage medium for analyzing wafer defects. Specifically, it provides a hierarchical, directional, and interpretable intelligent sampling mechanism with defect clusters as the basic sampling unit. Under limited detection resources, it significantly improves the accuracy and representativeness of sampling, thereby improving the efficiency of subsequent generalization of defect types in wafers.

[0046] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0047] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating a wafer defect identification method provided in an embodiment of this application, such as... Figure 1 As shown in the figure, this application provides a method for identifying wafer defects.

[0048] S11, acquire multiple defects on the wafer surface and the location information of each defect.

[0049] Specifically, a full-wafer scan is performed on the surface of a wafer that has completed one or more process steps using semiconductor optical inspection equipment to obtain inspection results containing multiple anomalies (i.e., defects). In actual manufacturing scenarios, the number of defects on a single wafer typically ranges from tens to thousands, and their distribution exhibits high randomness and local clustering.

[0050] The precise location coordinates of each defect in the wafer's global coordinate system are extracted from the detection results. For example, rectangular coordinates with the wafer center as the origin are used as the location information of each defect. The location information of all defects together constitutes a complete defect location information set, which is used for subsequent analysis steps such as spatial distance calculation, clustering, and special shape recognition.

[0051] S12, based on the location information of each defect, perform first-level clustering on the multiple defects and extract several defect clusters with special shapes.

[0052] In some embodiments, in order to effectively identify a small number of but ordered sporadic defects and improve early sensitivity to process disturbances, first-level clustering is used to fuse morphological priors and discreteness comparisons. Each defect cluster of a certain type simultaneously satisfies the following conditions: (a) The spatial arrangement of defects within the cluster conforms to the preset morphological standard.

[0053] That is, the defect distribution within each defect cluster conforms to the preset morphological standard conditions, such as a straight line or a circle with a specific radius.

[0054] (b) For any target defect within a cluster, the intra-cluster dispersion of the target defect and defects in the same cluster within the first neighborhood is less than the background dispersion of defects in the non-cluster within the second neighborhood, and the difference between the intra-cluster dispersion and the background dispersion is greater than or equal to a preset dispersion difference value.

[0055] The dispersion is calculated using location information (such as coordinates) to determine the average Euclidean distance between a defect and other defects within a certain range (such as a first neighborhood range and a second neighborhood range). A smaller value indicates a more dense clustering of defects within the range. Correspondingly, the preset dispersion difference is set according to the process noise level, such as 30-100 micrometers.

[0056] That is, by comparing the intra-cluster dispersion of the target defect in the first neighborhood with the background dispersion in the second neighborhood, dispersion verification is achieved, so that a class of defect clusters meets the dispersion condition of "internal compactness and external sparseness", avoiding misjudging high-density randomly clustered defects as a class of defect clusters.

[0057] In some embodiments, the first neighborhood range refers to a local region (e.g., circular or approximately circular) defined in the wafer plane coordinate system centered on a target defect within the cluster, used to evaluate the distribution density relationship between the target defect and its neighboring defects in the same cluster. Specific values ​​can be set according to the structural characteristics of typical special morphologies. For example, for linear defects, the radius r1 of the first neighborhood range can be set to 1.5 times the average spacing between adjacent defects in the same cluster to ensure effective coverage of adjacent defect points within the cluster.

[0058] Correspondingly, the second neighborhood is also a local region (e.g., circular or approximately circular) defined in the wafer plane coordinate system centered on the target defect. It is used to evaluate the distribution density between the target defect and surrounding non-cluster background defects (such as random noise). Specific values ​​can be set according to process noise levels, wafer process nodes, or historical defect statistics.

[0059] In some embodiments, the second neighborhood range is greater than or equal to the first neighborhood range, so that the second neighborhood range is large enough to reflect the typical density characteristics of local background noise and avoids introducing interference caused by long-distance irrelevant defects.

[0060] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating a defect cluster identification result and a traditional defect sampling result provided in an embodiment of this application. For example... Figure 2As shown in the figure, the inspection results of a wafer (or a local area of ​​a wafer) are illustrated. Multiple marked points (including red and blue markers) represent detected defects. Based on the location information of each defect, multiple defects are clustered using a first-level clustering method. The figure exemplarily shows two extracted arc-shaped defect clusters: E and F. Defect cluster E presents as a relatively long arc with a small number of sparsely distributed defects. Its incomplete shape and poor continuity indicate that the anomaly is in an early development stage, possibly caused by slight equipment degradation, and has not yet formed a significant cluster. Defect cluster F, on the other hand, appears as a shorter arc with densely packed and regularly arranged defects. Its complete shape and strong spatial continuity indicate that the anomaly has entered a mature stage, and the degree of equipment degradation is relatively high.

[0061] In some embodiments, the region where the defect cluster is located is dynamically determined based on the spatial distribution of all defects within the cluster. This region will be used for overlap determination, dispersion calculation, etc., in the embodiments of this application. For example, the region where the defect cluster is located is determined based on the coordinates of the defects within the cluster, using a minimum bounding rectangle or polygonal contour, or a density-based boundary fitting method. Figure 2 As shown, the area where the defect cluster is located is the area filled with light green within the red box.

[0062] In some embodiments, first-level clustering can be achieved by a pre-trained image segmentation model. The method further includes: acquiring multiple first-class defect cluster samples, each first-class defect cluster sample including a wafer defect distribution map and a defect cluster type label; training an image segmentation model based on the first-class defect cluster samples, the image segmentation model being used to output defect clusters that conform to the preset morphology standard and dispersion condition from the input wafer defect distribution map.

[0063] Specifically, defect distribution maps from multiple historical wafers are collected as training samples. Each sample contains the location coordinates of all defects on the wafer, and process experts label defect clusters with special shapes with corresponding shape category labels (such as straight lines, arcs, rings, fan shapes, etc.), thus constructing a training dataset (i.e., a sample of a type of defect cluster). Based on the training dataset, an image segmentation model (e.g., a convolutional neural network CNN or U-Net architecture) is trained, enabling the model to automatically learn and master the preset shape standards corresponding to each special shape from the input defect distribution maps, including its geometric shape, orientation, symmetry, and spatial arrangement. In actual analysis, the defect distribution map of a new wafer is input into the trained image segmentation model, and the model directly extracts a type of defect cluster with special shapes.

[0064] In some embodiments, the defect location data is geometrically fitted based on preset morphological standards (such as straight lines, rings, fan shapes, etc.), for example, using RANSAC (Random Sample Consensus) or Hough transform algorithms, to detect whether there is an ordered arrangement conforming to a specific geometric model. Only when the fitting residual is below a set threshold (e.g., 0.5 micrometers) and the number of defects within a cluster meets a minimum size requirement (e.g., no less than 3 points) is the set retained as a candidate first-class defect cluster. For example, when 10 defect points are highly collinear along the 0° direction and the distance from each point to the fitted line is less than 0.5 μm, it is identified as a linear candidate first-class defect cluster, thereby effectively capturing sparse but ordered defect patterns caused by early equipment anomalies and avoiding missed detection by traditional density clustering methods due to their small number.

[0065] Furthermore, the dispersion within the candidate first-class defect cluster and the background dispersion are calculated. Candidate first-class defect clusters that meet the dispersion criteria are output as valid first-class defect clusters. The first neighborhood range is defined as the region where the candidate first-class defect cluster is located, used to evaluate the density of defect distribution within the cluster. The second neighborhood range is defined as a larger region formed by extending the first neighborhood range outward by a preset distance (e.g., expanding by 1.5 to 2 times the width), used to evaluate the density of background defects surrounding the cluster. By comparing the two, the extracted first-class defect clusters not only conform to morphological norms but are also significantly distinguishable from random noise in local space.

[0066] In some embodiments, an iterative clustering strategy based on seed growth and morphology guidance is employed to more accurately construct a class of defect clusters. S12 further includes S121 to S123.

[0067] S121, select an initial seed defect from the plurality of defects, and determine the candidate expansion direction or topology according to the preset morphological standard.

[0068] Specifically, an initial seed defect is randomly selected from multiple defects, and its possible expansion direction or topology is inferred based on preset morphological criteria (such as linear, cyclic, etc.). For example, if the seed is located at the starting point of a suspected straight line, the candidate expansion direction is along the positive and negative directions of that line; if it is the starting point of a cyclic shape, the topology is a circular path around a certain center point.

[0069] S122, along the candidate expansion direction or topology, search for candidate defects to be added within a third neighborhood. The third neighborhood refers to a local search area defined in the wafer plane coordinate system centered on the end defect or seed defect of the current cluster, and can be set according to the expected extension step size for a specific shape.

[0070] S123, Perform discreteness verification for each candidate defect, including: Based on the location information, the intra-cluster dispersion d is calculated. in Background dispersion d out ; If d is satisfied in <d out And d in -d out If the difference is greater than or equal to α, where α is a preset dispersion difference, then the candidate defect is included in the current defect cluster, and the process returns to step S121 to continue expansion (e.g., returning to S121 to continue growth with the newly added defect as the starting point). If there are no candidate defects that meet the conditions on the current expansion path, the expansion of the defect cluster is terminated so that the final defect cluster not only conforms to the morphological requirements, but is also significantly different from random background noise in the local space.

[0071] In some embodiments, the method further includes: S13, performing secondary clustering on the plurality of defects to generate several clusters of two types of defects with aggregated states.

[0072] In some embodiments, the generation of two types of defect clusters must simultaneously satisfy the conditions of quantity and spatial compactness. Each type of defect cluster satisfies the following conditions: (c) the number of defects in the cluster is greater than or equal to a preset minimum number; (d) the spatial distance between any two defects in the cluster does not exceed a preset neighborhood radius.

[0073] The preset minimum number and preset neighborhood radius can be dynamically set according to the physical characteristics of the wafer manufacturing process, process noise, process node, spatial resolution of the inspection equipment, and historical defect distribution data. For example, the preset minimum number can be set to 5 to 20 defects, and the preset neighborhood radius can be set to 50 to 300 micrometers, so that the second type of defect cluster has statistical and analytical value.

[0074] In some embodiments, the implementation of secondary clustering can refer to relevant density clustering algorithms (such as DBSCAN, OPTICS, or adaptive grid clustering), which will not be elaborated here.

[0075] It should be understood that primary clustering and secondary clustering are two parallel clustering processes with different objectives: primary clustering focuses on morphological structure and adopts a broader dispersion standard (such as the dispersion within a cluster being significantly lower than the background dispersion). Through geometric fitting or model matching, it identifies ordered defect clusters that conform to a preset morphological standard (such as straight lines, rings, fan shapes, etc.) in spatial arrangement. Secondary clustering, on the other hand, focuses on defect density and adopts a stricter dispersion standard (such as the spatial distance between any two defects within a cluster not exceeding a preset neighborhood radius). It detects the aggregation of local regions and does not involve any morphological compliance verification.

[0076] Therefore, the same defect may belong to both a Class I defect cluster (due to its geometrically regular arrangement) and a Class II defect cluster (due to its high defect density in a local area), or it may only meet one of the criteria and thus belong to either a Class I or Class II defect cluster. This multi-perspective, multi-criteria clustering mechanism simultaneously identifies both early, weak but ordered systematic anomalies (such as sparse defect clusters generated in the early stages of equipment aging) and sudden, high-density process disturbances (such as clumping caused by particulate contamination), thereby achieving a more comprehensive diagnosis of wafer defects and effectively avoiding missed detections of critical anomalies or false alarms due to reliance on a single clustering strategy.

[0077] like Figure 2 As shown, defect cluster E has low local density and high dispersion, which does not meet the stricter dispersion criteria used in secondary clustering. Therefore, it cannot be identified as a second-class defect cluster by secondary clustering. Primary clustering uses a broader dispersion criterion and a preset morphological criterion (such as arc shape) as its core. Defect cluster E can still be successfully extracted as a first-class defect cluster, thus capturing the weak signals in the early stages of equipment degradation. In contrast, defect cluster F has dense defects and regular arrangement, which meets the morphological criteria of primary clustering and may also meet the dispersion criteria of secondary clustering. It may be identified by both primary and secondary clustering, forming an overlapping cluster.

[0078] In some embodiments, the spatial overlap between a first-class defect cluster and a second-class defect cluster is analyzed to determine the classification of each defect cluster, thereby adapting the corresponding response strategy. The method further includes: S14. Based on the spatial relationship between Class I defect clusters and Class II defect clusters, determine the classification type of each defect cluster. The classification types include: independent Class I clusters, independent Class II clusters, and overlapping clusters. S15, based on the type of defect cluster, invoke the corresponding defect response strategy to respond to the defect cluster.

[0079] Specifically, if a certain type of defect cluster has no intersection with any type of defect cluster, it is classified as an independent type of cluster; if a certain type of defect cluster has no intersection with any type of defect cluster, it is marked as an independent type of cluster; if a certain type of defect cluster has an intersection with a certain type of defect cluster, they are merged and determined to be overlapping clusters.

[0080] In some embodiments, the existence of overlap is determined based on the number of shared defects and / or the region overlap rate among defect clusters. For example, if the number of shared defects is less than a preset sharing threshold (e.g., 2) and / or the region overlap rate is less than a preset overlap rate (e.g., 10%), no overlap is determined; if the number of shared defects is greater than or equal to the preset sharing threshold and / or the region overlap rate is not less than the preset overlap rate, overlap is determined. The specific values ​​of the preset sharing threshold and preset overlap rate can be flexibly set according to process noise levels, process nodes, engineering experience, etc.

[0081] It should be understood that defects in independent Class I clusters only meet specific morphological criteria and do not fall into any high-density areas. They usually reflect sporadic, orderly, but not yet large-scale early anomalies that occur in the early stages of equipment aging. Defects in independent Class II clusters only meet local high-density aggregation conditions and lack structured arrangement, corresponding to sudden process disturbances (such as particulate contamination). Overlapping clusters, on the other hand, may have both ordered morphology and high-density characteristics in some or all of their parts, which may correspond to severe or complex process anomalies, such as the existence of multiple process root causes.

[0082] Thanks to the broad dispersion standard employed in primary clustering, it can effectively capture sporadic, incomplete, but directional or structured early-stage weak anomalies generated in the early stages of equipment aging. Secondary clustering, on the other hand, uses a stricter dispersion standard to identify mature cluster defects that have developed to a certain scale and have high local density. By fusing the results of the two clustering methods and determining their classification, the development stage, physical cause, and risk level of the anomaly can be accurately distinguished.

[0083] Based on this, differentiated response strategies are invoked as needed. For example, predictive maintenance is initiated for early morphological anomalies (i.e., independent type 1 clusters) before yield loss; emergency review and process interception are triggered for high-density clusters (i.e., independent type 2 clusters) to prevent the spread of batch defects; and root cause fusion analysis and resource priority scheduling are performed for complex risk areas (i.e., overlapping clusters).

[0084] Under the constraint of limited detection resources, the system achieves synergistic optimization of early warning sensitivity and high-risk response accuracy, overcoming the problems of missed detection, false alarms or response lag caused by traditional methods relying on single density or morphology criteria, and significantly improving the comprehensiveness and practicality of wafer defect diagnosis.

[0085] In some embodiments, S15 includes: S151, for an independent cluster, monitoring the defect change trend of corresponding regions in multiple consecutive wafers under the same process step; determining the expansion direction and rate of process anomalies based on the defect change trend to generate observation-level alerts.

[0086] In some embodiments, the defect change trend includes at least one of spatial location change (such as the offset or expansion of the defect spatial location), quantity change (i.e., the increase or decrease in the number of defects, such as from 3 to 8), or density change (i.e., the increase or decrease in the distribution density of defects in the region).

[0087] Specifically, continuous monitoring of defect evolution in a region (e.g., near a 30° azimuth angle at the wafer edge) of a single cluster within multiple wafers (e.g., from two consecutive batches) under the same process steps determines defect trends. Analyzing these trends allows for the inference of the expansion direction of process anomalies (e.g., diffusion from the wafer edge to the center) and the rate of deterioration (e.g., two new defect points per wafer), thereby assessing the degree of equipment degradation or process drift. Based on this, observation-level alerts are generated, notifying engineers to enhance monitoring, schedule preventative maintenance, or adjust process parameters, rather than immediately shutting down the system. This effectively balances yield assurance and production efficiency, avoiding production line interruptions due to over-responding to early, weak signals, while ensuring a sufficient intervention window before anomalies escalate into serious problems.

[0088] In some embodiments, S15 includes: S152, for independent binary clusters, outputting a review-level prompt to trigger at least one of the following inspection processes: operating environment, material batch, or transport path, for targeted investigation of potential process disturbance sources. It should be understood that since independent binary clusters exhibit localized high-density aggregation but lack specific geometric morphology, their causes are usually related to sudden, non-systematic anomalies, such as a sudden increase in cleanroom particulate concentration, impurities in material batches such as photoresist or cleaning fluid, contamination of the wafer transport robotic arm, or scratches caused by path deviation. Therefore, the review-level prompt does not directly point to equipment aging or structural failures, but rather guides engineers to review variable environmental and consumable factors.

[0089] In some embodiments, the review-level prompt can be associated with the process chamber number, material batch number, transport path record, and environmental monitoring log of the current wafer, and recommend performing at least one inspection process, such as checking whether the particle count in the working environment exceeds the standard, verifying whether the material batch has passed quality access, or checking whether there are foreign objects or positioning deviations on the transport track.

[0090] In some embodiments, S15 includes: for overlapping clusters, determining overlapping regions within the cluster that simultaneously belong to both a type I defect cluster and a type II defect cluster; calculating the degree of defect identification overlap based on the number of defects in the overlapping regions (also known as the number of shared defects), and determining the predicted deterioration risk level of the current special morphology defect; if the deterioration risk level reaches a preset intervention threshold, generating an intervention level prompt to trigger at least one intervention process in equipment calibration or preventive maintenance; if the deterioration risk level does not reach the preset intervention threshold, executing S151 or S152 based on the type I defect cluster and the type II defect cluster in the overlapping clusters respectively.

[0091] Specifically, defects in the overlapping region possess both ordered morphology and high density characteristics. Based on the number of defects in the overlapping region and / or their proportion to the total size of the first-class defect cluster and the total size of the second-class defect cluster, the degree of defect identification overlap can be calculated to determine the development stage of the anomaly (such as early, middle or late stage) and the level of deterioration risk (such as low, medium or high).

[0092] For example, if the degree of overlap is high (such as more than 80% of Class I defect points falling into Class II defect clusters), on the one hand, it may be caused by the continuous deterioration of a single equipment aging problem. The anomaly has evolved from early sporadic and orderly weak signals to high-density clustered events, which is in the middle and late stages and has a high risk of deterioration. On the other hand, the defects in the overlapping clusters may be related to multiple process steps, which indicates that the cause is complex and there are multiple root causes. It belongs to a composite anomaly with a high risk of deterioration, and it is necessary to initiate multi-dimensional root cause analysis and cross-module collaborative investigation.

[0093] For example, if the degree of overlap is low (such as only 20% of the first-class defect points falling into the second-class defect cluster), the anomaly is still in the early stage. The morphology has begun to appear but has not yet caused significant aggregation. This corresponds to slight equipment degradation or slight drift in process parameters, and the risk level of deterioration is low.

[0094] When the risk level of deterioration reaches the preset intervention threshold (e.g., high risk level) and / or the development stage reaches the preset stage (e.g., mid-to-late stage), an intervention-level prompt is automatically generated, and the corresponding proactive response process is triggered, including but not limited to equipment parameter calibration, key component replacement, chamber cleaning, or arranging preventive maintenance, in order to suppress the further spread of the anomaly and ensure the stability and yield consistency of the production line.

[0095] If the risk level of deterioration does not reach the preset intervention threshold and / or the development stage does not reach the preset stage, and if the development stage is still in the early stage, then the standard response strategies for Class I and Class II defect clusters are applied independently. For example, for the Class I defect cluster components, S151 is executed, supporting the output of observation-level prompts; for the Class II defect cluster components, S152 is executed, supporting the output of review-level prompts.

[0096] It should be understood that the embodiments of this application construct a differentiated response mechanism through review-level prompts, observation-level prompts, and intervention-level prompts. Observation-level prompts are for independent clusters of type I and are used to capture early degradation trends of equipment or processes (such as sparse linear defects caused by fixture aging). Their core function is continuous monitoring and early warning. They have low timeliness requirements, usually do not require immediate manual intervention, and do not interfere with normal production processes. Review-level prompts are for independent clusters of type II and are aimed at sudden and occasional disturbances (such as cleanroom particulate contamination, abnormal material batches, or deviations in conveying paths). They emphasize rapid investigation and source isolation, and require preliminary verification to be completed within a few hours. They have high timeliness but generally do not trigger shutdowns. Intervention-level prompts are used for high-risk scenarios that may exist in overlapping clusters. They require proactive scheduling of human and equipment resources to perform equipment calibration, replacement of key components, or preventive maintenance. If necessary, the relevant process modules may be suspended.

[0097] This tiered mechanism effectively avoids the problems of low-risk false alarms or high-risk omissions caused by traditional "one-size-fits-all" alarms, greatly improves the reliability of alarms, and avoids excessive intervention due to weak early signals.

[0098] In some embodiments, the method further includes: performing morphological similarity analysis on two types of defect clusters generated in multiple wafers; and generating a morphological update prompt when the morphological similarity of at least two types of defect clusters exceeds a preset similarity threshold.

[0099] Specifically, morphological similarity analysis is performed on the two types of defect clusters identified in multiple wafers (such as wafers in the same batch or in continuous production). When the morphological similarity of multiple two types of defect clusters is high and they appear repeatedly, it is speculated that they may originate from a new type of process anomaly or equipment failure that has not yet been covered by the existing morphological rule library. A morphological update prompt is automatically generated, triggering manual intervention for expert verification and annotation, thereby achieving the accuracy of identifying one type of defect cluster.

[0100] In some embodiments, if a certain type of defect cluster intersects with another type of defect cluster in space, a special combination state (such as linear and cyclic superposition, multi-directional straight line intersection, etc.) appears. Such a situation corresponds to more complex equipment failures or multi-source process disturbances, and there may be multiple process root causes, which are determined to be a composite type of defect cluster.

[0101] Please see Figure 3 , Figure 3 This is a schematic flowchart illustrating a wafer defect analysis method provided in an embodiment of this application, such as... Figure 3 As shown in the figure, this application provides a method for analyzing wafer defects.

[0102] S21, acquire multiple defects on the wafer surface and the location information of each defect.

[0103] S22, based on the location information of each defect, the multiple defects are clustered and several special defect clusters with special shapes are extracted.

[0104] In some embodiments, clustering is achieved through a pre-trained image segmentation model. The method further includes: acquiring multiple special defect cluster samples, each of the special defect cluster samples including a wafer defect distribution map and a defect cluster type label; training an image segmentation model based on the special defect cluster samples, the image segmentation model being used to identify and output special defect clusters from the input wafer defect distribution map.

[0105] Specifically, defect distribution maps from multiple historical wafers are collected as training samples. Each sample contains the location coordinates of all defects on the wafer, and process experts label defect clusters with special shapes with corresponding shape category labels (such as straight lines, arcs, rings, fan shapes, etc.), thus constructing a training dataset of special defect clusters (i.e., special defect cluster samples). Based on the training dataset, an image segmentation model (such as a convolutional neural network CNN or U-Net architecture) is trained, enabling the model to automatically learn and master the preset shape standard corresponding to each special shape from the input defect distribution maps. In actual analysis, the defect distribution map of a new wafer is input into the trained image segmentation model, and the model directly extracts special defect clusters with special shapes from it.

[0106] It should be understood that the labeling standards of the training samples directly affect the strictness of the trained image segmentation model in recognizing special defect clusters and the judgment conditions that it must meet. If the types of samples used for training are different, for example, there are differences between samples of a type of defect cluster and samples of special defect clusters, then the defect cluster results output by the same wafer after being processed by different models may also be different, thus enabling the recognition of a type of defect cluster and special defect clusters respectively.

[0107] In some embodiments, the clustering can be the first-level clustering in the foregoing embodiments, and correspondingly, the special defect cluster can be a type of defect cluster in the foregoing embodiments. In this case, each special defect cluster satisfies the following conditions: (a) the spatial arrangement of defects within the cluster conforms to a preset morphological standard; (b) for any target defect within the cluster, the intra-cluster dispersion of the target defect and defects in the same cluster within the first neighborhood is less than the background dispersion of defects in the same cluster within the second neighborhood, and the difference between the intra-cluster dispersion and the background dispersion is greater than or equal to a preset dispersion difference value.

[0108] At this point, all defect clusters in the training samples have undergone double verification, that is, they simultaneously satisfy conditions (a) and (b). The image segmentation model trained in this way will implicitly learn conditions (a) and (b) and inherit this double constraint during inference. Its output defect clusters have a low false alarm rate, belong to high confidence scenarios, and are highly suitable for scenarios that determine response strategies. This ensures that subsequent processing will not cause unnecessary intervention due to misjudgment, nor will it fail to respond in time due to missed judgment, thereby achieving accurate matching between response intensity and anomaly severity.

[0109] In other embodiments, the training samples focus on morphological diversity and coverage breadth, requiring only that defect clusters meet condition (a) without requiring the dispersion condition (b). The trained model then tends to more loosely identify and output special defect clusters with specific morphologies. While such identification results may contain a small number of false positives, their primary purpose is to construct sampling base units rather than directly triggering interventions, thus exhibiting higher fault tolerance. The falsely reported special defect clusters also contribute to achieving balanced sampling coverage, making them suitable for resource-rich scenarios.

[0110] In some embodiments, the special defect clusters include single-type defect clusters and composite defect clusters.

[0111] In some embodiments, the morphology of the single-type defect cluster corresponds to a unique process root cause, while the morphology of the composite defect cluster corresponds to multiple process root causes.

[0112] The specific morphology of a single type of defect cluster strongly corresponds to a single, underlying process cause. For example, given the current production line configuration, equipment status, and process window, if a linear defect has been verified by engineering to be caused solely by a fixture malfunction (i.e., other potential causes have been ruled out), the fixture malfunction can be inferred as the root cause. Furthermore, in high-confidence scenarios, the identification accuracy of specific defect clusters is relatively high, and the defect type and root cause can be directly inferred based on the morphological recognition results without relying on manual image review.

[0113] The special morphology of a complex defect cluster corresponds to the combined effect of multiple process factors, or is caused by any one of multiple potential process root causes under specific conditions. Its causes are ambiguous or coupled.

[0114] On the one hand, under different production line configurations, equipment conditions, or process windows, the same or similar special morphologies may be caused by a variety of independent process problems. For example, the same type of highly collinear defect arrangement may be caused by fixture wear, conveyor arm scratching, or foreign objects inside the chamber dragging on the wafer surface.

[0115] On the other hand, certain special shapes have been verified by engineering experience or historical data analysis as the result of the combined effects of multiple anomalies. Their overall form may be a combination of two or more superimposed, intersecting, or nested forms. For example, the composite type I defect clusters in the aforementioned embodiments are non-standard combinations formed by the spatial intersection of two type I defect clusters (e.g., a straight line intersecting an arc). During the construction of training samples, such composite defect clusters can be uniformly labeled as a whole and independently identified and output.

[0116] S23, include the remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster.

[0117] Specifically, the remaining defects that are not classified into any special defect cluster are uniformly included in the sporadic defect cluster. These defects usually do not have obvious spatial patterns, and their causes are mostly non-systematic factors such as random process noise, false alarms in detection, or environmental disturbances.

[0118] S24. For each defect cluster, determine the corresponding image sampling quantity according to the type of the defect cluster, and perform individual sampling and image acquisition of defects within the defect cluster based on the image sampling quantity.

[0119] Specifically, for each identified defect cluster, the corresponding image sampling quantity is dynamically determined based on its type (e.g., single defect cluster, compound defect cluster, sporadic defect cluster, etc.). Independent and directional image sampling is performed within each defect cluster, that is, representative defect points are selected only within the local area where the defect cluster is located, according to the determined sampling quantity. For example, defects at various radial positions within the sampling area are selected, and high-resolution detection equipment is used to acquire images of the selected defects.

[0120] Please see Figure 2 and Figure 4 , Figure 4 This is a schematic diagram of another defect sampling result provided in the embodiments of this application, wherein red markers represent unsampled defects and blue markers represent sampled defects.

[0121] like Figure 2 As shown, in the traditional random sampling mode, sampling points are evenly distributed in various radial areas of the wafer. Although it can cover the defect distribution in different locations, it has significant limitations: most sampling points only fall on background noise or sporadic defects, and fail to effectively extract defects in special defect clusters (such as arc-shaped defect clusters E and F). This results in key anomalies not being fully captured for image analysis and identification, increasing the risk of missing important information and leading to a decrease in the accuracy of subsequent generalization.

[0122] In this embodiment, a cluster-based balanced sampling strategy is adopted, that is, each identified special defect cluster is treated as an independent sampling unit. For example... Figure 4As shown, the blue markers are concentrated within the identified special defect clusters, such as the arc-shaped clusters E and F, as well as some sporadic defect clusters. This avoids repeated inspections of low-risk areas. In particular, targeted image acquisition and analysis were performed on early, subtle but engineering-significant special morphological defects, fundamentally reducing the risk of missing key anomalies.

[0123] It should be understood that by clustering to accurately extract clusters of defects with special morphological characteristics such as geometric shape, directionality, or symmetry, systematic anomalies are separated from a massive number of random defects, while unmatched remaining defects are uniformly grouped into sporadic defect clusters. Based on this, each defect cluster serves as an independent sampling unit, and the sampling strategy changes from the traditional "random selection on the entire wafer" to "balanced coverage by cluster." Even rare special morphological defects can be included in the sampling range as a whole, fundamentally avoiding the risk of missing key anomalies due to the randomness of global sampling, and also avoiding the risk of resampling and detection due to incomplete coverage and inaccurate generalization.

[0124] In some embodiments, the number of image samples is allocated differently according to the defect cluster type. S24 includes: allocating a first sampling number to the composite defect cluster, allocating a second sampling number to the single defect cluster, and the first sampling number is greater than the second sampling number; allocating a third sampling number to the sporadic defect cluster, and the third sampling number is greater than the first sampling number.

[0125] Specifically, a first sampling quantity is allocated to complex defect clusters because their complex morphology, multiple potential root causes, and strong coupling require a higher sampling density to support subsequent root cause analysis. A smaller second sampling quantity is allocated to single-type defect clusters because their causes are clear and their morphology is regular, and usually only a small number of representative images are needed to verify the hypothesized root cause and confirm the process problem. The largest third sampling quantity is allocated to sporadic defect clusters because they consist of a large number of isolated and disordered defect points. Increasing the sample size improves statistical reliability and thus enhances the accuracy of subsequent defect type generalization calculations.

[0126] The specific sampling quantity can be dynamically adjusted based on process sensitivity, historical false alarm rate, and equipment testing resources. For example, the first sampling quantity can be set to 5-10; the second sampling quantity can be set to 0-3; and the third sampling quantity can be set to 10-20 or determined according to a fixed percentage (such as 30%) of the total number of defects.

[0127] In some embodiments, S24 includes: S241, calculating the remaining sampling quantity based on the required sampling quantity for each special defect cluster and a preset upper limit; S242, using the remaining sampling quantity as the image sampling quantity for the sporadic defect cluster.

[0128] Specifically, based on the sum of the sampling quantities allocated to each type of special defect cluster, combined with the preset upper limit for the total number of images detected per single wafer (i.e., the preset upper limit), the remaining sampling quantity available for sporadic defect clusters is calculated. This remaining quantity is directly used as the image sampling quantity for sporadic defect clusters, and a corresponding number of defect points are selected within these clusters for image acquisition. This ensures that high-priority special defect clusters receive priority detection resources, while simultaneously covering sporadic defects as much as possible without exceeding the limits.

[0129] The preset upper limit refers to the maximum number of image inspection points allowed to be performed on a single wafer, used to control inspection costs and timeliness. The specific value can be dynamically adjusted according to equipment inspection resources, process layer criticality, and batch yield risk. For example, it can be increased to 50 in high-risk process nodes, while it can be reduced to 30 in mature and stable processes.

[0130] This improves the efficiency of defect type classification in subsequent generalized wafers while enabling flexible allocation and efficient utilization of resources.

[0131] S25, based on the acquired images, defect type identification is performed, and the defect clusters are individually generalized and classified according to the identification results.

[0132] Specifically, based on the high-resolution inspection images collected through sampling, fine-grained defect type identification is performed on the sampled defect points within each defect cluster, such as distinguishing specific physical or process defect categories like particulate contamination, mechanical scratches, photolithographic bridging, and metal residue. This can be accomplished, for example, by automated optical inspection equipment or manual visual inspection.

[0133] For each defect cluster, generalization classification is performed, which infers the defect composition of the entire cluster based on the distribution ratio of each type of defect in the sampling results. For example, if a specific defect cluster contains 20 defects, and 5 defects are sampled, with 2 identified as type A and 3 as type B, then the generalization classification result is: approximately 40% of the defects in the cluster are type A and 60% are type B, thus achieving a statistical reconstruction of the defect types of the entire cluster.

[0134] It should be understood that since special-shaped defects usually originate from specific equipment or process mechanisms, their causes are highly consistent, and therefore, defect types within a cluster are often highly homogeneous. Under this premise, even if only a small number of samples (such as 3-5) are drawn, the overall composition can be reflected with high confidence, balancing the accuracy of generalization classification and the cost of detection.

[0135] In some embodiments, to improve detection efficiency and resource utilization, the method further includes: when the morphology of any single-type defect cluster matches a preset morphology standard higher than a preset matching threshold, setting the second sampling quantity corresponding to the single-type defect cluster to zero, or not performing steps S24 to S25 for the single-type defect cluster; updating the sampling quantity required for each special defect cluster, and recalculating the remaining sampling quantity.

[0136] Specifically, when the morphology of a single defect cluster matches a preset morphology standard with a preset matching threshold, it indicates that its structure is highly typical, its cause is clear, and its confidence level is extremely high. The root cause of the process can be directly determined based on the morphological criteria, eliminating the need for subsequent image sampling (S24) and defect type identification and generalization (S25). The total number of samples required for each special defect cluster is updated accordingly, and the remaining number of samples allocated to sporadic defect clusters is recalculated. This allows the saved detection resources to be dynamically allocated to other defect clusters with higher uncertainty, optimizing overall resource utilization.

[0137] The matching degree between the morphology of a single-type defect cluster and the preset morphological standard can be quantified by the confidence score output by the image segmentation model during the recognition process. This confidence score reflects the model's certainty in judging that the current defect cluster conforms to a certain preset morphology (such as a straight line, a ring, etc.). Therefore, the higher the confidence score, the closer its geometric structure is to the template, and the higher the matching degree.

[0138] The preset matching threshold value can be set based on the statistical distribution of historical high-confidence samples and can be dynamically calibrated according to the current process stability and the production line's tolerance for false alarms. For example, in critical process layers with extremely high requirements for yield and reliability, where risk control is strict, the preset matching threshold can be increased to allow only a very small number of defect clusters with high confidence to be exempted from image verification, thereby reducing false judgments.

[0139] In some embodiments, to ensure the statistical representativeness of sporadic defect clusters and the accuracy of generalization calculation, when the remaining sampling quantity for sporadic defect clusters is less than the preset minimum sampling quantity, the second sampling quantity corresponding to the single-type defect cluster is set to zero, or steps S24 to S25 are not performed for the single-type defect cluster; the sampling quantity required for each special defect cluster is updated, and the remaining sampling quantity is recalculated.

[0140] Specifically, when the sampling quantity of sporadic defect clusters is too small to support an effective assessment of background noise, priority is given to ensuring coverage of sporadic defects. The second sampling quantity corresponding to a single-type defect cluster is temporarily set to zero, and subsequent image sampling (S24) and defect type identification and generalization (S25) are skipped. The total sampling quantity required for each special defect cluster is updated accordingly, and the remaining sampling quantity allocated to sporadic defect clusters is recalculated to increase the sample size allocated to sporadic defect clusters.

[0141] The preset minimum sampling number refers to the minimum number of image detection points required to ensure the effectiveness of statistical analysis of sporadic defect clusters. The specific value can be flexibly set according to the criticality of the process layer, noise level, etc., such as setting it to 8.

[0142] In some embodiments, when the remaining sampling quantity for sporadic defect clusters is less than the preset minimum sampling quantity, a target single-type defect cluster with a morphology matching degree higher than the preset morphology standard is obtained, and the second sampling quantity corresponding to the target single-type defect cluster is set to zero; the sampling quantity required for each special defect cluster is updated, and the remaining sampling quantity is recalculated.

[0143] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating another defect sampling result provided in an embodiment of this application. For example... Figure 5 As shown, when steps S24 to S25 are not performed on defect cluster F or the sampling quantity is set to zero, there are no blue marker points inside it, that is, no image sampling points are assigned to it.

[0144] In some embodiments, the method further includes: if the identification result of a single type defect cluster is inconsistent with the preset unique process root cause, then the single type defect cluster is reclassified as a composite defect cluster, and steps S24 to S25 are re-executed.

[0145] Specifically, after image sampling and defect type identification are completed for a single-type defect cluster, if the actual identification result (such as the physical type of the defect or the inferred root cause) is inconsistent with the originally preset single process root cause of that morphology, it is determined that the cluster is not a typical single-cause event, and there may be morphological coincidence, multi-source interference, or unknown faults. At this time, the cluster is reclassified as a composite defect cluster, and steps S24 to S25 are repeated, that is, according to the higher sampling requirements of composite defect clusters, more image samples are collected, and a more comprehensive generalization classification and root cause analysis are performed based on the updated data.

[0146] In some embodiments, a mapping database of morphology and process root causes can be pre-established to support intelligent attribution based on the specific morphology of defect clusters. The mapping database stores association rules between various special defect clusters (including single-type and composite defect clusters) and their potential process causes. For example, a linear defect can be mapped to multiple root causes such as "fixture malfunction," "transfer arm scratch," or "foreign object dragging in the chamber." The mapping relationships can be continuously updated based on historical yield analysis, equipment maintenance records, and expert experience.

[0147] In practical applications, once a specific defect cluster is identified, the relational database can be queried in real time to output one or more high-probability root cause hypotheses. These hypotheses are then weighted and sorted in conjunction with the current production line status (such as equipment operating parameters, material batches, and process steps) to provide engineers with accurate diagnostic suggestions, thereby improving the efficiency of anomaly response and the accuracy of root cause localization.

[0148] Please see Figure 6 , Figure 6 This is a schematic block diagram of a wafer defect analysis device provided in an embodiment of this application. The wafer defect analysis device can be configured in a server to perform the aforementioned wafer defect identification, wafer defect analysis method and specific implementation steps of the method.

[0149] like Figure 6 As shown in the figure, this application embodiment also provides a wafer defect analysis device 30, the device comprising; The defect acquisition module 301 is used to acquire multiple defects on the wafer surface and the location information of each defect; The clustering module 302 is used to cluster the multiple defects based on the location information of each defect, and extract several special defect clusters with special shapes. The sporadic defect module 303 is used to include remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; The sampling acquisition module 304 is used to determine the corresponding image sampling quantity for each defect cluster according to the type of the defect cluster, and to perform individual sampling and image acquisition of defects within the defect cluster based on the image sampling quantity. The defect identification module 305 is used to identify the defect type based on the acquired image and perform generalization classification on the defect clusters separately according to the identification results.

[0150] For example, the special defect clusters include single-type defect clusters and composite defect clusters; wherein, the morphology of the single-type defect cluster corresponds to a unique process root cause, and the morphology of the composite defect cluster corresponds to multiple process root causes.

[0151] For example, the sampling acquisition module 304 also includes: The first quantity allocation submodule is used to allocate a first sampling quantity to the composite defect cluster and a second sampling quantity to the single defect cluster, wherein the first sampling quantity is greater than the second sampling quantity. The second quantity allocation submodule is used to allocate a third sampling quantity to the sporadic defect clusters, and the third sampling quantity is greater than the first sampling quantity.

[0152] For example, the sampling acquisition module 304 also includes: The first remaining quantity submodule is used to calculate the remaining sampling quantity based on the sampling quantity required for each special defect cluster and the preset quantity limit; The sporadic sampling quantity submodule is used to use the remaining sampling quantity as the image sampling quantity of the sporadic defect cluster.

[0153] For example, the wafer defect analysis apparatus 30 further includes: The sampling skip module is used to set the second sampling quantity corresponding to the single-type defect cluster to zero when the matching degree between the shape of any single-type defect cluster and the preset shape standard is higher than the preset matching threshold; and / or, the quantity reset module is used to set the second sampling quantity corresponding to the single-type defect cluster to zero when the remaining sampling quantity for sporadic defect clusters is less than the preset minimum sampling quantity. The quantity update module is used to update the required sampling quantity for each special defect cluster and recalculate the remaining sampling quantity.

[0154] For example, the wafer defect analysis apparatus 30 further includes: The type update module is used to reclassify the single-type defect cluster as a composite defect cluster and re-execute steps S24 to S25 if the identification result of the single-type defect cluster is inconsistent with the preset unique process root cause.

[0155] For example, the wafer defect analysis apparatus 30 further includes: The sample acquisition module is used to acquire multiple special defect cluster samples, each of which includes a wafer defect distribution map and a defect cluster type label. The model training module is used to train an image segmentation model based on the special defect cluster samples. The image segmentation model is used to identify and output special defect clusters from the input wafer defect distribution map.

[0156] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the above-described apparatus and its modules and units can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0157] The methods and apparatus of this application can be used in a wide variety of general-purpose or special-purpose computing system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer terminal devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0158] Please see Figure 7 , Figure 7This is a schematic block diagram illustrating the structure of a computer device according to an embodiment of this application. The computer device may be a terminal device or a server.

[0159] For example, the above-described method and apparatus can be implemented as a computer program, which can be used in, for example... Figure 7 It runs on the computer device shown.

[0160] like Figure 7 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0161] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any method for identifying wafer defects, a method for analyzing wafer defects, and the specific implementation steps of those methods.

[0162] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0163] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any method for identifying wafer defects, method for analyzing wafer defects, and specific implementation steps of the method.

[0164] This network interface is used for network communication, such as sending assigned tasks.

[0165] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0166] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: S11, acquire multiple defects on the wafer surface and the location information of each defect; S12, based on the location information of each defect, perform first-level clustering on the multiple defects, and extract several defect clusters with special shapes, so that each defect cluster simultaneously satisfies the following conditions: (a) The spatial arrangement of defects within the cluster conforms to a preset morphological standard, which includes at least one of geometric shape, orientation, or symmetry; (b) For any target defect within a cluster, the intra-cluster dispersion of the target defect and defects in the same cluster within the first neighborhood is less than the background dispersion of defects in the non-cluster within the second neighborhood, and the difference between the intra-cluster dispersion and the background dispersion is greater than or equal to a preset dispersion difference, wherein the second neighborhood is greater than or equal to the first neighborhood.

[0167] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: S21, acquire multiple defects on the wafer surface and the location information of each defect; S22, based on the location information of each defect, the multiple defects are clustered and extracted to extract several special defect clusters with special shapes; S23, Incorporate the remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; S24. For each defect cluster, determine the corresponding image sampling quantity according to the type of the defect cluster, and perform individual sampling and image acquisition of defects within the defect cluster based on the image sampling quantity. S25, based on the acquired images, defect type identification is performed, and the defect clusters are individually generalized and classified according to the identification results.

[0168] For example, the processor is used to run a computer program stored in the memory, and is also used to implement the steps and specific implementation steps of the wafer defect identification method and wafer defect analysis method provided in any embodiment of this application, which will not be repeated here.

[0169] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement the steps and specific implementation steps of the wafer defect identification method and wafer defect analysis method provided in any of the embodiments of this application.

[0170] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0171] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered 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. A method for analyzing wafer defects, characterized in that, The method includes: S21, acquire multiple defects on the wafer surface and the location information of each defect; S22, based on the location information of each defect, the multiple defects are clustered to extract several special defect clusters with special shapes; wherein, the special defect clusters simultaneously satisfy the following conditions: (a) The spatial arrangement of defects within the cluster conforms to a preset morphological standard, which includes at least one of geometric shape, orientation, or symmetry; (b) For any target defect within a cluster, the intra-cluster dispersion of the target defect and defects in the same cluster within the first neighborhood range is less than the background dispersion of defects in the non-cluster range within the second neighborhood range, and the difference between the intra-cluster dispersion and the background dispersion is greater than or equal to a preset dispersion difference value, wherein the second neighborhood range is greater than or equal to the first neighborhood range. S23, Incorporate the remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; S24. For each defect cluster, determine the corresponding image sampling quantity according to the type of the defect cluster, and perform individual sampling and image acquisition of defects within the defect cluster based on the image sampling quantity. S25, based on the acquired images, defect type identification is performed, and the defect clusters are individually generalized and classified according to the identification results.

2. The method according to claim 1, characterized in that, The special form includes at least one of the following: straight line, arc, ring, fan shape, block shape, wavy shape, and a combination of at least two identical or different forms.

3. The method according to claim 1, characterized in that, The special defect clusters include single-type defect clusters and composite defect clusters; wherein, the morphology of the single-type defect cluster corresponds to a unique process root cause, and the morphology of the composite defect cluster corresponds to multiple process root causes. S24 includes: A first sampling quantity is allocated to the composite defect cluster, and a second sampling quantity is allocated to the single defect cluster, wherein the first sampling quantity is greater than the second sampling quantity; A third sampling quantity is allocated to the sporadic defect clusters, and the third sampling quantity is greater than the first sampling quantity.

4. The method according to claim 3, characterized in that, S24 includes: S241, calculate the remaining sampling quantity based on the required sampling quantity for each special defect cluster and the preset upper limit of the quantity; S242, the remaining sampling quantity is used as the image sampling quantity of the sporadic defect cluster.

5. The method according to claim 3 or 4, characterized in that, The method further includes: When the morphology of any single type of defect cluster matches the preset morphology standard higher than the preset matching threshold, and / or when the remaining sampling quantity for sporadic defect clusters is less than the preset minimum sampling quantity, the second sampling quantity corresponding to the single type of defect cluster is set to zero. Update the required sampling quantity for each specific defect cluster and recalculate the remaining sampling quantity.

6. The method according to claim 1, characterized in that, The method further includes: If the identification result of a single type of defect cluster is inconsistent with the preset unique process root cause, the single type of defect cluster is reclassified as a composite defect cluster, and steps S24 to S25 are repeated.

7. The method according to claim 1, characterized in that, The method further includes: Multiple special defect cluster samples are obtained, and each special defect cluster sample includes a wafer defect distribution map and a defect cluster type label; An image segmentation model is trained based on the special defect cluster samples. The image segmentation model is used to identify and output special defect clusters from the input wafer defect distribution map.

8. An analytical apparatus for wafer defects, characterized in that, The device includes: The defect acquisition module is used to acquire multiple defects on the wafer surface and the location information of each defect; The clustering module is used to cluster the multiple defects based on the location information of each defect, and extract several special defect clusters with special shapes; wherein, the special defect clusters simultaneously satisfy the following conditions: (a) The spatial arrangement of defects within the cluster conforms to a preset morphological standard, which includes at least one of geometric shape, orientation, or symmetry; (b) For any target defect within a cluster, the intra-cluster dispersion of the target defect and defects in the same cluster within the first neighborhood range is less than the background dispersion of defects in the non-cluster range within the second neighborhood range, and the difference between the intra-cluster dispersion and the background dispersion is greater than or equal to a preset dispersion difference value, wherein the second neighborhood range is greater than or equal to the first neighborhood range. The sporadic defect module is used to include remaining defects that have not been classified into any special defect cluster into the sporadic defect cluster; The sampling acquisition module is used to determine the corresponding number of image samples for each defect cluster according to the type of the defect cluster, and to perform individual sampling and image acquisition of defects within the defect cluster based on the number of image samples. The defect identification module is used to identify the type of defect based on the acquired images and to perform generalization classification on the defect clusters individually according to the identification results.

9. A computer device, characterized in that, The device includes: Memory, used to store computer programs; A processor for executing the computer program and, in executing the computer program, implementing the wafer defect analysis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the wafer defect analysis method as described in any one of claims 1 to 7.

Images