Machine learning based pattern grouping method

Abstract

A pattern grouping method can include receiving an image of a first pattern, generating a first fixed-dimension feature vector using trained model parameters applied to the received image, and assigning a first bucket ID to the first fixed-dimension feature vector. The method can also include creating a new bucket ID for the first fixed-dimension feature vector in response to determining that the first pattern does not belong to one of a plurality of buckets corresponding to defect patterns, or mapping the first fixed-dimension feature vector to the first bucket ID in response to determining that the first pattern belongs to one of the plurality of buckets corresponding to defect patterns.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to U.S. Application 62 / 697,898, filed on July 13, 2018, which is incorporated herein by reference in its entirety. Technical Field

[0003] The embodiments provided herein relate to a pattern grouping method, and more specifically, to a machine learning-based pattern grouping method, which is particularly useful for defect review during semiconductor manufacturing operations. Background Technology

[0004] In the manufacturing process of integrated circuits (ICs), unfinished or finished circuit components are inspected to ensure they are manufactured according to the design and free of defects. Inspection systems utilizing optical microscopes or charged particle (e.g., electron) beam microscopes (such as scanning electron microscopy (SEM)) can be employed. As the physical dimensions of IC components continue to shrink, the accuracy and yield of defect detection become increasingly important. However, the imaging resolution and throughput of inspection tools struggle to keep pace with the ever-reducing feature sizes of IC components.

[0005] As technology nodes shrink to, for example, 10nm, some inspection tools may generate a large number of harmful defects (i.e., false positives). For instance, in some inspection systems, 90% of identified defects may prove to be harmful. Therefore, defect re-inspection becomes increasingly critical. Defect re-inspection can be utilized by re-examining previously identified defects or suspicious locations on the wafer.

[0006] To meet the demands of high-volume and high-yield manufacturing processes, operators need to use a graphical user interface (GUI) to inspect wafers, chips, or masks and identify pattern defects as early as possible. This GUI displays various patterns of the observed object in the display section of the casting equipment. Unfortunately, identifying pattern defects can consume a significant amount of the operator's time, thereby reducing throughput. Summary of the Invention

[0007] In some embodiments of this disclosure, a grouping method is provided. The method may include: receiving an image of a first pattern; generating a first fixed-dimensional feature vector using trained model parameters; and assigning a first bucket identifier (ID) to the first fixed-dimensional feature vector. The model parameters may be based on the received image. For example, the model parameters may include data of the feature pattern, such as SEM image data of detected defects or pattern data from a wafer design plan at locations corresponding to the detected defects. The method can be used for defect pattern grouping in semiconductor manufacturing processes.

[0008] Defect patterns can include images of a wafer that include features associated with one or more defects. Defect patterns can include various types of information, including scanned raw image data, processed images, imaging features identified during defect inspection, and design patterns corresponding to those imaging features.

[0009] Model parameters can be trained, for example, by obtaining multiple images of multiple patterns with assigned bucket IDs and training model parameters for a deep learning network. Training can also include applying parameters to a single polygon located at the center of one of the multiple images. Attached Figure Description

[0010] Figure 1 This is a schematic diagram illustrating an exemplary electron beam inspection (EBI) system according to an embodiment of the present disclosure.

[0011] Figure 2 The illustrations are consistent with the embodiments of this disclosure. Figure 1 A schematic diagram of an exemplary electron beam tool, which is part of an exemplary electron beam inspection system.

[0012] Figure 3 This is a block diagram illustrating an exemplary defect review system consistent with embodiments of the present disclosure.

[0013] Figure 4 This is a schematic diagram illustrating an exemplary pattern similarity measurement method for pattern grouping.

[0014] Figure 5 This is a schematic diagram illustrating an exemplary feature extraction architecture for machine learning-based pattern grouping, consistent with embodiments of this disclosure.

[0015] Figure 6 The illustrations are consistent with embodiments of this disclosure and are derived from... Figure 5 A schematic diagram of an exemplary fixed-dimensional feature vector for a patterned image processed by a machine learning feature extraction architecture.

[0016] Figure 7 This is a schematic diagram illustrating an exemplary correlation between fixed-dimensional feature vectors, 3D representations of pattern groups, and GDS information of defect pattern images consistent with embodiments of this disclosure.

[0017] Figure 8 This is a flowchart illustrating an exemplary method consistent with embodiments of this disclosure. Detailed Implementation

[0018] Reference will now be made to exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings, wherein, unless otherwise indicated, the same numerals in the different drawings denote the same or similar elements. The embodiments set forth in the following description of the exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, these embodiments are merely examples of apparatuses, systems, and methods consistent with aspects relevant to the subject matter as described in the appended claims. For example, although some embodiments are described in the context of utilizing electron beams, this disclosure is not limited thereto. Other types of charged particle beams can be similarly applied. Furthermore, other imaging systems, such as optical imaging, light detection, X-ray detection, etc., can be used.

[0019] The increased computing power of electronic devices (while reducing their physical size) can be achieved by significantly increasing the packaging density of circuit components (such as transistors, capacitors, diodes, etc.) on IC chips. For example, a thumb-sized IC chip in a smartphone can contain over 2 billion transistors, each less than 1 / 1000th the size of a human hair. Therefore, it's not surprising that semiconductor IC manufacturing is a complex and time-consuming process with hundreds of individual steps. Even an error in one step can drastically affect the functionality of the final product. Even a "fatal defect" can cause equipment failure. The goal of manufacturing processes is to improve the overall yield of the process. For example, to achieve a 75% yield in a 50-step process, each individual step must have a yield greater than 99.4%, and if the yield of an individual step is 95%, the overall process yield will drop to 7%.

[0020] While high process yields are required in IC chip manufacturing facilities, maintaining high wafer output, defined as the number of wafers processed per hour, is also crucial. Both high process yields and high wafer output can be affected by the presence of defects, especially when operator intervention is involved to re-inspect these defects. Therefore, detecting and identifying micron- and nanometer-sized defects using inspection tools such as SEM is essential for maintaining both high yields and low costs.

[0021] In the augmented imaging system described herein, defect identification and classification can be performed with improved output and accuracy. The process can be enhanced by reducing reliance on operator intervention, for example, by eliminating the need for the operator to determine the similarity of the extracted pattern and classify it as a known defect type. The augmented imaging system can decompose defects into one or more features, each corresponding to one or more attributes of the defect pattern. Using these features, the augmented imaging system can provide an improved method for identifying and classifying defects. Patterns that deviate slightly from the identified shape can still be captured and described, for example, by representing the presence of specific features with a degree of determinism. This approach can achieve better accuracy, efficiency, and speed in defect identification and classification.

[0022] Patterns can be processed using convolution, pooling, and subsampling to allow for the analysis of individual feature attributes. Feature analysis allows for the generation of feature vectors. These feature vectors can then be used for pattern grouping. Minor deviations in orientation, scale, or occlusion may not significantly affect pattern processing, allowing similar patterns to still be grouped together. Furthermore, when defect images include features beyond the central pattern, features of the defect pattern can be extracted. Additionally, running a linear classifier may require fewer computational resources compared to corresponding pattern similarity measures used for pattern grouping.

[0023] As used herein, unless otherwise expressly stated, the term "or" covers all possible combinations unless practicable. For example, if a statement declares that a database may include A or B, then unless otherwise expressly stated or practicable, the database may include A, or B, or A and B. As a second example, if a statement declares that a database may include A, B, or C, then unless otherwise expressly stated or practicable, the database may include A, or B, or C, or A and B, or A and C, or B and C, or A and B and C.

[0024] Additional objects and advantages of the disclosed embodiments will be set forth in part in the description which follows, and will also be apparent in part from the description, or may be learned by practice of the embodiments. The objects and advantages of the disclosed embodiments may be realized and obtained by means of the elements and combinations set forth in this disclosure. However, such exemplary objects and advantages are not necessarily required to achieve the exemplary embodiments of this disclosure, and some embodiments may not achieve any of the stated objects and advantages.

[0025] Now for reference Figure 1 The illustration depicts an exemplary electron beam inspection (EBI) system 100 consistent with embodiments of this disclosure. The EBI system 100 can be used for imaging. Figure 1As shown, the EBI system 100 includes a main chamber 101, a loading / locking chamber 102, an electron beam tool 104, and an equipment front-end module (EFEM) 106. The electron beam tool 104 is positioned within the main chamber 101. The EFEM 106 includes a first loading port 106a and a second loading port 106b. The EFEM 106 may include (multiple) additional loading ports. The first loading port 106a and the second loading port 106b receive a front-opening wafer transport cassette (FOUP), which contains wafers (e.g., semiconductor wafers or wafers made of (multiple) other materials) or samples to be inspected (wafers and samples may be collectively referred to herein as “wafers”). A “batch” is a plurality of wafers that can be loaded for batch processing.

[0026] One or more robotic arms (not shown) in EFEM 106 can transport the wafer to loading / locking chamber 102. Loading / locking chamber 102 is connected to a loading / locking vacuum pump system (not shown) that removes gas molecules from loading / locking chamber 102 to achieve a first pressure below atmospheric pressure. After reaching the first pressure, one or more robotic arms (not shown) can transport the wafer from loading / locking chamber 102 to main chamber 101. Main chamber 101 is connected to a main chamber vacuum pump system (not shown) that removes gas molecules from main chamber 101 to achieve a second pressure below the first pressure. After reaching the second pressure, the wafer is inspected by electron beam tool 104. Electron beam tool 104 can be a single-beam system or a multi-beam system. Controller 109 is electrically connected to electron beam tool 104. Controller 109 can be a computer configured to perform various controls of EBI system 100. Although controller 109 is in Figure 1 The controller 109 is shown outside the structure including the main chamber 101, the loading / locking chamber 102 and the EFEM 106, but it should be understood that the controller 109 may be part of the structure.

[0027] Figure 2 An imaging system 200 according to an embodiment of the present disclosure is illustrated. Figure 2 The electron beam tool 104 can be configured for use in the EBI system 100. Although Figure 2 An electron beam tool 104 is shown as a single-beam inspection tool that can scan one location of the wafer 230 at a time using only one main electron beam; however, embodiments of this disclosure are not limited thereto. For example, the electron beam tool 104 could also be a multi-beam inspection tool that uses multiple main electron beams to scan multiple locations on the wafer 230 simultaneously.

[0028] System 200 can be used to inspect wafer 230 on a sample stage and includes an electron beam tool 104, as discussed above. System 200 also includes an image processing system 199, which includes an image acquisition unit 120, a storage device 130, and a controller 109. Image acquisition unit 120 may include one or more processors or circuitry, such as circuitry of one or more processors or other circuitry. For example, image acquisition unit 120 may include a computer, server, mainframe, terminal, personal computer, any kind of mobile computing device, or a combination thereof. Image acquisition unit 120 can be connected to detector 244 of electron beam tool 104 via a medium such as an electrical conductor, fiber optic cable, portable storage medium, infrared (IR), Bluetooth, Internet, wireless network, radio, or a combination thereof. Image acquisition unit 120 can receive signals from detector 244 and can construct images. Therefore, image acquisition unit 120 can acquire images of wafer 230. Image acquisition unit 120 can also perform various post-processing functions, such as generating contours, overlaying indicators on the acquired images, etc. Image acquisition unit 120 can be configured to perform adjustments such as brightness and contrast of the acquired images. Storage device 130 can be a storage medium, such as a hard disk, cloud storage device, random access memory (RAM), or other types of computer-readable storage. Storage device 130 can be coupled to image acquisition unit 120 and can be used to save scanned raw image data as initial and post-processed images. Image acquisition unit 120 and storage device 130 can be connected to controller 109. In some embodiments, image acquisition unit 120, storage device 130, and controller 109 can be integrated into a single control unit.

[0029] In some embodiments, the image acquisition unit 120 may acquire one or more images of a sample based on imaging signals received from the detector 244. The imaging signals may correspond to a scanning operation for imaging charged particles. The acquired image may be a single scanned raw image comprising multiple imaging regions. The image may be stored in the storage device 130. The image may be an initial image that can be divided into multiple regions. Each region may include an imaging area containing features of the wafer 230.

[0030] In some embodiments, a computer system may be provided that can identify defects in a wafer image and classify defects into categories according to their type. For example, once a wafer image is acquired, it can be transmitted to the computer system for processing. Figure 3 This is a schematic diagram of a defect review system 300 consistent with embodiments of the present disclosure.

[0031] Reference Figure 3The defect review system 300 may include a wafer inspection system 310, an automatic defect classification (ADC) server 320, and a knowledge recommendation server 330 electrically coupled to the ADC server 320. The wafer inspection system 310 may be relative to... Figure 1 The EBI system 100 is described. It should be understood that the ADC server 320 and the knowledge recommendation server 330 may be part of the EBI system 100 or located outside of the EBI system.

[0032] The wafer inspection system 310 can be any inspection system that generates inspection images of the wafer. For example, the wafer can be a semiconductor wafer substrate or a semiconductor wafer substrate having one or more epitaxial layers or processing films. The wafer inspection system 310 can be any currently available or under development wafer inspection system. Embodiments of this disclosure do not limit the specific type of wafer inspection system 310. Such a system can generate wafer images with resolution to observe key features on the wafer (e.g., less than 20 nm).

[0033] The ADC server 320 may include a communication interface 322 electrically coupled to the wafer inspection system 310 to receive wafer images. The ADC server 320 may also include a processor 324 configured to analyze the wafer images and detect and classify defects appearing in the wafer images, and may use a defect knowledge file in the analysis, detection, or classification. The defect knowledge file may be manually provided to the ADC server 320 by an operator. Alternatively, according to some embodiments of this disclosure, the defect knowledge file may be automatically provided to the ADC server 320 by a knowledge recommendation server 330.

[0034] For example, the knowledge recommendation server 330 may be electrically coupled to the ADC server 320. The knowledge recommendation server 330 may include a processor 332 and a storage device 334. The processor 332 may be configured to construct multiple defect knowledge files and store them in the storage device 334. The multiple defect knowledge files may contain information related to various types of defects generated during different stages of the wafer manufacturing process. These stages may include, but are not limited to, photolithography, etching, chemical mechanical polishing (CMP), or interconnect formation processes.

[0035] Processor 332 can be configured to construct multiple defect knowledge files based on multiple defect patch images. These multiple defect patch images can be generated by wafer inspection tools (such as...) Figure 2 The defect patch image is generated by the electron beam tool 104 shown in the figure. The defect patch image can be a small image (e.g., 34×34 pixels) containing a portion of the defect in the wafer. The defect patch image can be centered on the defect and can include adjacent pixels of the defect.

[0036] The processor 332 can be trained via a machine learning process to construct a knowledge file related to a specific type of defect based on multiple defect patch images of that type of defect. For example, the processor 332 can be trained to construct a knowledge file related to a broken wire defect generated during interconnect formation based on multiple defect patch images of a broken wire defect.

[0037] The processor 332 can also be configured to search for knowledge files that match the wafer image included in the received request in response to a request for knowledge recommendations from the ADC server 320, and provide the knowledge files to the ADC server 320.

[0038] Storage device 334 can store an ADC data center containing multiple defect knowledge files related to various types of defects generated during different stages of the wafer manufacturing process. These multiple defect knowledge files in the ADC data center can be created by the processor 332 of the knowledge recommendation server 330. Alternatively, a portion of the defect knowledge files in storage device 334 can be pre-configured by a user or an external computer system and can be pre-loaded into storage device 334.

[0039] The defect knowledge file may include general information about a single type of defect. This general information may include patch images and characteristic parameters (e.g., size, edge roughness, depth, height, etc.) to be used later to classify the single type of defect. Alternatively, according to some embodiments of this disclosure, the defect knowledge file may include general information about multiple types of defects present in the same processing layer of a wafer. A single processing layer may be, for example, a substrate layer, an epitaxial layer, a thin film layer, a photoresist layer, an oxide layer, a metal interconnect layer, etc.

[0040] Now for reference Figure 4 The figure is a schematic diagram illustrating an aspect of an exemplary pattern similarity measurement method for pattern grouping. The pattern similarity measurement method compares the geometric features of an imaged defect pattern with data patterns stored in a defect classification server to group the defect patterns into groups with the highest similarity.

[0041] Pattern data can be determined based on charged particle imaging. For example, a charged particle beam device can be controlled to image a wafer. Imaging may include scanning the wafer to image at least a portion of the wafer. One or more scanned raw images of the imaging area can be obtained. Defect inspection can be performed on the scanned raw images. For example, one or more features on the wafer can be identified. One or more features may include potential defects, critical areas, hot spots, other regions of interest, etc. Further processing may be performed, including image processing, defect mapping, providing a list of locations on the wafer corresponding to the identified potential defects, etc. In some embodiments, defect review may be performed. Defect review may include imaging the areas identified in the defect inspection step to analyze the identified potential defects in more detail, for example, at a higher resolution. In some embodiments, defect review may include analyzing the scanned raw images, for example, using techniques different from those used in the defect inspection step or techniques with different parameter sets. Techniques used for defect review may include, for example, using image processing algorithms.

[0042] Pattern data can be determined corresponding to identified defects. Pattern data can be extracted from the wafer design plan based on location information collected during defect inspection. The wafer design plan can be pre-stored. For example, the wafer design plan can be a graphical representation of features on the wafer surface. The wafer design plan can be based on a pattern layout used to construct the wafer. For example, the wafer design plan can correspond to a mask used to manufacture the wafer. The wafer design plan can be stored in a database. Individual features can be extracted corresponding to identified defects. For example, pattern data including geometric information of the initial feature design can be collected based on the xy coordinates of the defect location identified during defect inspection. Each identified defect can have a corresponding initial feature design.

[0043] Pattern data can be extracted by overlaying feature outlines onto the raw image scanned from the wafer. Image processing techniques can be applied to generate polygons corresponding to the imaged features. Pattern data can also be extracted by overlaying feature outlines (such as wafer design plans) onto a graphical representation of the wafer.

[0044] In pattern grouping systems employing pattern similarity measurement methods, imaged patterns can be processed to extract the geometric features of defect patterns. These geometric features can include the number of sides, the number of angles, dimensions, shape, or any combination thereof. Pattern data (such as Graphic Data System (GDS) information) can be derived from these geometric features and represented by one or more polygons drawn with solid lines, for example, as overlaid on... Figure 4The image pattern shown is an image of a defect grouping system. The pattern grouping system may have a defect classification server with a knowledge file storing GDS information of defect patterns. The pattern grouping system may then continue to compare the GDS information of the extracted geometric features of the defect patterns with the defect pattern GDS information representing multiple defect patterns stored in the defect classification server.

[0045] Each comparison between the extracted geometric features of a defect pattern and the GDS information of one defect pattern from a plurality of defect patterns stored in a defect classification server can generate a similarity parameter value. The value of the similarity parameter can depend on the similarity of the geometric features between the compared defect patterns. In some embodiments, a higher value indicates a higher degree of similarity shared between the received defect pattern and the stored defect patterns, and a higher probability that the received defect pattern belongs to a group of stored defect patterns. For example, polygon 410 in pattern image 400 can be compared with... Figure 4 The 12 defect patterns shown are compared one by one. Each comparison returns a similarity value greater than 0 and less than or equal to 1, where 0 indicates no similarity and 1 indicates identical. After completing 12 comparisons, the defect pattern associated with the comparison that returned the highest value can be determined as the pattern group to which the received defect pattern belongs.

[0046] Methods based on pattern similarity measurements may have drawbacks. Due to several limitations, the performance of techniques based on similarity parameters may not be optimal. For example, one limitation may be a heavy reliance on the definition of similarity, where the value is calculated based on geometric differences in the properties of GDS polygonal patterns. Even small changes in pattern properties or polygonal deformation can lead to significantly different results. Therefore, the calculated similarity value may depend heavily on the settings of the pattern properties, which may be influenced to some extent by empirical decisions. Additionally, similar patterns with different orientations, scales, or occlusions (partial pattern appearances) may not be grouped into the same pattern groups. Furthermore, defective patterns may not be correctly grouped when the defective image has a central pattern surrounded by other patterns or pattern patches, or when it comprises a combination of multiple GDS patterns. Further, the runtime complexity of pattern similarity measurement methods used for pattern grouping can be relatively high. For example, the runtime complexity can be approximated as O(N²), where O is the function order and N is the input size, which can be measured in bits.

[0047] Now for reference Figure 5This figure is a schematic diagram illustrating an exemplary feature extraction architecture for machine learning-based pattern grouping consistent with embodiments of the present disclosure. The feature extraction architecture may include a convolutional neural network 500. In some embodiments, other forms of classification and pattern recognition may be employed. For example, cluster analysis, object detection, and masking may be employed. Other types of supervised learning systems based on machine learning may be similarly applied. The feature extraction architecture for machine learning-based pattern grouping may be pre-trained before generating pattern features.

[0048] In some embodiments, a linear classifier network of a deep learning architecture can be used as a starting point for training and building a feature extraction architecture for pattern grouping. A large number of GDS pattern images may have been collected and can constitute a large dataset of existing defect patterns. These GDS pattern images can be adjusted to reflect the center pattern group labels. The values ​​of the model parameters in each layer of the linear classifier network of the deep learning architecture can be learned from the large dataset of existing defect patterns reflecting the center pattern groups. Due to the large amount of GDS pattern data, the pattern image data in each layer of the linear classifier network can be automatically labeled by the pattern grouping system, rather than manually. Each label in each layer can represent the attributes of the center polygon of the defect pattern. Utilizing GDS information, the training of the linear classifier network of the deep learning architecture can be completed automatically with very little manual input.

[0049] A machine learning model can be used, which may include a deep neural network topology with multiple layers. For example, the architecture of such a network may include an input, a first convolution, a first pooling, a second convolution, a second pooling, one or more hidden layers, a startup layer, and an output. The input 501 of network 500 may include a pattern image that includes identified defects. The output 509 of network 500 may include a feature vector.

[0050] Depending on the nature of the defect pattern, each layer of the architecture can have a different number of generated subsamples. For example, as Figure 5 As shown, after the first convolution operation, fewer than ten subsamples can be generated in the first pooling 510. However, after the second convolution operation, the second layer can have more than ten subsamples generated in the second pooling 520. In some embodiments, variations between layers can be introduced by the complexity of the geometric features of the defect pattern. Defect patterns with more geometric information can have a higher probability of generating more subsamples. For example, complex defect patterns can exhibit various subshapes that can be decomposed and analyzed into individual properties.

[0051] After the model parameters have been determined through training on a large number of GDS pattern images, a trained linear classifier network of a deep learning architecture can process new pattern images in the feature generation step. In this step, a pattern image with GDS data is passed through multiple layers of the trained network with defined model parameters to generate a fixed-dimensional feature vector at the end of the architecture. At each layer of convolutional computation, the GDS data of the pattern image can be analyzed against the architecture to determine the presence of pattern attributes and, if present, their geometric location. Each received pattern image with GDS data can be processed layer-by-layer through computation across multiple layers. Therefore, a vector representation of the input image can be generated by the trained linear classifier network.

[0052] Now for reference Figure 6 The figure illustrates an embodiment consistent with this disclosure, derived from a method based on... Figure 5 This is a schematic diagram of an exemplary fixed-dimensional feature vector representation of a patterned image processed by a machine learning network's feature extraction architecture. The feature vector can include multiple values ​​representing the magnitude, intensity, determinism, etc., of each of a plurality of features. For example, for multiple features, the degree of determinism of a patterned image can be evaluated to determine the presence of a defect. Figure 6 In this diagram, the X-axis comprises four features. A feature can correspond to, for example, a line-to-line break. The determinacy of a feature within the pattern can be represented numerically. Determinacy can also be represented through various forms of graphical display (such as color, fill patterns, or different degrees of shading). For example, as... Figure 6 As shown, multiple patterns can be evaluated to determine the certainty of the presence of each specific feature in a particular feature represented by the filling pattern.

[0053] After processing the pattern image through the operations of each layer of the network, quantitative analysis can be performed by dividing each dimension of the feature vector into a finite number of buckets. Buckets can include members of the same type that have already been distinguished and classified based on certain criteria. For example, buckets can correspond to defect categories. Training the network can involve classifying GDS pattern data into multiple buckets using a training dataset. Training can continue until a certain degree of convergence is achieved.

[0054] Each feature vector can be mapped to a bucket identifier (ID). The convolution result can be transformed into a fixed-dimensional vector representing the pattern features. Fixed-dimensional feature vectors (e.g., one-dimensional vectors) can be used as bucket IDs. For example, in... Figure 6In this model, each vertical bar along the X-axis can represent a feature pattern. Each vertical bar can be a one-dimensional vector. Each vertical bar can include, for example, multiple values ​​along the Y-axis, where different values ​​are represented by different codes. For example, a code can be represented by a fill pattern. The feature vector can be output for a specific input pattern (Pi) and can be converted into a bucket ID 601. Due to the different values ​​or different combinations of values, each feature vector may be unique and may represent a unique group of defective patterns. Because similar pattern images share similar feature distributions, similar feature distributions can be mapped to the same bucket ID. Therefore, it can be determined that feature vectors with the same combination of values ​​belong to the same group of patterns. In some embodiments, it can be determined that a pattern with feature vectors that do not fall into any bucket is a harmful defect.

[0055] Although one-dimensional vectors have been discussed, it should be understood that multi-dimensional vectors can be applied similarly. For example, in addition to the presence or absence of features, other properties (such as the location and orientation of features) can also be included and represented by values ​​in a multi-dimensional vector. Thus, various attributes of defect pattern features can be captured and vectorized.

[0056] Grouping the GDS information of pattern images offers various advantages, such as better accuracy, efficiency, and speed. Defect patterns can be decomposed into vectors with multiple features, each corresponding to one or more attributes of the defect pattern. Therefore, instead of comparing the entire pattern, the classification network can utilize individual features. Thus, slight deviations from the identified pattern can still be captured and described, for example, by representing the presence of specific features to a defined degree.

[0057] Now for reference Figure 7 The figure is a schematic diagram illustrating an exemplary correlation between fixed-dimensional feature vectors, 3D representations of pattern groups, and GDS information of defect pattern images consistent with embodiments of the present disclosure.

[0058] like Figure 7 As shown, extracted GDS information from a defect pattern image 700 can be provided. Pattern image 700 may include patterns 70001, 70002, 70003, ..., 70010. Pattern image 700 may be represented in black and white, with multiple polygons representing various defects. In some embodiments, the pattern image may be represented in color. A graph 710 can be generated that graphically represents multiple feature vectors generated from pattern image 700. The lines representing the feature vectors may be displayed in multiple colors. Feature vectors can be used for pattern grouping.

[0059] For example, one or more defects can be identified during a defect inspection operation. In some embodiments, defect inspection may include identifying multiple defect candidates and distinguishing actual defects from harmful defects among these candidates. Pattern data may be extracted corresponding to defects or defect candidates. Pattern data may include portions of the scanned raw image corresponding to imaging features identified in the defect inspection, corresponding design patterns, etc. Multiple two-dimensional images of a predetermined window size may be present for analysis. In some embodiments, the GDS information of the defect pattern image may have been processed to represent the properties of the central polygon of the defect pattern. In some other embodiments, the GDS information of the defect pattern image may not have been processed in this way and may represent the properties of the central and peripheral polygons of the defect pattern image. Multiple defects may exist, including defects of different types and locations on the sample. Figure 7 The example shows ten separate defects.

[0060] Feature extraction architectures can be built and trained to extract multiple features that indicate defects. For example, in Figure 7 In this model, defect pattern grouping can be performed based on 32 features of the image pattern. It should be understood that the number of features to be extracted from different pattern images can vary. Features can be determined based on training data. The number of convolutional and pooling layers can vary to map to the feature extraction architecture. In a fixed-dimensional feature vector 720, there are 32 features along the horizontal axis (X-axis) to represent ten defect patterns shown in the defect pattern image 700 along the vertical axis (Y-axis). Different codes can be used to represent different values ​​of the features in the XY plane. For example, in... Figure 7 In the diagram, a denser fill pattern can represent the presence of a feature (or a high certainty of the presence of a feature), while a lighter fill pattern can represent the absence of a feature (or a high certainty of the absence of a feature).

[0061] Taking defect pattern 70010 as an example, this pattern has 32 values ​​along the X-axis for 32 features, where the 32 values ​​are represented by similar or different fill patterns. However, similar fill patterns for different features can be used to represent different values ​​or value ranges. For example, features 5 and 10 of defect pattern 70010 both have dense fill patterns, but they can represent significantly different values ​​or value ranges for the corresponding features, or they can represent similar values ​​or value ranges for the corresponding features.

[0062] Taking feature 20 as an example, this feature has ten coded blocks along the Y-axis, where each of the ten coded blocks can represent a value or value range of a defect pattern associated with that block. However, different defect patterns can have different or similar values ​​or value ranges of the feature. It should be understood that the value or value range of a single feature may not be used to identify a defect pattern, but combinations or sequences of values ​​or value ranges of all features can be used in the architecture to identify defect patterns. Such combinations or sequences of values ​​or value ranges of all features can be fixed-dimensional vectors, for example, they can be used as one-dimensional (1D) vectors for bucket IDs.

[0063] Figure 7 The three-dimensional (3D) representation of the pattern grouping 710 can be used to visually represent the idea of ​​1D vector bucket IDs, which are combinations or sequences of values ​​or value ranges representing the defect pattern. Each wavy line can be used to reflect the defect pattern. For example, the feature vector can be represented by a polynomial function with a degree equal to the number of features. Each wavy line has multiple (e.g., 32) features. Therefore, a polynomial of an appropriate degree can be used to represent the wavy line. The value or value range of each of these 32 features can be represented as follows: Figure 7 The values ​​shown are normalized, with all values ​​normalized to the range of zero to one. Different defect patterns can have similar or different values ​​or ranges of values ​​for the features. Combinations or sequences of values ​​or ranges of values ​​for all features can be used to identify defect patterns and distinguish them from other feature patterns.

[0064] Now for reference Figure 8 , Figure 8 A flowchart illustrating an exemplary determination method consistent with embodiments of this disclosure is shown. The controller can be programmed to implement this method. Figure 8 The flowchart includes one or more blocks. The controller may be coupled to the charged particle beam device. The controller may control the operation of the charged particle beam device. The controller may be configured to receive information collected from the detectors of the charged particle beam device and may be configured to perform processing functions. For example, controller 109 may be configured to perform... Figure 8 All or part of the methods.

[0065] In step S101, the method may begin. Charged particle imaging may be performed to acquire an image of the wafer. Imaging may include scanning the wafer to image it, which may include scanning a portion of the wafer or the entire wafer. One or more raw scanned images of the imaging area may be acquired and stored. In step S102, defect inspection may be performed on the acquired images. In step S102, features on the wafer may be identified. Features may include defects or potential defects. Defect inspection may include performing image analysis on the acquired images of the wafer surface using image processing techniques. Defect identification may include determining the location on the wafer. Step S102 may include wafer mapping. In some embodiments, defect re-inspection or other forms of verification may be performed after initial defect identification.

[0066] In step S103, pattern data can be determined corresponding to the identified defects. Pattern data can be extracted from the wafer design plan based on the location information collected in step S102. The wafer design plan can be stored in advance. The wafer design plan can be based on the wafer's GDS representation. Therefore, the extracted pattern data can be GDS information.

[0067] In step S104, training can be performed. In some embodiments, pre-training occurs before or after S102, or before or after S103, among other things. During pre-training, one or more random or selected regions of the GDS data can be “cut out” and used for pre-training. “Cut out” means that all GDS data at all hierarchical levels within the selected region are identified and merged into a GDS cell containing all data from the “cut out” region and no data from outside the “cut out” region. The location of the cut out region is identified, such as a center point, and the location or cut out region is automatically labeled or otherwise identified with the shape type of the region. Examples of shape types include contacts, vias, lines, line corners, etc. These cut out regions of the GDS, along with their labels, are used to pre-train a linear classifier network.

[0068] Training can also occur at S104 and can utilize the pre-trained linear classifier network discussed above. Training in S104 may include: acquiring pattern data for multiple defects and classifying the patterns into multiple buckets. The buckets may correspond to known defects. The training dataset may be stored in advance. In some embodiments, data obtained from step S103 may be added to the training dataset, or only the data from step S103 may be used to perform training. Training may continue until a certain level of convergence is reached. In this way, a classifier network can be constructed. The classifier network may be a linear classifier. The parameters of the classifier network can be determined based on the training. For example, multiple vectors corresponding to features associated with known defects can be determined.

[0069] In step S105, classification can be performed. Classification may include, for example, running the classifier network constructed in previous steps on new data using the pre-trained linear classifier network discussed above, or by utilizing the classification network of S104. For instance, the new pattern obtained in step S103 can be input into the classifier network. The input pattern can be a defect pattern image that can pass through multiple layers of the classifier network. This can include convolution, subsampling, pooling, etc. When running the classifier, as in step S106, a feature vector output can be obtained. In step S106, a feature vector representing the input pattern can be generated.

[0070] In step S107, pattern grouping may be performed. Pattern grouping may include classifying defects into multiple categories. Categories may be defect types, such as particles, concavities, scratches, crystal defects, haze, voids, and thinning. In some embodiments, categories may be specific to certain geometries. For example, a category may be a four-finger terminal type defect. Furthermore, in some embodiments, categories are not limited to geometries. Features may be single-array or multi-array vectors. Determining the group to which a pattern belongs may be based on feature vectors generated for each pattern. For example, a pattern may be assigned to a group based on a bucket ID, where the bucket ID may be based on the feature vector. The feature vector of a pattern may be converted into a bucket ID.

[0071] Multiple bucket IDs can be stored in a database. If a new pattern's bucket ID corresponds to a bucket ID that already exists in the database, the pattern can be associated with a defect type and grouped accordingly. In some embodiments, if a bucket ID for an imaging pattern does not exist in the database, a new group can be formed. Therefore, it can be determined that a pattern does not belong to any existing group, and a new defect group is created corresponding to the pattern's bucket ID. The bucket ID can be added to the database corresponding to the defect's bucket ID.

[0072] Pattern grouping may include comparing multiple input patterns to each other. For example, it may be determined that two or more patterns have the same bucket ID. In response, the two or more patterns may be assigned to a group. The newly created group may correspond to a defect. The bucket IDs of the two or more patterns may be added to a database corresponding to the bucket IDs of the defect. Furthermore, in some embodiments, when multiple patterns are input, a flag may be generated if all patterns are grouped into the same bucket. Additionally, a flag may be generated if all patterns are found to have different bucket IDs. In response to generating a flag, an indication may be displayed to the user, or further processing may be performed, such as resampling the multiple patterns.

[0073] In some embodiments, pattern grouping may include determining that a pattern corresponds to a harmful defect if the pattern's feature vector does not correspond to any predetermined vector associated with a real defect. For example, a pattern may be considered a harmful defect if its bucket ID does not match any bucket ID associated with a defect.

[0074] Pattern grouping may further include dividing patterns into main groups and subgroups. In some embodiments, buckets may correspond to sub-features of defects. Patterns may be divided into multiple buckets corresponding to sub-features. Further grouping may then be performed based on the sub-features. For example, it may be determined that a pattern having a certain combination of sub-features may correspond to a bucket in the main group. Because similar pattern images may share similar feature distributions, it may be useful to determine the feature distribution based on multiple buckets.

[0075] In step S108, data can be displayed to the user. The data display may include an organized presentation of defects based on pattern grouping results. For example, a wafer may be displayed where all defects of a certain type belong to the same pattern group, while other defects are suppressed. Organized display based on pattern grouping can be beneficial, for example, in identifying systematic defects.

[0076] In step S109, the process can end. Further processing can then be performed. Alternatively, the process can return to S101 and be repeated under the same or different imaging conditions.

[0077] In some embodiments, the detector may communicate with a controller that controls the charged particle beam system. The controller may instruct components of the charged particle beam system to perform various functions, such as controlling charged particle sources to generate a charged particle beam and controlling deflectors to scan the charged particle beam across a sample. The controller may also perform post-processing functions, brightness / contrast adjustment, image subdivision, image processing, contour generation, overlaying indicators onto acquired images, etc. The controller may combine, for example, the functions of image acquisition unit 120 and controller 109. Additionally, the controller may include a storage device, such as storage unit 130. A non-transitory computer-readable medium may be provided storing instructions for a processor of controller 109 to perform charged particle beam inspection, run a classifier network, perform pattern grouping, or other functions and methods consistent with this disclosure. Common forms of non-transitory media include, for example, floppy disks, floppy disks, hard disks, solid-state drives, magnetic tape or any other magnetic data storage media, CD-ROMs, any other optical data storage media, any physical media with a perforated pattern, RAM, PROMs and EPROMs, FLASH-EPROMs or any other flash memory, NVRAMs, caches, registers, any other memory chips or cassette tapes and their networking versions.

[0078] The following terms may also be used to describe the embodiments.

[0079] 1. A method for group mapping, comprising:

[0080] Receive the image of the first pattern;

[0081] A first fixed-dimensional feature vector is generated using the trained model parameters, which are based on the received image; and

[0082] Assign a first bucket identifier (ID) to the first fixed-dimensional feature vector.

[0083] 2. According to the method of Clause 1, assigning the first bucket ID to the first fixed-dimensional feature vector further includes:

[0084] In response to determining that the first pattern does not belong to one of the multiple buckets corresponding to the defect pattern, a new bucket ID is created for the first fixed-dimensional feature vector.

[0085] 3. According to the method of Clause 1, assigning the first bucket ID to the first fixed-dimensional feature vector further includes:

[0086] In response to determining that the first pattern belongs to one of a plurality of buckets corresponding to the defect pattern, the first fixed-dimensional feature vector is mapped to the first bucket ID.

[0087] 4. The method according to Clause 2 or Clause 3, wherein the defect pattern includes GDS information associated with the defect.

[0088] 5. The method according to Clause 4, wherein the defect pattern includes information derived from the GDS information, including the number of sides, the number of angles, the size, the shape, or a combination thereof.

[0089] 6. The method according to any one of clauses 1 to 3, wherein the fixed-dimensional feature vector is a one-dimensional feature vector.

[0090] 7. The method according to any one of clauses 1 to 3, wherein the trained model parameters are obtained by:

[0091] Obtain multiple images of multiple patterns with assigned bucket IDs; and

[0092] Train the model parameters for the deep learning network.

[0093] 8. According to the method of Clause 7, the trained model parameters are further obtained by:

[0094] The parameters for the deep learning network application are used to locate a single polygon in the center of one of a plurality of images.

[0095] 9. A method for pattern grouping, comprising:

[0096] The first feature in the acquired image of the identified sample;

[0097] Determine the pattern data corresponding to the first feature;

[0098] The feature vector corresponding to the first feature is determined by a classifier; and

[0099] The pattern group of the first feature is determined based on the feature vector.

[0100] 10. The method according to Clause 9, wherein the pattern data includes a sample design plan.

[0101] 11. The method according to Clause 9, wherein the classifier is a linear classifier, the method further comprising:

[0102] The linear classifier is trained using multiple defect patterns with multiple vectors corresponding to defect features.

[0103] 12. The method pursuant to Clause 11 also includes:

[0104] The plurality of defect patterns are obtained from the storage device.

[0105] 13. The method pursuant to Clause 11 also includes:

[0106] Identify a plurality of first features in the image; and

[0107] Determine the pattern data corresponding to the plurality of first features.

[0108] The plurality of defect patterns include one or more of the first features.

[0109] 14. The method pursuant to any one of clauses 9 to 13 further includes:

[0110] Convert the feature vector into a bucket ID; and

[0111] Based on the bucket ID, the first feature is classified into the pattern group among multiple groups.

[0112] 15. The method pursuant to any one of clauses 9 to 14 further includes:

[0113] When the feature vector of the first feature does not match any feature vector stored in the database, a new pattern group for the first feature is determined.

[0114] 16. The method pursuant to any one of clauses 9 to 14 further includes:

[0115] When the feature vector of the first feature matches a feature vector stored in the database, the first feature is added to the group associated with the feature vector.

[0116] 17. A system for pattern grouping, comprising:

[0117] Charged particle beam device, including detector;

[0118] An image acquisition device includes circuitry for receiving a detection signal from a detector and constructing an image including a first feature; and

[0119] A controller having at least one processor and a non-transitory computer-readable medium, the non-transitory computer-readable medium including instructions that, when executed by the processor, cause the system to perform the following operations:

[0120] Determine the pattern data corresponding to the first feature;

[0121] The feature vector corresponding to the first feature is determined by a classifier; and

[0122] The pattern group of the first feature is determined based on the feature vector.

[0123] 18. The system according to Clause 17, wherein the pattern data includes a sample design plan.

[0124] 19. A system pursuant to Clause 17, wherein the classifier is a linear classifier, and the instruction further causes the system to perform the following operations:

[0125] The linear classifier is trained using multiple defect patterns with multiple vectors corresponding to defect features.

[0126] 20. The system according to Clause 19, wherein the plurality of defect patterns are obtained from a storage device.

[0127] 21. A system pursuant to Clause 19, wherein the instructions further cause the system to perform the following operations:

[0128] Identify a plurality of first features in the image; and

[0129] Determine the pattern data corresponding to the plurality of first features.

[0130] The plurality of defect patterns include one or more of the first features.

[0131] 22. A system pursuant to Articles 17 to 21, wherein said instructions also cause said system to perform the following operations:

[0132] Convert the feature vector into a bucket ID; and

[0133] The pattern group is classified into multiple groups based on the bucket ID of the first feature.

[0134] 23. A system pursuant to clauses 17 to 22, wherein said instructions also cause said system to perform the following operations:

[0135] When the feature vector of the first feature does not match any feature vector stored in the database, a new pattern group is determined for the first feature.

[0136] 24. The system pursuant to Articles 17 to 22, wherein said instructions further cause said system to perform the following operations:

[0137] When the feature vector of the first feature matches a feature vector stored in the database, the first feature is added to the group associated with the stored feature vector.

[0138] 25. A non-transitory computer-readable medium storing an instruction set executable by one or more processors of a system to cause the system to perform a method, the method comprising:

[0139] Acquire images of the samples;

[0140] The first feature in the image is identified;

[0141] Determine the pattern data corresponding to the first feature;

[0142] A feature vector corresponding to the first feature is determined using a linear classifier; and

[0143] The pattern group of the first feature is determined based on the feature vector.

[0144] According to Clause 25, the computer-readable medium wherein the pattern data includes a sample design plan.

[0145] 26. The computer-readable medium pursuant to Clause 25, wherein said instruction set also causes the system to perform the following operations:

[0146] The linear classifier is trained using multiple defect patterns with multiple vectors corresponding to defect features.

[0147] 27. The computer-readable medium pursuant to Clause 27, wherein the plurality of defect patterns are obtained from a storage device.

[0148] 28. The computer-readable medium pursuant to Clause 27, wherein said instruction set also causes the system to perform the following operations:

[0149] Identify a plurality of first features in the image; and

[0150] Determine the pattern data corresponding to the plurality of first features.

[0151] The plurality of defect patterns include one or more of the first features.

[0152] 29. A computer-readable medium according to any one of clauses 25 to 29, wherein said instruction set also causes the system to perform the following operations:

[0153] Convert the feature vector into a bucket ID; and

[0154] Based on the bucket ID, the first feature is classified into the pattern group among multiple groups.

[0155] 30. A computer-readable medium according to any one of clauses 25 to 30, wherein said instruction set also causes the system to perform the following operations:

[0156] When the feature vector of the first feature does not match any feature vector stored in the database, a new pattern group is determined for the first feature.

[0157] 31. A computer-readable medium according to any one of clauses 25 to 30, wherein said instruction set also causes the system to perform the following operations:

[0158] When the feature vector of the first feature matches a feature vector stored in the database, the first feature is added to the group associated with the stored feature vector.

[0159] 32. A method for group mapping, comprising:

[0160] A sample-based graph data system (GDS) pre-trained linear classifier network, wherein the pre-training includes:

[0161] Identify the region-associated portion of the GDS.

[0162] Generate label data for the portion of the GDS that indicates the location of the region and the shape type of the polygon data associated with the portion of the GDS, and

[0163] The linear classifier network is pre-trained based on the portion of the GDS and the labeled data;

[0164] Receive the image of the first pattern;

[0165] A first fixed-dimensional feature vector is generated using the trained model parameters, which are based on the received image; and

[0166] Assign a first bucket identifier (ID) to the first fixed-dimensional feature vector.

[0167] 33. The method according to Clause 33, wherein the location of the area is the center point of the area.

[0168] 34. The method according to Clause 33, wherein the area is a cut-out area.

[0169] 35. The method according to Clause 33, wherein the GDS of the sample is the GDS of an integrated circuit to be manufactured on a wafer.

[0170] The block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer hardware / software products according to various exemplary embodiments of the present disclosure. In this regard, each block in the diagrams may represent some arithmetic or logical operation that can be implemented using hardware such as electronic circuitry. A block may also represent a module, code segment, or portion of code that includes one or more executable instructions for implementing a specified logical function. It should be understood that in some alternative implementations, the functions indicated in the blocks may not occur in the order shown in the drawings. For example, depending on the functionality involved, two blocks shown consecutively may be executed or implemented substantially in parallel, or the two blocks may sometimes be executed in reverse order. Some blocks may also be omitted. For example, in some embodiments, step S104 may be omitted. This may be because, if sufficient training has already been performed, the classifier network can be used directly without further training.

[0171] It should also be understood that each block in the block diagram, and combinations thereof, can be implemented by a system based on dedicated hardware that performs the specified function or action, or by a combination of dedicated hardware and computer instructions. While the invention has been described in conjunction with various embodiments, other embodiments of the invention will be apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The specification and examples are intended to be considered exemplary only, wherein the true scope and spirit of the invention are indicated by the following claims.

Claims

1. A method for group mapping, comprising: Receive the image of the first pattern; A first fixed-dimensional feature vector is generated using the trained model parameters, which are based on the received image; as well as Assign a first bucket identifier (ID) to the first fixed-dimensional feature vector. The allocation of the first bucket ID to the first fixed-dimensional feature vector also includes: In response to determining that the first pattern does not belong to one of the multiple buckets corresponding to the defect pattern, a new bucket ID is created for the first fixed-dimensional feature vector.

2. The method according to claim 1, wherein assigning the first bucket ID to the first fixed-dimensional feature vector further comprises: In response to determining that the first pattern belongs to one of a plurality of buckets corresponding to the defect pattern, the first fixed-dimensional feature vector is mapped to the first bucket ID.

3. The method of claim 1, wherein the defect pattern includes GDS information associated with the defect.

4. The method of claim 3, wherein the defect pattern comprises information derived from the GDS information, the information comprising the number of edges, the number of angles, size, shape, or a combination thereof.

5. The method according to claim 1, wherein the fixed-dimensional feature vector is a one-dimensional feature vector.

6. The method of claim 1, wherein the trained model parameters are obtained by: Obtain multiple images of multiple patterns with assigned bucket IDs; and Train the model parameters for the deep learning network.

7. The method of claim 6, wherein the trained model parameters are further obtained by: The parameters for the deep learning network application are used to locate a single polygon in the center of one of a plurality of images.

8. The method according to claim 1, further comprising: Sample-based graph data system (GDS) pre-trained linear classifier network.

9. The method of claim 8, wherein the pre-training of the linear classifier network comprises: Identify the region-associated portion of the GDS. Label data is generated for the portion of the GDS, the label data indicating the location of the area and indicating the shape type of the polygon data associated with the portion of the GDS, and The linear classifier network is pre-trained based on the portion of the GDS and the labeled data.

10. A system for pattern grouping, comprising: Charged particle beam device, including detector; An image acquisition device includes circuitry for receiving a detection signal from the detector and constructing an image including a first feature; as well as A controller having at least one processor and a non-transitory computer-readable medium, the non-transitory computer-readable medium including instructions that, when executed by the processor, cause the system to perform the following operations: Determine the pattern data corresponding to the first feature; The feature vector corresponding to the first feature is determined by a classifier; as well as The pattern group of the first feature is determined based on the feature vector. The classifier includes trained model parameters for use in the method according to claim 1.

11. The system of claim 10, wherein the pattern data includes a sample design plan.

12. The system of claim 10, wherein the classifier is a linear classifier, and the instruction further causes the system to perform the following operations: The linear classifier is trained using multiple defect patterns with multiple vectors corresponding to defect features.

13. The system of claim 12, wherein the plurality of defect patterns are retrieved from the storage device.

14. A non-transitory computer-readable medium storing an instruction set, the instruction set being executable by one or more processors of a system to cause the system to perform a method, the method comprising: Acquire images of the samples; Identify the first feature in the image; Determine the pattern data corresponding to the first feature; A feature vector corresponding to the first feature is determined using a linear classifier; as well as The pattern group of the first feature is determined based on the feature vector. The linear classifier includes trained model parameters for use in the method according to claim 1.

Images