JP2021140693A - Method for detecting defect on test piece and its system - Google Patents

Abstract

To provide a system for detecting a defect on a test piece, and a method.SOLUTION: A method partitions one or more portions of a first die, receives one or more noise maps indicating a noise distribution in a second image which is captured with respect to one or more portions of a second die, and performs segmentation to each noise map at a runtime. In the segmentation, a prescribed noise map is aligned with a region, each region is associated with noise data which are aligned in the region, and a score with respect to the prescribed region is calculated on the basis of the noise data which are associated with at least the region. The method acquires one set of the segments corresponding to one or more regions, respectively, which are associated with the same segmentation label by making each region associated with one segmentation label out of one set of the segmentation labels indicating noise level which is defined in advance on the basis of the score.SELECTED DRAWING: Figure 2

Description

The subject matter of the present disclosure relates generally to the field of investigation of specimens, and more specifically to methods and systems for detecting defects on specimens.

Today's demand for high densities and performance associated with very large integrations of manufactured devices reduces features to micron or smaller, increases transistor and circuit speeds, and improves reliability. Is needed. As semiconductor processes evolve, pattern dimensions such as line widths and other types of marginal dimensions are also constantly shrinking. This is also called a design rule. Such requirements require the formation of device features with high accuracy and uniformity, and forming such device features includes both finished and / or unfinished devices. Even when the device is in the form of a semiconductor wafer, it is necessary to monitor the manufacturing process, including frequent and detailed inspection of the device.

As used herein, the term "test piece" refers to all types of wafers, masks, and other structures used to manufacture semiconductor integrated circuits, magnetic heads, flat panel displays, and other semiconductor products. , These combinations and / or parts should be broadly interpreted.

Unless otherwise stated, the term "investigation" as used herein should be broadly construed to include the detection and / or classification of all types of defects within an object. Investigation is provided, for example, by using a non-destructive investigation tool during or after the production of the object to be investigated. As a non-limiting example, the survey process uses one or more survey tools to provide scans (single or multiple scans), sampling, review, measurement in connection with an object or part thereof. , Classification, and / or other actions can be included. Similarly, the survey can be provided prior to the manufacture of the object to be surveyed and can include, for example, generating a survey recipe. It should be noted that unless otherwise stated, the term "investigation" or its derivatives as used herein is not limited in terms of the size of the area inspected, the speed or resolution of the scan, or the type of investigation tool. .. Various non-destructive investigation tools include, but are not limited to, optical tools, scanning electron microscopes, atomic force microscopes, and the like.

The survey process can include multiple survey steps. During the manufacturing process, these investigation steps can be performed multiple times, for example, after the production or processing of a particular layer. In addition or otherwise, each survey step can be repeated multiple times, for example for different wafer positions or with different survey settings for the same wafer position.

As a non-limiting example, a two-step procedure can be used in run-time investigations, for example, a specimen inspection followed by a review of sampled defects. During the inspection step, the surface of the test piece or portion thereof (eg, area of interest, hotspot, etc.) is typically scanned at relatively high speed and / or low resolution. The captured inspection image is analyzed to detect defects and obtain their location and other inspection attributes. In the review step, at least some images of defects detected during the inspection phase are typically captured at relatively slow and / or high resolution, thereby at least some classification of defects, and optional selection. Allows other analyzes. In some cases, both stages can be performed by the same inspection tool, and in some other cases, these two stages are performed by different inspection tools.

Investigation generally involves detecting light or electrons from a wafer by directing light or electrons to the wafer to generate some output (eg, an image, signal, etc.) to the wafer. After the output is generated, defect detection is typically performed by applying defect detection methods and / or algorithms to the output. In most cases, the goal of the study is to suppress the detection of unwanted material and noise on the wafer while providing high sensitivity to defects of interest.

Improving the sensitivity of defect detection is needed in the art.

According to a particular aspect of the subject matter disclosed herein, a computerized system for defect detection on a specimen, comprising a processing unit operably connected to an investigation tool, the processing unit being a memory and a memory. With a processor operably coupled to the processing unit, the processing unit is configured to perform partitioning on each part of one or more parts of the first die of the test piece. Partitioning for a given part of the parts is i) the image data that characterizes the given part, thereby producing multiple regions in the image space, and ii) characterizes the given part. Executed based on at least one of the design data thereby resulting in multiple regions in the design space, the processing unit was captured against one or more parts of the second die of the test piece. Receiving one or more noise maps showing the noise distribution on one or more second images from the research tool at runtime, the first die and the second die being characterized by the same design data. It is further configured to receive, receive, and perform segmentation for each of one or more noise maps, and segmentation for a given noise map of one or more noise maps. Is to calculate a score for each region of multiple regions, a given noise map is aligned with multiple regions, and each region is associated with noise data aligned within that region. Scores for a given region of multiple regions are calculated, at least based on the noise data associated with that region, and each region is given a different noise level based on the calculated score. Associates with one of a set of predefined segmented labels indicating, thereby obtaining a set of segments corresponding to one or more regions associated with the same segmented label. A computerized system is provided in which a set of segments can be used for defect detection on a specimen based on a given noise map.

In addition to the above features, the system according to this aspect of the subject matter disclosed herein provides any one or more of the features (i)-(xii) listed below that are technically possible. Can be included in combination or permutation.
(I) Partitioning can be performed based on design data, with multiple regions within the design space corresponding to one or more die regions within a given portion having the same design pattern, respectively. Can be a design group of.
(Ii) The system further comprises a research tool configured to capture image data, including a first image representing a given portion. Partitioning can be performed on the basis of image data, with multiple regions in the image space being valued at corresponding positions in the attribute space specified by a set of attributes that characterize the first image. Based on that, it can be obtained on the first image.
(Iii) The second die can be a different die than the first die, and the research tool can display one or more second images representing one or more parts of the second die. It can be further configured to capture at runtime and provide one or more noise maps showing the noise distribution on one or more second images.
(Iv) The first die can be a reference die used to perform compartmentalization, the second die can be an inspection die, and compartmentalization can be performed at the setup stage. Can be done.
(V) The first die can be a reference die used to inspect the second die, the second die can be an inspection die, and compartmentalization can be performed at runtime. can.
(Vi) The second die can be the first die and one or more second images are one or more captured for one or more parts of the first die. It can be the first image of, and the partitioning can be done at runtime.
(Vii) The investigation tool can be an inspection tool configured to scan the specimen to capture image data and one or more second images.
(Viii) A set of attributes can be selected from a bank of candidate attributes containing one or more predefined attributes and one or more attributes generated using machine learning.
(Ix) A set of attributes can contain one or more attributes generated using machine learning, and the processing unit uses a machine learning model to generate one or more attributes. Can be further configured.
(X) Machine learning models can be trained by generating training attributes and predicting noise using the training attributes, comparing the predicted noise to the reference noise generated by the defect detection algorithm. The machine learning model is optimized so that the trained machine learning model can characterize the first image and generate one or more attributes representing its spatial pattern showing different noise levels. Can be transformed into.
(Xi) Associate each region with one segmented label involves rating the calculated score for each region and grouping multiple regions into a set of segments based on this rating. Can be done.
(Xii) Performing defect detection can include configuring a detection threshold for each segment.

According to another aspect of the subject matter disclosed herein, a computerized method of defect detection on a test piece, compartmentalized for each portion of one or more parts of a first die of the test piece. The partitioning of a given part of one or more parts is i) image data that characterizes a given part, thereby producing multiple regions in the image space. And ii) Performing partitioning, which is performed based on at least one of the design data that characterizes a given part, thereby giving rise to multiple regions in the design space, and the first of the test pieces. Receiving one or more noise maps showing the noise distribution on one or more second images captured for one or more parts of the two dies from the investigative tool at runtime. One or more noises, including receiving one die and a second die characterized by the same design data, and performing segmentation for each of one or more noise maps at runtime. Segmentation of a map for a given noise map is the calculation of a score for each region of a plurality of regions, where a given noise map is aligned with the regions and each region is its own. Associated with noise data aligned within a region, the score for a given region of multiple regions is calculated, calculated, and calculated based on at least the noise data associated with that region. Based on the score, each region is associated with one of a set of predefined segmented labels indicating different noise levels, and thus one or more associated with the same segmented label. A computerized method is provided in which a set of segments can be used for defect detection on a specimen based on a given noise map, performed by obtaining a set of segments corresponding to each of the regions of NS.

This aspect of the subject matter disclosed is any desired combination or permutation of any of the features (i)-(xii) listed above with respect to the system, with the necessary modifications. Can be included in.

According to another aspect of the subject matter disclosed herein, in a non-temporary computer-readable storage medium that, when executed by a computer, tangibly implements a program of instructions that causes the computer to perform a method of defect detection on a specimen. Thus, the method is to perform partitioning on each part of one or more parts of the first die of the test piece, for a given part of one or more parts. Partitioning is i) image data that characterizes a given part, thereby creating multiple areas in the image space, and ii) characterizes a given part, thereby creating multiple areas in the design space. Performing partitioning, which is performed based on at least one of the resulting design data, and one or more captured for one or more parts of the second die of the specimen. Receiving one or more noise maps showing the noise distribution on the second image from the research tool at runtime, with the first die and the second die being characterized by the same design data. Segmentation for a given noise map in one or more noise maps involves performing segmentation for each of one or more noise maps at runtime. To calculate the score for a region, a given noise map is aligned with multiple regions, each region is associated with noise data aligned within that region, and given of the regions. A set of predefined scores for each region is calculated based on at least the noise data associated with that region, and each region shows a different noise level based on the calculated score. Performed by associating one of the segmented labels with a segmented label, thereby retrieving a pair of segments corresponding to one or more regions associated with the same segmented label. Segments provide a non-temporary computer-readable storage medium that can be used for defect detection on specimens based on a given noise map.

This aspect of the subject matter disclosed is any desired combination or permutation of any of the features (i)-(xii) listed above with respect to the system, with the necessary modifications. Can be included in.

Embodiments will be described below with reference to the accompanying drawings only as non-limiting examples so that the invention can be understood and how the invention can be practiced in practice.

FIG. 3 is a block diagram of a system of defect detection on a specimen according to a particular embodiment of the subject matter disclosed herein. It is a schematic flow chart of defect detection on a test piece according to a specific embodiment of the subject matter disclosed herein. It is an exemplary schematic of a design group according to a particular embodiment of the subject matter disclosed herein. FIG. 6 is a schematic block diagram that trains a machine learning model to be able to generate attributes that represent spatial patterns that indicate different noise levels for a particular embodiment of the subject matter disclosed herein. It is a figure which shows an example of the inspection image by a specific embodiment of the subject disclosed herein and a set of attributes which characterize an inspection image. It is an exemplary schematic of the aligned noise map and design data according to a particular embodiment of the subject matter disclosed herein. It is a figure which shows the example of segmentation by a specific embodiment of the subject matter disclosed herein. FIG. 5 illustrates an example of using segmentation in a run-time study according to a particular embodiment of the subject matter disclosed herein.

The following detailed description describes a number of specific details to provide a thorough understanding of the present invention. However, it will be appreciated by those skilled in the art that the subject matter disclosed herein can be practiced without these specific details. In other examples, well-known methods, procedures, components, and circuits are not described in detail in order not to obscure the subject matter disclosed herein.

Unless otherwise stated, throughout the discussions herein, "execute," "partition," "capture," "receive," "calculate," "calculate," as will be apparent from the discussion below. Align, provide, associate, generate, get, superimpose, scan, use, apply, train, train By using terms such as "rating" and "composing", it refers to the behavior and / or process of a computer that manipulates and / or transforms data into other data, such data being electronic quantities, etc. It is understood that expressed as a physical quantity of and / or such data represents a physical object. The term "computer" is used as a non-limiting example of any type having data processing power, including a computerized system for detecting defects on specimens and parts thereof disclosed in the present application, and processing units within the system. Broadly interpreted to include hardware-based electronic devices in.

The terms "non-temporary memory" and "non-temporary storage medium" as used herein are broadly construed to include any volatile or non-volatile computer memory suitable for the subject matter disclosed herein. sea bream.

The term "defect" as used herein should be broadly construed to include any type of anomaly or unwanted feature or void formed on or within a specimen.

The term "design data" as used herein should be broadly construed to include any data that indicates the hierarchical physical design (layout) of a test piece. Design data can be provided by each designer and / or derived from physical design (eg, by complex simulations, simple geometry and Boolean operations, etc.). Design data can be provided in different formats, such as GDSII format, OASIS format, as a non-limiting example. The design data can be presented in vector format, grayscale intensity image format, or the like.

Unless otherwise stated, it is understood that the particular features of the subject matter disclosed herein, which are described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, the various features of the subject matter disclosed herein that are described in the context of a single embodiment can be provided separately or in any suitable partial combination. The following detailed description describes a number of specific details to provide a thorough understanding of the method and equipment.

With this in mind, note FIG. 1 showing a block diagram of a defect detection system on a specimen according to a particular embodiment of the subject matter disclosed herein.

The system 100 shown in FIG. 1 can be used for defect detection on a test piece (eg, a wafer, a die on a wafer, and / or a portion thereof). As mentioned above, the term "test piece" as used herein refers to all types of wafers, masks used to manufacture semiconductor integrated circuits, magnetic heads, flat panel displays, and other semiconductor products. , Reticles, and other structures, combinations and / or parts thereof. According to certain embodiments, the test strips used herein can be selected from the group including wafers, reticles, masks, integrated circuits, and flat panel displays (or at least some of them).

For purposes of illustration only, certain embodiments described below are provided with respect to dies and wafers. The embodiments also apply to other types, sizes, and representations of test pieces.

According to certain embodiments, the system 100 may include one or more survey tools 120, or may be operably connected to one or more survey tools 120. As used herein, the term "investigation tool" is a non-limiting example of imaging, scanning (single or multiple scanning), sampling, review, provided in connection with a specimen or portion thereof. It should be broadly interpreted to include any tool that can be used in research-related processes, including measurement, classification, and / or other processes. One or more survey tools 120 may include one or more inspection tools and / or one or more review tools. In some cases, at least one of the survey tools 120 scans the test piece (eg, the entire wafer, the entire die, or a portion thereof) and scans the inspection image (typically) for the detection of potential defects. It can be an inspection tool configured to capture (at relatively high speed and / or low resolution). In some cases, at least one of the investigation tools 120 will capture at least some review images of the defects detected by the inspection tool to see if the potential defects are actually defects. It can be a review tool configured in. Such review tools are typically configured to inspect die fragments one at a time (typically at relatively slow and / or high resolution). The inspection tool and the review tool can be different tools located in the same or different positions, or a single tool operating in two different modes. In some cases, at least one survey tool can have weighing capability.

Although the scope of the present disclosure is not limited in any way, it should be noted that the survey tool 120 can be implemented as various types of inspection machines such as an optical imaging machine and an electron beam inspection machine.

According to certain embodiments, the survey tool 120 can be configured to capture image data, including one or more images for one or more parts of the die of the test piece. The survey tool 120 can be further configured to acquire one or more noise maps showing the noise distribution on one or more images. These images can be obtained from different research modalities and the present disclosure is not limited by the inspection and weighing techniques used to generate the images. In some embodiments, the survey tool 120 can be an inspection tool configured to scan the specimen to capture one or more images. As described in more detail with respect to FIG. 2, in some cases, the captured image of the specimen is processed to generate a defect map showing suspicious locations on the specimen that are likely to be defect of interest (DOI). (For example, it can be processed by an image processing module, the function of which can be incorporated within the research tool 120 or processing unit 102, or implemented as a stand-alone computer). The defect map is also referred to herein as a noise map because DOIs are relatively rare and most of the suspicious defects reflected in the defect map are more likely to be noise or false alarms.

As used herein, the term "defect of interest (DOI)" refers to any real defect to be detected that the user is interested in. For example, any "killer" defect that can cause a yield loss can be shown as a DOI, and in contrast, annoying types of defects are also real defects but can affect yield. Not, and therefore should be ignored.

The term "noise" as used herein refers to any unwanted or uninteresting defects (also called non-DOI or annoyances), as well as various variations during inspection (eg, process variations, color variations, machinery). It should be widely interpreted as including random noise caused by (such as target and electrical fluctuations).

The system 100 may include a processing unit 102, which is operably connected to the input / output interface 126 and the survey tool 120. The processing unit 102 is a processing circuit configured to provide all the processing necessary to operate the system 100, which will be further described in detail below with reference to FIG. The processing unit 102 includes a processor (not shown separately) and a memory (not shown separately). The processor of the processing unit 102 can be configured to execute some functional modules in response to computer-readable instructions executed on the non-temporary computer-readable memory contained within the processing unit. Hereinafter, such a functional module is referred to as being included in the processing unit 102.

Functional modules contained within the processing unit 102 may include partition modules 104 and segmentation modules 106 that are operably connected to each other. The compartment module 104 can be configured to perform compartmentalization for each portion of one or more portions of the first die of the test piece. Partitioning for a given portion of one or more portions of the first die of the test piece i) image data that characterizes the given portion, thereby producing multiple regions in the image space. , And ii) Can be performed based on at least one of the design data that characterizes a given portion, thereby giving rise to multiple regions in the design space.

According to certain embodiments, in some cases the system 100 can be operably connected to the design data server 110 (eg, CAD server) via the hardware-based I / O interface 126. The design data server 110 is configured to store and provide design data that characterizes the specimen. The design data of the test piece can be in the form of a physical design layout of the test piece (eg, CAD clip), a raster image, and a simulated image derived from the design layout. According to certain embodiments, the input / output interface 126 can be configured to receive design data from the design data server 110 that characterizes a given portion of one or more parts of the die. As described in more detail below with reference to FIGS. 2 and 3, when partitioning is performed based on design data, multiple regions within the design space are given portions having the same design pattern. It can be a plurality of design groups corresponding to one or more of the die areas.

The segmentation module 106 examines one or more noise maps showing the noise distribution on one or more second images captured for one or more parts of the second die of the test piece. It can be configured to receive from the tool at runtime. The first die and the second die are characterized by the same design data. The segmentation module 106 can be further configured to perform segmentation for each of the one or more noise maps. Specifically, segmentation for a given noise map in one or more noise maps involves calculating a score for each region in the plurality of regions, and a given noise map is plural. Each region is associated with noise data aligned within that region. In some embodiments, the score for a given region of the plurality of regions can be calculated at least based on the noise data associated with that region.

After the scores have been calculated, segmentation of one or more noise maps for a given noise map is based on the calculated score and each region is predefined by a set indicating different noise levels. It can include associating one of the segmented labels with a segmented label, thereby obtaining a set of segments corresponding to one or more regions associated with the same segmented label. One set of segments constitutes segmented data that can be used for further investigation of the specimen. As an example, segmented data can be provided to the investigation tool 120 and / or any other investigation tool to detect defects on the specimen based on a given noise map. In some cases, segmented data can be used to adjust a given noise map (eg, by recalculating at least certain pixels in the difference and / or grade image). Defect detection can be performed on the test piece based on the generated noise map. As another example, the segmented data can be used by the processing unit to perform defect detection on the test piece based on a given noise map (in this case, the processing unit 102 is a defect detection module (in this case). (Not shown in FIG. 1) can be further provided). In some embodiments, the segmented data can include a segmented layout that can be sent to a computer-based graphical user interface (GUI) 124 to draw the results. Segmentation will be described in more detail below with reference to FIG.

According to certain embodiments, the system 100 may include a storage unit 122. The storage unit 122 can be configured to store any data required to operate the system 100, such as data related to the inputs and outputs of the system 100, as well as intermediate processing results produced by the system 100. As an example, the storage unit 122 can be configured to store images and / or derivatives thereof created by the research tool 120. Therefore, one or more images can be retrieved from the storage unit 122 and provided to the processing unit 102 for further processing. Alternatively, the storage unit 122 may be configured to store the design data of the test piece, such design data may be retrieved from the storage unit 122 and provided as input to the processing unit 102. Can be done. As an addition or alternative, the storage unit 122 can be configured to store partitioning results, i.e., multiple areas, such partitioning results being retrieved from the storage unit 122 and further to the processing unit 102. Can be provided for processing.

In some cases, system 100 is configured to store one or more external data repositories (and / or derivatives thereof) created by the research tool 120 and / or the design data server 110. It can be operably connected (not shown in 1). The image data, noise map, and / or design data of the test piece can be provided to the processing unit 102 for further processing.

In some embodiments, the system 100 may optionally include a computer-based graphical user interface (GUI) 124 configured to allow user-specified input associated with the system 100. For example, a visual representation of the test piece (eg, by a display that forms part of GUI 124) may be presented to the user, including design data and / or image data of the test piece. Through the GUI, the user can be provided with the option of defining specific operating parameters. The user can also see the segmentation results as well as other operation results such as defect detection results on the GUI.

FIG. 1 shows that the survey tool 120 is implemented as part of the system 100, but in certain embodiments, the functions of the system 100 can be implemented as a stand-alone computer and work with the survey tool 120. Note that it can be operably connected to the survey tool 120 as such. In such cases, the image data of the test piece can be received directly from the survey tool 120 or via one or more intermediate systems and can be provided to the processing unit 102 for further processing. In some embodiments, each feature of the system 100 can be incorporated, at least in part, into one or more survey tools 120, thereby facilitating and enhancing the functionality of the survey tool 120 in the survey-related process. can do. In such cases, components of the system 100, or at least a portion thereof, can form part of the survey tool 120. As an example, the partition module 104 and / or the segmentation module 106 can be implemented as part of the survey tool 120 or can be incorporated as part of the survey tool 120. As another example, the processing unit 102 and the storage unit 122 can form a part of the processing unit and the storage unit of the investigation tool 120, respectively, and the input / output interface and GUI of the investigation tool 120 are the input / output interface 126 and the GUI. It can function as GUI 124.

The teachings of the subject matter disclosed herein are not constrained by the system shown in FIG. It will be readily appreciated by those skilled in the art that it can be carried out in any suitable combination with the ware.

The system shown in FIG. 1 can be implemented in a distributed computing environment, and the aforementioned functional modules shown in FIG. 1 can be distributed to several local and / or remote devices over a communication network. Note that you can link. The research tool 120, storage unit 122, and GUI 124 are shown in FIG. 1 as being part of system 100, but in some other embodiments, at least some of the aforementioned units are system 100. It should be further noted that it can be implemented as being external to the system and can be configured to operate in data communication with the system 100 via the input / output interface 126.

With reference to FIG. 2 below, a schematic flow chart of defect detection on a test piece according to a particular embodiment of the subject matter disclosed herein is shown.

Compartmentalization can be performed for each portion of one or more portions of the first die of the test piece (202) (eg, by compartmentalization module 104 provided within the processing unit 102). Specifically, partitioning for a given part of one or more parts i) image data that characterizes a given part, thereby creating multiple regions in the image space, and ii. ) It can be performed based on at least one of the design data that characterizes a given part, thereby giving rise to multiple regions in the design space. In some embodiments, one or more parts can refer to one or more blocks of a die. The blocks can be of various sizes and dimensions, for example 100 * 200, 100 * 1000, 200 * 2000 pixels in image space, and the present disclosure should be construed as limited by their unique implementation. Note that it is not.

According to certain embodiments, in some cases compartmentalization can be performed based on design data. In such cases, the regions are obtained as multiple design groups within the design space, each corresponding to one or more die regions (within a given portion) having the same design pattern.

In the above case, the design data of a given part or one or more parts, or the design data of the first die can be received (eg, from the design data server 110 by the input / output interface 126). As mentioned above, the design data can be in the form of a physical design layout (eg, a CAD clip), a raster image, and a simulated image derived from the design layout. Die (or parts thereof) design data can include various design patterns of unique geometry and arrangement. A design pattern can be defined as being composed of one or more structural elements each having a geometry including contours (eg, one or more polygons).

In some embodiments, the received design data can represent multiple design groups, each corresponding to one or more die regions having the same design pattern. Clustering of design groups (that is, division of CAD data into a plurality of design groups) can be executed in advance, and design group information can be stored in the design data server 110 in advance. Therefore, the design data acquired by the system 100 can already include the design group information. In some cases, the design data received by the system 100 can include only design group information (eg, design coordinates of different design groups) without physical design layout (eg, CAD clip) information. In some other cases, the design data received by the system 100 can include both design group information and specific design layout information.

In these embodiments, the partitioning into a plurality of areas can be performed according to the design group information, for example, the plurality of areas can correspond to a plurality of design groups. Note that design patterns can be considered "same" when they are the same, or when they are very correlated or similar to each other. Various similarity measures and algorithms can be applied to align and cluster similar design patterns, and the present disclosure is limited by any specific measure used to derive the design group. Should not be interpreted.

In some embodiments, optionally, when the physical design layout of the die (or part thereof) is received from the design data server, the processing unit 102 of system 100 may perform clustering of the design groups. can. It should be noted that the partitioning described with reference to block 202 can be performed at the setup stage (ie, pre-production / runtime) or at the runtime stage.

With reference to FIG. 3 below, an exemplary schematic of a design group according to a particular embodiment of the subject matter disclosed herein is shown.

Design data for the die (or portion thereof) is schematically shown in FIG. 3 for illustration purposes only. Different types of "trees" represent different design patterns on the design data. After clustering / grouping similar design patterns, the design data is divided into four design groups 302, 304, 306, and 308, respectively, corresponding to one or more die regions having the same design pattern. Note that in some design groups, the design patterns are not exactly the same or identical, but have a high degree of similarity. For example, in design group 304, it can be seen that the patterns in the two regions on the left and the patterns in the region on the right are slightly different (eg, in opposite directions). As mentioned above, in some cases, the design data received by the system 100 can be in the form of a representation on the left side of FIG. 3, including grouping information as well as specific design layouts and patterns. In some other cases, the design data can be in the form of a representation on the right side of FIG. 3 containing only grouping information (eg, the position of the group in design coordinates). In this example, the regions obtained from the compartmentalization correspond to the four design groups 302, 304, 306, and 308.

It should be noted that the example shown in FIG. 3 is for illustration purposes only and should not be considered to limit the disclosure in any way. The actual design pattern of the die can be much more complex, so the grouping of similar design patterns can vary and in some cases can be much more complex than this example. It will be easily understood by those skilled in the art.

Continuing with the description of block 202 of FIG. 2, according to a particular embodiment, the partitioning is based on image data that characterizes a given portion of one or more portions of the first die. It can be done (for example, instead of the design data mentioned above). In such cases, the image data captured by the survey tool 120 may include a first image representing a given portion. The regions are obtained in the image space and are on the first image based on the values of the corresponding positions in the attribute space specified by the set of attributes that characterize the first image. To be acquired. In such cases, partitioning can be performed at the configuration stage or otherwise at runtime, depending on the various scenarios described in detail below with reference to block 204. Since the partitioning is performed on each of the one or more parts, the survey tool is an image data containing one or more first images representing one or more parts of the first die. Can be configured to capture. One advantage of using image data rather than design data is that there is no need to acquire and / or process the design data of the specimen. Because the design data is not always available, for example a particular customer may not be willing to provide such data, and in addition, especially consider the superposition between the design data and the image data. Then, the processing of the design data may be computationally complicated and difficult.

In some further embodiments, compartmentalization can optionally be performed based on both image data and design data.

As a continuation of the description of FIG. 2, the runtime segmentation process will be described below. One or more noise maps can be received at runtime (eg, by processing unit 102 from survey tool 120) (204), and one or more noise maps can be one or more of the second dies of the specimen. The noise distribution on one or more second images captured for the plurality of parts is shown. The first die and the second die are characterized by the same design data.

According to some embodiments, the second die can point to a different die than the first die, and one or more second images are at runtime (eg, by the research tool 120). It can be captured and can represent one or more parts of the second die. The noise map may be provided as showing the noise distribution on the second image (eg, by the research tool 120).

According to the embodiments described above, in some cases the second die can refer to an inspection die (ie, the die to be inspected at production / runtime) and the first die performs compartmentalization. It can refer to a test die or reference die used to do so, and partitioning is a pre-processing operation intended to be responsible for run-time segmentation. In such a case, the partitioning based on the image data described above with reference to block 202 will result in different image data (ie, test die or reference die) from the inspection image (ie, the second image) captured at runtime. Since it is based on the image of), it can be executed at the setting stage. In such cases, one of the advantages is to reduce the computational time and resources required for run-time inspection and / or detection, thereby improving system performance.

In some other cases, the second die can refer to the inspection die and the first die refers to the reference die used for inspection of the second die (eg, die-reference detection method). be able to. As an example, in the die-die detection method, the reference die can be an adjacent die of the inspection die on the wafer. In such a case, the partitioning based on the image data described above with reference to block 202 is based on the image data captured at runtime (eg, the image of the adjacent die) and can be performed at runtime. .. In such cases, one of the advantages is that the reference and inspection dies are adjacent / adjacent to each other, so that the first image (eg, multiple regions on the first image) and the second It is to make the alignment / overlay process with an image (eg, the noise map corresponding to the second image) easier, computationally easier and cheaper.

According to some further embodiments, the second die and the first die can refer to the same die. Thus, the one or more second images are actually the same as the one or more first images captured for one or more parts of the first die. In such cases, the second die and the first die can refer to the same inspection die. In such a case, the partitioning based on the image data described above with reference to block 202 can be performed at runtime because it is based on the inspection image captured at runtime. In such cases, one of the advantages is to avoid performing alignment / overlay between the first and second images, which can be computationally cumbersome and expensive as described above. Is.

In some embodiments, the investigation tool is configured to scan the specimen to capture image data (ie, a first image) and / or one or more second images. can do. As mentioned above, in some cases the inspection tool can be configured to scan the specimen with a unique scanning configuration. The scanning configuration is defined in various ways, such as lighting conditions, polarization, noise level per area (area is defined, for example, based on user / customer information or as related to a design pattern. It can include configuring one or more of the parameters such as), the detection threshold per area, and the noise intensity calculation method per area. In some cases, the inspection tool is specifically composed of sensitive parameters that allow sensitive scanning of the test piece. As an example, a sensitive scan can refer to sensitivity to spatial patterns that exhibit different noise levels and characteristics, so inspection images and / or noise maps obtained from a sensitive scan are indicated by such patterns and patterns. Noise information can be reflected. For example, a sensitive scan can be used to obtain a first image, so the compartmentalization can take into account more pattern information obtained from this scan.

According to a particular embodiment, a noise map showing the noise distribution on the inspection image (ie, the second image) can be obtained. In some cases, the noise map can be obtained using the detection threshold.

The noise map can be generated in various forms (eg, by means of a detection module and / or an image processing module, its functionality can be incorporated within the investigation tool 120 or processing unit 102). In some embodiments, the noise map can be generated by applying the detection threshold directly to the pixel values of the captured inspection image. In some other embodiments, the test image of the test piece can be further processed to generate a noise map. Different inspection and detection methods can also be applied to the processing of inspection images and the generation of noise maps, and the present disclosure is not limited by the particular detection techniques used herein. For illustration purposes only, some examples of defect detection and noise map generation based on inspection images will be described below.

In some embodiments, one or more reference images can be used for defect detection for each inspection image. Reference images can be obtained in various forms, and the number of reference images used herein and the method of obtaining such images should not be construed as limiting this disclosure in any way. In some cases, one or more reference images can be captured from one or more dies of the same test piece (eg, adjacent dies of inspection dies). In some other cases, one or more reference images are one or more dies of another specimen (eg, a second specimen that is different from the current specimen but shares the same design data). Can include one or more images captured from. As an example, in the die-history (D2H) test method, a test image can be captured from the current test piece at the current time (eg, t = t'), and one or more reference images can be baseline time. (For example, previous time t = 0) can include one or more previous images captured from one or more dies on the second specimen. In some further embodiments, the one or more reference images can include at least one simulated image representing a given die of the one or more dies. As an example, a simulated image can be generated based on die design data (eg, CAD data).

In some embodiments, at least one difference image can be generated based on the difference between the pixel values of the inspection image and the pixel values derived from one or more reference images. Optionally, at least one grade image can be generated based on at least one difference image. A grade image can consist of pixels having corresponding pixel values in the difference image and values calculated based on a predefined difference normalization factor. The predefined difference normalization coefficient can be determined based on the behavior of the normal population of pixel values and can be used to normalize the pixel values of the difference image. As an example, the pixel grade can be calculated as the ratio of the corresponding pixel value of the difference image to the predefined difference normalization coefficient. Noise maps can be generated by determining the location of suspicious defects (noise) using detection thresholds based on at least one difference image or at least one grade image.

In some embodiments of the subject matter disclosed herein, the detection threshold can include all noise information such that the noise map is reflected on at least one difference image or at least one grade image. It can be a threshold of zero.

The acquired noise map can show the noise distribution on the second image. In some embodiments, the noise distribution can include one or more noise characteristics of the noise in the noise map revealed by the detection process, such as the location of the noise (eg, on the inspection image). In addition, the noise property can further include at least one of the intensity and size of the noise. In some embodiments, the noise in the noise map can include pattern-related noise. This type of noise is related to the local density and complexity of the design pattern with which the noise is associated. The noise map can also include other types of noise, such as noise caused by investigative tools (eg shot noise), process variability, and color variability.

As mentioned above, when partitioning is performed on the basis of image data, within the image space, i.e., on the first image, within the attribute space specified by the set of attributes that characterize the first image. Multiple regions are obtained based on the value of the corresponding position of. According to a particular embodiment, a set of attributes is selected from a bank of candidate attributes containing one or more predefined attributes and one or more attributes generated using machine learning. Can be done. As an example, predefined attributes can be determined based on noise analysis results obtained from previous inspection and detection processes. In some cases, some of the predefined attributes can be obtained from theoretical modeling of existing noise sources, such as process variability, measurement noise, and algorithm noise. Some of the predefined attributes can be obtained from the spatial decomposition of fixed (data-independent) or data-dependent image-based features. In some cases, some of the predefined attributes can be obtained from known layers. Optionally, some attributes can be provided by the customer based on customer data.

According to certain embodiments, a set of attributes can include one or more attributes generated using machine learning. One or more attributes can be generated using a machine learning model (eg, by processing unit 102). The machine learning model can be trained by extracting training attributes that characterize the training image and using the extracted training attributes to predict noise in the training image. The predicted noise is then compared to the reference noise generated by the defect detection algorithm and feedback (eg, by loss function) is provided to the machine learning model to tune and optimize the model. After the model has been trained (for example, the difference between the predicted noise and the reference noise is within range), this model is used in production / runtime to generate / extract attributes for the input inspection image. The generated attributes can represent spatial patterns that show different noise levels within the input image.

With reference to FIG. 4 below, a schematic block diagram for training a machine learning model to be able to generate attributes representing spatial patterns indicating different noise levels according to a particular embodiment of the subject matter disclosed herein. It is shown.

For purposes of illustration, the machine learning (ML) model 400 includes attribute neural networks (shown as attribute convolutional neural network (CNN) 404 in FIG. 4) and decision neural networks (shown as decision CNN406) operably connected to each other. Can include. To train the ML model 400, a training image 402 is provided as an input to the model 400 each time. The training image is shown in FIG. 4 with dimensions of x * y. The attribute CNN404 can be configured to extract a set of training attributes (eg, a set containing N attributes) that characterizes the training image. In fact, each of the N attributes can be an attribute map (ie, x * y dimensions) that actually has the same dimensions as the training image, so for each pixel in the image, in the attribute map. Corresponds to the attribute value of. In other words, each pixel in the inspection image has N corresponding attribute values from its respective attribute map. Therefore, a set of N attributes is actually a dimension of x * y * n as shown in FIG.

A set of N attributes is provided as input to the decision CNN406. The determination CNN406 can be configured to predict noise in the inspection image using only one set of extracted attribute candidates. As an example, the predicted noise 408 can be in the form of a grade image (or difference image) corresponding to the above-mentioned training image showing noise / defect characteristics in the inspection image. For the purpose of evaluating the predicted noise, the training image 402 and one or more reference images 410 are subjected to the usual defect detection algorithm 412 (eg, die-reference detection method) in parallel and are also grade images (or grade images (or). The detection result in the form of (difference image) can be used as the reference noise 414 and compared with the predicted noise 408. Reference noise is actually used as a ground truth for evaluating noise prediction. This comparison result can be provided to the ML model 400 as feedback (eg, by loss function) so that the ML model 400 can learn to generate better attributes because it has better noise prediction results. The model can be tuned and optimized. Such a training process can be repeated with different training images until the model is properly trained (eg, the predicted noise is sufficiently close to the reference noise indicated by the comparison results).

After the ML model 400 has been trained, the trained attribute CNN404 is capable of generating a set of attributes that characterize each input inspection image, and the generated attributes are the noise in each input inspection image. Represents a spatial pattern that indicates a level. Therefore, the trained attribute CNN404 can be set or used at runtime for attribute generation for each input inspection image (eg, first image). These attributes provide a good representation of the spatial pattern indicating the noise level in each input inspection image, and are based on such attributes to perform the partitioning described above and the partitioned areas for such patterns. Can be obtained. Specifically, partitioning can be done based on the value of the corresponding position in the attribute space specified by a set of attributes.

As an example, attribute values for different positions on each attribute map can be binned, and bins from different attribute maps can be combined to form a partitioning result, i.e., multiple regions. However, the present disclosure is not limited to the specific method of performing compartmentalization. In addition to or instead of the above, other possible methods of partitioning the image based on a set of attributes can also be used.

With reference to FIG. 5, an example of an inspection image according to a particular embodiment of the subject matter disclosed herein and a set of attributes that characterize the inspection image is shown.

An inspection image 502 showing a portion of the die is shown. A set of attributes 504, 506, and 508 that characterize the inspection image 502 is generated. A set of attributes is generated, for example, by selecting from a bank of candidate attributes containing one or more predefined attributes and one or more attributes generated using machine learning, as described above. can do. For purposes of illustration and description, attribute 504 can indicate a circular pattern contained within the image, attribute 506 can indicate a pattern of horizontal lines contained within the image, and attribute 508 can indicate within the image. The pattern of the included vertical lines can be shown. As an example, these attributes can be generated by passing the inspection image through the corresponding feature extraction filter. The 510 demonstrates compartmentalization results showing multiple areas marked with different gray levels depending on the attached grayscale bar. As mentioned above, these regions can be generated in various ways based on a set of attributes. An example shown in this figure is to combine a set of attributes (eg, overlay a set of attributes on top of another and combine the values at the corresponding positions) to form a compartmentalized area. ..

It should be noted that the above examples of patterns, attributes, and compartmentalization are for illustration purposes only and should not be construed as limiting this disclosure in any way. For example, in some cases, the generated attributes may, or instead of, indicate other spatial patterns of different types and sizes, such as rectangular patterns, triangular patterns, and so on. In some cases, certain attributes, especially when generated by machine learning models, are not always able to demonstrate a direct representation of spatial patterns (for example, their correlation with spatial patterns is human). It may not be visible). However, when the machine learning model is so specially trained, these generated attributes can also indicate the noise level in the inspection image and are therefore within the scope of the present disclosure.

With reference to FIG. 2 again, segmentation for each of one or more noise maps can be performed at runtime (206) (eg, by segmentation module 106). Specifically, calculating the score for each region of the plurality of regions (208), and based on the calculated score, each region is divided into a set of predefined segments showing different noise levels. Segmentation of one of the labels Performing segmentation for a given noise map of one or more noise maps (ie, any given noise map) by associating it with a label (210). Can be done.

To calculate the score (208), a given noise map is such that each region of a plurality of regions falls within the noise data aligned within that region (ie, within the region obtained from the alignment). It needs to be aligned with multiple regions so that it can be associated with (noise data from the noise map). The score for a given region of a plurality of regions can be calculated at least based on the noise data associated with that region.

As mentioned above with reference to block 202, multiple regions can be obtained in the design space or image space based on the data used for partitioning. If multiple regions are multiple design groups in a design space corresponding to one or more die regions having the same design pattern, then each given design group of the multiple design groups is the design group. A given noise map needs to be aligned with the design data so that it can be associated with the noise data in the corresponding die region. Multiple regions are in the image space (eg, these regions are based on the value of the corresponding position in the attribute space specified by the set of attributes that characterize the first image, the first image. If obtained above and the first image is different from the second image), then the given noise map needs to be aligned with the image data (ie, the first image).

Alignment (similar to any of the above cases) can be performed at different stages. According to certain embodiments, the alignment can be performed in advance, eg, at the setup stage, by a different system, and the alignment result can be received by the system 100 for further processing. .. In some other embodiments, the alignment can be performed by the processing unit 102 at runtime by overlaying the noise map on multiple regions. The superposition process is performed according to any suitable alignment algorithm known in the art (eg, described in US Patent Application Publication No. 2007/0280527, US Patent Application Publication No. 2013/204569, etc.). can do.

As an example, when multiple regions are in the design space, the noise map in the inspection coordinates (the coordinates in the inspection space are called the inspection coordinates) is superimposed on the design data (for example, CAD clip), thereby the inspection space. You can get the design data coordinates in. Scanning conditions (eg, lighting), as well as imperfections, shifts, and obvious errors in the scanning process, electrical circuits printed on the wafer between the inspection coordinates of the noise map and the corresponding positions in the design coordinates. There are likely to be some differences due to various reasons, such as errors in the manufacture of the. Position calibration data can be generated showing the global (eg, average) offset between the noise map and the design data and / or multiple offsets associated with their particular region of interest or pattern or object. Optionally, the position calibration data can include data structures that specify their respective offsets for each object of interest (or groups thereof). The position calibration data can be stored in the memory contained in the processing unit 102 or the storage unit 122.

With reference to FIG. 6, an exemplary schematic of the aligned noise map and design data according to a particular embodiment of the subject matter disclosed herein is shown.

Using the superposition algorithm described above, the design group derived in FIG. 3 is aligned with the noise map. The noise or noise data in the noise map is shown in FIG. 6 as black dots 601 with different sizes located within different locations. The size can indicate the strength of the noise signal or the actual spatial size of the noise. For example, as described above in the detection process, if the noise map is generated based on the grade image, the intensity of the noise in the noise map can be represented by the pixel values corresponding to the noise in the grade image. After alignment, the four design groups 302, 304, 306, and 308 are associated with noise data (eg, represented by black dots) that fall within the die region corresponding to the design group, respectively. Aligned design groups with that associated noise data are marked as 602, 604, 606, and 608 and are inputs to the segmentation process as described below with reference to blocks 208 and 210. Provided as. In some cases, after alignment, the aligned design group with noise data can be located within the inspection coordinates.

Although the alignment described above with reference to FIG. 6 is exemplified as between the design data and the noise map, the regions 302, 304, 306, and 308 are similarly image space (ie, first). Note that a partitioned area within an image) can be represented and the alignment described above can also be applied between the noise map and the area within the first image.

Continuing the description of FIG. 2, as described above, the score for a given region among the plurality of regions can be calculated based on at least the noise data associated with that region. As an example, the score can be calculated as the noise density within a given area. The noise density can be calculated, for example, as the ratio of the amount of noise associated with a given area to the area of that area. As another example, the score can be derived based on the maximum / minimum pixel values shown in the noise data within that region.

According to certain embodiments, the score can be calculated based on the noise data associated with a given region and the defect budget allocated for the area of the given region. In the inspection and detection process, a total defect budget is typically assigned to the entire die. Total defect budget refers to the total amount of desired defect candidates expected to be detected after the inspection and detection process. Assuming the DOI has a uniform distribution on the die, the total defect budget for the entire die can be allocated to multiple regions of the die, eg, divided according to the area of those regions. For example, the defect budget allocated for a given area can be calculated as the product of the total defect budget for the entire die and the ratio of the area of the given area to the area of the entire die.

In some embodiments, a noise histogram can be created for a given region based on the noise data associated with that region. The score for a given region can be calculated as a threshold by applying the defect budget assigned to that given region on the noise histogram. As an example, a noise histogram can be created as a relationship between the number of pixels (y-axis) and the pixel values (x-axis) in a noise map indicating the intensity of noise / defects (eg, grade). By applying the assigned defect budget to the histogram, a threshold can be derived that separates the amount of suspicious DOI equal to the defect budget from the rest. This threshold can be used as a score for a given area.

Note that the above is a possible example of calculating the score, and other suitable methods may be applied instead of or in addition to the above.

In some cases, the result of compartmentalization, ie, duplication between multiple regions, may exist. This can be caused, for example, by overlapping spatial relationships between specific structures within different layers of specimens when the region is a design group, or, for example, the region is acquired based on image data. If so, it can be caused by the unique design of the partitioning algorithm. Therefore, after the noise map is aligned with multiple regions, when associating the noise data with those regions or calculating the score for those regions, consider the noise data that falls within the overlapping area between the different regions. There is a need. As an example, noise data located within the overlapping area between two (or more) regions should be counted only once within one region, and therefore, for example, when calculating the score for such regions. , It is necessary to determine which area this noise data actually belongs to and exclude that area from other overlapping areas.

After the scores are calculated for each region, each region is associated with one of a set of predefined segmentation labels that indicate different noise levels, based on the calculated score. (210), thereby obtaining a set of segments, each corresponding to one or more regions associated with the same segmented label. In some embodiments, a set of segments can constitute segmented data, which can be used for further investigation of the specimen. As an example, segmented data can be provided to survey tool 120 and / or any other survey tool or processing unit to perform defect detection on specimens based on a given noise map (212). .. In some embodiments, the segmented data can include a segmented layout that can be sent to a computer-based graphical user interface (GUI) 124 to draw the results.

In some embodiments, multiple regions can be rated according to their score in order to perform an association between the regions and the segmented labels, and based on this rating, these regions are rated as 1. It can be grouped into pairs of segments (ie, rated areas can be subdivided into a predefined set of segments based on their rating). As an example, a set of predefined segmented labels can include three labels indicating low noise, high noise, and ultra-high noise noise levels. By segmentation, each region can be labeled as one of a low noise region, a high noise region, and an ultra-high noise region, depending on the rating of each score out of all the scores. Note that the disclosure is not limited by the number of pre-defined segmented labels as a set.

The segmented data can be used in a variety of ways for further investigation of the specimen (eg, to perform defect detection on the specimen based on a given noise map). As an example, segmented data can be used to construct detection thresholds for each die segment. For example, a die segment labeled ultra-high noise can have a higher threshold than a die segment labeled low noise. As another example, segmented data can be used to construct the calculation of differential and / or grade images. For example, pixel values in differential and / or grade images can be normalized or adapted depending on the noise level of different segments.

With reference to FIG. 7, an example of segmentation according to a particular embodiment of the subject matter disclosed herein is shown below.

An inspection image 702 showing a portion of the die is shown (corresponding to inspection image 502 in FIG. 5). Two examples of segmentation results 704 and 706 are shown, each with two inspection images sharing the same design (eg, optionally taken from the same part in different dies sharing the same design data). For simplicity, FIG. 7 shows only one inspection image 702). As shown in both 704 and 706, two segments marked in black and white, respectively, are acquired, with the black segment pointing to the low noise segment and the white segment pointing to the high noise segment. These segments can be acquired based on the compartmentalization result 510 shown in FIG. For example, the multiple regions acquired in the partition result 510 can be aligned with the noise map acquired at runtime, and the score is calculated for each region according to the above description with reference to block 208. be able to. Each region can be assigned to a segmented label (low noise or high noise) based on the score according to the teachings described above with reference to block 210. As an example, the difference between the two segmentation results 704 and 706 can be obtained from different noise maps obtained from the survey tools corresponding to the two inspection images.

With reference to FIG. 8 below, an example of using segmentation in a run-time study according to a particular embodiment of the subject matter disclosed herein is shown below.

As shown in FIG. 8, two figures are shown for comparison purposes, the figure on the left shows the noise histogram 802 generated and used in run-time defect detection without segmentation, and the figure on the right shows the noise histogram 802. Two noise histograms 808 and 810 are shown that are generated in place of one noise histogram and are used in run-time defect detection using segmentation. Specifically, the two noise histograms 808 and 810 correspond to the two segments acquired via the runtime segmentation process described above (eg, the two segments shown in FIG. 7). Thus, instead of applying a single detection threshold 804 on the noise histogram 802, two different detection thresholds 812 and 814 can be assigned to the two histograms 808 and 810, respectively. In this example, the segment corresponding to the noise histogram 810 is relatively noisy than the segment corresponding to the noise histogram 808. Therefore, a higher threshold 814 is applied to the noise histogram 810 and a lower threshold 812 is applied to the noise histogram 808. By applying different thresholds to different segments, different detection sensitivities can be achieved for segments with different noise levels, which can improve overall detection sensitivity and defect detection rate. In particular, in this example, the DOI806 within the less noisy segment (shown on the left), which was previously ignored in the original defect detection process without segmentation, has a relatively low threshold applied within this segment. Therefore, it can be detected here (shown in the figure on the right) (for example, because the intensity (eg, grade) of this DOI is relatively low compared to the noise in the high noise segment).

According to a particular embodiment of the subject matter disclosed herein, a run-time segmentation process is proposed, where segmentation of specimens (eg, dies or parts thereof) is a noise map obtained in the run-time inspection and detection process. Run on. This is an advantage over segmentation performed during the setup stage on one or more test or reference dies that provide only statistical / average noise information. In the case of setup stage segmentation, after the segmentation is determined based on these dies, this becomes part of the inspection recipe that remains "quiet" (ie, immutable) and is specific to a particular die. Used at runtime to inspect each die, regardless of the noise characteristics of. In comparison, the run-time segmentation proposed herein is based on actual noise information within a particular inspected die (ie, run-time noise map) and is therefore die-specific, ie segmentation. It is dynamic and can vary from die to die, drawing more accurate segmentation results for specific dies, thereby improving detection sensitivity.

In addition, in some embodiments, as described above, in preparation for segmentation, partitioning of specimens into regions is performed without design data (ie, without CAD) based solely on image data. can do. This not only makes segmentation more adaptable and easy to use (especially when design data is not available), but was previously used to handle CAD and / or CAD related to image data. It is also possible to save the calculation time and resources that have been used.

In some further embodiments, partitioning can be performed based on specific image attributes generated using machine learning techniques. The machine learning model used in this embodiment is unique as described above so that it is possible to characterize a given inspection image to generate one or more attributes that represent its spatial pattern showing different noise levels. Can be trained in the following ways. This makes it possible to perform partitioning without CAD data while still providing accurate results.

It should be noted that the machine learning models referred to herein can be implemented for different machine learning architectures. As a non-limiting example, in some cases the machine learning model can be a deep learning neural network (DNN), and the layers of the DNN are convolutional neural network (CNN) architecture, recurrent neural network architecture, recurrence. It can be systematized according to the type neural network architecture, GAN architecture, and the like. It should be noted that the teachings of the subject matter disclosed herein are not constrained by a particular machine learning architecture.

The flow chart shown in FIG. 2 has been described with reference to the elements of the system 100, but it is by no means constrained and its actions can be performed by elements other than those described herein. Please also note.

It should be understood that the present invention is not limited to the details described herein or in the description set forth in the drawings in its application. Other embodiments are possible, and the invention can be implemented and practiced in a variety of ways. Therefore, it should be understood that the terms and terminology used herein are for explanatory purposes and should not be considered limiting. Therefore, the concepts on which this disclosure is based can be readily used as the basis for designing other structures, methods, and systems for carrying out some of the purposes of the subject matter disclosed herein. Those skilled in the art will understand.

It will also be appreciated that the system according to the invention can be implemented, at least in part, as a well-programmed computer. Similarly, the invention also contemplates computer programs that are computer readable to perform the methods of the invention. The present invention further contemplates a non-transitory computer-readable storage medium that tangibly implements a program of instructions that can be executed by a computer to perform the methods of the invention.

Those skilled in the art will be able to apply various modifications and modifications to the embodiments of the invention described herein without departing from the scope of the invention as defined in the appended claims. Will be easily understood.

100 System 102 Processing Unit 104 Partition Module 106 Segmentation Module 110 Design Data Server 120 Survey Tool 122 Storage Unit 124 Graphical User Interface 126 Input / Output Interface 302 Design Group 304 Design Group 306 Design Group 308 Design Group 400 Machine Learning (ML) Model 402 Training image 404 Attribute convolutional neural network 406 Decision neural network 408 Predicted noise 410 Reference image 412 Defect detection algorithm 414 Reference noise 502 Inspection image 504 Attribute 506 Attribute 508 Attribute 510 Partitioning result 601 Noise or noise data 602 Design group 604 Design group 606 Design group 608 Design group 702 Inspection image 704 Segmentation result 706 Segmentation result 808 Noise histogram 810 Noise histogram

Claims (20)

A computerized system for detecting defects on specimens,
The investigation tool comprises a processing unit operably connected, said processing unit comprising a memory and a processor operably coupled to the memory, said processing unit.
It is configured to perform compartmentalization for each portion of one or more portions of the first die of the test piece, and the compartmentalization for a given portion of said one or more portions. , I) Image data that characterizes the given portion and thereby produces multiple regions in the image space, and ii) Characterizes the given portion and thereby produces multiple regions in the design space. Performed based on at least one of the design data to be made
The investigation said that the processing unit showed one or more noise maps showing the noise distribution on one or more second images captured for one or more parts of the second die of the test piece. Receiving from the tool at runtime, the first die and the second die being characterized by the same design data,
The segmentation for a given noise map of the one or more noise maps is further configured to perform segmentation for each of the one or more noise maps.
To calculate a score for each region of the plurality of regions, the given noise map is aligned with the plurality of regions, and each region is associated with noise data aligned within the region. And that the score for a given region of the plurality of regions is calculated, at least based on the noise data associated with the region.
Based on the calculated score, each region is associated with one of a set of predefined segmentation labels indicating different noise levels, thereby associating one with the same segmentation label. Performed by retrieving a set of segments, each corresponding to one or more regions,
A computerized system in which the set of segments can be used for defect detection on the specimen based on the given noise map. The partitioning is performed based on the design data, and the plurality of areas in the design space correspond to one or more die areas in the given part having the same design pattern. The computerized system according to claim 1, which is a group. Further comprising said investigation tool configured to capture said image data including a first image representing said given portion, said partitioning is performed based on said image data and said in image space. According to claim 1, a plurality of regions are acquired on the first image based on the values of the corresponding positions in the attribute space specified by the set of attributes that characterize the first image. The computerized system described. The second die is a different die than the first die, and the investigative tool displays the one or more second images representing said one or more parts of the second die. The computerized system of claim 3, further configured to capture at runtime and provide the one or more noise maps showing the noise distribution on the one or more second images. 4. The first die is a reference die used to perform the compartmentalization, the second die is an inspection die, and the compartmentalization is performed in the setup stage, claim 4. Described computerized system. 4. The fourth die, wherein the first die is a reference die used to inspect the second die, the second die is an inspection die, and the compartmentalization is performed at runtime. Computerized system. The second die is the first die, and the one or more second images are one or more captured with respect to the one or more portions of the first die. The computerized system of claim 3, which is a first image, wherein the partitioning is performed at runtime. The computerized system according to claim 1, wherein the investigation tool is an inspection tool configured to scan the test piece to capture the image data and the one or more second images. The third aspect of claim 3, wherein the set of attributes is selected from a bank of candidate attributes containing one or more predefined attributes and one or more attributes generated using machine learning. Computerized system. Such that the set of attributes includes the one or more attributes generated using machine learning and the processing unit uses the machine learning model to generate the one or more attributes. The computerized system according to claim 9, further comprising. The machine learning model is trained by generating training attributes and predicting noise using the training attributes, comparing the predicted noise to the reference noise generated by the defect detection algorithm. The machine learning model is designed so that the trained machine learning model can generate the one or more attributes that represent its spatial pattern that characterizes the first image and exhibits different noise levels. The computerized system according to claim 10, which is optimized. The association of each region with one segmented label comprises rating the calculated score for each region and grouping the plurality of regions into the set of segments based on the rating. , The computerized system according to claim 1. The computerized system of claim 1, wherein performing the defect detection constitutes a detection threshold for each segment. A computerized method of defect detection on a specimen performed by a processing unit operably connected to an investigation tool, wherein the processing unit comprises a memory and a processor operably coupled to the memory. The above method
Performing a compartmentalization for each portion of the first die of the test piece, the compartmentalization for a given portion of the one or more portions. i) Image data that characterizes the given portion and thereby produces multiple regions in the image space, and ii) characterizes the given portion and thereby produces multiple regions in the design space. Performing partitioning, which is performed based on at least one of the design data,
One or more noise maps showing the noise distribution on one or more second images captured for one or more parts of the second die of the test piece from the survey tool to runtime Receiving, where the first die and the second die are characterized by the same design data,
The segmentation for a given noise map of the one or more noise maps comprises performing segmentation for each of the one or more noise maps at runtime.
It is to calculate the score for each region of the plurality of regions, the given noise map is aligned with the plurality of regions, and each region is associated with noise data aligned within the region. And that the score for a given region of the plurality of regions is calculated, at least based on the noise data associated with the region.
Based on the calculated score, each region is associated with one of a set of predefined segmentation labels indicating different noise levels, thereby associating one with the same segmentation label. Performed by retrieving a set of segments, each corresponding to one or more regions,
A computerized method in which the set of segments can be used for defect detection on the test piece based on the given noise map. The investigation tool further comprises capturing the image data, including a first image representing the given portion, the partitioning being performed on the basis of the image data, and the plurality of said in the image space. 14. The 14th aspect, wherein the region is acquired on the first image based on the value of the corresponding position in the attribute space specified by the set of attributes that characterize the first image. Computerization method. The second die is a die different from the first die, and the method captures the one or more second images representing said one or more parts of the second die. 15. The computer of claim 15, further comprising capturing at runtime by a research tool and providing the one or more noise maps showing the noise distribution on the one or more second images. How to make it. 16. The first die is a reference die used to perform the compartmentalization, the second die is an inspection die, and the compartmentalization is performed in the setup stage, claim 16. Described computerization method. The set of attributes includes one or more attributes generated using machine learning, further comprising the method using a machine learning model to generate the one or more attributes. , The computerization method according to claim 15. The machine learning model is trained by generating training attributes and predicting noise using the training attributes, comparing the predicted noise to the reference noise generated by the defect detection algorithm. The machine learning model is designed so that the trained machine learning model can generate the one or more attributes that represent its spatial pattern that characterizes the first image and exhibits different noise levels. The computerization method according to claim 18, which is optimized. A non-temporary computer-readable storage medium that, when executed by a computer, tangibly executes a program of instructions that causes the computer to perform a method of detecting defects on a test piece.
Performing a compartmentalization for each portion of the first die of the test piece, the compartmentalization for a given portion of the one or more portions. i) Image data that characterizes the given portion and thereby produces multiple regions in the image space, and ii) characterizes the given portion and thereby produces multiple regions in the design space. Performing partitioning, which is performed based on at least one of the design data,
Receiving at runtime a noise map showing the noise distribution on one or more second images captured for one or more parts of the second die of the test piece. , The first die and the second die are characterized by the same design data, receiving and
The segmentation for a given noise map of the one or more noise maps comprises performing segmentation for each of the one or more noise maps at runtime.
It is to calculate the score for each region of the plurality of regions, the given noise map is aligned with the plurality of regions, and each region is aligned in the noise data. To calculate, the score for a given region of the associated regions is calculated based on at least the noise data associated with the region.
Based on the calculated score, each region is associated with one of a set of predefined segmentation labels indicating different noise levels, thereby associating one with the same segmentation label. Performed by retrieving a set of segments, each corresponding to one or more regions,
A non-temporary computer-readable storage medium in which the set of segments can be used for defect detection on the specimen based on the given noise map.

Images