Unsupervised defect segmentation
By combining classifier training with signal attributes and context attributes in the semiconductor chip inspection system, identifying and segmenting defect groups are solved, and defect segmentation in the prior art is achieved, and defect identification and classification with high accuracy is achieved.
Patent Information
- Application Number
- CN201980087510.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-01-08
- Filing Date
- 2019-12-30
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2039-12-30
AI Technical Summary
Existing defect inspection systems face segmentation challenges with a large number of data points when identifying and classifying defects on semiconductor wafers, especially in the case of design rules shrinkage and noise floor change, segmentation inadequate based solely on signal or context attributes Or invalid, it is difficult to effectively identify defect groups.
By using a processor controller in a defect verification system to receive a verification data set containing signal attributes and context attributes, classify the data based on the location in the signal space, and train the classifier to segment the data into a defect group, combining signal attributes and Context attributes identify defect groups with overlapping distributions.
Unsupervised and automatic defect segmentation is achieved, the accuracy of defect discovery, classification and disturbance point filtering is improved, the selection of optical devices and the optimization of manufacturing processes is promoted, and the limitations of human intervention and data structured are overcome.
Smart Images

Figure CN113260855B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates generally to defect inspection and, more particularly, to unsupervised segmentation of defects. Background Art
[0002] Defect inspection systems identify and / or classify defects on semiconductor wafers. A given semiconductor wafer may include hundreds of chips, each containing thousands of components of interest, and each component of interest may have millions of instances on a given layer of the chip. Therefore, inspection systems may generate a large number of data points on a given wafer. Therefore, properly segmenting these data points into different groups to efficiently generate useful information is an ongoing challenge for defect classification. For example, defect segmentation may facilitate various aspects of the manufacturing process (including but not limited to defect discovery, defect classification, disturber point filtering, or selection of optical devices) to achieve a desired level of performance.
[0003] Defect segmentation is typically performed as a post-inspection step and is often performed with the assistance of human experts (e.g., supervised segmentation). However, the demands for ever-shrinking devices have resulted in increased demands on inspection systems. For example, as design rules shrink, the noise floor associated with the defect of interest relative to the difference between disturber inspection data points may also shrink. Therefore, segmentation based solely on attributes or contextual attributes of the signal generated at the defect site may be insufficient or ineffective. Further, the number of attributes that can be used for defect segmentation may make supervised segmentation ineffective or impractical. Therefore, it would be desirable to provide a system and method for addressing disadvantages such as those identified above. Summary of the invention
[0004] An inspection system according to one or more illustrative embodiments of the present disclosure is disclosed. In one illustrative embodiment, the system includes a controller having one or more processors. In another illustrative embodiment, the controller receives an inspection data set associated with inspection of one or more samples from a defect inspection system, wherein the inspection data set associated with a defect includes values of two or more signal attributes and values of one or more context attributes. In another illustrative embodiment, the signal attribute represents a signal generated by the defect inspection system in response to a defect. In another illustrative embodiment, a context attribute of the one or more context attributes represents one or more additional characteristics of the defect. In another illustrative embodiment, the controller labels each of the inspection data sets with a class label based on a respective position of each of the inspection data sets in a signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space. In another illustrative embodiment, the controller segments the plurality of inspection data sets into two or more defect groups by training a classifier using the values of the one or more context attributes and corresponding class labels for the inspection data sets, wherein the two or more defect groups are identified based on the trained classifier.
[0005] An inspection system according to one or more illustrative embodiments of the present disclosure is disclosed. In one illustrative embodiment, the system includes a defect inspection system. In another illustrative embodiment, the defect inspection system includes an illumination source for generating one or more illumination beams. In another illustrative embodiment, the defect inspection system includes a detector for capturing radiation emitted from a sample in response to the one or more illumination beams. In another illustrative embodiment, the system includes a controller communicatively coupled to the detector. In another illustrative embodiment, the controller receives an inspection data set associated with inspection of one or more samples from a defect inspection system, wherein the inspection data set associated with a defect includes values of two or more signal attributes and values of one or more context attributes. In another illustrative embodiment, the signal attribute represents a signal generated by the defect inspection system in response to a defect. In another illustrative embodiment, the context attribute represents one or more additional characteristics of the defect. In another illustrative embodiment, the controller labels each of the inspection data sets with a class label based on a respective position of each of the inspection data sets in a signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space. In another illustrative embodiment, the controller segments the verification data set into two or more defect groups by training a classifier using the values of the one or more contextual attributes and corresponding class labels for the verification data set, wherein the two or more defect groups are identified based on the trained classifier.
[0006] A method for segmenting metrology defects according to one or more illustrative embodiments of the present disclosure is disclosed. In one illustrative embodiment, the method includes receiving an inspection data set associated with inspection of one or more samples from a defect inspection system, wherein the inspection data set associated with the defect includes values of two or more signal attributes and values of one or more context attributes. In another illustrative embodiment, the signal attribute represents a signal generated by the defect inspection system in response to the defect. In another illustrative embodiment, the context attribute represents one or more additional characteristics of the defect. In another illustrative embodiment, the method includes labeling each of the inspection data sets with a class label based on a respective position of each of the inspection data sets in a signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space. In another illustrative embodiment, the method includes segmenting the inspection data set into two or more defect groups by training a classifier using the values of the one or more context attributes and corresponding class labels for the inspection data set, wherein the two or more defect groups are identified based on the trained classifier.
[0007] It should be understood that both the foregoing summary and the following detailed description are exemplary and explanatory only and do not necessarily limit the invention as claimed.The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the summary, serve to explain the principles of the invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Several advantages of the present disclosure may be better understood by those skilled in the art by referring to the accompanying drawings, in which:
[0009] Figure 1A is a conceptual view illustrating an inspection system according to one or more embodiments of the present disclosure.
[0010] Figure 1B is a conceptual view illustrating an optical defect inspection system according to one or more embodiments of the present disclosure.
[0011] Figure 1C is a conceptual diagram of a defect inspection system configured as a particle beam metrology tool according to one or more embodiments of the present disclosure.
[0012] Figure 2 is a flow chart illustrating steps performed in a method for segmenting a defect according to one or more embodiments of the present disclosure.
[0013] Figure 3 is a flow chart illustrating sub-steps associated with labeling each of the test data sets with a class label based on the respective position of each of the test data sets in signal space in accordance with one or more embodiments of the present disclosure.
[0014] Figure 4A is a plot of inspection data associated with a plurality of defects mapped into a two-dimensional signal space based on two signal attributes in accordance with one or more embodiments of the present disclosure.
[0015] Figure 4B is a plot of a signal space partitioned into a set of non-overlapping polygons, each with a different class label, in accordance with one or more embodiments of the present disclosure.
[0016] Figure 5A is a plot illustrating inspection data including defect groups having overlapping signal clouds in signal space that are distinguishable based on contextual attributes in accordance with one or more embodiments of the present disclosure.
[0017] Figure 5B According to one or more embodiments of the present disclosure Figure 5A Histogram of the test data as a function of the context attributes in .
[0018] Figure 6 According to one or more embodiments of the present disclosure, by using Figure 5A A decision tree is generated by training a decision tree classifier based on the contextual attribute information of the test data in and the corresponding category labels associated with the non-overlapping regions of the signal space.
[0019] Figure 7 is a decision tree associated with verification data including 11 context attributes, only 6 of which are used to segment the verification data, according to one or more embodiments of the present disclosure. DETAILED DESCRIPTION
[0020] Reference will now be made in detail to the disclosed subject matter illustrated in the accompanying drawings. The present disclosure has been particularly shown and described with respect to certain embodiments and specific features thereof. The embodiments set forth herein are to be considered illustrative and not restrictive. The embodiments set forth herein are to be considered illustrative and not restrictive. Those of ordinary skill in the art will readily appreciate that various changes and modifications in form and details may be made without departing from the spirit and scope of the present disclosure.
[0021] Embodiments of the present disclosure relate to systems and methods for unsupervised and automatic segmentation of defects. In this regard, defect segmentation can occur without the assistance of a user (e.g., a human expert) or without the use of training data. It should be recognized that unsupervised defect detection can overcome limitations associated with human intervention during segmentation, can promote consistent segmentation, and can lead to tool competence.
[0022] It should be recognized herein that a defect inspection system may generate an inspection data set associated with each defect (or potential defect) on a sample. This inspection data set may include various inspection data points, each of which may be applicable in defect segmentation techniques. For example, the inspection data set may include one or more signal attributes or attributes of a signal measured by the defect inspection system that is associated with a defect (e.g., associated with a location on the sample corresponding to the defect), such as, but not limited to, the energy, magnitude, or polarity of the measured signal. The signal attributes may further include parameters associated with a defect detection algorithm, such as, but not limited to, spotlikeness or multi-die automatic thresholding (MDAT) offset values. By way of another example, the inspection data set may include one or more context attributes associated with the defect. For example, the context attributes may include information associated with a relationship between a defect and a surrounding area, such as, but not limited to, the relative brightness (e.g., polarity) of the defect relative to the surrounding area, the relative or absolute location of the defect on the sample, the proximity of the defect to a feature or area of interest, the brightness of a reference image used to detect the defect, or the roughness associated with a reference image used to detect the defect. The context attributes may further include parameters associated with the defect detection algorithm, such as, but not limited to, MDAT grayscale or MDAT noise level.
[0023] Embodiments of the present disclosure are directed to performing unsupervised segmentation by structuring the inspection data as a supervised classification problem, training a supervised classifier (e.g., supervised learning technique), and segmenting the inspection data based on patterns or structures identified during training. In this regard, any of a wide range of supervised learning techniques may be used to identify different defect groups based on multi-dimensional patterns between signal attributes and context attributes of the inspection data. Thus, supervised learning techniques may be utilized to provide unsupervised defect segmentation.
[0024] It should be recognized herein that inspection data may be segmented based on signal attributes. For example, inspection data may be mapped to a multi-dimensional signal space defined by two or more signal attributes and segmented into different defect groups based on location in the signal space. For example, a clustering algorithm may be used to identify clusters of inspection data sets in the signal space to identify defect groups with similar signal attributes. However, it may be the case that segmentation based solely on signal attributes may not be sufficient to identify all desired defect groups (e.g., defect groups with different noise characteristics, different degrees of impact on performance, or the like). For example, multiple defect types may have overlapping distributions in the signal space. Further, as manufacturing technology moves toward smaller design rules, the signal-to-noise ratios of multiple types of defects may become increasingly similar, such that identification of a particular DOI from other signals (e.g., a disturber signal) may become increasingly challenging.
[0025] It should be further recognized that the inspection data can be segmented based on context attributes. For example, the inspection data can be plotted in a histogram according to one or more context attributes and segmented based on the dips in the histogram. However, defect segmentation based solely on context attributes may similarly be insufficient to distinguish desired defect groups.
[0026] Additional embodiments of the present disclosure relate to segmenting inspection data based on both signal attributes and context attributes. In some embodiments, inspection data is labeled using class labels (e.g., class labels suitable for use in supervised learning algorithms) based on locations in signal space. For example, a multidimensional signal space may be divided into a set of non-overlapping regions, each of which may be assigned a unique class label. Further embodiments of the present disclosure relate to training a supervised classifier based on context attributes and corresponding class labels and segmenting inspection data into defect groups based on the trained classifier. Thus, a classifier may be trained to correlate patterns in the context attributes of the inspection data with specific locations within the signal space associated with the assigned class labels. These patterns in the context attributes may then be used to segment the inspection data into defect groups corresponding to separable regions in the signal space. In this regard, the trained classifier may utilize context attributes to identify separate defect groups that may have overlapping distributions in the signal space. Thus, highly granular defect groups may be identified, which may facilitate defect discovery, defect classification, disturbing point filtering, optical device selection, and the like.
[0027] Further, reducing the multi-dimensional signal space to a single dimension (set of class labels) enables the use of any supervised learning technique, such as (but not limited to) decision tree technology, nearest neighbor technology, regression technology, Bayesian technology, support vector machine technology, similarity technology or neural network technology.
[0028] Defect segmentation as disclosed herein may be performed by any combination of systems. For example, defect segmentation may be performed by an inspection system adapted to inspect one or more samples for defects. By way of another example, defect segmentation may be performed by one or more dedicated controllers communicatively coupled to at least the inspection system. In this regard, at least a portion of the defect segmentation operation may be offloaded from the inspection tool.
[0029] Additional embodiments of the present disclosure relate to providing inspection data including defect segmentation data as feedback or feed-forward data for any process step, including but not limited to defect discovery, defect classification, disturber point filtering, recipe control, or optics selection. For example, inspection data including defect segmentation information about one or more samples may be utilized to generate an inspection recipe (e.g., including definition of inspection locations, setup parameters for inspection, or the like) for one or more additional samples. By another example, inspection data including defect segmentation information about samples within a batch may be provided as feedback data for process steps for additional samples within the batch or additional batches. By further example, inspection data including defect segmentation information about samples within a batch may be provided as feed-forward data for subsequent process steps associated with the same samples in the batch.
[0030] Figure 1Ais a conceptual view illustrating an inspection system 100 according to one or more embodiments of the present disclosure. In one embodiment, the inspection system 100 includes a defect inspection subsystem 102. For example, the defect inspection subsystem 102 may generate an image of one or more portions of a sample. In another embodiment, the inspection system 100 includes a controller 104. In another embodiment, the controller 104 includes one or more processors 106 configured to execute process instructions maintained on a memory device 108 (e.g., a memory). In this regard, the one or more processors 106 of the controller 104 may perform any of the various process steps described throughout the present disclosure. For example, the controller 104 may (but is not limited to): receive defect inspection data associated with any number of samples; divide a signal space defined by any number of signal attributes within the inspection data into different regions; assign a class label to the inspection data based on which region of the signal space the inspection data maps to; train a classifier (e.g., a supervised classifier, a supervised learning technique, or the like) to correlate context attributes with locations in the signal space based on the class label; or identify defect groups based on the trained classifier.
[0031] The one or more processors 106 of the controller 104 may include any processing element known in the art. In this sense, the one or more processors 106 may include any microprocessor type device configured to execute algorithms and / or instructions. In one embodiment, the one or more processors 106 may be composed of a desktop computer, a main computer system, a workstation, an image computer, a parallel processor, or any other computer system (e.g., a network computer) configured to execute a process (which is configured to operate the inspection system 100), as described throughout this disclosure. It should be further recognized that the term "processor" may be broadly defined to cover any device having one or more processing elements that execute process instructions from a non-transitory memory device 108. Further, the steps described throughout this disclosure may be implemented by a single controller 104 or alternatively multiple controllers. In addition, the controller 104 may include one or more controllers housed in a common housing or multiple housings. In this way, any controller or combination of controllers may be separately packaged as a module suitable for integration into the inspection system 100. Further, the controller 104 may analyze the data received from the detector 126 and feed the data to additional components within the defect inspection subsystem 102 or external to the inspection system 100 .
[0032] The memory device 108 may include any storage medium known in the art that is suitable for storing process instructions that can be executed by the associated one or more processors 106. For example, the memory device 108 may include a non-transitory memory medium. By way of another example, the memory device 108 may include, but is not limited to, a read-only memory, a random access memory, a magnetic or optical storage device (e.g., a disk), a tape, a solid-state drive, and the like. It should be further noted that the memory device 108 may be housed in a common controller housing with the one or more processors 106. In one embodiment, the memory device 108 may be remotely located relative to the physical location of the one or more processors 106 and the controller 104. For example, the one or more processors 106 of the controller 104 may access a remote memory (e.g., a server) that is accessible via a network (e.g., the Internet, an intranet, and the like). Therefore, the above description should not be interpreted as limiting the present invention but is merely illustrative.
[0033] Reference Figure 1B and 1C , the various components of the inspection system 100 are described in more detail.
[0034] The defect inspection subsystem 102 may include any type of metrology device known in the art. For example, the defect inspection subsystem 102 may illuminate the sample using an illumination beam and may further collect radiation emitted from the sample in response to the illumination beam. The illumination beam may include any type of illumination beam suitable for probing the sample, such as, but not limited to, a light beam (e.g., photons), an electron beam, or an ion beam. Further, the radiation emitted from the sample may include photons, electrons, ions, neutral particles, or the like. Thus, the defect inspection subsystem 102 may include an optical metrology tool, an electron beam metrology tool, a scanning electron microscope (SEM), an ion beam metrology tool, or the like.
[0035] The defect inspection subsystem 102 may further operate in a direct imaging mode or a scanning imaging mode. For example, the defect inspection subsystem 102 operating in the direct imaging mode may illuminate a portion of the sample that is larger than the system resolution and capture an image of the illuminated portion of the sample on a detector. The captured image may be any type of image known in the art, such as, but not limited to, a bright field image, a dark field image, a phase contrast image, or the like. Further, the captured images may be stitched together (e.g., by the defect inspection subsystem 102, by the controller 104, or the like) to form a composite image of the sample. By way of another example, the defect inspection subsystem 102 operating in the scanning mode may scan the focused beam across the sample and capture radiation and / or particles emitted from the sample at one or more measurement angles on one or more detectors. The focused beam may be scanned across the sample by modifying the beam path (e.g., using a galvanometer, piezoelectric mirror, or the like) and / or by translating the sample through a focal volume of the focused beam. The captured radiation may be stitched together (eg, by the defect inspection subsystem 102, by the controller 104, or the like) to form a composite image of the sample.
[0036] Figure 1B 1 is a conceptual diagram illustrating an optical defect inspection subsystem 102 according to one or more embodiments of the present disclosure. In one embodiment, the defect inspection subsystem 102 includes an illumination source 110 configured to generate an illumination beam 112. The illumination beam 112 may include light of one or more selected wavelengths, including, but not limited to, ultraviolet (UV) radiation, visible radiation, or infrared (IR) radiation.
[0037] The illumination source 110 may be any type of illumination source known in the art suitable for generating the optical illumination beam 112. In one embodiment, the illumination source 110 comprises a broadband plasma (BBP) illumination source. In this regard, the illumination beam 112 may comprise radiation emitted by a plasma. For example, the BBP illumination source 110 may comprise (but need not comprise) one or more pumping sources (e.g., one or more lasers) configured to focus pump light into a volume of gas, thereby causing energy to be absorbed by the gas in order to generate or maintain a plasma suitable for emitting radiation. Further, at least a portion of the plasma radiation may be used as the illumination beam 112. In another embodiment, the illumination source 110 may comprise one or more lasers. For example, the illumination source 110 may comprise any laser system known in the art capable of emitting radiation in the infrared, visible, or ultraviolet portions of the electromagnetic spectrum.
[0038] In another embodiment, the illumination source 110 directs an illumination beam 112 to a sample 114 disposed on a sample stage 116 via an illumination path 118. The illumination path 118 may include one or more lenses 120 or additional optical components 122 suitable for modifying and / or adjusting the illumination beam 112. For example, the one or more optical components 122 may include, but are not limited to, one or more polarizers, one or more filters, one or more beam splitters, one or more diffusers, one or more homogenizers, one or more apodizers, or one or more beam shapers. In another embodiment, the defect inspection subsystem 102 includes an objective lens 124 for directing the illumination beam 112 onto the sample 114.
[0039] In another embodiment, the defect inspection subsystem 102 includes a detector 126 configured to capture radiation (e.g., sample radiation 128) emitted from the sample 114 through the collection path 130. For example, the detector 126 can receive an image of the sample 114 provided by an element in the collection path 130 (e.g., the objective 124, one or more collection lenses 132, or the like). The collection path 130 can further include any number of collection optical elements 134 for directing and / or modifying the illumination collected by the objective 124, including, but not limited to, one or more filters, one or more polarizers, one or more apodizers, or one or more beam stops.
[0040] The detector 126 may include any type of optical detector known in the art suitable for measuring radiation received from the sample 114. For example, the detector 126 may include, but is not limited to, a CCD detector, a TDI detector, a photomultiplier tube (PMT), an avalanche photodiode (APD), or the like. In another embodiment, the detector 126 may include a spectral detector suitable for identifying wavelengths of radiation emitted from the sample 114. In another embodiment, the defect inspection subsystem 102 may include multiple detectors 126 (e.g., associated with multiple beam paths generated by one or more beam splitters to facilitate multiple metrology measurements by the defect inspection subsystem 102).
[0041] In one embodiment, Figure 1B , the defect inspection subsystem 102 may include a beam splitter 136 oriented so that the objective lens 124 can simultaneously direct the illumination beam 112 to the sample 114 and collect radiation emitted from the sample 114. In another embodiment, the incident angle of the illumination beam 112 on the sample 114 is adjustable. For example, the path of the illumination beam 112 through the beam splitter 136 and the objective lens 124 can be adjusted to control the incident angle of the illumination beam 112 on the sample 114.
[0042] In another embodiment, the defect inspection subsystem 102 is communicatively coupled to a controller 104 of the inspection system 100. In this regard, the controller 104 may be configured to receive data including, but not limited to, metrology images.
[0043] Figure 1C 1 is a conceptual view of a defect inspection subsystem 102 configured as a particle beam metrology tool according to one or more embodiments of the present disclosure. In one embodiment, the illumination source 110 includes a particle source (e.g., an electron beam source, an ion beam source, or the like) such that the illumination beam 112 includes a particle beam (e.g., an electron beam, a particle beam, or the like). The illumination source 110 may include any particle source known in the art that is suitable for generating the illumination beam 112. For example, the illumination source 110 may include, but is not limited to, an electron gun or an ion gun. In another embodiment, the illumination source 110 is configured to provide a particle beam with tunable energy. For example, the illumination source 110 including an electron source may, but is not limited to, provide an accelerating voltage in the range of 0.1 kV to 30 kV. As another example, the illumination source 110 including an ion source may, but need not, provide an ion beam with an energy in the range of 1 keV to 50 keV.
[0044] In another embodiment, the illumination path 118 includes one or more particle focusing elements (e.g., lens 120 or the like). For example, the one or more particle focusing elements may include, but are not limited to, a single particle focusing element or one or more particle focusing elements forming a composite system. In another embodiment, the one or more particle focusing elements include an objective lens 124 configured to direct the illumination beam 112 to the sample 114. Further, the one or more particle focusing elements may include any type of electron lens known in the art, including, but not limited to, electrostatic, magnetic, single potential, or dual potential lenses. It should be noted herein that Figure 1C The description of the imaging metrology tool depicted in and the associated description above are provided for illustrative purposes only and should not be construed as limiting. For example, the defect inspection subsystem 102 may include any excitation source known in the art suitable for generating metrology data about the sample 114. In another embodiment, the defect inspection subsystem 102 includes two or more particle beam sources (e.g., electron beam sources or ion beam sources) for generating two or more particle beams. In a further embodiment, the defect inspection subsystem 102 may include one or more components (e.g., one or more electrodes) configured to apply one or more voltages to one or more locations of the sample 114. In this regard, the defect inspection subsystem 102 can generate voltage contrast imaging data.
[0045] In another embodiment, the defect inspection subsystem 102 includes one or more particle detectors 126 for imaging or otherwise detecting particles emitted from the sample 114. In one embodiment, the detector 126 includes an electron collector (e.g., a secondary electron collector, a backscattered electron detector, or the like). In another embodiment, the detector 126 includes a photon detector (e.g., a light detector, an x-ray detector, a scintillation element coupled to a photomultiplier tube (PMT) detector, or the like) for detecting electrons and / or photons from the sample surface.
[0046] Figure 2 is a flow chart illustrating steps performed in a method 200 for segmenting a defect in accordance with one or more embodiments of the present disclosure. Applicants should note that the embodiments and enabling techniques described previously herein in the context of the inspection system 100 should be interpreted as extending to the method 200. However, it should be further noted that the method 200 is not limited to the architecture of the inspection system 100.
[0047] In one embodiment, method 200 includes a step 202 of receiving inspection data including values of two or more signal attributes and values of one or more context attributes associated with a defect. For example, the inspection data may be received from a defect inspection system (e.g., inspection system 100 or the like) based on inspection of one or more samples. Further, the inspection data may include an inspection data set associated with each identified defect, wherein each inspection data set includes corresponding values of signal and context attributes associated with the defect.
[0048] Signal attributes may include those associated with the response of a defect inspection system (such as, but not limited to, Figure 1B 1C). For example, considering the inspection system 100, the signal attributes may include, but are not limited to, measurable characteristics associated with sample radiation 128 from the sample 114 in response to the illumination beam 112, such as, but not limited to, the energy, magnitude, wavelength, or polarization of the sample radiation 128. By way of another example, the signal attributes may include parameters associated with a defect detection algorithm. For example, a defect detection algorithm (such as, but not limited to, MDAT technology) may be utilized to determine whether a defect exists (or whether a potential defect exists) at a given location based on the sample radiation 128 at the location. Thus, the signal attributes may include parameters associated with the defect detection algorithm, such as, but not limited to, pointiness or MDAT offset values.
[0049] The context attributes may include background content information associated with the identified defect or potential defect. For example, the context attributes may include information associated with the relationship between the defect and the surrounding area, such as (but not limited to) the relative brightness of the defect relative to the surrounding area (e.g., polarity), the relative or absolute position of the defect on the sample, the proximity of the defect to the feature or area of interest, the brightness of the reference image used to detect the defect, or the roughness associated with the reference image used to detect the defect. The context attributes may further include parameters associated with the defect detection algorithm, such as (but not limited to) MDAT grayscale or MDAT noise level.
[0050] In another embodiment, method 200 includes a step 204 of labeling at least some of the test data set with class labels based on location in a signal space defined by two or more signal attributes, wherein each class label corresponds to a region of the signal space.
[0051] A signal space may be defined based on two or more signal attributes associated with the inspection data. In this regard, each inspection data set associated with a defect may be mapped to a point in the signal space. In one embodiment, each inspection data set associated with a defect is labeled with a class label based on the location of the inspection data set in the signal space.
[0052] Figure 3 204 , according to one or more embodiments of the present disclosure, is a flowchart illustrating sub-steps associated with labeling each of the test data sets with a category label based on the corresponding position of each of the test data sets in the signal space.
[0053] In one embodiment, step 204 may include a sub-step 302 of mapping a plurality of validation data sets in a selected signal space.
[0054] Figure 4A FIG. 4 is a plot 400 of inspection data 402 (e.g., generated by inspection system 100 or the like) associated with a plurality of defects mapped into a two-dimensional signal space 404 based on two signal attributes (Signal Attribute 1 and Signal Attribute 2) according to one or more embodiments of the present disclosure. It should be understood that the inspection data 402 is Figure 4A The mapping in the two-dimensional signal space illustrated in is provided for illustrative purposes only and should not be construed as limiting. In a general sense, the inspection data 402 may be mapped into a multi-dimensional space of any number of dimensions based on any number of selected signal attributes.
[0055] Further, if Figure 4A As shown in FIG. 4 , the test data 402 may (but need not) be clustered into one or more signal clouds 406 in the signal space 404. Figure 4A4. The inspection data 402 in FIG. 4 is grouped into at least a first signal cloud 406a and a second signal cloud 406b. Such signal clouds 406 may (but need not) correspond to different defect groups. However, as previously described but not shown herein, defect segmentation based solely on the signal clouds 406 in signal space (e.g., by clustering algorithms, principal component analysis, or the like) may not be sufficient to identify defect groups of interest. For example, different defect groups with different noise characteristics, impact on production, impact on performance, or the like may have overlapping signal clouds 406. However, defect segmentation as described herein may facilitate identification and separation of defect groups of interest in the signal clouds 406 based on contextual attributes.
[0056] In another embodiment, step 204 may include a sub-step 304 of dividing the signal space into a set of polygons. In this regard, the test data set may be distributed between at least two polygons. The signal space 404 may be divided into polygons using any method known in the art, such as (but not limited to) into a multidimensional grid, Delaunay triangulation, or tessellation. Further, sub-step 304 may include (but need not include) dividing the signal space 404 into a set of non-overlapping polygons, where each polygon represents a unique region of the signal space 404.
[0057] The multi-dimensional signal space 404 may generally be divided into polygons of any shape. For example, the two-dimensional signal space 404 illustrated in FIG. 4 may be divided into a grid of polygons 408 (here, rectangles). By way of another example, but not shown, Figure 4A The two-dimensional signal space 404 illustrated in FIG. 4 is divided into polygons 408 of any shape, including but not limited to triangles, pentagons, hexagons, or the like. Figure 4A 4 is divided into polygons 408 of different shapes and / or sizes. It should be understood that a signal space 404 having more than two dimensions (e.g., as defined by more than two signal attributes) may be similarly divided into polygons 408 of any shape or size suitable for representing the signal space 404.
[0058] In another embodiment, step 204 may include sub-step 306 of labeling each inspection data set associated with the defect with a class label based on the position in the signal space 404. For example, step 204 may include labeling each inspection data set associated with the defect with a class label corresponding to the polygon 408 containing the corresponding inspection data set in the signal space 404.
[0059] Figure 4BFIG. 4 is a plot 410 of a signal space 404 divided into a set of non-overlapping polygons 408 (here, rectangles) each having a different class label, according to one or more embodiments of the present disclosure. For example, using a numerical identifier label in the range from 1 to 285 Figure 4B polygon 408. However, it should be understood that Figure 4B The use of numbers as class labels and the division of the signal space 404 illustrated in FIG. 4 are provided for illustrative purposes only and should not be construed as limiting. For example, a set of class labels may include any one-dimensional label or identifier set known in the art, including, but not limited to, numbers, letters, or words. Further, a class label need not be assigned to each polygon 408 associated with the signal space 404. In one embodiment, a unique polygon 408 that includes at least one inspection data set associated with a defect is assigned a class label.
[0060] In another embodiment, the method 200 includes a step 206 of segmenting the inspection data set into defect groups by training a classifier using the values of the context attributes and the corresponding class labels, wherein two or more defect groups are identified based on the trained classifier.
[0061] Step 206 may include training any type of classifier known in the art based on the contextual attributes and the class labels to identify different defect groups. For example, step 206 may include training a supervised classifier (e.g., a supervised learning technique) based on the contextual attributes and the class labels to identify different defect groups. A supervised classifier may typically identify patterns between input parameters and class labels based on training data in order to predict a class label for a new input based on the input parameters associated with the new input. In one embodiment, step 206 includes using contextual attributes from an inspection data set associated with the defect as input to a supervised classifier and class labels associated with corresponding polygons in signal space 404 as training data for the supervised classifier. Thus, training a supervised classifier may (e.g., based on polygons 408 associated with class labels) identify patterns between contextual attributes and locations in signal space 404 that may be used to segment the inspection data into different defect groups. In this regard, supervised classification techniques may use contextual attributes to identify defect groups that may have overlapping signal clouds 406 and / or different noise characteristics in signal space 404.
[0062] Figure 5A5 is a plot 500 illustrating inspection data including defect groups having overlapping signal clouds 406 in a signal space 502 that are distinguishable based on context attributes in accordance with one or more embodiments of the present disclosure. The plot 500 includes inspection data 504 mapped in a three-dimensional space including two signal attributes and one context attribute. As illustrated in the plot 500, the inspection data 504 may be segmented into a first segment 506 having context attribute values less than or equal to 15,754.5 and a second segment 508 having context attribute values greater than 15,754.5.
[0063] However, when segmenting based solely on signal attributes or context attributes, it may be impossible to distinguish the first segment 506 from the second segment 508. For example, when observing entirely in signal space 502 (e.g., by projecting the test data 504 onto a Figure 5A When viewed entirely in the background content domain (e.g., by projecting the verification data 504 onto the plane defined by the context attribute and either signal attribute 1 or signal attribute 2), the signal clouds 406 overlap, making it possible to indistinguish the first segment 506 from the second segment 508. By way of another example, when viewed entirely in the background content domain (e.g., by projecting the verification data 504 onto the plane defined by the context attribute and either signal attribute 1 or signal attribute 2), the signal clouds 406 also overlap, making it possible to indistinguish the first segment 506 from the second segment 508. Further, Figure 5B 5 is a histogram 510 of the test data 504 as a function of context attributes according to one or more embodiments of the present disclosure. Figure 5B As illustrated in FIG. 5 , the number of inspection data sets (eg, the number of defects) within the inspection data 504 is evenly distributed according to the context attributes, such that the first segment 506 and the second segment 508 are indistinguishable.
[0064] However, using context attributes and Figure 5A A supervised classifier trained by associating corresponding class labels with positions in signal space 502 defined by signal attribute 1 and signal attribute 2 in the signal space 502 can distinguish between first segment 506 and second segment 508. For example, defects with context attribute values less than or equal to 15,754.5 can be mapped to positions in signal space 502 defined by a circle centered at approximately 0 and having a radius of approximately 1, while defects with context attribute values greater than 15,754.5 can be mapped to positions in signal space 502 defined by a circle centered at approximately 0 and having a radius of approximately 0.5.
[0065] Step 206 may incorporate any supervised classifier known in the art for segmenting inspection data. For example, step 206 may incorporate one or more machine learning techniques, including, but not limited to, decision tree techniques, nearest neighbor techniques, regression techniques, Bayesian techniques, support vector machine techniques, similarity techniques, or neural network techniques. Thus, the cutting lines associated with segmenting the inspection data into different defect groups may reflect the structure of the trained classifier. For example, training a decision tree classifier may generate a decision tree with cutting points associated with values of context attributes that define branches of the tree. Further, the resulting decision tree may define ranges of various context attributes that lead to a common class label or a group of class labels with a particular probability. Thus, step 206 may include segmenting the inspection data into different defect groups based on the ranges of the context attributes associated with the decision tree. Further, the defect groups identified by the decision tree may correspond to different, potentially overlapping signal clouds in signal space.
[0066] Figure 6 The decision tree 600 is generated by training a decision tree classifier using contextual attribute information of the test data 504 and corresponding class labels associated with non-overlapping regions of the signal space 502 according to one or more embodiments of the present disclosure. Figure 6 , a single cut point associated with a context attribute value of 15,754.5 is identified. Therefore, defects with context attribute values less than or equal to 15,754.5 are segmented into first segments 506 and defects with context attribute values greater than 15,754.5 are segmented into second segments 508, so that first segments 506 and second segments 508 can be accurately and effectively distinguished.
[0067] It should be understood that Figures 5A to 6 and associated descriptions are provided for illustrative purposes only and should not be construed as limiting. For example, as previously described herein, an inspection data set may generally include any number of signal attributes and context attributes. Thus, the specific example of two signal attributes and one context attribute is merely illustrative. By way of another example, a situation may be defined by a segment defined by a trained classifier without incorporating each context attribute. In this regard, excluding context attributes may not be required to define a defect group. Figure 7 7 is a decision tree 700 associated with inspection data including 11 context attributes, only 6 of which (reference roughness, reference brightness, MDAT noise, radial position (PosRadius), test roughness, and test brightness) are used to segment the inspection data according to one or more embodiments of the present disclosure. Based on the decision tree 700, 14 defect groups are identified using different ranges of context attributes.
[0068] Furthermore, the structure of the trained classifier may vary based on the type of classifier. Figure 6 and 7 The decision tree illustrated in is provided for illustrative purposes only and should not be construed as limiting. However, it should be recognized herein that supervised learning techniques should segment the test data into groups based on the identified ranges of at least some of the contextual attributes.
[0069] The method 200 may further include utilizing the segmentation information as feedback or feed-forward data for any process step, including but not limited to defect discovery, defect classification, disturber point filtering, recipe control, or optics selection. For example, segmentation information about samples within a batch may be provided as feedback data for process steps for additional samples within the batch or additional batches. By way of further example, segmentation information about samples within a batch may be provided as feed-forward data for subsequent process steps associated with the same samples in the batch.
[0070] In one embodiment, segmentation information can be used to facilitate discovery of new defects. For example, defect segmentation based on signal and contextual attributes as described herein can enable detection of certain defects of interest that may otherwise be hidden within a signal cloud of other defects and can provide significant improvements over traditional segmentation techniques.
[0071] In one embodiment, the segmentation information may be used to facilitate classification of defects into any type of classification groups, such as (but not limited to) physical structure of the defect, likelihood of affecting performance, or likelihood of failure. For example, defect segmentation based on signal and contextual attributes as described herein may enable classification of defects based on noise characteristics of the defects in a group (e.g., associated with the size of the signal cloud).
[0072] In another embodiment, segmentation information can be used to facilitate optics selection. For example, it may be desirable to select and configure optics in an inspection system based on the separation between selected defects of interest and other defects identified in a given optics mode. In some embodiments, a receiver operating characteristic (ROC) curve generated based on segmentation information generated using signal and context attributes as described herein can facilitate the selection of optics for efficient and sensitive inspection.
[0073] In another embodiment, inspection data including defect segmentation information for one or more samples may be utilized to generate an inspection recipe (eg, including definitions of inspection locations, setup parameters for inspection, or the like) for one or more additional samples.
[0074] In another embodiment, segmentation information can be used to facilitate disturber point filtering. For example, after selecting an optics mode, a disturber point filter can be included in the inspection recipe to distinguish and isolate selected defects of interest from additional identified defects on the sample.
[0075] The subject matter described herein sometimes illustrates different components contained within or connected to other components. It should be understood that such depicted architectures are exemplary only, and in fact many other architectures that achieve the same functionality can be implemented. In a conceptual sense, any arrangement of components used to achieve the same functionality is effectively "associated" so that the desired functionality is achieved. Therefore, any two components combined in this article to achieve a specific functionality can be considered to be "associated" with each other so that the desired functionality is achieved without regard to architecture or intermediate components. Similarly, any two components so associated can also be considered to be "connected" or "coupled" to each other to achieve the desired functionality, and any two components that can be so associated can also be considered to be "coupleable" to each other to achieve the desired functionality. Specific examples of coupleable include (but are not limited to) physically interactable and / or physically interacting components and / or wirelessly interactable and / or wirelessly interacting components and / or logically interactable and / or logically interacting components.
[0076] It is believed that the present disclosure and many of its attendant advantages will be understood from the foregoing description, and it will be appreciated that various changes may be made to the form, construction, and arrangement of components without departing from the disclosed subject matter or sacrificing all of its significant advantages. The forms described are merely illustrative, and the following claims are intended to cover and encompass such changes. Further, it is to be understood that the invention is defined by the appended claims.
Claims
1. A testing system, comprising: A controller comprising one or more processors configured to execute process instructions configured to cause the one or more processors to: receiving, from a defect inspection system, a plurality of inspection data sets associated with inspection of one or more samples, wherein inspection data sets of the plurality of inspection data sets associated with a defect include values of two or more signal attributes and values of one or more context attributes, wherein signal attributes of the two or more signal attributes represent signals generated by the defect inspection system in response to a defect, and wherein context attributes of the one or more context attributes represent one or more additional characteristics of the defect; labeling each of the test data sets with a class label from a set of class labels based on a respective position of each of the test data sets in a signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space; and segmenting the plurality of verification data sets into two or more defect groups by training a classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of verification data sets, wherein the two or more defect groups are identified based on the trained classifier, wherein the one or more processors label each of the verification data sets with the class label by: mapping the plurality of verification data sets in the signal space; dividing the signal space into a set of polygons, wherein the test data set is distributed between at least two polygons of the set of polygons; and labeling each of the plurality of test data sets using a class label from the set of class labels corresponding to a polygon from the set of polygons containing a corresponding test data set.
2. The inspection system of claim 1 , wherein training the classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of inspection data sets comprises: Training the classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of inspection data sets using a supervised learning technique, wherein identifying the two or more defect groups from the trained classifier comprises: The two or more defect groups are identified based on at least one of a learned function or a structure of the trained classifier.
3. The inspection system of claim 2, wherein the supervised learning technique comprises: Machine learning technology.
4. The inspection system of claim 2, wherein the supervised learning technique comprises: At least one of decision tree technology, nearest neighbor technology, regression technology, Bayesian technology, support vector machine technology, similarity technology or neural network technology.
5. The inspection system of claim 1, wherein the set of class labels corresponds to non-overlapping regions of the signal space.
6. The inspection system of claim 1, wherein the set of category tags comprises: A one-dimensional representation of the signal space defined by the two or more signal attributes.
7. The inspection system of claim 1 , wherein labeling each of the inspection data sets with a class label from the set of class labels based on a respective position of each of the inspection data sets in a signal space defined by the two or more signal attributes comprises: The set of class labels is generated corresponding to non-overlapping regions of the signal space defined by the two or more signal attributes.
8. The inspection system of claim 1 , wherein dividing the signal space into a set of polygons comprises: The signal space is divided into the set of polygons by dividing the signal space into a regular multi-dimensional grid.
9. The inspection system of claim 1 , wherein dividing the signal space into a set of polygons comprises: The signal space is partitioned into the set of polygons using at least one of Delogne triangulation or tessellation.
10. The inspection system of claim 1, wherein the two or more defect groups correspond to one or more contiguous groups of polygons in the set of polygons associated with the signal space defined by the two or more signal attributes.
11. The inspection system of claim 1 , wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: One or more defects are classified based on the two or more defect groups.
12. The inspection system of claim 1, wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: An inspection recipe is generated based on at least one of the two or more defect groups.
13. The inspection system of claim 12, wherein the defect inspection recipe provides for detection of selected defects of concern.
14. The inspection system of claim 12, wherein the defect inspection recipe provides disturber point filtering.
15. The inspection system of claim 1, wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: An optics mode is selected based on the two or more defect groups.
16. The inspection system of claim 15, wherein selecting the optics mode based on the two or more defect groups comprises: The optics mode is selected based on one or more separations between the one or more selected defects of interest and additional defects.
17. A testing system comprising: A defect inspection system comprising: an illumination source configured to generate one or more illumination beams; and a detector configured to capture radiation emitted from the sample in response to the one or more illumination beams; and a controller communicatively coupled to the detector, the controller comprising one or more processors configured to execute process instructions configured to cause the one or more processors to: receiving, from a defect inspection system, a plurality of inspection data sets associated with inspection of one or more samples, wherein inspection data sets of the plurality of inspection data sets associated with a defect include values of two or more signal attributes and values of one or more context attributes, wherein signal attributes of the two or more signal attributes represent signals generated by the defect inspection system in response to a defect, and wherein context attributes of the one or more context attributes represent one or more additional characteristics of the defect; Labeling each of the test data sets with a class label from a set of class labels based on a respective position of each of the test data sets in a signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space, wherein the one or more processors label each of the test data sets with the class label by: Mapping the plurality of test data sets in the signal space; dividing the signal space into a set of polygons, wherein the test data sets are distributed between at least two polygons of the set of polygons; and labeling each of the plurality of test data sets using a class label from the set of class labels corresponding to a polygon from the set of polygons containing a corresponding test data set; and The plurality of verification data sets are segmented into two or more defect groups by training a classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of verification data sets, wherein the two or more defect groups are identified based on the trained classifier.
18. The inspection system of claim 17, wherein the illumination source comprises: Optical illumination source.
19. The inspection system of claim 18, wherein the illumination source comprises: Broadband plasma illumination source.
20. The inspection system of claim 18, wherein the wavelengths of the one or more illumination beams include: At least one of an extreme ultraviolet wavelength, a vacuum ultraviolet wavelength, a deep ultraviolet wavelength, or an ultraviolet wavelength.
21. The inspection system of claim 17, wherein the illumination source comprises: Particle beam illumination source.
22. The inspection system of claim 21, wherein the illumination source comprises: Electron beam illumination source.
23. The inspection system of claim 17, wherein training the classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of inspection data sets comprises: Training the classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of inspection data sets using a supervised learning technique, wherein identifying the two or more defect groups from the trained classifier comprises: The two or more defect groups are identified based on at least one of a learned function or a structure of the trained classifier.
24. The inspection system of claim 23, wherein the supervised learning technique comprises: Machine learning technology.
25. The inspection system of claim 23, wherein the supervised learning technique comprises: At least one of decision tree technology, nearest neighbor technology, regression technology, Bayesian technology, support vector machine technology, similarity technology or neural network technology.
26. The inspection system of claim 17, wherein the set of class labels correspond to non-overlapping regions of the signal space.
27. The inspection system of claim 17, wherein labeling each of the inspection data sets with a class label from the set of class labels based on a respective position of each of the inspection data sets in a signal space defined by the two or more signal attributes comprises: The set of class labels is generated corresponding to non-overlapping regions of the signal space defined by the two or more signal attributes.
28. The inspection system of claim 17, wherein dividing the signal space into a set of polygons comprises: The signal space is partitioned into the set of polygons by partitioning the signal space using at least one of a multidimensional grid, a Delogne triangulation, or a tessellation.
29. The inspection system of claim 17, wherein the two or more defect groups correspond to one or more contiguous groups of polygons in the set of polygons associated with the signal space defined by the two or more signal attributes.
30. The inspection system of claim 17, wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: One or more defects are classified based on the two or more defect groups.
31. The inspection system of claim 17, wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: An inspection recipe is generated based on at least one of the two or more defect groups.
32. The inspection system of claim 31, wherein the defect inspection recipe provides disturber point filtering.
33. The inspection system of claim 17, wherein the one or more processors are further configured to execute process instructions that cause the one or more processors to: An optics mode is selected based on the two or more defect groups.
34. A method for segmenting a metrology defect, comprising: receiving, from a defect inspection system, a plurality of inspection data sets associated with inspection of one or more samples, wherein inspection data sets of the plurality of inspection data sets associated with a defect include values of two or more signal attributes and values of one or more context attributes, wherein signal attributes of the two or more signal attributes represent signals generated by the defect inspection system in response to a defect, and wherein context attributes of the one or more context attributes represent one or more additional characteristics of the defect; Labeling each of the test data sets with a class label from a set of class labels based on the corresponding position of each of the test data sets in the signal space defined by the two or more signal attributes, wherein each class label corresponds to a region of the signal space, wherein labeling each of the test data sets with the class label comprises: mapping the multiple test data sets in the signal space; dividing the signal space into a set of polygons, wherein the test data sets are distributed between at least two polygons of the set of polygons; and labeling each of the multiple test data sets with a class label from the set of class labels corresponding to a polygon from the set of polygons containing the corresponding test data set; and The plurality of verification data sets are segmented into two or more defect groups by training a classifier using the values of the one or more context attributes and corresponding class labels from the set of class labels for the plurality of verification data sets, wherein the two or more defect groups are identified based on the trained classifier.
