Digital Non-Linear Neural Network-Based Image Filtering for Semiconductor Applications

Abstract

A method and system for determining information about a specimen location are provided. One system includes a computer subsystem configured to input one or more images about a specimen location into a region-based neural network configured to perform digital non-linear filtering of said one or more images to generate a filtered image about the specimen location. The computer subsystem is also configured to determine information about the specimen location from the filtered location.

Description

Technology Field

[0001] The present invention generally relates to a method and system for determining information regarding specimen location. Specific embodiments described herein relate to a method and system for improving defect detection and / or nuisance filtering in optical wafer inspection using digital nonlinear neural network-based image filtering. Background Technology

[0002] The following descriptions and examples are not to be recognized as prior art by their inclusion in this section.

[0003] Manufacturing semiconductor devices, such as logic devices and memory devices, typically involves processing a specimen, such as a semiconductor wafer, using a number of semiconductor manufacturing processes to form various features and multiple levels of the semiconductor device. For example, lithography is a semiconductor manufacturing process that typically involves transferring a pattern onto a resist arranged on a semiconductor wafer. Additional examples of semiconductor manufacturing processes include, but are not limited to, chemical mechanical polishing, etching, deposition, and ion implantation. A number of semiconductor devices may be manufactured within an arrangement on a semiconductor wafer and then separated into individual semiconductor devices.

[0004] Inspection processes are used at various stages during the semiconductor manufacturing process to detect defects on specimens in order to obtain higher yields and, consequently, higher profits. Inspection has always been an important part of manufacturing semiconductor devices. However, as the dimensions of semiconductor devices decrease, inspection is becoming even more critical for the successful manufacturing of acceptable semiconductor devices, as smaller defects can cause the device to fail.

[0005] Defect review generally involves re-detecting defects identified by the inspection process and generating additional information about the defects at higher resolution using high-magnification optical systems or scanning electron microscopes (SEMs). Therefore, defect review is performed at discrete locations on the specimen where defects were detected by inspection. The higher-resolution data regarding defects generated by defect review is more suitable for determining defect attributes such as profiles, roughness, and more precise size information. Defects can generally be classified more accurately by defect type based on the information determined by defect review compared to inspection.

[0006] The metrology process is also used to monitor and control the process at various stages during the semiconductor manufacturing process. Unlike the inspection process, where defects are detected on a specimen, the metrology process is used to measure one or more characteristics of a specimen that cannot be determined using currently used inspection tools. For example, the metrology process is used to measure one or more characteristics of a specimen, such as the dimensions of features formed on the specimen during the process (e.g., line width, thickness, etc.), so that the performance of the process can be determined from one or more characteristics. Additionally, if one or more characteristics of the specimen are unacceptable (e.g., if they fall outside a predetermined range for the characteristic(s), the measurement of one or more characteristics of the specimen can be used to modify one or more parameters of the process so that additional specimens produced by the process have acceptable characteristic(s).

[0007] The measurement process also differs from the defect review process in that, unlike the defect review process where defects detected by inspection are revisited during the defect review, the measurement process can be performed at locations where no defects are detected. In other words, unlike defect review, the location where the measurement process is performed on a specimen can be independent of the results of the inspection process performed on the specimen. In particular, the location where the measurement process is performed can be selected independently of the inspection results. Furthermore, since the location on the specimen where the measurement is performed can be selected independently of the inspection results, the location where the measurement process is performed can be determined before the inspection process is performed on the specimen, unlike defect review, where the location on the specimen to be reviewed cannot be determined until the inspection results for the specimen are generated and available for use.

[0008] Methods and systems configured to perform the yield-related processes described above are often developed by first discovering the best possible hardware configuration for generating images, data, measurements, signals, etc., for a specimen. Once the hardware configuration is established, the hardware parameters that are optimal for the process are selected. The selection of hardware parameters can significantly influence how images, data, measurements, signals, etc., respond to the specimen and how well they can be used to determine information about the specimen.

[0009] Sometimes, the best possible hardware configuration and associated parameters cannot produce an ideal output to determine information about the specimen. For example, the best possible hardware configuration and parameters may still generate images containing a significant amount of noise that can interfere with defect detection, measurement, and defect review. In such cases, methods and systems may be configured to attempt to improve the image before using it to determine information about the specimen. Many different types of post-processing can be performed on images, data, measurements, signals, etc., generated for the specimen by tool hardware. For example, some broadband optical inspection tools currently use only a limited set of predefined digital linear filter templates.

[0010] While post-processing tool hardware output can often improve the ability to determine information about the specimen, new methods for improving images are constantly being explored due to some current existing limitations of the tool / method and additional challenges anticipated in the future. For example, the limited set of predefined digital linear filter templates currently used by some tools may not be sufficient to increase detection sensitivity or to reduce the nuisance rate for challenging inspections and defects. In particular, for relatively small challenging defects, optical signal detection typically occurs at or below the sensitivity limits of optical wafer inspection tools, resulting in excessive noise or nuisance counting rates. Available processes for recording and auto-linear difference image filters often do not provide a satisfactory solution to address the problem of relatively high nuisance or relatively low defect of interest (DOI) capture rates.

[0011] Therefore, it would be advantageous to develop a system and method for filtering images to determine information about specimens that do not have one or more of the disadvantages described above.

[0012] The following description of various embodiments should by no means be interpreted as limiting the gist of the appended claims.

[0013] One embodiment relates to a system configured to determine information about a specimen location. The system includes a computer subsystem configured to input one or more images about a specimen location into a region-based neural network configured to perform digital non-linear filtering of one or more images to generate a filtered image about the specimen location. The computer subsystem is also configured to determine information about the specimen location from the filtered image. The system may be further configured as described herein.

[0014] Another embodiment relates to a computer-implemented method for determining information regarding a specimen location. The method comprises input and determination steps performed by a computer subsystem as described above. Each of the steps of the method may be performed as further described herein. The method may include any other step(s) of any other method(s) described herein. The method may be performed by any of the systems described herein.

[0015] Another embodiment relates to a non-transient computer-readable medium storing program instructions executable on a computer system for performing a computer-implemented method for determining information about a specimen location. The computer-implemented method may include steps of the method described above. The computer-readable medium may also be configured as described herein. Steps of the computer-implemented method may be performed as further described herein. Additionally, the computer-implemented method in which program instructions are executable may include any other step(s) of any other method(s) described herein. Brief explanation of the drawing

[0016] Further advantages of the present invention will become apparent to those skilled in the art who benefit from the following detailed description of preferred embodiments with reference to the accompanying drawings: FIGS. 1 and 2 are schematic diagrams illustrating a side view of an embodiment of a system configured as described in this specification; FIG. 3 is a schematic diagram illustrating an embodiment of the configuration of a cascaded filter neural network for image filtering; FIG. 4 is a schematic diagram illustrating a difference image for a defect of interest (DOI) and nussons, a filtered image of the difference image generated by an embodiment described herein, and a scanning electron microscope (SEM) image sketch illustrating the defect location within the difference image; FIG. 5 includes a plot illustrating an ROC curve showing that when used for image filtering, the embodiment described herein suppresses background noise pixels by 2.5 orders of magnitude for the same true positive value compared to the unfiltered case; FIG. 6 is a flowchart illustrating an example of use case for the image filtering embodiment described herein in a post-detection stage; FIG. 7 includes an example of a plot illustrating a DOI capture rate versus Newson rate curve for a wafer after an example of filtering described herein is applied, and a DOI capture rate versus Newson rate curve after an example of post-detection filtering described herein is applied to a highly imbalanced data set; FIG. 8 is a block diagram illustrating an embodiment of a non-transient computer-readable medium storing program instructions for enabling a computer system to perform a computer implementation method described in this specification. Although various modifications and alternative forms of the invention are possible, specific embodiments thereof are illustrated by way of example in the drawings and described in detail herein. The drawings may not be drawn to actual scale. However, the drawings and the detailed description thereof are not intended to limit the invention to the specific forms disclosed; on the contrary, the intention is to encompass all modifications, equivalents, and alternatives that fall within the spirit and scope of the invention as defined by the appended claims. Specific details for implementing the invention

[0017] Now, returning to the drawings, note that they are not drawn to scale. In particular, the scale of some elements in the drawings has been greatly exaggerated to emphasize the characteristics of those elements. It is also noted that the drawings are not drawn at the same scale. Elements depicted in more than one drawing that may be similarly configured are indicated using the same reference numerals. Unless otherwise indicated in this specification, any element described and depicted may include any suitable commercially available element.

[0018] Generally, the embodiments described herein are configured to determine information regarding specimen locations. More specifically, the embodiments provide a method and system for improving defect detection and nuance filtering in optical wafer inspection by using digital nonlinear neural network-based filtering. The embodiments described herein provide a method and system for improving and optimizing defect detection in optical wafer inspection by applying nonlinear digital filtering called AiryNet™. Applying this filter results in higher sensitivity for targeted defect types. The embodiments may also be used to increase the detection rate of defects of interest (DOI) and / or simultaneously decrease the detection rate of nuance events. The embodiments described herein also provide a method for predicting and localizing defect locations on optical images.

[0019] As used herein, the term "nusense" (sometimes used interchangeably with "nusense defect") is defined as an event that is generally not of concern to the user and / or is detected on a specimen but is not actually an actual defect on the specimen. A non-defective nusense may be detected as an event due to non-defective noise sources on the specimen (e.g., grain of metal wire on the specimen, signals from an underlaying layer or material on the specimen, line edge roughness (LER), relatively small critical dimension (CD) variation of patterned properties, thickness variation, etc.), and / or due to the marginality of the inspection system itself or the limitations of the configuration of the inspection system used for inspection.

[0020] As used herein, the term “Defect of Interest (DOI)” is defined as a defect detected on a specimen and is, in fact, an actual defect on the specimen. Therefore, since the user is generally interested in how many actual defects there are on the specimen being inspected and what kind of actual defects they are, the DOI is of interest to the user. In some situations, the term “DOI” is used to refer to a subset of all actual defects on the specimen, which includes only the actual defects of interest to the user. For example, there may be multiple types of DOIs on any given specimen, and one or more of them may be of greater interest to the user than one or more other types.

[0021] In some embodiments, the specimen is a wafer. The wafer may include any wafer known in the semiconductor field. While some embodiments may be described herein in relation to wafers or wafers, the embodiments are not limited to specimens in which they may be used. For example, the embodiments described herein may be used in specimens such as, for example, reticles, flat panels, personal computer (PC) boards, and other semiconductor specimens.

[0022] An embodiment of a system configured to determine information about a specimen location is illustrated in FIG. 1. In one embodiment, the system includes an inspection subsystem (100) configured to generate an image of a specimen comprising at least one of one or more images of a specimen location used by an embodiment as described herein. In another embodiment, at least one of the one or more images generated about a specimen location is an optical wafer image generated by the inspection subsystem. In FIG. 1, the inspection subsystem is configured as an optical-based inspection subsystem. However, the inspection subsystem may be configured as an electron beam or charged particle beam-based inspection subsystem.

[0023] Generally, the inspection subsystem described herein includes at least an energy source and a detector. The energy source is configured to generate energy directed toward the specimen. The detector is configured to detect energy from the specimen and generate an output that responds to the detected energy.

[0024] In a light-based inspection subsystem, the energy directed toward the specimen includes light, and the energy detected from the specimen includes light. For example, as illustrated in FIG. 1, the inspection subsystem includes an illumination subsystem configured to direct light toward the specimen (14). The illumination subsystem includes at least one light source, for example, a light source (16). The illumination subsystem is configured to direct light toward the specimen at one or more angles of incidence, which may include one or more angles of inclination and / or one or more angles of perpendicularity. For example, as illustrated in FIG. 1, light from the light source (16) passes through an optical element (18) and is then directed toward the specimen (14) at an angle of inclination through a lens (20). The angle of inclination may include any suitable angle of inclination, which may vary, for example, depending on the characteristics of the specimen and the defect to be detected in the specimen.

[0025] The illumination subsystem may be configured to direct light toward a specimen at different angles of incidence. For example, the inspection subsystem may be configured to change one or more parameters of one or more elements of the illumination subsystem so that light may be directed toward the specimen at an angle of incidence different from that shown in FIG. 1. In one such example, the inspection subsystem may be configured to move the light source (16), optical element (18), and lens (20) so that light may be directed toward the specimen at different oblique angles of incidence as well as at a vertical (or nearly vertical) angle of incidence. The illumination subsystem may have any other suitable configuration known in the art to direct light toward the specimen at one or more angles of incidence sequentially or simultaneously.

[0026] The lighting subsystem may also be configured to direct light having different characteristics toward the specimen. For example, the optical element (18) may be configured as a spectral filter, and the characteristics of the spectral filter may be changed in various different ways (e.g., by replacing one spectral filter with another) so that light of different wavelengths may be directed toward the specimen at different times.

[0027] The light source (16) may include a broadband plasma (BBP) light source. In this way, the light generated by the light source and directed toward the specimen may include broadband light. However, the light source may include any other suitable light source, such as any suitable laser known in the art, configured to generate light at any suitable wavelength(s). Additionally, the laser may be configured to generate monochromatic or nearly monochromatic light. In this way, the laser may be a narrowband laser. The light source may also include a multicolor light source that generates light at a number of discrete wavelengths or wavelength bands.

[0028] Light from an optical element (18) can be focused onto a specimen (14) by a lens (20). Although the lens (20) is illustrated in FIG. 1 as a single refractive optical element, in practice the lens (20) may include a plurality of refractive and / or reflective optical elements combined to focus light from the optical element onto the specimen. The illumination subsystem illustrated in FIG. 1 and described herein may include any other suitable optical element (not illustrated). Examples of such optical elements include, but are not limited to, polarizing components(s), spectral filters(s), spatial filters(s), reflective optical elements(s), apodizer(s), beam splitters(s), aperture(s), etc., which may include any such suitable optical element known in the art. Additionally, the system may be configured to change one or more elements of the illumination subsystem based on the type of illumination to be used for inspection.

[0029] The inspection subsystem also includes a scanning subsystem configured to change the position on the specimen where the light is directed and the light is detected, and possibly to allow the light to be scanned over the specimen. For example, the inspection subsystem may include a stage (22) on which the specimen (14) is placed during inspection. The scanning subsystem may include any suitable mechanical and / or robotic assembly (including the stage (22)) configured to move the specimen so that the light can be directed to different positions on the specimen and detected from different positions. Additionally or alternatively, the inspection subsystem may be configured such that one or more optical elements of the inspection subsystem perform partial scanning of the light over the specimen so that the light can be directed to different positions on the specimen and detected from different positions. The light may be scanned over the specimen in any suitable manner, such as in a serpentine-like path or a spiral path.

[0030] The inspection subsystem further includes one or more detection channels. At least one of the detection channel(s) includes a detector configured to detect light from a specimen due to illumination of the specimen by the system and to generate an output in response to the detected light. The inspection subsystem illustrated in FIG. 1 includes two detection channels, one of which is formed by a collector (24), an element (26), and a detector (28), and the other of which is formed by a collector (30), an element (32), and a detector (34). The two detection channels are configured to collect and detect light at different collection angles. In some examples, the two detection channels are configured to detect scattered light, and the detection channel is configured to detect light scattered from the specimen at different angles. However, one or more of the detection channels may be configured to detect another type of light (e.g., reflected light) from the specimen.

[0031] In FIG. 1, two detection channels are shown positioned on the plane of the paper, and the illumination subsystem is also shown positioned on the plane of the paper. Thus, in this embodiment, the two detection channels are positioned on the plane of incidence (e.g., in the center). However, one or more of the detection channels may be positioned outside the plane of incidence. For example, the detection channels formed by the collector (30), the element (32), and the detector (34) may be configured to collect and detect light scattered outside the plane of incidence. Thus, such detection channels may generally be referred to as "side" channels, and such side channels may be at the center of a plane substantially perpendicular to the plane of incidence.

[0032] FIG. 1 illustrates an embodiment of an inspection subsystem comprising two detection channels, but the inspection subsystem may comprise a different number of detection channels (e.g., only one detection channel or two or more detection channels). The detection channel formed by the collector (30), the element (32), and the detector (34) may form a single lateral channel as described above, and the inspection subsystem may comprise an additional detection channel (not illustrated) formed as another lateral channel located opposite the plane of incidence. Thus, the inspection subsystem may comprise a collector (24), an element (26), and a detector (28), and may comprise a detection channel configured to collect and detect light at a scattering angle(s) perpendicular or near-perpendicular to the surface of the specimen, which is centered on the plane of incidence. Thus, this detection channel may generally be referred to as the "top" channel, and the inspection subsystem may also comprise two or more lateral channels configured as described above. In this way, the inspection subsystem may include at least three channels (i.e., one top channel and two side channels), and each of the at least three channels is configured to collect light at a scattering angle different from each of the other collectors.

[0033] As further explained above, one or more of the detection channels may be configured to detect scattered light. Thus, the inspection subsystem illustrated in FIG. 1 may be configured for dark field (DF) inspection. However, the inspection subsystem may also or alternatively include detection channel(s) configured for bright field (BF) inspection. Thus, the inspection subsystem described herein may be configured for DF only, BF only, or both DF and BF inspection. Although each of the collectors is illustrated in FIG. 1 as a single refractive optical element, each of the collectors may include refractive optical element(s) and / or reflective optical element(s).

[0034] One or more detection channels may include any suitable detector known in the art, such as a photomultiplier tube (PMT), a charge coupled device (CCD), and a time delay integration (TDI) camera. The detectors may include non-imaging detectors or imaging detectors. If the detectors are non-imaging detectors, each detector may be configured to detect specific characteristics of scattered light, such as intensity, but may not be configured to detect characteristics such as, for example, a function of position within the imaging plane. Thus, the output generated by each of the detectors within each of the detection channels may be a signal or data, but not an image signal or image data. In such cases, the computer subsystem may be configured to generate an image of the specimen from the non-imaging output of the detector. However, in other examples, the detector may be configured as an imaging detector configured to generate an imaging signal or image data. Accordingly, the inspection subsystem may be configured to generate images in a number of ways.

[0035] The computer subsystem (36) is coupled to the detector of the inspection subsystem in any suitable manner (e.g., via one or more transmission media that may include "wired" and / or "wireless" transmission media) so that the computer subsystem can receive the output generated by the detector. The computer subsystem (36) may be configured to perform a number of functions using the output of the detector as further described herein. The computer subsystem (36) may be further configured as described herein.

[0036] Computer subsystems (and other computer subsystems described herein) may also be referred to as computer system(s) in this specification. Each computer subsystem(s) or system(s) described herein may take various forms, including personal computer systems, image computers, mainframe computer systems, workstations, network devices, internet devices, or other devices. Generally, the term “computer system” may be broadly defined to include any device having one or more processors that execute instructions from a memory medium. Computer subsystem(s) or system(s) may also include any suitable processor known in the art, such as, for example, a parallel processor. Additionally, computer subsystem(s) or system(s) may include a computer platform having high-speed processing and software as a standalone or networked tool.

[0037] If the system includes one or more computer subsystems, different computer subsystems may be combined with each other so that images, data, information, instructions, etc., can be transmitted between the computer subsystems. For example, a computer subsystem (36) may be combined to the computer system (102) by any suitable transmission medium, which may include any suitable wired and / or wireless transmission medium known in the art, as shown by the dashed line in FIG. 1. Two or more of these computer subsystems may also be effectively combined by a shared computer-readable storage medium (not shown).

[0038] In an electron beam inspection subsystem, the energy directed toward the specimen contains electrons, and the energy detected from the specimen contains electrons. In such an embodiment illustrated in FIG. 2, the inspection subsystem includes an electron column (122), and the inspection subsystem includes a computer subsystem (124) coupled thereto. The computer subsystem (124) may be configured as described above. Additionally, this inspection system may be coupled to one or more other computer subsystems in the same manner as described above and illustrated in FIG. 1.

[0039] As also illustrated in FIG. 2, the electron column comprises an electron beam source (126) configured to generate electrons focused onto a specimen (128) by one or more elements (130). The electron beam source may include, for example, a cathode source or an emitter tip, and one or more elements (130) may include, for example, a gun lens, an anode, a beam limiting aperture, a gate valve, a beam current selecting aperture, an objective lens, and a scanning subsystem, all of which may include any such suitable elements known in the art.

[0040] Electrons returned from the sample (e.g., secondary electrons) can be focused to a detector (134) by one or more elements (132). One or more elements (132) may include a scanning subsystem, for example, which may be the same scanning subsystem contained within the element(s) (130).

[0041] The electronic column may include any other suitable elements known in the art. Additionally, the electronic column may also be configured as described in U.S. Patent No. 8,664,594, granted to Jiang et al. on April 4, 2014; U.S. Patent No. 8,692,204, granted to Kojima et al. on April 8, 2014; U.S. Patent No. 8,698,093, granted to Gubbens et al. on April 15, 2014; and U.S. Patent No. 8,716,662, granted to MacDonald et al. on May 6, 2014, which are combined by reference as fully described herein.

[0042] In FIG. 2, an electron column is illustrated as being configured such that electrons are directed toward a specimen at an oblique angle of incidence and scattered from the specimen at another oblique angle, but the electron beam may be directed toward and scattered from the specimen at any suitable angle. Additionally, the electron beam inspection subsystem may be configured to use multiple modes to generate output to the specimen as further described herein (e.g., different illumination angles, collection angles, etc.). The multiple modes of the electron beam inspection subsystem may differ in any output generation parameter of the inspection subsystem.

[0043] The computer subsystem (124) may be coupled to the detector (134) as described above. The detector may detect electrons returned from the surface of the specimen to form an electron beam image of the specimen (or other output for the specimen). The electron beam image may include any suitable electron beam image. The computer subsystem (124) may be configured to perform any step(s) described herein. A system including the inspection subsystem shown in FIG. 2 may be further configured as described herein.

[0044] FIGS. 1 and 2 are provided herein to generally illustrate the configuration of an inspection subsystem that may be included in the system embodiments described herein. Clearly, the inspection subsystem configuration described herein may be modified to optimize the performance of the inspection subsystem as is typically done when designing commercial inspection systems. Additionally, the system described herein may be implemented using existing inspection systems, such as tools commercially available from KLA in Milpitas, California (for example, by adding the functionality described herein to an existing inspection system). For some of these systems, the method described herein may be provided as an optional functionality of the inspection system (e.g., in addition to other functionalities of the inspection system). Alternatively, the inspection system described herein may be designed "from scratch" to provide a completely new inspection system.

[0045] Although the inspection subsystem has been described above as being an optical or electron beam inspection subsystem, the inspection subsystem may be an ion beam inspection subsystem. Such an inspection subsystem may be configured as shown in FIG. 2, except that the electron beam source may be replaced with any suitable ion beam source known in the art. Additionally, the inspection subsystem may include any other suitable ion beam system, such as those included in commercially available focused ion beam (FIB) systems, helium ion microscopy (HIM) systems, and secondary ion mass spectroscopy (SIMS) systems.

[0046] The inspection subsystem may be configured to generate an output of a specimen having multiple modes, for example, an image. Generally, a "mode" is defined by a parameter value of the inspection subsystem used to generate an image of the specimen (or an output used to generate an image of the specimen). Thus, the modes may differ in value for at least one of the parameters of the inspection subsystem (excluding the location on the specimen where the output is generated). For example, the modes may differ in one or more adjustable parameters of the inspection subsystem (e.g., illumination polarization(s), angle(s), wavelength(s), etc.; detection polarization(s), angle(s), wavelength(s), etc.). The inspection subsystem may be configured to scan the specimen in different modes in the same scan or different scans, for example, depending on the ability to use multiple modes to scan the specimen simultaneously.

[0047] In a similar manner, the electron beam subsystem may be configured to generate an image having two or more modes, which can be defined by the values ​​of the parameters of the electron beam subsystem used to generate an image of the specimen. Thus, the modes may have different values ​​for at least one of the electron beam parameters of the electron beam subsystem. For example, different modes may use different angles of incidence for illumination.

[0048] In another embodiment, the system includes an instrumentation subsystem. In an additional embodiment, the system includes a defect inspection subsystem. For example, an embodiment of the inspection subsystem illustrated in FIGS. 1 and 2 may be modified in one or more parameters to provide different imaging capabilities depending on the application in which it is used. In one such example, the inspection subsystem may be configured to have higher resolution when used for instrumentation rather than for inspection. That is, the embodiment of the inspection subsystem illustrated in FIGS. 1 and 2 describes some general and various configurations for an imaging subsystem that can be customized in a number of ways that will be obvious to those skilled in the art to produce systems with different imaging capabilities somewhat suitable for different applications.

[0049] In this way, the imaging subsystem may be configured to generate an output suitable for re-detecting defects on a specimen in the case of a defect review system and measuring one or more characteristics of a specimen in the case of a measurement system. In an embodiment of a defect review system, the computer subsystem (124) shown in FIG. 2 may be configured to re-detect defects on a specimen (128) by applying a defect re-detection method to the output generated by the detector (134) and possibly determining additional information about the re-detected defect using the output generated by the detector. In an embodiment of a measurement system, the computer subsystem (36) shown in FIG. 1 may be configured to determine one or more characteristics of a specimen (14) using the output generated by the detectors (28 and / or 34).

[0050] As mentioned above, the inspection subsystem may be configured to scan energy (e.g., light, electron) over the physical version of the specimen to generate an image of the physical version of the specimen. In this way, the inspection subsystem may be configured as a "real" subsystem rather than a "virtual" subsystem. However, the storage medium (not shown) and computer subsystem(s) (102) illustrated in FIG. 1 may be configured as a "virtual" system. In particular, the storage medium and computer subsystem(s) may be configured as a "virtual" inspection system as described in U.S. Patent No. 8,126,255, granted to Bhaskar et al. on February 28, 2012, and U.S. Patent No. 9,222,895, granted to Duffy et al. on December 29, 2015, which are all incorporated by reference as fully specified herein. The embodiments described in this specification may be further configured as described in these patents.

[0051] A computer subsystem is configured to input one or more images of a specimen location into a region-based neural network configured to perform digital non-linear filtering of one or more images to generate a filtered image of the specimen location. In some embodiments, the system includes one or more components executed by the computer subsystem, and one or more components include a region-based neural network. Such system embodiments may also include or not include an inspection subsystem. In this way, the embodiments described herein may be performed "on-tool." For example, as illustrated in FIG. 1, the system may include one or more components (104) executed by the computer subsystem (36) (and / or computer system(s) (102)), and one or more components include a region-based neural network (NN) (106). The computer subsystem (36) (and / or computer system(s) (102)) may input an image generated for the specimen (14) (by the inspection subsystem (100) and / or one or more computer subsystems) into the region-based neural network. The region-based NN (106) performs digital non-linear filtering of one or more images to generate a filtered image for the specimen location, which may then be transmitted to a computer subsystem (36) (and / or computer system(s) (102)). Subsequently, the computer subsystem may be configured to perform one or more functions described herein with the filtered image.

[0052] A computer subsystem may be configured to acquire one or more images, which may or may not include generating one or more images using an inspection subsystem configured as described herein, for example. Alternatively, the computer subsystem may acquire image(s) from a storage medium in which the image(s) are stored or from another method or system that generates the image(s). In this way, the computer subsystem may acquire image(s) by retrieving or receiving image(s) from another method or system. Thus, one system may generate images as described herein, and the system described herein may generate filtered images using a region-based neural network and determine information regarding specimen locations from the filtered images as further described herein. However, one system may be configured to perform all or at least some of these functions.

[0053] In one embodiment, the region-based neural network is configured as an all-convolutional region-based neural network having multiple input image channels. Accordingly, the supervised nonlinear image filtering technique described herein may be based on digital nonlinear filtering using an all-convolutional region-based neural network having multiple input channels. In this specification, "region-based neural network" is also referred to simply as "neural network," "neural net," "AiryNet™," "AiryNet™ neural network," etc. The neural network is "region-based" in that not all pixels within the input image are labeled during training or inference. Instead, as further described herein, different subregions within the input image (e.g., defective and non-defective subregions) may be labeled or labeled to be trained by the neural network during inference (runtime). Image filtering is "digital" in that it is not performed by tool hardware, but by computer components (neural networks) in the post-image generation phase.

[0054] In another embodiment, the region-based neural network is composed of approximately 1,000 or fewer training parameters. In a further embodiment, the region-based neural network includes a predetermined limited number of layers (e.g., only 3 layers, fewer than 10 layers, etc.). In a further embodiment, the region-based neural network includes only a predetermined limited number of layers of stacked digital expansion convolution filters having expansion factors ranging from 0 to some predetermined limited number (e.g., 3 layers, fewer than 10 layers, etc.). For example, compared to a traditional deep learning (DL) neural network configuration having more than 1M training parameters and tens of layers, the neural network configuration described herein is designed to significantly reduce the number of training parameters to 1,000 or fewer and a predetermined limited number (e.g., 3), which still allows for nonlinearity and gain stability when trained on a highly imbalanced dataset having substantially only a few labeled data points (< 100) in a DOI class. Of course, the neural network configuration described herein is the best possible identified by the inventors, but variations of such neural network configuration may also be suitable for the applications described herein.

[0055] FIG. 3 illustrates the configuration of a cascaded filter neural network for image filtering. This configuration consists of only three layers of stacked digital expansion convolution filters, with the expansion factor increasing from 0 to a certain limited number. Additionally, the neural network includes a relatively shallow stack of convolution filters and a subsequent non-linear activation. Input image(s) within one or more channels (not shown) may be input to a first stack (300) of convolution layers having a non-linear activation (304) following a receptive FOV (302). The output of the non-linear activation (304) is input to a second stack (308) of convolution layers, followed by a non-linear activation (310). A non-linear activation (316) may follow the final stack (314) of convolution layers, followed by a region max pooling layer (320). The region max pooling layer can generate outputs (322 and 324). Output (322) corresponds to the noise region illustrated by the lighter gray area in the region max pooling layer (320), and output (324) corresponds to the detection region illustrated by the darker gray area in the region max pooling layer. In this way, output (322) may be a Newsons classification, and output (324) may be a defect / Newsons classification.

[0056] To construct AiryNet™ for the pre-detection stage, the network may first be trained on a set of relatively large images (e.g., work dumps, typically 4K × 1K pixels) or on smaller image patches (e.g., 32 x 32 pixels, 64 x 64 pixels, etc.) from lot results. For the post-detection case, training may be implemented using patch images from lot results. Image patches from lot results may include one or more sets of test, reference, and mask images for each detected defect location. In this way, training of the neural network may be performed using training images such as an inspection mask image (326) showing the detection spot (328) (an image generated by an inspection tool that masks out the non-defective parts of the difference image). A training image (or at least a defective part of the training image) can be expanded to generate an image (330) showing an expanded detection spot (332) that can be used for training in place of or in combination with the image (326). For example, a detected defect spot within a mask image can be expanded / inflated (this can be done in any suitable manner known in the art). In this way, the defect area in the training image can be defined using one or more of the expanded inspection mask image, defect review (e.g., scanning electron microscope (SEM)) image if available, and design information if available. Then, during training, the nuance classification and defect / nuance classification generated by the neural network (i.e., outputs (322 and 324)) can be compared with the label of the training image, and one or more parameters of the neural network can be changed until the classification matches the label for the training image.

[0057] Therefore, training may be performed in a supervised manner, which means that the images used for training are labeled in some way, for example, by a user or by another method or system having known capabilities, to detect defects on a specimen and / or to separate the detected defects into DOIs and nusens. In this way, the embodiments described herein provide a supervised nonlinear image filtering technique for defect detection and nusens reduction for optical wafer inspection and possibly other applications described herein. Training of the neural network may be performed in any suitable manner known in the art. Generally, as briefly described above, training may include inputting training images into the neural network and modifying one or more parameters of the neural network until the neural network output matches the label assigned to the training image. Of course, training may be much more complex than this, and the neural network may be trained in any other suitable manner known in the art.

[0058] In a further embodiment, the computer subsystem is configured to train multiple models with different combinations of predetermined hyperparameter values, identify which of the multiple models produces the result with the highest predefined metric (e.g., accuracy or other performance-related scores or metrics), and select the identified one of the multiple models as a region-based neural network. For example, the neural network training procedure may also be performed using multiple parameters referred to as hyperparameters. These hyperparameters include, but are not limited to, regularization factors, learning rates, network configurations, data augmentation, dropout rates, etc. While the values ​​of some parameters are pre-set / fixed or provided by the user and are not optimized during training iterations, other parameters are substantially sensitive to the training data, and the optimization of such parameters may be performed using hyperparameter search. A brute force method may be deployed for hyperparameter search, whereby the computer subsystem trains multiple models by trying all possible combinations of given hyperparameter values ​​and then, for example, searches for the best model with the highest accuracy. Then, the best model identified by hyperparameter search can be selected as the final model.

[0059] In one embodiment, the computer subsystem is configured to train a region-based neural network with a training dataset containing training images, and not all pixels within the training images are labeled. For example, the advantage of the region-based training approach stems from the fact that not all pixels within the input images need to be accurately annotated / labeled. In another embodiment, the computer subsystem is configured to train a region-based neural network with a training dataset containing training images, and different regions within the training images are labeled. For each image, different regions, e.g., defective subregions and non-defective subregions, can be automatically determined based on the original default detection method. For example, the defective regions and Newsons regions for the output image can be defined based on available SEM images or detection locations determined by known good tools and methods. Subsequently, the network can be trained using the regions to improve convergence and stability.

[0060] In this way, neural network filter coefficients can be optimized using a region-based supervised machine learning (ML) method based on image samples for DOIs and Newsense defects. That is, each training image can be divided into two subregions: a defect and a noise subregion. The defect subregion can be obtained by expanding the defect detection spot extracted from the mask image generated for the defect (which can be performed in any suitable manner known in the art). Subsequently, the expanded / inflated defect region is designated as the defect subregion, and the surrounding or residual region is designated as the noise subregion. Thus, the defect subregion and the noise subregion are different regions of the same image. The noise region and the defect subregion can be further adjusted using region information. The noise region and the defect subregion can also be adjusted by removing pixels from these subregions based on design information or image(s) if available. Maximum signal values ​​for the defect region and the surrounding noise region are extracted from each image. Then, these signal values ​​are classified (binded) into a positive class in the case of a DOI and a negative class in the case of a Newsense or surrounding noise event. The cost function used for optimization can be the negative log likelihood. Using stochastic gradient descent, neural network filters can be optimized to increase the separation between signals within the positive and negative classes.

[0061] In one embodiment, a computer subsystem is configured to train a region-based neural network using a training dataset comprising training images and labels for one or more defective subregions and one or more non-defective subregions within said training images. Such an embodiment also includes applying weights to one or more of the training images and labels to alter the contribution between the positive and negative classes in the cost function used for training. Further such embodiments include applying weights to one or more of the training images and labels to alter the contribution between different positive classes or different negative classes in the cost function used for training. For example, defect weights may be applied to data points to balance the contribution between the positive and negative classes within the cost function. Additionally, defect weights may be adjusted within each class to enhance or reduce the contribution of a selected subset of defects. Furthermore, defect weights may be further adjusted within the negative class to favor local defect signal enhancement or DOI-Newsense separation when the optical properties of the Newsenses and ambient noise events differ.

[0062] In an additional embodiment, a region-based neural network is configured to label at least one of one or more defective subregions and one or more non-defective subregions within a filtered image. For example, the AiryNet™ technique predicts defect locations by outputting an image having a filtered defect signal. For any one specimen location, the output image has the same resolution as the input image(s) showing the filtered defect signal and its location, as illustrated in FIG. 4. Images (400 and 402) are difference images for DOIs and nusens, respectively, showing multiple negative polarity (dark) spots that obscure defect localization. Filtered images (404 and 406) are generated by the neural network from images (400 and 402), respectively. The filtered image (intensity map) predicts and resolves multiple defect locations (in image (404)) and shows strong signals for two DOIs (in image (404)) and weak signals for nusens (in image (406)). Images (408 and 410) are SEM image sketches corresponding to images (400 and 402), respectively. The aligned corresponding SEM image sketches show two defects (a light line at the center of image (408)) and no defects (in image (410)). Thus, the defect locations in the SEM image sketches match the locations in the predicted images.

[0063] Accordingly, FIG. 4 illustrates defect images before and after the application of AiryNet™. Filtering removes ambient noise, enhances defect signals, and accurately reveals multiple DOI locations on the image (404) that match the location of the corresponding SEM image sketch (408). In the case of Newsons, filtering suppresses signals below the threshold level. FIG. 5 illustrates the effect of AiryNet™ filtering on ambient noise pixels. For the same 100% DOI true positive rate, the neural network suppresses background noise pixels by 2.5 orders of magnitude compared to the case where no filter is used on the test image. To determine the performance of the trained AiryNet™, a verification test using AiryNet™ filtering can be performed on another wafer or on the same wafer but in an area not used for AiryNet™ training.

[0064] The ROC curve of FIG. 5 illustrates that when AiryNet™ is used for filtering, background noise pixels are reduced by 2.5 times compared to the case where no filter is used for the same true positive value. In particular, FIG. 5 shows the ROC curve (500) generated after AiryNet™ filtering and the ROC curve (502) generated without such filtering. FIG. 5 also shows an example of a DOI image (504) and a Newsons image (506) generated by AiryNet™, and a DOI image (508) and a Newsons image (510) filtered by AiryNet™ to generate the DOI image (504) and Newsons image (506), respectively. As can be seen by comparing the DOI image and Newsons image before and after AiryNet™ filtering, AiryNet™ filtering significantly suppresses background pixel noise, enabling a much better ROC curve.

[0065] In some embodiments, the computer subsystem is configured to train and validate a region-based neural network using a cross-validation method. In one such embodiment, the initial stage of the cross-validation method generates cross-validated attributes for the training images, and subsequent stages of the cross-validation method are performed using cross-validated attributes combined with one or more attributes generated by other supervised attribute-based filtering methods. For example, in the case of a relatively limited training dataset (e.g., when the number of DOIs is relatively small), it is better to avoid strictly splitting the dataset into a training subset and a test subset. One of the reasons for this is to avoid discarding model predictions for relatively sparse data points within the training subset due to overfitting.

[0066] In such cases, AiryNet™ can be trained and validated on the same set of defects using a cross-validation method. For example, the dataset can be partitioned into n-folds. (n-1) folds are used for training, and the remaining 1 fold is used for validation to generate predictions. This process is repeated multiple times at each time using a different fold as the validation set. In this way, cross-validated predictions are generated for the entire dataset. Subsequently, the cross-validated attributes from AiryNet™ can be used in conjunction with other attributes available in the inspection tool and fed to the next stage of supervised attribute-based filtering available in the inspection tool.

[0067] In some embodiments, the computer subsystem is configured to select one or more images for a specimen location to be input into a region-based neural network from a plurality of images available for the specimen location. For example, the neural network may include a plurality of input image channels, and the images input into the neural network for any one specimen location may be selected by the computer subsystem. Typically, the same type of images may be input into the neural network for all specimen locations being inspected, measured, etc. In one such example, the computer subsystem may select one or more types of images to be input into the neural network during recipe setup, and then these image(s) may be input into the neural network for all locations on the specimen. However, in the same process, the image(s) input into the neural network for different specimen locations may be different.

[0068] In another embodiment, one or more images input to a region-based neural network include at least one of a difference image, a reference image, a test image, and a design image for a specimen location. For example, the neural network may include image channel(s) or some combination thereof for at least one of the difference image, the reference image, the test image, and the design image. Even if multiple input channels (images) are used, AiryNet™ generates only a single output image, which may also be referred to as a probability density map.

[0069] The images input to the neural network may vary depending on the configuration of the neural network and the images used to train it. For example, the images input to the neural network may include only difference images generated by a defect detection algorithm. However, the images input to the neural network may include only test images generated by an inspection subsystem. For any single specimen location or detected defect, the images input to the neural network may also include multiple images, which may contain two or more of the test images, reference images, and difference images. The images may have any suitable size, which may vary depending on the imaging hardware configuration and possibly the neural network architecture. Additionally, for any single specimen location or detected defect, the images input to the neural network may be, for example, relatively small patch images or much larger frames or working images, depending on the application. The input to the neural network may also include other information about the specimen location, such as design clips, which may be a relatively small part of the design for the specimen corresponding to the specimen location.

[0070] The images input to the neural network may include only specimen images generated by the inspection subsystem in which defects are detected. Defect detection may be performed on specimen images generated by the inspection subsystem in any suitable manner known in the art. For example, defect detection may include applying a defect detection method to the specimen image or a difference image generated therefrom. In the simplest configuration, defect detection may include applying a threshold value to the image (or difference image) and determining that any image having a signal exceeding the threshold value corresponds to a potential defect. Potential defects may include any type of defect present and detected on the specimen, as well as any nusons incorrectly detected as potential defects, which must therefore be separated from actual defects on the specimen. Defect detection may also be performed by applying any suitable defect detection algorithm to the specimen image, such as the MDAT algorithm used by some inspection tools commercially available from KLA.

[0071] In some cases, the embodiments described herein may be configured to perform defect detection. However, another system or method may be configured to perform defect detection on an image, and then only the image in which a defect has been detected may be input to a neural network by the embodiments described herein.

[0072] However, in other cases, the images input to the neural network may include arbitrary specimen images generated by the inspection subsystem for the specimen. For example, the neural network may be configured to perform digital non-linear filtering on all images generated for the specimen. In this way, the images may be input to the neural network without performing any kind of prior defect detection on the images.

[0073] In another embodiment, a region-based neural network is trained for a first type of DOI, and a computer subsystem is configured to input at least one image of a specimen location into an additional region-based neural network configured to perform digital non-linear filtering of at least one image to generate an additionally filtered image of the specimen location, and the additional region-based neural network is trained for a second type of DOI. For example, if multiple DOI types may exist on a wafer, multiple AiryNet™ neural networks may be trained, and each AiryNet™ may be trained for each defect type. These multiple AiryNet™ neural networks may be added to an inspection recipe and may be used to calculate multiple attributes to improve separability for different defect types. Each of the AiryNet™ neural networks may have the configuration described herein and may be trained as further described herein using only different training data sets (for different DOIs).

[0074] The computer subsystem is also configured to determine information regarding specimen locations from the filtered image. For example, AiryNet™ does not perform defect detection or Newsense filtering itself. Instead, AiryNet™ merely outputs a filtered image. The filtered image can then be used as input to other algorithms or methods, including applications further described herein.

[0075] In one embodiment, determining information involves applying a defect detection method to a filtered image. Accordingly, one of the ways in which the neural network filtering described herein may be used is inline during inspection in a pre-detection stage. For example, after AiryNet™ is trained, AiryNet™ may be applied inline during detection processing for image filtering. In this stage, the neural network generates a filtered image. For the pre-detection case, the generated filtered image is fed into a defect detection algorithm or method. Such defect detection may be performed as further described above or in any other suitable manner known in the art. That is, the filtered image may be used in the same way as any other image input into any suitable defect detection method.

[0076] In another embodiment, determining information involves applying a defect detection method to a filtered image and at least one other image generated for the specimen location. Accordingly, the neural network filtering described herein may be used inline during inspection at a pre-detection stage in different ways. For example, the filtered image may be used alone for defect detection, or may be used as a complementary image or substitute for one of the images in the currently used detection method. In one such example, both the filtered image and the original difference image may be input to the defect detection method. In another example, the original reference image and the filtered image used to generate the difference image for the specimen location may be input to the defect detection method. The other image described herein may also be input to defect detection using the filtered image.

[0077] AiryNet™ can also be used in post-detection. In some embodiments, determining information involves determining one or more attributes of specimen locations from filtered images and performing Newsons filtering based on one or more attributes. For example, in this stage, a neural network is trained on image patches from lot results or combinations of lot results and is then applied to an input set of images for each defect in the test lot results to calculate attributes. These attributes are then used for Newsons filtering. Thus, the neural network filtering described herein can be used inline during inspection or offline in the post-detection stage.

[0078] Similar to the pre-detection stage, the neural network generates a filtered image. For the post-detection case, neural network filtering is applied to inspection image patches in the lot results to generate a filtered image. Attribute(s) are calculated from defect and noise subregions of the filtered image and are then used for Newsense filtering or defect classification for Newsense filtering. Attribute(s) can be generated by extracting maximum signals from defect and noise subregions within the filtered image generated by AiryNet™. Subsequently, these attribute(s), rather than the filtered image, may be used for Newsense filtering along with other currently generated defect and / or image attributes.

[0079] AiryNet™ filtering can also be used for both the pre-detection and post-detection stages on the same wafer. For example, a filtered image generated for defect detection can also be used to calculate attributes for Newsons filtering. In this way, the same AiryNet™ filter can be trained and used to generate filtered images for both applications. However, if it is reasonable to generate different AiryNet™ filters for different applications (e.g., when different images are used for defect detection versus Newsons filtering attribute calculation), different AiryNet™ filters can be trained and used to filter images for defect detection or Newsons filtering purposes on the same wafer.

[0080] AiryNet™ may also be trained on a single wafer and subsequently used on wafers with the same design and at the same process stage. In other words, AiryNet™ filters may be specific to the design and process stage, but not necessarily specific to the wafer. Thus, a single trained AiryNet™ may be used for the same inspection recipe performed on multiple wafers. Additionally, in some cases, the same trained AiryNet™ may be useful for different specimen designs and / or process stages, or different AiryNet™ filters may be trained and used for specimens with different designs or specimens at different stages during device manufacturing. Furthermore, it may be possible for an AiryNet™ trained for one use case to be used as an initial network configuration to be retrained to generate a different AiryNet™ for another use case. Such retraining may be performed in the same manner as the training described herein.

[0081] FIG. 6 illustrates a use case for AiryNet™ filtering in a post-detection stage. As illustrated in step (600), the method may include a step of defining a goal or reason for determining post-detection attributes. As illustrated in FIG. 6, one example of a goal may be to improve DOI-Newsense separation for logic devices with a relatively high Newsense rate. Obviously, this is merely one example of a goal that may be defined for the embodiments described herein. In step (602), the method includes the step of sampling DOIs and Newsenses from lot results and the step of training a neural network on the lot result sampled patches (to improve DOI-Newsense separation). Sampling may be performed in any suitable manner, and training may be performed as further described herein. In step (604), the method includes running a neural network on lot result patches, extracting attributes (e.g., peak signals of defect and noise regions), and using these attributes as attributes for further filtering using a defect classifier. These steps may be performed as further described in this specification.

[0082] FIG. 7 illustrates the DOI capture rate versus Newsons rate curve (700) for a wafer after AiryNet™ filtering has been applied. AiryNet™ filtering was applied to a highly imbalanced dataset (the number of DOIs and Newsons is 122 and 5700, respectively). At a 50% Newsons rate corresponding to a 50x Newsons reduction factor, the DOI capture rate reaches 90%. That is, while advantageously maintaining 90% of the DOI, the cascaded filtering performed by AiryNet™ reduces the Newsons count by 50 times (50x) compared to size-based filtering.

[0083] FIG. 7 also illustrates the results of a stability test including ROC curves (702) corresponding to multiple optimizations for a randomly selected 50% of DOIs and Newsons. The remaining defects were used for testing. In FIG. 7, the ROC curve labeled as the cascade filter ROC curve is for the AiryNet™ filtering results, and the ROC curve labeled as the magnitude-based ROC curve in FIG. 7 is for filtering using only defect signal magnitudes. Newsons count fluctuations are significantly suppressed when applying AiryNet™ filtering, which exhibits good stability.

[0084] A computer subsystem may generate a result that may include the result of any of the steps described herein. The computer subsystem may be configured to store the result on any suitable computer-readable storage medium. The storage medium may include any storage medium described herein or any other suitable storage medium known in the industry. After the result is stored, the result may be accessed from the storage medium and used by any of the method or system embodiments described herein, formatted to be displayed to a user, and used by another software module, method or system, etc.

[0085] The results may include a trained AiryNet™ neural network, defect detection results generated by the trained AiryNet™ neural network, Newsons filtering results generated by the trained AiryNet™ neural network, or any other information generated by the embodiments described herein. The results may be generated by a computer subsystem in any suitable manner. The results may have any suitable form or format, such as a standard file type. The computer subsystem may generate and store the results so that the results are used by the computer subsystem and / or another system or method to perform one or more functions using the AiryNet™ neural network and / or to perform one or more functions using the results generated by the AiryNet™ neural network. For example, the AiryNet™ neural network (or its filename and location) may be stored in a recipe, such as an inspection recipe, and then used to generate a filtered image used to detect defects on the specimen or on Newsons filtering defects. In such circumstances, the computer subsystem may be configured to generate a new inspection recipe or modify an existing inspection recipe.

[0086] In another example, defects detected based on filtered images generated by an AiryNet™ neural network, or Newsons filtering results generated based on filtered images generated by an AiryNet™ neural network, may be stored in a lot result file and may be used for one or more other functions described herein. In such cases, the computer subsystem may generate inspection results partially generated using an AiryNet™ neural network, which may include information regarding the detected defects, such as the detected defects, size, detection score, information on defect classification such as class labels or IDs, for example, or information regarding the detected defects such as defect IDs, locations, etc., of any such suitable information known in the art bounding box. Results for defects may be generated by the computer subsystem in any suitable manner. Results for defects may have any suitable form or format, such as standard file types. The computer subsystem generates and stores results so that the results may be used by the computer subsystem and / or another system or method to perform one or more functions on a specimen or another specimen of the same type.

[0087] The results and information generated by performing an inspection on a specimen may be used in various ways by the embodiments and / or other systems and methods described herein. Such functions include, but are not limited to, modifying processes, such as manufacturing processes or steps performed or to be performed on the inspected specimen or another specimen in a feedback or feedforward manner. For example, a computer subsystem may be configured to determine one or more modifications to the processes performed or to be performed on the inspected specimen as described herein, based on detected defect(s). Modifications to the processes may include any suitable modifications to one or more parameters of the processes. The computer subsystem preferably determines such modifications so that defects may be reduced or prevented on another specimen where the modified process is performed, defects on the specimen may be corrected or eliminated in another process performed on the specimen, defects may be compensated in another process performed on the specimen, and so on. The computer subsystem may determine such modifications in any suitable manner known in the art.

[0088] Subsequently, such changes may be transmitted to a semiconductor manufacturing system (not shown) or a computer subsystem and a storage medium (not shown) accessible to the semiconductor manufacturing system. The semiconductor manufacturing system may or may not be part of the system embodiments described herein. For example, the computer subsystem and / or inspection subsystem described herein may be coupled to the semiconductor manufacturing system through one or more common elements, such as, for example, a housing, a power supply, a specimen processing device or mechanism. The semiconductor manufacturing system may include any semiconductor manufacturing system known in the art, such as a lithography tool, an etching tool, a chemical-mechanical polishing (CMP) tool, a deposition tool, etc.

[0089] Each of the embodiments of the system described above may be combined together into a single embodiment. In other words, unless otherwise stated in this specification, none of the system embodiments are mutually exclusive with any other embodiment.

[0090] In one such embodiment, a computer subsystem is configured to train a region-based neural network with a training data set including training images, at least one defective subregion or at least one non-defective subregion is labeled in each training image, not all pixels within the training image are labeled, and the region-based neural network is configured to label at least one of one or more defective subregions and one or more non-defective subregions within a filtered image. This embodiment may be configured and performed as further described herein.

[0091] In one such embodiment, determining information includes applying a defect detection method to a filtered image. In another such embodiment, determining information includes the step of determining one or more attributes of a specimen location from a filtered image and performing Newsons filtering based on one or more attributes. These steps may be performed as further described herein.

[0092] In additional such embodiments, at least one defective subregion or at least one non-defective subregion in each of the training images is labeled based solely on information regarding at least one location of the defective or non-defective in at least one other image generated for the specimen. In some such embodiments, information regarding at least one location of the defective or non-defective is determined from at least one different image having a scale different from that of the training images. In additional embodiments, information regarding at least one location of the defective or non-defective is determined from at least one different image without aligning at least one other image to any of the training images. In additional such embodiments, the training images and one or more images input to the region-based neural network are generated by an inspection subsystem, and at least one different image is generated by a defect review subsystem.

[0093] The "different images" described above may or may not be ground truth information as the term is commonly used, but they may serve as ground truth for labeling training images for the embodiments described herein. For example, the different images may be arbitrary images that are not subject to inference after neural network training, and may also be different from the training images. Thus, the different images may be ground truth type images, and their labels generated by ground truth type methods or systems, such as SEM defect inspection tools, may be used to label training images as described herein. However, the different images may simply be detection spots extracted from mask images generated by inspecting a specimen or other specimen using other known good defect detection methods. As further described herein, the advantage of using any of such different images for training image labeling is a result of the region-based characteristics of the embodiments described herein.

[0094] Accordingly, the embodiments described above have some significant differences from the neural network training and inference currently used in applications described herein. For example, currently used neural network detection methods (e.g., those that may be used in electron beam type inspection systems) would require labeling or annotating every pixel for neural network training. For optical inspection, pixel labeling typically originates from SEM images and requires an additional and often time-consuming set of steps, namely, acquiring the SEM images and subsequently scaling and aligning the defect locations from the SEM to optical images as pixel labels for the optical images, and accurate transfer. Furthermore, for various reasons, SEM images may not be available for neural network training image labeling, which prevents defect location labeling at the image level.

[0095] Accordingly, a significant difference between the embodiments described herein is that they do not require or rely on accurate labeling at the pixel level for neural network training. Only binary labels, such as DOIs or nuances (one label per set of input defect images), are sufficient for training. The absence of accurate labeling at the pixel level is compensated for by the fact that only the approximate location of the defects needs to be known to train the neural network embodiments presented herein, which is achieved using a region-based approach. Defect images can be subdivided into defect / noise subregions using only detection spots extracted from mask images. It has been shown that training the neural network using only binary labels and subregions together is sufficient to predict defect locations. Clearly, therefore, the embodiments described herein provide significant advantages for training and inference, including but not limited to performing much faster and cheaper training without requiring defect review type images for labeling training images.

[0096] The advantages of the embodiments described herein are provided by a number of important novel features. One of these features is that the embodiments enable a neural network-based defect filtering and detection method that does not require labeling or annotating all pixels for AiryNet™ detection or filtering schemes. Another such feature is the ability to select different input channels / images and select pixels belonging to different regions for training and inference. An additional novel feature is the ability to attach multiple models and compute multiple attributes that can improve separability for different defect types. An additional novel feature is the ability to sequentially train AiryNet™ with other supervised filtering attribute-based methods using cross-validation methods. The embodiments may also be implemented as part of existing optical inspection software packages for currently used tools or in any newly developed inspection tool. The embodiments may be implemented on different types of inspection tools, including but not limited to BF tools and DF tools.

[0097] Another embodiment relates to a computer-implemented method for determining information about a specimen location. The method comprises the step of inputting one or more images of a specimen location into a region-based neural network configured to perform digital non-linear filtering of one or more images to generate a filtered image of the specimen location. The method also comprises the step of determining information about the specimen location from the filtered image.

[0098] Each of the steps of the method may be performed as further described herein. The method may include any other step(s) that may be performed by the inspection subsystem, computer subsystem, and / or region-based neural network described herein. Additionally, the method described above may be performed by any system embodiment described herein.

[0099] Additional embodiments relate to a non-transient computer-readable medium storing program instructions executable on a computer system for performing a computer-implemented method for determining information about a specimen location. One such embodiment is illustrated in FIG. 8. In particular, as illustrated in FIG. 8, the non-transient computer-readable medium (800) includes program instructions (802) executable on a computer system (804). The computer-implemented method may include any step(s) of any method(s) described herein.

[0100] For example, program instructions (802) implementing a method such as that described herein may be stored on a computer-readable medium (800). The computer-readable medium may be a storage medium such as, for example, a magnetic disk or optical disk, a magnetic tape, or any other suitable non-transient computer-readable medium known in the art.

[0101] Program instructions may be implemented in any of the various methods, particularly including procedure-based techniques, component-based techniques, and / or object-oriented techniques. For example, program instructions may be implemented using ActiveX controls, C++ objects, JavaBeans, Microsoft Foundation Classes ("MFC"), Streaming SIMD Extension (SSE), or other technologies or methodologies, as desired.

[0102] The computer system (804) may be configured according to any embodiment described in this specification.

[0103] Further modifications and alternative embodiments of various aspects of the present invention will be apparent to those skilled in the art from this description. For example, a method and system for determining information regarding specimen locations are provided. Accordingly, this description should be interpreted merely as illustrative and is intended to teach those skilled in the art a general manner of carrying out the present invention. It should be understood that the forms of the invention shown and described herein are to be taken as currently preferred embodiments. Elements and materials may be substituted with those exemplified and described herein, parts and processes may be reversed, and specific attributes of the present invention may be utilized independently, as will be apparent to those skilled in the art after taking advantage of this description of the present invention. As described in the following claims, the elements described herein may be modified without departing from the spirit and scope of the present invention.

Claims

Claim 1 A system configured to determine information about a specimen includes a computer subsystem, and said computer subsystem: One or more images of a specimen location are input into a region-based neural network configured to perform digital non-linear filtering of the one or more images to generate a filtered image of the specimen location; A system configured to determine information about a specimen, configured to determine information about the specimen location from the filtered image above. Claim 2 In claim 1, the computer subsystem is also configured to train the region-based neural network with a training data set including training images, and is configured to determine information about a specimen such that not all pixels within the training images are labeled. Claim 3 In claim 1, the computer subsystem is also configured to train the region-based neural network with a training data set including training images, and is configured to determine information about a specimen in which different regions within the training images are labeled. Claim 4 A system configured to determine information about a specimen, wherein, in claim 1, the region-based neural network is also configured to label at least one of one or more defect subregions and one or more non-defect subregions within the filtered image. Claim 5 A system configured to determine information about a specimen, wherein, in claim 1, the computer subsystem is also configured to train the region-based neural network with a training data set including training images, at least one defective sub-region or at least one non-defective sub-region is labeled in each of the training images, and not all pixels in the training images are labeled, and the region-based neural network is also configured to label at least one of one or more defective sub-regions and one or more non-defective sub-regions in the filtered image. Claim 6 A system configured to determine information about a specimen, wherein determining the information in paragraph 5 includes applying a defect detection method to the filtered image. Claim 7 A system configured to determine information about a specimen, wherein, in paragraph 5, determining the information includes determining one or more attributes of the specimen location from the filtered image and performing nuisance filtering based on the one or more attributes. Claim 8 A system configured to determine information about a specimen, wherein, in paragraph 5, the at least one defect sub-region or at least one non-defect sub-region within each of the training images is labeled based solely on information about at least one location of the defect or non-defect within at least one other image generated for the specimen. Claim 9 A system configured to determine information about a specimen, wherein, in claim 8, information about the at least one location of the defect or the non-defect is determined from the at least one other image having a different scale from the training image. Claim 10 A system configured to determine information about a specimen, wherein, in claim 8, information about the at least one location of the defect or the non-defect is determined from the at least one other image without aligning the at least one other image with any of the training images. Claim 11 A system configured to determine information about a specimen, wherein, in claim 8, the training image input to the region-based neural network and the one or more images are generated by an inspection subsystem, and the at least one other image is generated by a defect review subsystem. Claim 12 A system configured to determine information about a specimen, wherein, in claim 1, the region-based neural network is also configured as a convolutional region-based neural network having a plurality of input image channels. Claim 13 In claim 1, the region-based neural network is also a system configured to determine information about a specimen, wherein the region-based neural network is composed of approximately 1,000 or fewer training parameters. Claim 14 A system configured to determine information about a specimen, wherein the region-based neural network in claim 1 comprises a predetermined limited number of layers. Claim 15 A system configured to determine information about a specimen, wherein the region-based neural network comprises a predetermined limited number of stacked digital expansion convolution filters having a dilation factor ranging from 0 to a predetermined limited number. Claim 16 A system configured to determine information about a specimen, wherein the computer subsystem is also configured to select one or more images of the specimen location input to the region-based neural network from a plurality of images available for the specimen location. Claim 17 A system configured to determine information about a specimen, wherein, in claim 1, the region-based neural network is trained for a first type of defect of interest, and the computer subsystem is also configured to input at least one image of the specimen location into an additional region-based neural network configured to perform digital non-linear filtering of the at least one image to generate an additional filtered image of the specimen location, and the additional region-based neural network is trained for a second type of defect of interest. Claim 18 In claim 1, the computer subsystem is also configured to determine information about a specimen, wherein the computer subsystem is configured to train and validate the region-based neural network using a cross-validation method. Claim 19 A system configured to determine information about a specimen, wherein, in paragraph 18, the initial stage of the cross-validation method generates cross-validated attributes for a training image, and the subsequent stage of the cross-validation method is performed using the cross-validated attributes combined with one or more attributes generated by another supervised attribute-based filtering method. Claim 20 A system configured to determine information about a specimen, wherein, in claim 1, the one or more images input to the region-based neural network include at least one of a difference image, a reference image, a test image, and a design image for the specimen position. Claim 21 A system configured to determine information about a specimen, wherein, in claim 1, the computer subsystem is also configured to train the region-based neural network with a training dataset comprising training images and labels for one or more defective subregions and one or more non-defective subregions within the training images, and to apply weights to one or more of the training images and labels to change the contribution between the positive class and the negative class within the cost function used for the training. Claim 22 A system configured to determine information about a specimen, wherein the computer subsystem also trains the region-based neural network with a training dataset comprising training images and labels for one or more defective subregions and one or more non-defective subregions within the training images, and applies weights to one or more of the training images and labels to change the contribution between different positive classes within the cost function used for the training. Claim 23 A system configured to determine information about a specimen, wherein the computer subsystem also trains the region-based neural network with a training dataset comprising training images and labels for one or more defective subregions and one or more non-defective subregions within the training images, and applies weights to one or more of the training images and labels to change the contribution between different speech classes within the cost function used for the training. Claim 24 A system configured to determine information about a specimen, wherein the computer subsystem is also configured to train a plurality of models with different combinations of predetermined hyperparameter values, identify which of the plurality of models produces a result having the highest predetermined metric, and select the identified one of the plurality of models as the region-based neural network. Claim 25 A system configured to determine information about a specimen, wherein determining the information includes applying a defect detection method to the filtered image in the first paragraph. Claim 26 A system configured to determine information about a specimen, wherein determining the information includes applying a defect detection method to the filtered image and at least one other image generated for the specimen location in claim 1. Claim 27 A system configured to determine information about a specimen, wherein determining the information includes determining one or more attributes of the specimen location from the filtered image and performing Newsons filtering based on the one or more attributes. Claim 28 A system configured to determine information about a specimen, further comprising an inspection subsystem configured to generate at least one of the one or more images of the specimen location in claim 1. Claim 29 A system configured to determine information about a specimen, wherein, in claim 1, at least one of the one or more images generated for the specimen location is an optical wafer image generated by an inspection subsystem. Claim 30 A non-transient computer-readable medium storing program instructions executable on a computer system for performing a computer-implemented method for determining information about a specimen location, wherein the computer-implemented method comprises: a step of inputting one or more images about a specimen location into a region-based neural network configured to perform digital non-linear filtering of the one or more images to generate a filtered image about the specimen location; and a step of determining information about the specimen location from the filtered image. Claim 31 A computer-implemented method for determining information about a specimen location, comprising: inputting one or more images about a specimen location into a region-based neural network configured to perform digital non-linear filtering of said one or more images to generate a filtered image about said specimen location; and determining information about said specimen location from said filtered image, wherein the input and determination are performed by a computer subsystem.

Images